Document [original]

Identifying customer usage profiles of

two-wheeled vehicles

vorgelegt von

M.Sc.

Christian Gorges

geb. in Potsdam

von der Fakultät V – Verkehrs- und Maschinensysteme

der Technischen Universität Berlin

zur Erlangung des akademischen Grades

Doktor der Ingenieurwissenschaften

– Dr.-Ing. –

genehmigte Dissertation

Promotionsausschuss:

Vorsitzender: Prof. Dr.-Ing. Steffen Müller

Gutachter: Prof. Dr.-Ing. Robert Liebich

Gutachter: Prof. Dr.-Ing. Dieter Schramm

Tag der wissenschaftlichen Aussprache: 19.11.2018

Berlin 2018

Acknowledgements

This research has been funded by BMW’s PhD ProMotion programme. This support

is gratefully acknowledged.

First, I would like to thank Dr. Kemal Öztürk, expert in durability and load analysis

at BMW Motorrad, Germany, who supervised this thesis. Thank you for the great

support, the many constructive and fruitful discussions, keeping my back free from

day-to-day business, accepting me as an equally valued employee, the unconditional

trust with the direction of my project, and your personal engagement for making

this research high quality.

A further thank you belongs to Prof. Dr. Robert Liebich at the Department of

Engineering Design and Product Reliability at the Institute of Technology Berlin,

Germany. Thank you for giving me the possibility of an industry-sponsored PhD

programme and the courage for the first cumulative PhD thesis at your chair. Your

knowledgeable and useful hints and comments supported me in quickly publishing

the first results at relevant journals.

Additionally, I would like to thank Prof. Dr. Dieter Schramm, Head of the De-

partment Chair Mechatronics and Head of the Department Mechanical Engineering

at the University of Duisburg-Essen, Germany, for reviewing this thesis and giving

useful comments for improving the discussion chapter.

Finally, I would also like to thank my family and my friends for their great support.

Munich, November 2018

Christian Gorges

Preface

This cumulative PhD thesis contains the manuscripts of the following publications:

[1] – C. Gorges, K. Öztürk and R. Liebich. ‘Customer loads of two-wheeled vehicles’.

In: Vehicle System Dynamics 55.12 (2017), pp. 1842–1864

[2] – C. Gorges, K. Öztürk and R. Liebich. ‘Road classification for two-wheeled

vehicles’. In: Vehicle System Dynamics 56.8 (2018), pp. 1289–1314

[3] – C. Gorges, K. Öztürk and R. Liebich. ‘Impact detection using a machine

learning approach and experimental road roughness classification’. In: Mechanical

Systems and Signal Processing 117 (2019), pp. 738–756

iii

Summary

This cumulative PhD thesis shows the development of methods for identifying cus-

tomer usage profiles of two-wheeled vehicles utilising the vehicle’s onboard sensors.

It comprises three papers that have been published to relevant journals. At present,

regarding the automotive industry, customer usage profiles are mostly unknown in

durability engineering and the vehicle development process. The detailed knowledge

about this crowd-sourced data would improve vehicle design targets and enable a

virtual load acquisition. Therefore, it is desirable to identify customer usage and

customer loads for every vehicle. The first paper presents a model-based customer

load acquisition system that calculates the occurring wheel forces. Therefore, the

current road slope and the vehicle mass are estimated using a Kalman filter. The

resulting wheel forces are subsequently counted with the rainflow method. The valid-

ation was achieved by the comparison of measurements with wheel-load transducers.

The second publication presents a three-part road classification system: first, a curve

estimator was developed for identifying and classifying road curves; second, the road

slope was utilised for counting the hilliness of a given road; and third, a modular

road profile estimator was developed for classifying the road roughness according to

ISO 8608. The approach uses the vehicle’s transfer functions to estimate the road

excitation from the resultant vehicle motions. The third publication experimentally

validates the road roughness classification method by comparing the results to laser-

scanned road profiles. The comparison shows that even rough roads are detected

correctly within a short time span. In addition, an impact detection strategy was

developed using a supervised machine learning approach. A study of the six most

popular classification algorithms was achieved for detecting mild and severe special

events. The combination of the road roughness classification method and the impact

detection strategy enables a holistic field-data acquisition of customer usage profiles.

The methods presented are discussed in the context of the digital transformation

and the increasing value of data.

Zusammenfassung

Die vorliegende kumulative Dissertation zeigt die Entwicklung von Methoden zur

Identifizierung der Kundennutzungsprofile von Motorrädern unter Nutzung der vor-

handenen Onboard-Signale der serienmäßig verbauten Sensorik. Die Arbeit besteht

aus drei Veröffentlichungen, welche in einschlägigen Fachzeitschriften veröffentlicht

wurden. In der Betriebsfestigkeit und im Produktentstehungsprozess sind Kunden-

nutzungsprofile weitgehend unbekannt bzw. nicht aufgezeichnet. Eine detaillierte Be-

trachtung dieser Felddaten würde die Lastannahmen validieren und die Fahrzeugent-

wicklung verbessern. Es ist daher erstrebenswert, Kundennutzungsprofile von allen

Fahrzeugen zu erfassen und für die Entwicklung einzusetzen. Die erste Publikation

stellt eine modellbasierte Radkraftberechnung vor. Dafür wurde die aktuelle Steigung

der Fahrbahn und die Gesamtmasse des Fahrzeugs mit Hilfe eines Kalman-Filters

geschätzt. Die berechneten Radkräfte werden mittels der Rainflow-Zählung klassiert.

Eine Validierung wurde mit Hilfe von Radmessnaben durchgeführt. Die zweite Ver-

öffentlichung beschreibt eine dreiteilige Streckenprofil-Klassierung. Zuerst wurde ein

Klassierungsverfahren für die Kurvigkeit einer Strecke entwickelt. Anschließend wur-

de die Steigung der Fahrbahn zur Klassierung der Hügeligkeit verwendet. Abschlie-

ßend wurde ein modulares System zur Klassierung der Streckenrauheit nach ISO

8608 entwickelt. Die vorgestellte Methode verwendet die Übertragungsfunktionen

des Fahrzeugs, um mit Hilfe der resultierenden Bewegungen auf die Unebenheiten

der Strecke zu schließen. Die dritte Publikation zeigt die experimentelle Validierung

der Klassierung der Streckenrauheit durch den Vergleich mit ausgesuchten, laser-

vermessenen Streckenprofilen. Zusätzlich wurde ein Machine Learning Ansatz ver-

wendet, um milde und schädliche Sonderereignisse zu detektieren. Die Kombination

aus Rauheitsklassierung und Detektion von Sonderereignissen, wie z.B. Hindernis-

überfahrt, ermöglicht eine ganzheitliche Erfassung von Kundennutzungsprofilen. Die

vorgestellten Methoden wurden abschließend in den Kontext der digitalen Transfor-

mation und der steigenden Bedeutung von Daten gesetzt.

vii

Contents

1 Introduction 1

1.1 Durability in vehicle engineering . . . . . . . . . . . . . . . . . . . . 1

1.2 The need for customer usage profiles . . . . . . . . . . . . . . . . . . 3

1.3 Approach and experimental set-up . . . . . . . . . . . . . . . . . . . 6

1.4 Vehicle dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

1.5 A brief literature survey . . . . . . . . . . . . . . . . . . . . . . . . . 11

1.6 Overview of published papers . . . . . . . . . . . . . . . . . . . . . . 13

2 Customer loads of two-wheeled vehicles 15

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

2.2 Experimental set-up and data analysis . . . . . . . . . . . . . . . . . 18

2.3 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

2.4 Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

2.5 Summary and Conclusion . . . . . . . . . . . . . . . . . . . . . . . . 42

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

3 Road classification for two-wheeled vehicles 47

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

3.2 Experimental set-up . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

3.3 Road curve estimator . . . . . . . . . . . . . . . . . . . . . . . . . . . 50

3.4 Road slope classification . . . . . . . . . . . . . . . . . . . . . . . . . 53

3.5 Road profile estimator . . . . . . . . . . . . . . . . . . . . . . . . . . 54

3.6 Results and Validation . . . . . . . . . . . . . . . . . . . . . . . . . . 70

3.7 Summary and Conclusion . . . . . . . . . . . . . . . . . . . . . . . . 76

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78

4 Impact detection using a machine learning approach and experimental

road roughness classification 83

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84

4.2 Longitudinal road profiles . . . . . . . . . . . . . . . . . . . . . . . . 89

4.3 Measurement campaign . . . . . . . . . . . . . . . . . . . . . . . . . 91

4.4 Road roughness classification . . . . . . . . . . . . . . . . . . . . . . 98

Contents

4.5 Impact detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103

4.6 Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . 106

4.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116

5 Discussion 123

5.1 Summary of the developed methods . . . . . . . . . . . . . . . . . . 123

5.2 Implementation into the vehicle development process . . . . . . . . . 126

5.3 Statistical considerations . . . . . . . . . . . . . . . . . . . . . . . . . 131

5.4 Application to four-wheeled vehicles . . . . . . . . . . . . . . . . . . 135

5.5 Data as a resource . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137

5.6 Outlook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138

Bibliography 143

A Onboard signals 147

B Vehicle dynamics 151

C Kalman filter 157

List of Figures

1.1 Vehicle development process. . . . . . . . . . . . . . . . . . . . . . . 2

1.2 Classification of customer loads. . . . . . . . . . . . . . . . . . . . . . 3

1.3 Approach for the online identification of customer usage profiles. . . 6

1.4 Test motorcycle of type BMW R1200GS. . . . . . . . . . . . . . . . . 8

1.5 2D data-logging device and GPS-module. . . . . . . . . . . . . . . . 8

2.1 Distribution of customer loads, survey sampling, and structural strength. 17

2.2 Test motorcycle with the global reference coordinate system. . . . . . 18

2.3 Amplitude spectrum of vertical wheel forces. . . . . . . . . . . . . . . 19

2.4 Road slope estimation physics. . . . . . . . . . . . . . . . . . . . . . 21

2.5 Model of the driveline dynamics. . . . . . . . . . . . . . . . . . . . . 24

2.6 External forces acting on the motorcycle. . . . . . . . . . . . . . . . . 26

2.7 Motorcycle model with three rigid bodies. . . . . . . . . . . . . . . . 31

2.8 Flow chart of the customer load estimation model. . . . . . . . . . . 35

2.9 Validation of the road slope estimator at the proving ground. . . . . 37

2.10 Validation of the road slope estimator at the mountain track. . . . . 37

2.11 Validation of the mass estimator. . . . . . . . . . . . . . . . . . . . . 38

2.12 Validation of the longitudinal wheel forces. . . . . . . . . . . . . . . . 39

2.13 Validation of the resultant wheel forces in the Y-Z-plane. . . . . . . 41

2.14 Detailed extract of the resultant wheel forces. . . . . . . . . . . . . . 41

2.15 Rainflow matrices of the resultant rear-wheel forces in the Y-Z-plane. 42

3.1 Road curve estimation physics. . . . . . . . . . . . . . . . . . . . . . 52

3.2 Definition of longitudinal road profiles. . . . . . . . . . . . . . . . . . 56

3.3 PSD and straight line fit of an artificial road profile according to ISO

8608. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58

3.4 Pseudo random test track comprising road classes A–H and a detailed

extract of class C. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60

3.5 Full-vehicle model with four DOFs. . . . . . . . . . . . . . . . . . . . 61

3.6 Bodeplot of transfer functions for v= 15 m s−1.. . . . . . . . . . . . 65

3.7 Flow chart of the road profile estimation algorithm. . . . . . . . . . . 67

3.8 Examples of road profile classification. . . . . . . . . . . . . . . . . . 69

List of Figures

3.9 Results of the road curve estimator. . . . . . . . . . . . . . . . . . . 70

3.10 Properties of road curve No. 7. . . . . . . . . . . . . . . . . . . . . . 71

3.11 Results of the road curviness classification. . . . . . . . . . . . . . . . 72

3.12 Results of the road slope classification. . . . . . . . . . . . . . . . . . 73

3.13 Validation of the road profile estimator. . . . . . . . . . . . . . . . . 75

4.1 Classification of customer loads. . . . . . . . . . . . . . . . . . . . . . 85

4.2 Test tracks for the road roughness classification. . . . . . . . . . . . . 94

4.3 Road profile and histogram of test track No. 4. . . . . . . . . . . . . 95

4.4 Smoothed PSDs of selected test tracks. . . . . . . . . . . . . . . . . . 96

4.5 Road obstacles. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97

4.6 Derivation of reduced stiffness and damping coefficients. . . . . . . . 100

4.7 Lower and upper bound of spatial frequency ndepending on the ve-

locity v.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102

4.8 Schematic view of an impact. . . . . . . . . . . . . . . . . . . . . . . 103

4.9 Onboard signals during kerb crossing. . . . . . . . . . . . . . . . . . 104

4.10 Scatter plot of the training set. . . . . . . . . . . . . . . . . . . . . . 105

4.11 Results of the road roughness classification for test tracks No. 1–6. . 107

4.12 Results of the road roughness classification for test tracks No. 7–9. . 108

4.13 Decision surfaces for different classifiers. . . . . . . . . . . . . . . . . 111

4.14 Binary decision tree. . . . . . . . . . . . . . . . . . . . . . . . . . . . 113

4.15 Flowchart of the impact detection strategy. . . . . . . . . . . . . . . 116

5.1 Information value loop. . . . . . . . . . . . . . . . . . . . . . . . . . . 127

5.2 Durability test rig for motorcycles. . . . . . . . . . . . . . . . . . . . 130

5.3 Annual mileage of a population of vehicles. . . . . . . . . . . . . . . 131

5.4 Median air temperature of different regions. . . . . . . . . . . . . . . 133

5.5 Statistical treatment of distributed variables. . . . . . . . . . . . . . 134

5.6 Statistical derivation of personas. . . . . . . . . . . . . . . . . . . . . 139

5.7 Three components of a virtual load acquisition. . . . . . . . . . . . . 140

A.1 Block diagrams and signal flow of developed methods. . . . . . . . . 148

B.1 Quarter of Vehicle (QoV). . . . . . . . . . . . . . . . . . . . . . . . . 152

B.2 Half of Vehicle (HoV). . . . . . . . . . . . . . . . . . . . . . . . . . . 154

xii

List of Tables

1.1 Features of customer usage profiles. . . . . . . . . . . . . . . . . . . . 5

2.1 Required onboard signals for the wheel force calculation. . . . . . . . 32

2.2 Pseudo damage ratio. . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

3.1 Properties of the road classes according to ISO 8608 and Sayers and

Karamihas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60

3.2 Full-vehicle model properties. . . . . . . . . . . . . . . . . . . . . . . 63

3.3 Octave bands and geometric mean values for road classification. . . . 68

3.4 Properties of the classified road curves. . . . . . . . . . . . . . . . . . 71

3.5 Confusion matrix of the road profile estimator. . . . . . . . . . . . . 76

4.1 Properties of test tracks. . . . . . . . . . . . . . . . . . . . . . . . . . 93

4.2 Properties of measured special events. . . . . . . . . . . . . . . . . . 98

4.3 Results of the road roughness classification. . . . . . . . . . . . . . . 107

4.4 Properties of classification methods. . . . . . . . . . . . . . . . . . . 110

4.5 Confusion matrix of binary decision tree. . . . . . . . . . . . . . . . . 113

A.1 Overview of sensor signals and model signals. . . . . . . . . . . . . . 149

xiii

List of Abbreviations

Abbreviation Description

2D Debus & Diebold Meßsysteme GmbH

ABS Anti-lock Braking System

ACEM European Association of Motorcycle Manufacturers

ANFIS Adaptive Neuro-Fuzzy Inference System

ANN Artificial Neuronal Network

ARAS Advanced Rider Assistance Systems

ASTM American Society for Testing and Materials

BMW Bayerische Motoren Werke

CAN Controller Area Network

CBM Condition-Based Monitoring

CE Consumer Electronics

CMC Connected Motorcycle Consortium

COG Centre of Gravity

CPU Central Processing Unit

CRG Curved Regular Grid

DOF Degree of Freedom

DTC Dynamic Traction Control

eCall Emergency Call

ECU Electronic Control Unit

EKF Extended Kalman Filter

EU European Union

FFT Fast Fourier Transform

FIR Finite Impulse Response

GPS Global Positioning System

HIL Hardware-in-the-Loop

List of Abbreviations

HoV Half of Vehicle

ICA Independent Component Analysis

IIR Infinite Impulse Response

IMU Inertial Measurement Unit

IOT Internet of Things

IRI International Roughness Index

ISO International Organization for Standardization

KF Kalman Filter

LTI Linear Time Invariant

MIRA Motor Industry Research Association

ML Machine Learning

OBIS On Bike Information System

ODE Ordinary Differential Equation

QoV Quarter of Vehicle

PPV Positive Predictive Value

PSD Power Spectral Density

RLS Recursive Least Square

SIM Subscriber Identity Module

SVM Support Vector Machine

TPR True Positive Rate

UKF Unscented Kalman Filter

VMC Virtual Measurement Campaign

WEF World Economic Forum

xvi

1 Introduction

1.1 Durability in vehicle engineering

Durability is an essential discipline in engineering. Johannesson and Speckert [4]

offered a rather general definition of the term: “Durability is the capacity of an item

to survive its intended use for a suitable long period of time”. In order to evaluate

the durability of a product, the occurring loads during the product’s life need intense

investigation, namely through load analysis. According to Johannesson and Speckert,

load analysis in vehicle engineering comprises three steps: First, evaluating and

quantifying the customer service loads. Second, deriving design loads for vehicles,

sub-systems and components. Third, define verification loads and test procedures

for the verification of components, sub-systems and vehicles. Thus, the process of

durability engineering coincides with the well known v-model of product development

[5], see Figure 1.1. Initially, the product concept must be defined, which comprises

the class of vehicle, market segment, target cost, volume, size, weight, wheel base,

etc. After the initial concept phase passed, the vehicle engineering process consists

mainly of two phases: product design and product validation.

The product requirements must be defined at the beginning of the design pro-

cess. Thus, overall targets are defined for the physical properties of the product;

for example, performance, durability, safety, acoustics and vibration comfort. In the

case of durability, targets, in the form of occurring loads, are usually defined with

measurements from predecessors and load assumptions. Subsequently and follow-

ing the cascade down, design targets are derived for sub-systems and components;

for example, chassis, engine and suspension systems. After the first test parts are

produced, strength and durability tests are conducted on test rigs based on load

assumptions and simulations. The strength of these components is influenced by the

design, the manufacturing process and the material properties.

The validation of the components, the sub-systems, and systems follows the same

1 Introduction

Integration &

Implementation

Vehicle Testing

Level of maturity

Level of detail

Component

System

Vehicle

Design Validation

Requirements

Design &

Simulation System Testing

Initial Phase

Specs

Scope

Figure 1.1 – Vehicle development process.

cascade up. After the components were validated, sub-systems are evaluated; for ex-

ample, on different test rigs. Finally, after all sub-tests passed, the complete vehicle is

validated with endurance tests, which simulate a vehicle’s life in a short time period.

The test environment for endurance tests is designed close to the previously-defined

vehicle requirements and load assumptions; for example, a given mix of different road

types is driven with pre-defined velocity ranges. This simplified description of the

vehicle engineering process does not include the essential development loops and the

influence of virtual product development methods. For similar descriptions of the

v-model, see Johannesson and Speckert [4] and Pötter [6].

However, the pre-defined product requirements and load assumptions influence the

complete vehicle engineering process, beginning from the product design and ending

with the vehicle validation. In the past, measurements from predecessors and load

assumptions lead to incrementally improving products. Given that measurements are

cost- and time-consuming, and in terms of lightweight design, the product require-

ments should be defined as precise as possible to meet the customer requirements. At

the present, the customer requirements are mostly unknown. A well-defined product

requirement will only be optimal when the occurring loads are known in advance;

otherwise, the requirements remain as assumptions and are not optimal in sense of

1.2 The need for customer usage profiles

Figure 1.2 – Classification of customer loads.

durability and lightweight design. In this context, the present research addresses the

first and last step of the product development, which is, quantifying the requirements

and subsequently, testing the vehicle.

1.2 The need for customer usage profiles

The distribution of customer loads is often unknown. As mentioned above, it is

standard practice to estimate customer loads with survey sampling and measure-

ments with predecessors, where selected test vehicles are equipped with additional

measurement devices. It is common to choose customers or even test drivers who are

characterised by a forced driving style. Figure 1.2 illustrates the broad probability

distribution of customer loads in comparison to the narrower probability distribu-

tions of survey sampling, endurance tests and structural strength. Obviously, survey

sampling cannot reveal the complete distribution of customer loads. In addition, it

is not guaranteed, in which relation the selected drivers are with the distribution of

customers, which means more precisely, which quantile the measured loads repres-

ent. Moreover, survey sampling is expensive due to the additional equipment costs.

Furthermore, the endurance tests at the end of the vehicle development process,

cannot be assigned to specific quantiles without a detailed knowledge of the under-

lying distribution. These aspects raise the demand of revealing the customer load

distribution.

The problem in hand can be divided mainly into two steps: First, revealing the

1 Introduction

entire distribution of customer loads. Consequently, every customer would have

to be evaluated. Second, defining individual load targets and endurance tests for

the individual customer quantiles. This is more a statistical problem and differs

from manufacturer, product, and component and is not part of the present scope.

The current publication addresses the first task by presenting methods that collect

customer loads in real time by using the onboard sensors.

The distribution of customer loads is affected by variability, as illustrated in Fig-

ure 1.2. Load variability has different sources, as Johannesson and Speckert [4]

discussed. According to the authors, controlled variation is given by different vehicle

specifications, markets or regions. This controlled variation can be distinguished by

classification. In contrast, uncontrolled variation can only be handled by statistical

considerations. A population of the same type of vehicles will be subjected to a

various number of road irregularities, curves and obstacles, depending on their usage

and environment. In addition, different driving styles ranging from restrained to ag-

gressive increase the variability. This variation is even more present at two-wheeled

vehicles, since they are characterised by diverse application possibilities, compared

to passenger cars.

In the automotive industry and referring to Matz [7] and Pötter [6], customer loads

are divided in three categories: service loads, special events and misuse events, as

illustrated in Figure 1.2. This separation is characterised by statistical considerations

and is necessary for product liability and warranty. Service loads occur during the

normal use of the vehicle, which is called the intended purpose. They can be described

by a continuously distributed load spectra during the life of the vehicle. In the case

of a motorcycle, service loads comprise acceleration and brake manoeuvres, cornering

and loads that occur due to the roughness of the road surface. In addition to the

service loads, the intended purpose also includes the occurrence of special events,

which are rare compared to service loads. Special events induce a higher load on the

vehicle components and they are often characterised by impacts from sudden events;

for example, driving over a pothole. By definition, misuse events are not part of the

intended purpose, but they are also considered during the vehicle design process.

Figure 1.2 shows that the load severity of misuse events typically coincides with the

structural strength of the components. Thus, the components will be over exposed

and damaged. Misuse events are also often the consequence of impacts; for example,

1.2 The need for customer usage profiles

Table 1.1 – Features of customer usage profiles.

Classification method Continuous signal Event

Direct

Mileage Gear shifts

Velocity Controls

Engine speed

Throttle

Gear position

(Location & time)

Model-based

Vehicle loading[1]Curves[2]

Wheel forces[1]Special events[3]

Curviness[2]

Hilliness[2]

Road roughness[2,3]

riding against, or over a significant obstacle. The collection of both service loads and

special events is addressed by this publication.

The present research extends the term of customer loads to the more general

concept of customer usage profiles, because the methods presented capture more

than merely loads. Table 1.1 shows the components of customer usage profiles de-

pending on their classification method. Direct classification defines features that can

simply be classified from the sensor signals; for example, velocity and gear position.

In the case of a continuous signal, the classification is often realised using a time-at-

level counting in one or two dimensions. For more information about classification

methods, see Köhler et al. [8]. In contrast, model-based classification requires mul-

tiple sensor signals and an underlying model for calculating or estimating advanced

features and vehicle states. These model-based features are the focus of this PhD

thesis and they comprise vehicle loading, wheel forces, road curviness, road hilliness,

road roughness and the occurrence of special events. The superscripts in Table 1.1

indicate the publication, which present the respective methods. Direct classification

of sensor signals is state of the art in the automotive industry and is not part of the

present research. Furthermore, location- and time-based motion profiles are also not

part of the present research.

All of the mentioned aspects can be categorised as external loads. Internal loads

are also part of customer usage profiles; for example, vibrations induced by the en-

1 Introduction

ABS sensor

Acceleration

sensor Suspension

sensor

CAN bus Online processing

Methods for classification and counting

©BMW AG

Figure 1.3 – Approach for the online identification of customer usage profiles.

gine. However, they are not in the scope of this research. Johannesson and Speckert

[4] describe the external load environment with three aspects: longitudinal input,

which means braking and accelerating; transversal input, which means loads in-

duced by driving curves and; vertical input, which means loads induced due to road

roughness and dynamic wheel loading. All three external load inputs are addressed

by this cumulative PhD thesis to reveal complete customer usage profiles in terms

of durability.

1.3 Approach and experimental set-up

The number of onboard sensors increases based on the development of functions

of two-wheeled vehicles such as anti-lock braking system (ABS), dynamic traction

control (DTC), curve assistant, as well as driving assistance systems in the near

future. Consequently, the connected motorcycle comprises various signals that can

be accessed by the vehicle’s Controller Area Network (CAN) bus, as shown in Fig-

ure 1.3. Following, the main approach of this research is gathering the individual

components of customer usage profiles using the onboard sensors. Therefore, dif-

ferent methods were developed to estimate advanced vehicle states in real time; for

example, vehicle loading and wheel forces. Subsequently, the estimated values are

counted and stored in the vehicle’s electronic control units (ECUs). An online im-

plementation of the counting algorithms is mandatory, since the time series of the

CAN bus signals cannot be stored in the control units due to memory efficiency. In

1.3 Approach and experimental set-up

addition, various vehicle applications could use the revealed information for changing

vehicle settings and improving existing functions. Up to this point, a dataset will be

created which represent the individual customer-specific behaviour. It is also called

a customer usage profile on the small scale. The vehicles will send the collected clas-

sification results to the manufacturer; for example, during workshop appointments.

BMW Motorrad already released an eCall system in 2017, which comes along with

a built-in SIM card. Such a system could also be used to send the customer usage

profiles via mobile communication to the manufacturer. Subsequently, the customer

usage profiles can be derived offline from the overall distribution of the individual

customer usage components. As a result, these derived customer usage profiles rep-

resent the customer behaviour on the large scale; for example, for a given statistic,

which is often the median or the 99 % quantile. The definition of design targets and

product requirements is a statistical problem and is demonstrated in the discussion

(see Chapter 5). By implementing this type of field-data acquisition, almost every

customer will be evaluated, as these sensors and the respective signals are part of

almost all produced vehicles. Further information about the development of driving

assistance systems for two-wheeled vehicles can be found at ACEM1and the CMC2.

A motorcycle of the type BMW R1200GS was chosen as the test vehicle, since it is

the most sold motorcycle of BMW, and all required onboard sensors are available, see

Figure 1.4. This motorcycle was prepared with data-logging devices for experimental

tests and validation of the algorithms. Therefore, a CAN logging device from 2D3

was mounted together with a global positioning system (GPS) logging device, which

provides information about the position and the altitude of the vehicle. Figure 1.5

shows the data-logging device and the GPS module. Both devices are sufficient small

to mount them without restrictions on the motorcycle. The data-logging device needs

a power supply and a CAN signal as inputs. The data logging of vehicle internal

signals is standard practice at developing applications for ECUs. It ensures the

recording of all necessary information for a posterior offline evaluation of the vehicle

dynamics and the driven manoeuvres. In the present case, the signals recorded

were used to develop methods for collecting the individual customer usage profiles.

1ACEM, the European Association of Motorcycle Manufacturers, see www.acem.eu.

2CMC, the Connected Motorcycle Consortium, see www.cmc-info.net.

32D, Debus & Diebold Meßsysteme GmbH, see www.2d-datarecording.com.

1 Introduction

Figure 1.4 – Test motorcycle of type BMW R1200GS c

BMW AG.

Data-logging

device

GPS-module

Figure 1.5 – 2D3data-logging device and GPS-module.

This is possible, because the methods presented do not interact with the vehicle, or

more precisely, they are not designed to be feedback control loops. By contrast, the

development of applications that do interact with the vehicle, specific hardware-in-

the-loop (HIL) test beds would be required, but this is not the scope of the present

research. An overview of the recorded signals can be found in Appendix A.

After the onboard signals had been logged, they were imported in a Simulink R



environment. The discrete models use the same time step size as the vehicle’s onboard

system, which is set to ts= 0.01 s. Consequently, the offline test environment can

simulate the vehicle and its signals as they appeared during the test manoeuvres.

This enables the development of algorithms that in principle work in real time, and

1.4 Vehicle dynamics

can be implemented in new vehicles. However, it is not part of the present study

to evaluate specific hardware requirements for the implementation of the developed

methods. In general, this is merely a cost factor of memory and CPU capacity.

1.4 Vehicle dynamics

Vehicle dynamics is an essential discipline in engineering and it plays a fundamental

role at the development of motor vehicles. According to Schramm et al. [9], vehicle

dynamics deals “with the motional actions necessary for moving road vehicles and

their resulting forces under consideration of the natural laws”. It describes the re-

actions of a vehicle due to the actions of the driver and the interaction with the

road surface and the driving unit [10]. It is common to divide the problem of vehicle

dynamics into three subdomains: longitudinal, vertical, and lateral dynamics [11].

Longitudinal vehicle dynamics comprises:

•Acceleration and braking events

•Driving resistance forces due to air resistance, slope, friction and inertia

•Dynamic wheel loading due to longitudinal acceleration

•Transmission of traction and brake moments

Vertical vehicle dynamics is characterised by:

•Vertical oscillation due to the road excitation

•Assessment of comfort and safety due to vertical oscillation

•Dynamic wheel loading due to pitch and roll effects

Lateral vehicle dynamics evaluates the dynamic behaviour of the vehicle during

cornering and comprises:

•Lateral acceleration of the vehicle mass

•Yaw effects

•Dynamic wheel loading during cornering

The forces acting between the road surface and the tyre are of major importance

in vehicle dynamics, since all required loads for vehicle dynamics are transmitted

by this contact patch. The investigation of the wheel forces is thus often the first

1 Introduction

quantitative load assumption in terms of load analysis. A method for calculating the

wheel forces with the onboard signals is shown in the first publication [1]. Various

techniques for the modelling and simulation of vehicle dynamics exist, while they

differ in complexity and purpose. For example, simple vibration models evaluate

the comfort and safety of the vehicle due to the road excitation. Thus, they aim to

investigate vertical vehicle dynamics.

The most commonly-employed model is the Quarter of Vehicle (QoV) with two

degrees of freedom. It is made up of a sprung mass, which represents the vehicle

mass, and an unsprung mass, which represents the wheel mass. The QoV model is

often the very beginning of a deep model understanding in vertical vehicle dynamics.

It exists in a variety of modifications and is suitable for simple comfort analyses

and the evaluation and optimisation of the suspension characteristics. The extended

Half of Vehicle (HoV) model includes a second wheel on the same track and is

designed for investigating pitch effects and coupled oscillation phenomena. In the

case of a motorcycle, this model is already considered as a full-vehicle model, since

a motorcycle is a single-track vehicle. The major effects of vertical vehicle dynamics

can be investigated with such models, as shown in the second publication [2]. Further

information about the detailed modelling of two-wheeled vehicles can be found in

Cossalter [12]. More complex models are common in multi-body simulations, which

are composed of various rigid or elastic bodies and connection elements; for example,

joints, springs and dampers. These models enable a detailed investigation of all

subdomains of vehicle dynamics and the interaction between the single components.

Mathematical models and the numerical treatment can be found - for example -

in Schramm et al. [9]. Often, the first step in modelling vehicle dynamics is the

formulation of the equations of motion:

M¨

x(t) + C˙

x(t) + Kx(t) = F(t).(1.1)

The equations of motion can be obtained from free body diagrams, based on

Newton’s second law of motion. Mis defined as the mass matrix, Cis the damping

matrix, Kis the stiffness matrix, Fis the force vector of external forces, and xis

the position vector comprising the degrees of freedom. Once the equations of motion

are derived, they can be transformed to first-order ordinary differential equations

(ODEs). This makes a numerical treatment more applicable. In the case of linear

1.5 A brief literature survey

time-invariant system matrices, the linear state-space representation can be derived:

q(t) = Aq(t) + Bu(t),(1.2)

y(t) = Cq(t) + Du(t).(1.3)

Such a system is said to be linear and time-invariant, or LTI for short. The state

variables are gathered in the state vector q, the control variables in the control vector

u, and the measured signals in the measurement vector y. The constant matrices

A,B,C, and D, are called the system matrix, control matrix, measurement matrix

and direct matrix. This system representation is often used, because even complex

models can be described by these two equations. The use of state-space models is

presented in the first [1] and second publication [2]. Another most common and

useful method of representing an LTI system, is by its transfer function H(s), which

relates the output Y(s)to the input U(s)in the frequency domain. Therefore, the

Laplace transformation of the system equations needs to be derived:

H(s) = Y(s)

U(s)=L{y(t)}

L{u(t)}.(1.4)

The application of transfer functions is often used in signal processing and con-

trol theory and was employed by the second publication [2] for mapping the road

excitation (input) to the response of the vehicle (output). The equations of motion,

state-space models and the transfer functions of the frequently-used QoV and HoV

models are shown in Appendix B.

1.5 A brief literature survey

The present publication aims to collect customer usage profiles in terms of durability,

rather than monitoring the vehicle conditions. However, the methods presented are

similar to the model-based condition monitoring systems and they follow the same

approach. Condition-based monitoring (CBM) aims to observe a system’s condition

to provide prognosis and diagnosis of component degradation, and the detection of

in-service failures. Charles et al. [13] annotated that “Condition monitoring uses

some level of knowledge of the system of interest. This may be in the form of a

1 Introduction

model, expert system, experience, learnt behaviour, etc.”. In general, two different

techniques have been established: signal-based condition monitoring, and model-

based condition monitoring.

Signal-based condition monitoring is common at the observation of continuously

running machines and engines; for example, wind turbines, railway vehicles and

stationary turbines. It is characterised by the use of signal-processing techniques in

the time and frequency domain. Therefore, several sensors are mounted in the near

of critical parts; for example, rotor shafts and bearings. A comparison between the

current signals and the measurements of an accurate system, indicates malfunction

and machine deterioration. It is also possible to compare the current state with

simulated values of the proper system. The detection of malfunction is often realised

using threshold detection, trend analysis and spectral analysis. An example of signal-

based condition monitoring is the work of Mei and Ding [14]. They developed a

fault detection system of rail vehicle suspensions based on the cross-correlation of

acceleration signals. An example of vibrational analysis for the detection of faults in

engine bearings can be found in Tandon et al. [15]. Aliustaoglu et al. [16] developed

a tool wear condition monitoring using a sensor fusion model based on fuzzy logic.

Model-based condition monitoring is characterised by the comparison of the cur-

rent machine state to a simulated ideal system behaviour, based on a system model.

In general, model-based approaches are more complex, since they require a detailed

knowledge about the system, which is often dynamic, complex and non-linear. An

example of model-based systems can be found in the PhD thesis from Schlechtingen

[17]. He uses an ANFIS model to analyse the behaviour of a wind power plant.

Furthermore, Liu et al. [18] developed a recursive least square algorithm (RLS) in

combination with a Kalman filter and machine learning (ML) methods to detect a

vertical suspension fault on railway vehicles. Li et al. [19] utilised a linear model of a

railway vehicle to estimate suspension parameters for condition monitoring. An ex-

ample of an automotive application is the work of Börner et al. [20]. They discussed

the comparison of an ideal linear damping constant with the measured suspension

travel to detect deterioration of the suspension system. Various examples of CBM

exist and in times of industry 4.0 and data-driven decisions, many others will follow.

The “’Applied Condition Monitoring’ series of Haddar et al. [21] publishes the latest

applications of condition monitoring.

1.6 Overview of published papers

The methods presented within this PhD thesis can be classified as model-based

collection of customer usage profiles. Several of these systems have already been

published in vehicle engineering. Rupp et al. [22] used linear transfer functions

derived from previous measurements for estimating stresses on components with

acceleration signals as inputs. Müller [23] simulated operative stresses on critical

parts of commercial vehicles within a multi-body system model in real time. She used

measured acceleration signals as input for the simulation and counted the resultant

stresses online with the help of the rainflow counting method. In contrast to Müller,

Matz [7] directly calculated the wheel forces of passenger cars and then translated

these forces acting on components with kinematic transfer functions. As well as the

present publication, he also used onboard signals as inputs. The first publication [1]

of this cumulative PhD thesis transfers his method to the application on two-wheeled

vehicles.

1.6 Overview of published papers

Customer loads of two-wheeled vehicles

The first publication [1] presents methods for estimating the vehicle loading and the

current road slope to calculate the occurring wheel forces. Information about the

vehicle loading is especially important for two-wheeled vehicles, because the vehicle’s

empty weight is low in comparison to the additional loading comprising the driver,

passenger and luggage weight. Finally, the first publication shows that the wheel

forces can be calculated merely with the signals of the onboard sensors. A validation

of the simulated wheel forces was achieved by comparison with measured wheel forces

using wheel-load transducers. The knowledge about wheel forces is an improvement

for the product development, because they can be used do derive design loads and

verification loads for test rigs.

Road classification for two-wheeled vehicles

The second publication [2] aims to classify the driven roads in terms of curviness,

hilliness and road roughness. The study shows the vehicle-independent collection

of road properties using the onboard sensors. First, the driven curves are counted

and classified on a scale of curviness. Second, the road slope is classified in terms of

1 Introduction

hilliness. Third, a modular road roughness estimator is presented, which utilises the

vehicle’s transfer functions to estimate the current road roughness in terms of the ISO

8608 [24] classification. Information about the road characteristics helps defining test

and endurance tracks for product verification. This means that verification tracks

and tracks for survey samplings can be optimised for the actual customer behaviour.

For example, the Virtual Measurement Campaign (VMC) project from Speckert et

al. [25] aims at the planning of measurement campaigns. In addition, knowledge

about the road characteristics enables a virtual load acquisition, where customer

loads are simulated on virtual test tracks.

Impact detection using a machine learning approach and experimental

road roughness classification

The third publication [3] presents the experimental validation of the modular road

roughness estimator. Therefore, a measurement campaign was carried out, where

previously measured road surfaces were ridden to experimentally validate the pro-

posed method. The proposed method of road roughness classification was successfully

validated even on rough test tracks. In addition, the study presents a ML approach

to detect and classify impacts in terms of mild and severe special events. Knowledge

about the number and intensity of occurred special events during the product’s life

helps deriving design loads and enables a virtual load acquisition.

2 Customer loads of two-wheeled

vehicles

Customer loads of two-wheeled vehicles

Christian Gorgesa, Kemal Öztürka, Robert Liebichb

aBMW AG, Munich, Germany;bChair of Engineering Design and Product

Reliability, Berlin Institute of Technology, Berlin, Germany

(This is an Accepted Manuscript of an article published by Taylor & Francis in Vehicle System Dynamics

on 13/06/2017, available online: http: // www. tandfonline. com/ 10. 1080/ 00423114. 2017. 1335874 .)

Customer usage profiles are the most unknown influences in vehicle design targets and

they play an important role in durability analysis. This publication presents a customer

load acquisition system for two-wheeled vehicles that utilises the vehicle’s onboard signals.

A road slope estimator was developed to reveal the unknown slope resistance force with the

help of a linear Kalman filter. Furthermore, an automated mass estimator was developed

to consider the correct vehicle loading. The mass estimation is performed by an extended

Kalman filter. Finally, a model-based wheel force calculation was derived, which is based on

the superposition of forces calculated from measured onboard signals. The calculated wheel

forces were validated by measurements with wheel–load transducers through the comparison

of rainflow matrices. The calculated wheel forces correspond with the measured wheel forces

in terms of both quality and quantity. The proposed methods can be used to gather field

data for improved vehicle design loads.

Keywords: Customer loads, motorcycle dynamics, road slope estimation, mass estimation,

load acquisition, Kalman filter

2.1 Introduction

Two-wheeled vehicles are characterised by diverse application possibilities, given that

motorcycles have changed from means of transportation to general purpose vehicles,

at least in the modern world. Scooters and small-sized motorcycles remain the first

choice for personal mobility in urban traffic, whereas more specific vehicle segments

2 Customer loads of two-wheeled vehicles

such as Enduro, Sport, Cross, Tour, Roadster and Cruiser exist for individual pur-

poses. Since mass reduction and lightweight design play an important role in vehicle

engineering, the components are optimised to their individual design targets. Con-

sequently, every product segment needs its own requirements regarding durability

and operating strength. Furthermore, different markets with varying regional de-

mands influence product design [1]. The most unknown influences to determine

vehicle design targets are customer usage profiles. Detailed knowledge of customer

usage profiles improves design loads and ultimately the vehicle development process.

Customer usage profiles describe the unknown distribution comprising wheel forces,

vehicle loading, road profile characteristics, engine loads, brake events, and special

events. In this paper, customer loads are defined as wheel forces and vehicle load-

ing. Knowledge about the vehicle loading is particularly important for two-wheeled

vehicles and has a strong impact on the customer loads, given that the vehicle’s

empty weight is low in comparison to the additional loading comprising the driver,

passenger, and luggage weight.

At present, a common method to obtain field data is survey sampling, where a

certain number of test vehicles are equipped with additional measurement devices

to perform a measurement campaign with selected or randomly chosen customers.

However, survey sampling cannot reveal the complete probability distribution of

customer loads. Moreover, it is expensive due to the additional equipment costs.

Figure 2.1 illustrates the broad probability density function of customer loads in

comparison to the structural strength of the components and the narrow distribu-

tion of survey samplings. To reveal the entire distribution of customer loads, every

customer would have to be evaluated. Johannesson and Speckert [1] highlight that

there are two scales when discussing customer loads: on the small scale the profile of

a specific customer needs to be evaluated; while on the large scale, the problem is to

identify the severity of a population of customers. The final vehicle design loads in-

volve combining survey sampling, measurements on test tracks, and experience from

previous product designs.

The number of onboard sensors rises due to the increased functions of motorcycles,

such as an anti-lock braking system (ABS), dynamic traction control (DTC), or curve

assistant. Hence, the main approach of this study is to gather information from

every customer with the help of these onboard signals. Onboard signals are defined

2.1 Introduction

Customer loads

Structural strength

Survey sampling

Severity

Probability density

Figure 2.1 – Distribution of customer loads, survey sampling, and structural strength.

as signals that can be accessed by the vehicle’s Controller Area Network (CAN)

bus and are part of every production vehicle. This type of field data collection can

be classified as a model-based online monitoring system with integrated counting

of durability-related values. Online implies that the customer loads are estimated

while the vehicle is in use, meaning that there is no long-term logging of the signals.

Several of these systems have already been published in vehicle engineering. Müller

[2] simulated operative stresses on critical parts of commercial vehicles within a

multi-body system model in real time. She has used measured acceleration data as

input for the simulation and counted the resultant stresses online with the help of

the rainflow counting method. In contrast to Müller, Matz [3] directly calculated the

wheel forces of passenger cars and then translated these forces acting on components

with kinematic transfer functions. His approach also has used onboard signals as

inputs. Karlsson [4] has presented different methods to model loads for customer

usage profiles with the help of road classification.

In contrast to specific application-driven publications, the main contribution of

this study is to develop a holistic approach to collect customer loads with onboard

signals of two-wheeled vehicles. There is neither the ambition to realise a real-time

control system nor to intervene vehicle dynamics. Therefore, no specific hardware

requirements will be discussed. This paper is organised as follows. Section 2.2

describes the reference motorcycle with measurement equipment and analyses previ-

ously measured wheel forces which serve as reference values. Section 2.3 derives the

algorithms for road slope estimation, mass estimation, and the calculation of wheel

2 Customer loads of two-wheeled vehicles







Figure 2.2 – Test motorcycle with the global reference coordinate system c

BMW AG.

forces. Section 2.4 discusses and compares the results from simulation to their refer-

ence measurements to validate the method. Finally, Section 2.5 provides a summary

and conclusion about the developed methods.

2.2 Experimental set-up and data analysis

A motorcycle with data-logging devices was prepared for experimental tests and val-

idation of the algorithms. Onboard signals were logged during pre-defined tracks

for offline simulation of vehicle dynamics. Figure 2.2 shows the test and reference

motorcycle (BMW R1200GS) together with the global reference coordinate system.

The reference frame will not rotate around the roll axes during banking of the mo-

torcycle, which means that it is aligned with the road plane. When the motorcycle

is upright, it coincides with the vehicle’s coordinate system. The test motorcycle

has the following onboard signals available, which are required for the developed

algorithms:

•5 DOF Inertial Measurement Unit (IMU) to measure

–Acceleration in X,Y, and Zin the vehicle coordinate system.

–Angular velocity around X(roll) and Z(yaw) in the vehicle coordinate

system.

•wheel velocities,

•brake pressures,

2.2 Experimental set-up and data analysis

Frequency

Force

I. Domain II. Domain

Figure 2.3 – Amplitude spectrum of vertical wheel forces.

•spring deflections,

•model-based signals (e.g. engine torque and roll angle).

These signals were logged through the CAN bus. Additionally, the vehicle was

equipped with a Global Positioning System (GPS) logging device, which provides

information about the position and the altitude for later validation.

Wheel forces were measured with wheel–load transducers [5] during previous meas-

urement campaigns. These forces were used for data analysis and validation of the

wheel force calculation. Figure 2.3 shows a generic amplitude spectrum of vertical

wheel forces. The amplitude spectrum can be divided into two domains: the first

domain contains driver-induced forces and low-frequency path excitation from road

undulations; and the second domain contains track-induced forces from stochastic

road excitation. These two domains are separated by the crossover frequency fc. It

is well known that these two domains belong to the different modes of the motorcycle

[6]. The first domain contains bounce and pitch mode of the sprung mass, while the

second domain contains wheel hop modes of unsprung masses. Driver-induced forces

are defined as forces directly dedicated to driver manoeuvres, such as accelerating,

braking, and cornering. The main aspect of this analysis is to highlight the cros-

sover frequency fc, which is used to set up the filter frequency for the wheel force

calculation, see Section 2.3.4.

The logged signals from the previous measurement campaigns were used in a

Simulink R

model to simulate vehicle dynamics and validate the developed algorithms.

2 Customer loads of two-wheeled vehicles

The discrete model uses the same time step size as the vehicle’s onboard time step

size. This in principle enables an online application of the developed algorithms.

It is not part of the present study to evaluate specific hardware requirements for

implementation of the methods into existing or new production vehicles.

2.3 Methods

The wheel force calculation requires all resistance forces that act on the motorcycle.

Therefore, two unknown resistance forces had to be determined first: the slope res-

istance force and the inertial force. A road slope estimator based on a linear Kalman

filter (KF) was developed, which estimates the current road slope angle. The total

vehicle mass comprises the motorcycle and the loading, including the driver, pas-

senger, and luggage weight. The estimation of the total vehicle mass was developed

with the help of an extended Kalman filter. The mass estimator requires the road

slope angle and the traction force as input signals. A stiff driveline model was used

to calculate the traction force from the internal engine torque. Once the vehicle

mass and the road slope had been determined, the wheel force calculation was feas-

ible. The wheel forces were calculated by a superposition of forces calculated from

measured motions. Subsequently, the wheel forces were counted with the rainflow

method, which reduces the memory requirements and enables a classification of the

customer loads.

2.3.1 Road slope estimator

Estimation of the road slope is essential for the wheel force calculation, since road

slope can cause a major driving resistance force. Different methods for the estimation

of road slope have already been published. Boniolo et al. [7] utilised a 6-DOF IMU

to describe the state of the motorcycle with Euler angles. Moreover, they used

an extended KF to estimate the vehicle states. Since a 6-DOF IMU is not yet

part of the onboard signals, this method will not be adopted for the present work.

Vahidi et al. [8] estimated the road slope together with the mass of a heavy-duty

vehicle using a recursive least-square estimation with forgetting factor. Lingman and

Schmidtbauer [9] reported another example for this type of slope estimation, while

they used an extended KF. The aforementioned methods use driving resistance forces

2.3 Methods

𝑣𝑎𝑥

𝑔

Figure 2.4 – Road slope estimation physics.

and longitudinal dynamics for the estimation problem. This paper uses a similar

approach described by Corno et al. [10], because it is more suitable for the application

to motorcycles. The idea is that the gravitational acceleration is measured by the

IMU in longitudinal direction while riding up- or downhill, see Figure 2.4. Assuming

a rigid motorcycle, the estimated pitch angle is the road slope angle α. In contrast to

the other methods, this approach makes the estimation of the road slope independent

from the mass estimation. The road slope angle can be calculated directly with the

help of the inertial acceleration axand the vehicle acceleration ˙v, which is derived

from the wheel velocities. Hence, the road slope angle αis given by

α= arcsin ax−˙v

g.(2.1)

Since the input signals are affected by noise from both measurement and differen-

tiation, a linear KF [11] was implemented to estimate the road slope. Many applica-

tions have been realised with Kalman filtering, particularly in navigation problems,

as well as vehicle dynamics. Maybeck [12] describes the KF as a recursive data-

processing algorithm that minimises the error statistically [13]. To estimate the

road slope with the linear KF, the discrete linear difference equation needs to be

derived. Therefore, the problem was formulated in state-space representation. The

state vector xs∈Rencomprises the velocity vand Φ, as shown in Equation (2.2).

xs=(v

Φ),Φ = sin α. (2.2)

2 Customer loads of two-wheeled vehicles

The subscript ‘s’ indicates that the state vector is formulated to estimate the slope.

The substitution of Φ = sin αleads to a linear formulation of the system, which is

essential to apply the linear KF. The measurement vector zs∈Remis defined by the

measured velocity v, which can be derived from the rotational speed of the wheels

and is part of the onboard signals. Transformation of Equation (2.1) yields the

state-space representation of the problem, see Equation (2.3).

xs=(ax−gΦ

0),zs=v. (2.3)

To apply the linear KF, the explicit discrete time-invariant formulation of the prob-

lem must be derived. This is achieved by using the explicit Euler forward integration,

whereby sis the time step size:

xs|k=(vk

Φk)=(vk−1+s[ax|k−1−gΦk−1] + qs1|k−1

Φk−1+qs2|k−1).(2.4)

Process and measurement noise is represented by qsand rs, respectively. They are

assumed to be independent and uncorrelated with a normal white noise probability

distribution [13]:

p(qs)∼N(0,Qs),(2.5)

p(rs)∼N(0,Rs).(2.6)

Qsis the process noise covariance matrix and Rsis the measurement noise cov-

ariance matrix. Equation (2.4) is reformulated into the linear stochastic difference

Equation (2.7) with measurement Equation (2.8).

2.3 Methods

xs|k=[1−sg

0 1 ]

  

(vk−1

Φk−1)

  

xs|k−1

+[s



ax|k−1



us|k−1

+[1 0

0 1](qs1|k−1

qs2|k−1)

  

qs|k−1

,(2.7)

zs|k=[1 0]

  

(vk

Φk)



xs|k

+rs|k.(2.8)

The n×nlinear system matrix Asrelates the state xsat the previous time step

k−1to the state at the current time step k. The linear n×linput matrix Bsrelates

the control input vector us∈Relto the state xs, while the inertial acceleration ax

is defined as the control input. The linear m×nmeasurement matrix Hsrelates the

state xsto the measurement zs.

In comparison to the direct computation, see Equation (2.1), this formulation does

not require a differentiation of the velocity vto obtain the road slope. The meas-

urement noise covariance matrix was estimated from previous measurements of the

velocity vand was thus set to Rs= 0.01. The system covariance matrix Qswas

tuned empirically, which is a standard practice in KF application. In contrast to

the rigid motorcycle model, a real motorcycle has a degree of freedom around the

Y-axis (pitch). Therefore, an on/off logic was implemented to restrict the pitch in-

fluence on the road slope estimation. Since the test motorcycle has no pitch signal

available, the longitudinal acceleration was utilised to detect acceleration and brake

events. Thus, the road slope estimation pauses when a certain value of longitudinal

acceleration is exceeded. Furthermore, the model assumptions are only valid for mo-

torcycles without steering action. The algorithm thus also pauses when the measured

angular yaw rate exceeds a certain value. To sum up, the following restrictions were

formulated for the road slope estimator:

•Absolute value of the angular yaw rate is lower than a given threshold.

•Absolute value of the longitudinal acceleration is lower than a given threshold.

When at least one condition is violated, the algorithm holds the road slope estim-

ation until the conditions are true again. In the meantime, the last valid value of

the road slope is delivered for the mass estimator and the wheel force calculation.

2 Customer loads of two-wheeled vehicles

𝑇t

Engine & Clutch

Transmission

Rear wheel Cardan drive &

Differential

𝐼rr

𝜔rr

𝐹T

𝑇

𝑟

𝑇

𝑇f

𝑇

𝐼e, 𝜔e

𝐼f, 𝜔f, 𝑖f𝐼t, 𝑖t

Figure 2.5 – Model of the driveline dynamics.

Please note that these limitations are only valid for the road slope estimator. They

are independent of the mass estimator and the wheel force calculation.

2.3.2 Driveline model

The mass estimation and the wheel force calculation require the traction force FT.

Thus, the traction force was derived from the rotational equations of motion from

both the driveline and the rear-wheel dynamics, as illustrated in Figure 2.5. The

engine torque Teis provided by the engine electronic control unit (ECU). It is cal-

culated by the engine speed and the throttle position. The engine torque already

considers the necessary amount of slip to accelerate the vehicle, because the engine

torque is adapted to the current road conditions due to the DTC. The driveline is

assumed to be stiff, whereby driveline oscillation and torsional effects are neglected.

The transmission torque Ttcan be calculated by subtracting the rotational inertia

of the engine components Iefrom the engine torque Te, as shown in Equation (2.9).

Tt=Te−Ie˙ωe,(2.9)

Tf= (Tt−It˙ωe)it,(2.10)

Tw= (Tf−If˙ωf)if.(2.11)

The final drive torque Tfcan be calculated by subtracting the inertial losses of the

transmission parts Itfrom the transmission torque Tt. Furthermore, it is amplified by

the gear ratio it, see Equation (2.10). The wheel torque Twis derived from the final

2.3 Methods

drive torque Tfreduced by the inertial losses of the cardan drive and the differential

If. Furthermore, it is amplified by the fixed final drive ratio if, see Equation (2.11).

The traction force FTcan be derived from the rotational equation of motion of the

rear wheel, see Equation (2.12), where rrr is the dynamic rolling radius of the rear

wheel.

Irr ˙ωrr =Tw−FTrrr,(2.12)

˙ωrr =˙v

rrr

, u =itf

rrr

, itf =itif, µeff =efficiency parameter.(2.13)

Moreover, it is assumed that the rolling condition is valid. The gear ratio it,

the final drive ratio if, and the dynamic rear-wheel radius rrr are condensed into

the coefficient u. The rotational acceleration of the rear wheel ˙ωrr can be derived

from the rear-wheel velocity v. Mechanical losses of the driveline are reduced to

the efficiency parameter µeff, see Equation (2.13). Finally, the traction force FTcan

be partitioned into two parts: the steady-state traction force and the losses of the

traction force due to the driveline inertia, as shown in Equation (2.14).

FT=Tw−Irr ˙ωrr

rrr

=Teuµeff

  

Steady–state

−(Ie+It+If

+Irr

)u2

  

Rotational mass

˙v. (2.14)

These losses are reduced to an equivalent rotational mass, which is multiplied by

the rear-wheel acceleration ˙vto obtain a force component. The validation of the

traction force FTis made within the wheel force validation in Section 2.3.4. In the

case of acceleration, the longitudinal rear-wheel force is the traction force.

2.3.3 Mass estimator

As mentioned in the introduction, estimation of the vehicle mass is particularly im-

portant for motorcycles. The empty weight of two-wheeled vehicles is less compared

to passenger cars and thus the influence of the vehicle loading on customer loads

increases. Furthermore, the vehicle mass is essential for the wheel force calculation.

Several algorithms have already been published. Rozyn and Zhang [14] measured

2 Customer loads of two-wheeled vehicles

𝑣

𝑚

Figure 2.6 – External forces acting on the motorcycle.

the sprung mass response to estimate the inertial parameters of the vehicle, which

requires detailed knowledge about the suspension stiffness and the damping char-

acteristics. Lingman and Schmidtbauer [9] used longitudinal vehicle dynamics and

a KF that estimates both the road slope and the vehicle mass. Fathy et al. [15]

developed a recursive least-square model to estimate the vehicle mass. This study

uses an approach based on resistance forces and longitudinal dynamics for mass

estimation, as Ritzen et al. [16–18] proposed for heavy-duty vehicles. Figure 2.6

illustrates the longitudinal dynamics with external forces acting on the motorcycle.

The dynamic equation of motion can be solved for the vehicle mass m, see Equa-

tions (2.15)–(2.17).

m˙v=FT−FD−FS−FR,(2.15)

FD=1

2ρcxAv2=κv2, FS=mg sin α, FR=mgfrcos α, (2.16)

m=FT−κv2

˙v+g(sin α+frcos α).(2.17)

FDis the aerodynamic drag force and FTis the traction force acting on the rear

wheel, as derived in Equation (2.14). The aerodynamic coefficients were obtained

from measurements in a wind tunnel of BMW. They are substituted by the single

constant κ. The slope resistance force FSdepends on the road slope angle α, which is

estimated with the road slope estimator. The rolling resistance force FRis calculated

with a constant rolling resistance coefficient fr.

2.3 Methods

Since the input signals are affected by measurement noise as well as model uncer-

tainties from the road slope estimator and the driveline model, a direct calculation

of the vehicle mass is infeasible. For this reason, the estimation problem was formu-

lated with the help of an extended Kalman filter (EKF). The application of a linear

KF is infeasible because the process equations are nonlinear. The EKF can handle

nonlinear stochastic difference equations, such as Equations (2.18)–(2.19).

xk=f(xk−1,uk−1,qk−1),(2.18)

zk=h(xk,rk).(2.19)

Welch and Bishop [13] highlight that the EKF linearises around the current mean

and covariance and that the most interesting and successful applications have been

solved with EKFs. Jacobian matrices A,W,H, and Nare required for the linear-

isation. The basic operations for the EKF are the same as for the linear KF. To

define the problem in state-space representation, Equation (2.17) can be solved for

the vehicle acceleration ˙v, see Equation (2.20).

˙v=(FT−κv2)

m−g(sin α+frcos α).(2.20)

The state vector xmis defined by the velocity vand the reciprocal mass Θ, as

shown in Equation (2.21). The subscript ‘m’ indicates that the state vector is for-

mulated to estimate the vehicle mass.

xm=(v

Θ),Θ = 1

m.(2.21)

The substitution of Θ = 1/m leads to a more robust formulation of the estimation

problem. The measurement vector zm∈Remis defined by the measured velocity v.

Finally, the state-space representation is obtained, see Equations (2.22)–(2.24).

2 Customer loads of two-wheeled vehicles

xm=(ΘΓ −gΛ

0)with (2.22)

Γ = FT−κv2,Λ = sin α+frcos α, (2.23)

zm=v. (2.24)

The explicit discrete time-invariant formulation of the problem can be derived with

the use of explicit Euler forward integration, see Equations (2.25)–(2.29), where sis

the time step size.

xm|k=f(xm|k−1,um|k−1,qm|k−1) = (vk−1+s[Θk−1Γk−1−gΛk−1]

Θk−1+qm4|k−1)with

(2.25)

Γk−1=FT|k−1(1 + qm1|k−1)−κ(vk−1+qm2|k−1)2,(2.26)

Λk−1= sin (αk−1+qm3|k−1) + frcos (αk−1+qm3|k−1),(2.27)

um=(FT

α),(2.28)

zm|k=h(xm|k,rm|k) = [1 0]

  

(vk

Θk)

  

xm|k

+rm|k.(2.29)

While the process equations are nonlinear, the measurement equation remains

linear. The process noise qmis modelled as normally distributed white noise. The

noise influence qm1is modelled as a percentage amount of the traction force FT, since

the model uncertainties for the traction force increase with a higher engine torque.

The unknown wind speed is considered with the noise qm2. Additionally, the noises

qm3and qm4are added to the road slope angle αand the reciprocal vehicle mass Θto

consider model uncertainties. The measurement noise rmis added to the measured

velocity v. The control input vector umcomprises the traction force FTand the road

slope angle α. The Jacobian matrices Am,Wm,Hm, and Nmneed to be derived to

apply the EKF, see Equations (2.30)–(2.34).

2.3 Methods

Am[i,j]=∂f[i]

∂xm[j]

(xm|k−1,um|k−1,0),Wm[i,j]=∂f[i]

∂qm[j]

(xm|k−1,um|k−1,0),(2.30)

Hm[i,j]=∂h[i]

∂xm[j]

(xm|k,0),Nm[i,j]=∂h[i]

∂rm[j]

(xm|k,0),(2.31)

Am=[1−2κsvk−1Θk−1sΓk−1

0 1 ],(2.32)

Wm=[sΘk−1FT|k−1−2sκΘk−1vk−1−sg(cos αk−1−frsin αk−1) 0

0 0 0 1],(2.33)

Hm=[1 0],Nm= 1.(2.34)

The model assumes valid acceleration events that are suitable for the mass es-

timation. Hence, a second on/off logic was formulated for the mass estimator. The

restrictions are designed to be strict to identify evaluable acceleration events, which

makes the mass estimation more robust. For this reason, the algorithm pauses dur-

ing cornering of the motorcycle. Additionally, the engine torque model is only valid

for steady-state conditions of the engine. For the mass estimation, this means that

the traction force FTmust not change more than a given threshold. To sum up, the

conditions to identify valid acceleration events are as follows:

•Absolute value of the angular yaw rate is lower than a given threshold.

•Traction force is higher than a given threshold.

•Derivative of the traction force is lower than a given threshold.

The thresholds were evaluated empirically and as a consequence, they can differ for

other motorcycles and underlying models. If a condition is violated, the algorithm

holds the mass estimation until all conditions are true again. Despite these strict

conditions, the results can still vary for different acceleration events since several

other influences are not considered. Therefore, another linear filter was implemented

to calculate the running mean of the estimated mass. When the conditions are

violated or no converged mass estimate is available, the last valid mass estimate is

forwarded to the wheel force calculation. The start value for the algorithm is defined

2 Customer loads of two-wheeled vehicles

as a standard mass with a normal rider weight. Please note that these restrictions

are valid for the mass estimator only. Nevertheless, there is a continuous output of

the mass estimate for the subsequent wheel force calculation.

2.3.4 Wheel force calculation

A model-based wheel force calculation was developed, whereby the individual mech-

anical effects are calculated for the particular rigid bodies. This modular approach

makes a subsequent composition of the wheel forces possible. The method is based

on the following assumptions: A motorcycle is modelled with three rigid bodies for

sprung mass, rear unsprung mass, and front unsprung mass, as Cossalter described

in [19]. These models exist in a variety of publications, see e.g. [6,10,19,20]. They

differ in complexity and degrees of freedom for their individual application. The

sprung mass comprises the frame, the engine, and the rider. Additionally, parts of

the front and rear suspension system are counted to the sprung mass. The sprung

mass is lumped in the centre of gravity (COG). The rear unsprung mass comprises

the rear wheel, the rear brake, and parts of the rear suspension. The front unsprung

mass comprises the front wheel, the front brake, and parts of the front suspension.

Figure 2.7 shows the motorcycle model together with the three rigid bodies and the

global reference frame. Furthermore, the main geometric dimensions are illustrated:

wheelbase p, height of the COG hcog|0and perpendicular distance lcog of the COG

from the rear wheel Z-axis. The degrees of freedom of the rigid bodies were formu-

lated according to their equivalent onboard sensor. This means that every degree of

freedom is represented by a signal from the onboard sensors. The sprung mass has

the following four degrees of freedom:

•Displacement in Xand Z.

•Roll motion around tyre contact patch line.

•Yaw motion around vehicles Z-axis.

The unsprung masses have one degree of freedom in vertical direction, while the

wheels can rotate around their axes. Steering and rotation of the handlebars are

neglected, because they are not yet part of the onboard signals. This means that

every roll motion acts on the three bodies with the same amount. The illustrated

model considers no springs or dampers, because they are not required for the modular

2.3 Methods

𝑍𝑍

𝑍

𝑋

5ROO

<DZ

𝑝

ℎcog|0

𝑙cog

&2*

5HDUZKHHO )URQWZKHHO

𝜔rr 𝜔ft

𝑍

𝑋

𝑌

Figure 2.7 – Motorcycle model with three rigid bodies.

approach. The wheel forces are calculated from the measured responses of the par-

ticular rigid bodies, which means displacements, velocities, and accelerations. These

kinds of problems are defined as inverse dynamic problems. Table 2.1 shows the

required onboard signals for the wheel force calculation.

As introduced in Section 2.2, the crossover frequency fcis used to filter the meas-

ured accelerations to distinguish between forces acting on the sprung mass and forces

generated by the unsprung mass excitation. This means that accelerations measured

from the sprung mass are low-pass filtered, whereas accelerations measured from the

unsprung masses are high-pass filtered with the crossover frequency fc. This differ-

entiation guarantees that no excitation from the unsprung masses affects the force

calculation of the sprung mass and vice versa. As the geometric dimensions of the

COG change during banking of the motorcycle, they are defined as functions of the

roll angle φ, see Equations (2.35)–(2.36). The out-of-plane component dcog describes

the lateral distance from COG to the X-Z-plane. The measured acceleration in Zis

corrected for the same reason, see Equation (2.37).

Height of COG: hcog =hcog|0cos φ, (2.35)

Depth of COG: dcog =hcog|0sin φ, (2.36)

Acceleration COG in Z:az=az|raw cos φ. (2.37)

Dynamic equations of motion have been used to calculate the driving resistance

2 Customer loads of two-wheeled vehicles

Table 2.1 – Required onboard signals for the wheel force calculation.

Description Symbol Unit Filter type Source

Acceleration COG ax,az|raw m/s2Low-pass IMU

Angular rates COG ωx,ωy,ωzs−1Low-pass IMU

Velocity vm/s- ABS sensor

Angular wheel rates ωft,ωrr s−1- ABS sensor

Brake pressure pft,prr N/m2- ABS sensor

Spring deflection sft,srr mHigh-pass Suspension sensor

Road slope α◦- Road slope estimator

Mass mkg - Mass estimator

Traction force FTN- Engine & driveline model

Roll angle φ◦- Engine ECU

forces and the inertial forces from the measured onboard signals. For a modular

formulation of the final wheel force calculation, a subset of equations is derived, see

Equations (2.38)–(2.42). The centrifugal force can be derived from the equilibrium

of moments around the X-axis, see Equation (2.43).

Normal force: FN=mazcos α, (2.38)

Slope force: FS=mazsin α, (2.39)

Aerodynamic drag force: FD=1

2ρcxAv2=κv2,(2.40)

Inertial force in X:FI|X=max,(2.41)

Inertial moment around X:MRoll =Ix˙ωx,(2.42)

Centrifugal force: FC=FNtan φ−MRoll

hcog

.(2.43)

The forces from unsprung mass excitation can be derived with the measured spring

deflections, see Equations (2.44)–(2.45).

Front unsprung mass force: Fun|ft =mun|ft¨sft,(2.44)

Rear unsprung mass force: Fun|rr =mun|rr¨srr.(2.45)

The spring deflections are first high-pass filtered with the crossover frequency fc

2.3 Methods

and subsequently differentiated twice to obtain the unsprung mass acceleration ¨s.

The rolling resistance forces depend on the vertical wheel forces and the rolling

resistance coefficient fr, see Equations (2.46)–(2.47).

Rolling resistance force front: FR|ft =FZ|ftfr,(2.46)

Rolling resistance force rear: FR|rr =FZ|rrfr.(2.47)

The brake moments are calculated by multiplying the brake pressures pby a linear

brake coefficient c, which depends on the brake characteristics, e.g. quantity of brake

pistons and friction values, see Equations (2.48)–(2.49).

Brake moment of front wheel: MB|ft =cftpft,(2.48)

Brake moment of rear wheel: MB|rr =crrprr.(2.49)

Equilibrium of forces and moments is applied on the motorcycle model to obtain

the components of external wheel forces in Yand Z. The wheel forces in vertical

direction are calculated as follows:

FZ|ft =FNlcog

p−FShcog

p−FDhcog

p−FI|Xhcog

p+Fun|ft cos φ, (2.50)

FZ|rr =FN(p−lcog)

  

Steady

state

+FShcog

  

Slope

resistance

+FDhcog

  

Aerodynamic

resistance

+FI|Xhcog

  

Inertial

force X

+Fun|rr cos φ

  

Unsprung

masses

.(2.51)

The equations are not summarised to describe the particular components of the

wheel force calculation. The load distribution of the normal force FNaffects both

the front and rear-wheel vertical force depending on the longitudinal location lcog of

the COG and the wheelbase p. The other resistance forces are added to the vertical

wheel forces depending on the height hcog of the COG. The resistance and inertia

forces in negative Xdirection lead to a decrease in the vertical front-wheel force and

vice versa. The forces from unsprung mass excitation are added depending on the

roll angle φ. The centrifugal force FCaffects the lateral force components depending

2 Customer loads of two-wheeled vehicles

on the load distribution. Due to banking of the motorcycle, the resistance forces

act on the lateral force components depending on the out of plane component dcog.

The forces from unsprung mass excitation are added, respectively. Equations (2.52)–

(2.53) show the calculation of the lateral wheel-force components.

FY|ft =FClcog

p−FSdcog

p−FDdcog

p−FI|Xdcog

p+Fun|ft sin φ, (2.52)

FY|rr =FC(p−lcog)

  

Centrifugal

force

+FSdcog

  

Slope

resistance

+FDdcog

  

Aerodynamic

resistance

+FI|Xdcog

  

Inertial

force X

+Fun|rr sin φ

  

Unsprung

masses

.(2.53)

The longitudinal wheel forces are calculated by the rotational equations of motion,

see Equations (2.54)–(2.55). The traction force FTacting on the rear wheel is already

derived in Section 2.3.2 and the brake moments and the rolling resistance forces are

added to complete the rotational equations of motion.

FX|rr =FT−MB|rr

rrr

−FR|rr,(2.54)

FX|ft =−Ift ˙ωft

rft



Wheel

dynamics

−MB|ft

rft



Brake

force

−FR|ft



Rolling

resistance

.(2.55)

2.3.5 Rainflow counting method

Rainflow counting of both external and internal loads is a standard practice in durab-

ility analysis, as described in [1,21]. As it is based on the Masing memory rule, it is

a counting method with a direct relation to the physical background of the material

and fatigue damage assessment. The algorithm extracts hysteresis loops from the

measured or simulated load signals and stores them in a rainflow matrix. There are

different standards for the rainflow counting method, e.g. ASTM Standard [22] and

French ANFOR Standard [23]. In this paper, the four-point algorithm was imple-

mented according to Clormann and Seeger [24] to count the calculated wheel forces.

Once the rainflow matrices are collected, several other representations of hysteresis

loops can be derived, e.g. range-mean matrix or level crossing counting.

2.3 Methods

Mass estimator

CAN-Bus Preprocessor Road slope estimator

Logic

Predictor

Corrector

Logic

Predictor

Corrector

EKF

Filter

Differentiation

Correction

Wheel forces

Rainflow

𝑚

Buffer

Figure 2.8 – Flow chart of the customer load estimation model.

2.3.6 Whole model

Figure 2.8 shows the flow chart of the whole model to estimate the customer loads.

In summary, a preprocessor filters the measured onboard signals and calculates de-

rivatives for the following algorithms. A correction module ensures that no drifts and

offsets occur. It also takes into account the fact that geometric and inertial paramet-

ers change during time, e.g. the location of COG during banking of the motorcycle.

Additionally, the traction force is calculated based on the driveline model. Once the

signals are prepared, the road slope estimator computes the current road slope angle

αwith a linear KF. The road slope is essential for the mass estimation and the wheel

force calculation. The road slope estimator has its own limitations, although it still

provides a slope angle for the following algorithms at any time, as described in Sec-

tion 2.3.1. The mass estimator calculates an estimate for the vehicle mass mbased

on valid acceleration events with the help of an EKF. Once the vehicle mass value

has converged, it is forwarded to the wheel force calculation. In case the converged

estimate of the vehicle mass is unavailable, an initial mass for a full-fuelled motor-

cycle with a normal rider weight is provided, as described in Section 2.3.3. Since the

road slope estimator causes a delay, a signal buffer was implemented to synchronise

all of the input values for the wheel force calculation. As the algorithm itself has

2 Customer loads of two-wheeled vehicles

no limitations, the wheel force calculation continuously computes forces. The equa-

tions are valid under all driving conditions. Linear finite impulse response filters

(FIR) were implemented. They are designed to have a linear phase delay, which

makes the synchronisation of all required signals feasible. Hamming [25] provides

further information about digital filters. Finally, the four-point algorithm counts the

calculated wheel forces with the rainflow method.

2.4 Validation

2.4.1 Validation of the road slope estimator

Test runs were performed with the reference motorcycle at the proving ground of

BMW to validate the road slope estimator. The proving ground has a track including

artificial hills with pre-defined road slopes, where the algorithm was validated at

different velocities. Figure 2.9 shows the results from two runs together with the

associated road slopes of the proving ground, which are α= 12 %,−16 %,32 % and

−20 %. The road slope estimator works as expected at the proving ground. The

vehicle speed shows only a slight influence on the algorithm. A mountain track ride

was performed for a validation of the algorithm under real conditions with braking

and cornering manoeuvres. GPS devices were mounted on the motorcycle to measure

the absolute altitude for the validation. The mountain track was driven upwards and

downwards to validate both positive and negative slope angles. The estimated road

slope was integrated over the travelled distance to compare the calculated elevation

profile with the GPS altitude, as shown in Figure 2.10. The integration was made

with the start value from the measured altitude to provide the same height at the

start. The results show a close coincidence between the estimated and measured

elevation profile. Since the mountain track is within a forest and thus the GPS

signal can be corrupted, the spikes in the GPS altitude signal are potentially caused

by noise and GPS errors.

2.4 Validation

0 50 100 150 200 250 300 350 400

Distance (m)

-20

-10

Road slope (%)

Road slope estimator 25 km/h

Road slope estimator 45 km/h

Proving ground

Figure 2.9 – Validation of the road slope estimator at the proving ground.

2 4 6 8 10 12 14

Distance (km)

600

700

800

900

Altitude (m)

Uphill Downhill

Road slope estimator

GPS altitude

Figure 2.10 – Validation of the road slope estimator at the mountain track.

2 Customer loads of two-wheeled vehicles

0 5 10 15 20 25 30 35 40

Time (min)

100

200

300

400

500

Mass (kg)

Mass estimator

Solo

Passenger

Confidence interval ±5%

Figure 2.11 – Validation of the mass estimator.

2.4.2 Validation of the mass estimator

A test ride with a change in the vehicle mass was carried out to validate the mass

estimator, based upon the following setup: the rider drives the motorcycle about 20

min, before a passenger was added to the vehicle to analyse the convergence beha-

viour of the algorithm. The results from this test ride are shown in Figure 2.11. The

vehicle mass, the driver mass, and the passenger mass were measured in advance.

These true reference weights are illustrated with solid lines (msolo = 330 kg and

mpassenger = 424 kg). For validation of the convergence behaviour of the algorithm,

the start value was intentionally set up to an incorrect value of m0= 100 kg. The

mass estimation converged after eight min. The quality criterion for the mass estim-

ator is to converge within the confidence interval, which is defined to be ±5 % of the

true value. After 22 min, the passenger got onto the motorcycle, which can be seen

by the rise of the vehicle mass estimation. The mass estimation converged within

a few minutes and strongly depends on the quantity of valid acceleration events.

Several test rides were performed for empirical KF tuning.

2.4 Validation

100

120

Velocity (km/h)

Braking Accelerating

-1

FX|ft (kN)

Wheel force calculation

Measurement

65 70 75 80 85

Time (s)

-1

FX|rr (kN)

Figure 2.12 – Validation of the longitudinal wheel forces.

2.4.3 Validation of the wheel force calculation

For validation of the calculated wheel forces, several test rides with wheel–load trans-

ducers have been accomplished. Different manoeuvres were analysed to evaluate the

particular physical phenomena of the wheel force calculation. As described in Sec-

tion 2.3.4, four different phenomena affect the longitudinal wheel forces: traction

force, brake force, rolling resistance force, and inertial forces. All of these effects

appear together in the test manoeuvre, as can be seen in Figure 2.12.

The comparison shows a manoeuvre with longitudinal wheel forces for both the

front and rear wheel during braking and accelerating. For the braking part, the

rolling resistance force induces a small offset, while the brake force accounts for

the main part of the longitudinal front-wheel force FX|ft. This is overlaid by the

rotational wheel dynamics, which induces a high-frequency oscillation. For the ac-

celeration part, the traction force FTof the rear wheel is the main contributor to

2 Customer loads of two-wheeled vehicles

the longitudinal rear-wheel force FX|rr. When accelerating, gear shifting disrupts

the traction force, because opening of the clutch decouples the engine from the rear

wheel. Again, a high-frequency oscillation from the rotational wheel dynamics is

overlaid. To compare the calculated wheel forces in vertical and lateral direction

with the measurement, they need to be combined to a resultant force FY Z in the Y-

Z-plane. A coordinate transformation with the roll angle ϕis unsuitable in this case.

This is because the measured wheel forces actually act at the wheel hub, whereas the

calculated wheel forces act between the road surface and the tyre. For a validation

of the algorithm, the forces are thus summarised to resultant forces, as shown in Fig-

ure 2.13 for the same manoeuvre. The load transfer induces a rise of the resultant

front-wheel force FY Z|ft during braking. The acceleration manoeuvre causes the con-

trary and leads to a rise of the resultant rear-wheel force FY Z|rr. Besides the effect of

load transfer, the dynamics of the unsprung masses can be seen as a high-frequency

oscillation. Figure 2.14 shows an extract from t= 77 s −80 s in further detail. The

frequency of the calculated wheel force oscillation generally coincides with the meas-

ured wheel force oscillation. However, small deviations of the amplitudes exist. This

is due to the model assumptions and unaccounted effects. The comparison shows

that the wheel force calculation can reproduce the main physical phenomena. It is

common practice to compare the severity of the signals with counting methods, e.g.

the rainflow counting method. In order to achieve a diverse load profile, several test

runs were linked together.

The rainflow counting algorithm first discretises the signal levels of the load in

nequidistant bins and subsequently counts the hysteresis loops and stores them in

the rainflow matrix. Figure 2.15 shows the comparison of the rainflow matrices from

the measured and the calculated resultant rear-wheel forces in the Y-Z-plane. The

number of bins is chosen to n= 100. The axes represent the container number in

which the amplitude cycle starts (From) and ends (To). The colour represents the

quantity of counted cycles on a logarithmic scale. The rainflow matrices look similar

in their form and shape. This indicates that the magnitudes of the load oscillations

are correct. The comparison of the midpoints of the rainflow matrices also indicates

that the mean values coincides. A common practice in durability analysis is to reduce

the effect of severity to one comparable value, the so-called pseudo-damage number

d[1]. The algorithm is based on the Palmgren–Miner rule, which simply sums up

2.4 Validation

100

120

Velocity (km/h)

Braking Accelerating

FY Z|ft (kN)

Detail

65 70 75 80 85

Time (s)

FY Z|rr (kN)

Detail

Wheel force calculation

Measurement

Figure 2.13 – Validation of the resultant wheel forces in the Y-Z-plane.

FY Z |ft (kN)

77 77.5 78 78.5 79 79.5 80

Time (s)

FY Z |rr (kN)

Wheel force calculation

Measurement

Figure 2.14 – Detailed extract of the resultant wheel forces.

2 Customer loads of two-wheeled vehicles

1 20 40 60 80 100

100

From

Measurement

100

1000

10000

Quantity

1 20 40 60 80 100

100

From

Simulation

100

1000

10000

Quantity

Figure 2.15 – Rainflow matrices of the resultant rear-wheel forces in the Y-Z-plane.

Table 2.2 – Pseudo damage ratio.

Wheel force component dmes/dsim

FY Z|ft 0.9995

FY Z|rr 0.9708

FX|ft 1.0006

FX|rr 1.0187

the damage contributions of all cycles, see Equation (2.56).

d=∑

Sβ

i(2.56)

The counted amplitudes Sare exponentiated by β, which is the damage exponent.

The ratios of the pseudo-damage numbers for the different wheel force components

are given in Table 2.2. The ratios of the pseudo-damage numbers show that all of

the simulated wheel forces are within a confidence interval of ±5 % of the measured

wheel forces.

2.5 Summary and Conclusion

The objective of this research was to develop a system that reveals the unknown

customer loads of two-wheeled vehicles. The results show that a customer load

estimation for two-wheeled vehicles with onboard signals is possible. By combining

the methods from several studies, an independent and automated customer load

2.5 Summary and Conclusion

acquisition system was developed. This system comprises three main subsystems:

the road slope estimator, the mass estimator, and the wheel force calculation. To

estimate the unknown road slope, a linear KF was developed with an on/off logic.

The algorithm was tested successfully at a proving ground as well as on a mountain

track. With the knowledge of the road slope, one major resistance force was revealed.

An automated mass detection is essential for the wheel force calculation. For this

reason, an EKF was developed to estimate the vehicle mass. The algorithm is based

on longitudinal vehicle dynamics and takes acceleration events into account, which

were evaluated with an on/off logic. The algorithm was tested during several test

rides and under different loading configurations. The results show that the vehicle

mass estimation is possible within a confidence interval and that sudden changes in

the vehicle mass are considered after a few minutes. At the small scale, the knowledge

of the vehicle mass is important for the wheel force calculation and at the large scale

the distribution gives useful information for vehicle design and reliability targets.

The wheel force calculation was developed based on a superposition of forces de-

rived from the particular physical effects. The degrees of freedom of the three rigid

bodies were chosen so that every relevant onboard signal could be used. All input

signals for the wheel force calculation are synchronised with a signal buffer. The

wheel force calculation computes continuous results and is valid under all operating

conditions. The results show that the model covers the previously measured wheel

forces in both quality and quantity. The integrated rainflow counting algorithm re-

duces the time history of the loads to hysteresis cycles, which makes a comparison

between the customer service loads on the large scale possible. The developed wheel

force calculation can be used to replace classical wheel force measurements with

wheel–load transducers. Extra measurement equipment is no longer necessary and

thus mounting time and material is saved. The presented methods work with respect

to the model restrictions and are valid under normal operating conditions. Hence,

this system is able to reveal operating loads. Infrequent severe events often cause

the most damage to the parts and thus further research is planned to investigate the

identification of special events with onboard signals.

Detailed knowledge of the customer load distribution provides information for

product design targets and can be analysed for different product segments and dif-

ferent regional markets. Customer loads can be used for an incremental evolution of

2 Customer loads of two-wheeled vehicles

the design targets of existing products. Another benefit is the comparison between

product segments and a transformation to new vehicle projects. Additionally, test

rig loads can be derived from the customer load distribution. In general, customer

loads offer a variety of possibilities throughout the automotive business.

Since no time and location stamps are collected, an implementation of the system is

assumed to be uncritical in terms of data protection. The data are collected anonym-

ously without personal relation. The customer distribution is treated statistically to

derive values such as quantiles, which do not correspond to specific customers.

References

[1] P. Johannesson and M. Speckert, eds. Guide to load analysis for durability in

vehicle engineering. Chichester: Wiley, 2013, p. 434.

[2] L. Müller. ‘Mehrkörpermodell-basiertes Online Monitoring der Betriebsbeans-

pruchung am Beispiel eines Nutzfahrzeug-Demonstrators’. PhD thesis. Kaiser-

slautern: TU Kaiserslautern, 2011.

[3] C. Matz. ‘Online Berechnung von Fahrwerkskräften auf Basis von Onboard-

Sensorik’. PhD thesis. Clausthal: TU Clausthal, 2015.

[4] M. Karlsson. ‘Load modelling for fatigue assessment of vehicles – a statist-

ical approach’. PhD thesis. Göteborg, Sweden: Chalmers, Göteborg University,

2007.

[5] M. Kuchler and R. Schrupp. ‘Mehrkomponenten-Motorradmessnabe’. In: Meas-

uring and testing techniques in the vehicle industry. Nuremberg, Germany:

VDI, 2001 5 10–11 2001, pp. 249–278.

[6] M. Tanelli, M. Corno and S. M. Savaresi, eds. Modelling, simulation and control

of two-wheeled vehicles. Chichester: Wiley, 2014, pp. 1–348.

[7] I. Boniolo, S. Corbetta and S. M. Savaresi. ‘Attitude estimation of a motor-

cycle in a Kalman filtering framework’. In: 6th IFAC symposium advances in

automotive control. Munich, Germany: IFAC, 2010 7 12–14 2010, pp. 779–784.

[8] A. Vahidi, A. Stefanopoulou and H. Peng. ‘Recursive least squares with for-

getting for online estimation of vehicle mass and road grade: theory and ex-

periments’. In: Vehicle System Dynamics 43.1 (2005), pp. 31–55.

[9] P. Lingman and B. Schmidtbauer. ‘Road slope and vehicle mass estimation

using Kalman filtering’. In: Vehicle System Dynamics 37.1 (2002), pp. 12–23.

References

[10] M. Corno, P. Spagnol and S. M. Savaresi. ‘Road slope estimation in bicycles

without torque measurements’. In: IFAC Proceedings Volumes. Cape Town,

South Africa: IFAC, 2014 8 24–29 2014, pp. 6295–6300.

[11] R. E. Kalman. ‘A new approach to linear filtering and prediction problems’.

In: J Basic Eng 82.D (1960), pp. 35–45.

[12] P. S. Maybeck. Stochastics models, estimation, and control. 1st ed. New York:

Academic Press, 1979, p. 444.

[13] G. Welch and G. Bishop. An introduction to the Kalman filter. Technical report

TR 95-041. Chapel Hill (NC): Department of Computer Science, University of

North Carolina at Chapel Hill, 2006.

[14] M. Rozyn and N. Zhang. ‘A method for estimation of vehicle inertial paramet-

ers’. In: Vehicle System Dynamics 48.5 (2010), pp. 547–565.

[15] H. K. Fathy, D. Kang and J. L. Stein. ‘Online vehicle mass estimation using

recursive least squares and supervisory data extraction’. In: 2008 American

Control Conference. Seattle (WA): IEEE, 2008 6 11–13 2008, pp. 1842–1848.

[16] E. Ritzen. ‘Adaptive vehicle weight estimation’. MA thesis. Linköping, Sweden:

Department of Electrical Engineering, Linköping University, 1998.

[17] E. J. Holm. ‘Vehicle mass and road grade estimation using Kalman filter’. MA

thesis. Linköping, Sweden: Department of Electrical Engineering, Linköping

University, 2011.

[18] B. Lundin and A. Olsson. ‘Estimation of vehicle mass using an extended Kal-

man filter’. MA thesis. Gothenburg, Sweden: Department of Signals and Sys-

tems, Chalmers University of Technology, 2012.

[19] V. Cossalter. Motorcycle Dynamics. 2nd ed. [place unknown]: LULU, 2006,

p. 372.

[20] G. Savino, F. Giovannini, N. Baldanzini et al. ‘Real-time estimation of road–

tyre adherence for motorcycles’. In: Vehicle System Dynamics 51.12 (2013),

pp. 1839–1852.

[21] M. Köhler, S. Jenne, K. Pötter et al. Zählverfahren und Lastannahme in der

Betriebsfestigkeit. Heidelberg: Springer Berlin Heidelberg, 2012.

[22] ASTM International. Standard practices for cycle counting in fatigue analysis.

Standard No. E1049-85. West Conshohocken (PA): ASTM International, 2011.

[23] AFNOR. Fatigue sous sollicitations d’amplitude variable. Standard No. A03-

406. La Plaine Saint-Denis, France: AFNOR, 1993.

[24] U. H. Clormann and T. Seeger. ‘RAINFLOW - HCM. Ein Zählverfahren für

Betriebsfestigkeitsnachweise auf werkstoffmechanischer Grundlage’. In: Stahl-

bau 55.3 (1986), pp. 65–71.

2 Customer loads of two-wheeled vehicles

[25] R. W. Hamming. Digital filters. 3rd ed. Mineola (NY): Dover Publications,

Inc., 1998, p. 304.

3 Road classification for two-wheeled

vehicles

Road classification for two-wheeled vehicles

Christian Gorgesa, Kemal Öztürka, Robert Liebichb

aBMW AG, Munich, Germany;bChair of Engineering Design and Product

Reliability, Berlin Institute of Technology, Berlin, Germany

(This is an Accepted Manuscript of an article published by Taylor & Francis in Vehicle System Dynamics

on 14/12/2017, available online: http: // www. tandfonline. com/ 10. 1080/ 00423114. 2017. 1413197 .)

This publication presents a three-part road classification system that utilises the vehicle’s

onboard signals of two-wheeled vehicles. First, a curve estimator was developed to identify

and classify road curves. In addition, the curve estimator continuously classifies the road

curviness. Second, the road slope was evaluated to determine the hilliness of a given road.

Third, a modular road profile estimator has been developed to classify the road profile

according to ISO 8608, which utilises the vehicle’s transfer functions. The road profile es-

timator continuously classifies the driven road. The proposed methods for the classification

of curviness, hilliness, and road roughness have been validated with measurements. The road

classification system enables the collection of vehicle-independent field data of two-wheeled

vehicles. The road properties are part of the customer usage profiles which are essential to

define vehicle design targets.

Keywords: Road classification, road profile estimation, ISO 8608, curve detection, cus-

tomer usage profiles, motorcycle dynamics

3.1 Introduction

Customer usage profiles are a key factor in durability analysis and product design.

The detailed knowledge of customer usage profiles improves design loads and ul-

timately the vehicle development process. On the one hand they are essential to

define vehicle design targets as Gorges et al. [1] and Johannesson and Speckert [2]

3 Road classification for two-wheeled vehicles

discussed. On the other hand customer usage profiles enable a virtual load acquis-

ition, where customer loads are simulated on virtual test tracks. Johannesson and

Speckert [2] highlight that the customer usage distribution consists of three compon-

ents: customer usage, vehicle-independent road properties, and the vehicle model.

The vehicle-independent road properties are the main focus of the present research.

It has the objective to collect information about the driven road classes using the

onboard signals of two-wheeled vehicles.

Common methods to obtain customer loads are survey sampling and online- mon-

itoring systems. Survey sampling is expensive due to extra measurement equipment

and cannot reveal the entire customer load distribution. Müller [3] and Matz [4]

already developed model-based online monitoring system with integrated counting

of durability-related values. This has become possible because the number of on-

board sensors rises anyway due to the increased functions of motorcycles including

an anti-lock braking system, dynamic traction control, or curve assistant. Onboard

signals are defined as signals that can be accessed by the vehicle’s Controller Area

Network (CAN) bus. In addition, Gorges et al. [1] developed a customer load ac-

quisition system for two-wheeled vehicles that utilises the vehicle’s onboard signals

to gather wheel forces and vehicle loading. As a result, customer loads are revealed

that are directly associated with the fatigue of the vehicle. Hence, the loads are

vehicle-dependent. On the contrary, the present research has the objective to col-

lect vehicle-independent road properties. Karlsson [5] has already discussed methods

to derive customer loads from road classification but the question remains how the

distribution of road classes driven by the customers can be revealed.

Therefore, this publication presents methods to detect and evaluate road classes in

terms of curviness, hilliness, and most importantly road roughness, which is evaluated

in terms of ISO 8608 [6]. On the small scale, the knowledge of the current road class

enables further real-time applications, that is, active suspension systems or curve

warning systems. On the large scale, the distribution of driven road classes helps

understanding the customer usage and makes a virtual load acquisition feasible.

In detail, the distribution of curve characteristics gives useful information about

the lateral wheel forces and abrasion of the outer tyre flanks. The distribution of

travelled altitude differences improves design targets for the powertrain. Third, road

roughness is highly correlated to the severity of vertical wheel forces. Therefore, the

3.2 Experimental set-up

distribution of road roughness is an essential part of the customer usage profiles.

In summary, the information about driven road classes improves the vehicle design

targets and enables the design of artificial test tracks for a virtual product develop-

ment and validation. Furthermore, the overall road class distribution helps designing

real measurement campaigns according to the customer usage. Speckert et al. [7] de-

veloped the Virtual Measurement Campaign (VMC) which improves the derivation

of design loads by geo-referenced data. Real existing roads have been recorded in a

database and were scored within different classes for curviness, hilliness, and road

roughness. Together with the information of road classes driven by the customers,

specific measurement tracks could be selected that correspond with the the customer

distribution. In contrast to specific application-driven publications, the contribution

of this paper is the road classification with onboard signals of two-wheeled vehicles.

This paper is organised as follows. Section 3.2 describes the reference motorcycle

and the onboard measurement equipment. Section 3.3 describes the algorithm for

the curve detection and classification. Section 3.4 explains the counting and classi-

fication method of the road slope to obtain the elevation gain. Section 3.5 introduces

the evaluation of measured road profiles, the generation of pseudo-random road pro-

files, the utilised full-vehicle model and ultimately the estimation and classification

algorithm to evaluate the road profile. Section 3.6 presents the results of the de-

veloped methods and Section 3.7 provides a summary and conclusion.

3.2 Experimental set-up

A motorcycle (BMW R1200GS) with data-logging devices was prepared for exper-

imental tests and validation of the algorithms, as introduced in the previous work

[1]. The reference frame will not rotate around the roll axes during banking of the

motorcycle, which means that it is aligned with the road plane. When the motor-

cycle is upright, it coincides with the vehicle’s coordinate system. The following

onboard signals were logged during pre-defined routes and manoeuvres for an offline

simulation of the developed algorithms:

•vehicle velocity v,

•front and rear spring deflections sft,srr, and

•model-based signals (e.g. roll angle φ).

3 Road classification for two-wheeled vehicles

These signals were logged through the CAN bus. Additionally, the vehicle was

equipped with a Global Positioning System (GPS) logging device, which provided

information about the position and the altitude for subsequent validation. The

logged signals were imported in a Simulink R

model to simulate vehicle dynamics

and validate the developed algorithms. The discrete model uses the same time step

size as the vehicle’s onboard system, which is set to ts= 0.01 s. This in principle

enables an online application of the developed algorithms. It is not part of the

present study to evaluate specific hardware requirements for an implementation of

the methods into existing or new production vehicles.

3.3 Road curve estimator

The first part of the road classification system is the road curve estimator. The

knowledge about road curve characteristics has two major applications: First, the

estimation of the current curve properties helps driving assistant systems to detect

dangerous situations in which the rider exceeds the physical limitation of velocity

for a given curve, see, for example, Biral et al. [8]. They developed a curve warning

system, which is part of the SAFERIDER1project. The detection strategy is based

on the difference between the actual state of the vehicle and a previewed optimal safe

state, which is forecasted with the help of the GPS position. Therefore, a detailed

model of the motorcycle’s dynamic behaviour during cornering is essential for real-

time applications, see Cossalter et al. [9–11] and Tanelli et al. [12]. The second

purpose for estimating the current road curve properties is field data collection, which

is part of the main focus of this research. Compared to the detailed and more complex

models [9–12], the authors propose a simpler approach, which requires less onboard

sensors and is therefore more suitable for field data collection. According to Cossalter

[11], a two-wheeled vehicle during cornering can be described as a lumped mass with

physical properties mass m, velocity v, and banking angle φ, see Figure 3.1a. During

steady-state cornering and on the assumption of infinitely slim tyres and a negligible

steering angle, the equilibrium of moments around the X-axis can be applied to

obtain the curve radius rc, see Equations (3.1)–(3.3). The normal force FNappears

1SAFERIDER Consortium is a paradigm of cooperation between the users, the motorcycle

industry, the ARAS (Advanced rider assistance systems)/OBIS (On bike information systems),

subsystems suppliers and the Research and Academic world, see www.saferider-eu.org.

3.3 Road curve estimator

due to gravity. The centrifugal force FCen occurs due to the circular movement of

the motorcycle around the rotational axis, which is the centre of the curve. The

curvature κis defined as the inverse of the curve radius rcand is always positive.

FN=mg, (3.1)

FCen =FNtan |φ|=mv2

(3.2)

⇒rc=v2

gtan |φ|, κ =gtan |φ|

v2.(3.3)

The developed algorithm detects the beginning of a curve with the help of the

absolute value of the banking angle φ. There are two additional conditions. First,

the vehicle velocity must exceed a given threshold to omit the detection of banking

during parking. Once a curve is detected, the algorithm calculates the running mean

of the velocity ¯v, the running mean of the banking angle ¯φ, and the timespan ∆t

for the cornering manoeuvre. When the absolute value of the banking angle falls

below a given threshold, the mean curve radius ¯rcand the mean curve angle ¯γare

computed, see Figure 3.1b and Equation (3.4). Second, the mean curve angle ¯γmust

be >60◦to omit the detection of small curves and overtaking manoeuvres.

¯γ=¯v∆t

¯rc

,∆t=t1−t0.(3.4)

For the evaluation of the curve characteristics, preliminary assumptions must be

considered. The objective of the algorithm is to classify a curve with feasible para-

meters. Assuming a curve is designed as a circular arc, it can be described by the

curve radius and the curve angle. The obtained parameters mean curve radius ¯rcand

mean curve angle ¯γdescribe the curve in a geometrical manner, but they lead to a

loss of information due to the simplification of the curve as a circular arc. Instead, a

road curve is designed with clothoids at the beginning and at the end of the curve to

ensure a smooth derivative of the curvature κ. For this reason, the authors suggest to

establish a further parameter named curviness c, see Figure 3.1b and Equation (3.5).

3 Road classification for two-wheeled vehicles

𝑚, 𝑣

𝑟

𝐹N

𝐹Cen

𝑍

𝑌

𝑋

(a) Y-Z-plane

𝑟

𝑋

𝑌

𝑥0, 𝑡0

𝑙c𝑥1, 𝑡1

𝑍

(b) X-Y-plane

Figure 3.1 – Road curve estimation physics.

The curvature κis therefore integrated over the travelled curve path x:

∫0

κ(x)dxwith lc=x1−x0.(3.5)

The calculation of the curviness crequires the length lcof the curve path which

can easily be derived from the velocity once the curve is detected. The curviness

parameter chas the unit radian. It is equal to the curve angle γin case of an ideal

circular shaped curve since the curvature is also defined by

κ=⏐⏐⏐⏐

dγ

dx⏐⏐⏐⏐.(3.6)

Although the curviness parameter chas also the dimensions of an angle, it takes

the clothoids into account and is more suitable for distorted curves than the pure

calculation of the mean curve angle. Therefore, the curviness cdescribes the curve by

meanings of its curve angle, assuming the curve would have an ideal circular shape

with the same overall curvature as the real curve the rider went through. These

advantages justify the introduction of the curviness c.

In principle, the obtained parameters describing the curve characteristics can be

classified by diverse possibilities, depending on the purpose of classification and

memory requirements. The authors prefer a two-dimensional counting of mean curve

3.4 Road slope classification

radius over mean curve angle. This approach allows the identification of challenging

curves with a small curve radius and a large curve angle, for example, sharp bends.

The number of gradations of the data bins can be chosen by the users. The curviness

ccan also be classified into bins. To sum up, a single curve can be classified by its

mean curve radius ¯rc, mean curve angle ¯γand curviness c.

The classification of the curve by its characteristic parameters is a kind of event

detection and event classification. To score the curviness of a given road or road

segment, the travelled distance shall also be considered. Therefore, the authors

propose to integrate the curvature κnot just over the curve path, but also over a

standard distance of l= 1 km to compute the road curviness C, see Equation (3.7).

∫0

κ(x)dxwith l= 1 km.(3.7)

A higher amount of curves as well as the curviness cof the curves located within l=

1 km lead to a higher road curviness C. In contrast to the single event classification,

this approach allows a continuous classification of the road segments in terms of

curviness. In summary, the curve estimator is based on a lumped-mass model and

requires the velocity vand the roll angle φas input signals.

3.4 Road slope classification

The second part of the road classification system is the evaluation of the travelled

elevation gain. Gorges et al. [1] developed a road slope estimator, which estimates

the current road slope α(◦) of two-wheeled vehicles with the help of a linear Kalman

filter. They utilise the road slope to compute the slope resistance force and finally

the wheel forces to identify customer loads. It is also possible to utilise the current

road slope for a real-time application, that is, a downhill brake assistant. In the

present paper, the road slope is utilised to score the hilliness of a driven road. The

difference in altitude is defined as the elevation gain. The knowledge of the elevation

gain travelled by the customers helps for a better understanding of the customer

usage. In addition, it improves the design of virtual or real measurement campaigns.

For a differentiation between positive and negative elevation gain, the road slope α

3 Road classification for two-wheeled vehicles

is prior splitted into only positive and negative values, see Equation (3.8).

αp=⎧

⎨

⎩

αfor α≥0

0for α < 0

and αn=⎧

⎨

⎩

0for α≥0

αfor α < 0

.(3.8)

The road slope αis integrated over the distance xto determine the elevation gain

h, see Equation (3.9).

hp=∫sin αp(x)dxand hn=∫sin αn(x)dx. (3.9)

The elevation gain is accumulated separately for positive and negative sign of the

road slope in order to compute continuous elevation gains hpand hnfor positive and

negative values. This distinction makes it possible to differentiate between uphill and

downhill ride. The continuous integration of the road slope results in absolute values,

which describe the overall amount of elevation gain travelled by the motorcycle.

Similar to the road curviness Cit is convenient to score the road hilliness Hof road

segments. Hence, the positive and negative elevation gain for a standard distance of

l= 1 km is computed to obtain the road hilliness H, see Equation (3.10).

Hp=

∫0

sin αp(x)dxand Hn=

∫0

sin αn(x)dx, (3.10)

with l= 1 km.

The road hilliness His also accumulated separately for positive and negative el-

evation gain. It can be classified into bins to make an online classification and

subsequent counting feasible. In contrast to the accumulated values, this approach

allows a continuous classification of the road hilliness. In a nutshell, the proposed

method takes the velocity vand the road slope αas input signals.

3.5 Road profile estimator

The third part of the road classification system is the road profile estimator. Road

roughness causes vehicle vibrations and has a direct influence on vehicle wear, com-

3.5 Road profile estimator

fort, safety, and fuel consumption. In addition, the dynamic wheel forces induced by

road roughness cause road deterioration [13]. For this reason, the knowledge of road

roughness has various applications. Road manufacturers and public authorities are

interested in the road conditions due to maintenance reasons, see, for example, [14,

15]. Moreover, road roughness affects traffic safety and helps defining speed limits.

Vehicle engineers utilise the current road roughness for real-time applications, for

example, active suspension systems, as shown in [16–21]. The present study aims at

the evaluation of the road roughness to derive customer usage profiles in terms of

durability, as it is also the focus in [22,23].

Various studies have already been published to estimate the current road profile

from mechanical responses of the vehicle, using different techniques. Ngwangwa et

al. [15] and Yousefzadeh et al. [24] adopted an Artificial Neuronal Network (ANN) to

reconstruct and classify the road profile depending on the measured vehicle responses.

One disadvantage of the ANN method is that it requires high computational efforts

for an online application and a large set of training data. Imine et al. [25] and

Rath et al. [26] developed sliding mode observers to estimate the road profile. Other

methods of control theory have been applied by Doumiati et al. [18,19] and Tudón-

Martínez et al. [21], who used an adaptive observer with the Q-parameterisation

method. The methods of control theory require in general more onboard sensors

than the present study can provide. Kalman filters and augmented Kalman filters

have been utilised by Doumiati et al. [16], Yu et al. [17], Jeong et al. [27], and

Fauriat et al. [23]. Furthermore, an H∞observer was adopted by Tudón-Martínez

et al. [20]. These methods based on observer theory utilise a Quarter-of-Vehicle

model to estimate the road profile. This is unsuitable for motorcycle dynamics, since

two-wheeled vehicles have a distinct pitch mode that needs consideration. Burger

[22] formulated an inverse control problem to estimate the road profile and solved

it with the help of the control-constraints method, which requires the solution of

differential-algebraic equations. Mathematical optimisation techniques have been

applied by Harris et al. [28] and Nordberg [29]. Wavelet transformation of the

vehicle’s response signals and a subsequent Adaptive Neuro-Fuzzy Inference System

for the road classification has been developed by Qin et al. [30,31]. The application

of wavelets has also been accomplished by Solhmirzaei et al. [32]. The rather simple

but fast approach of estimating the road profile in the frequency domain with the

3 Road classification for two-wheeled vehicles

Lateral road profile

Longitudinal road profiles

𝑋

𝑌

𝑍

Figure 3.2 – Definition of longitudinal road profiles.

help of transfer functions has been published by González et al. [14] and Barbosa

[33–35].

The cited methods differ in complexity, objective, and computational cost. The

present study uses the approach of transfer functions [14,33–35] and extends it to

a full-vehicle model with a delayed rear-wheel excitation. The main disadvantage of

constant vehicle velocity is eliminated by applying the method just to a small time

span with velocity-dependent transfer functions, which is the novel contribution of

this research. The objective is to develop an algorithm that can estimate the current

road profile continuously with the onboard signals.

3.5.1 Road profile evaluation

Sayers and Karamihas postulate that “a road profile is a two-dimensional slice of the

road surface, taken along an imaginary line” [36]. If the line is following the road

direction, the profile is defined to be a longitudinal profile, as illustrated in Figure 3.2.

Longitudinal road profiles describe the roughness and texture of the road.

Road profilers evaluate and measure the longitudinal road profile. They exist in

several variations. In the 1960’s, the first inertial profilers developed by General Mo-

tors Research Laboratories [37] had a breakthrough to measure large road networks

at high speed. An accelerometer measures the inertial reference and a non-contacting

sensor measures the relative height, for example, a laser transducer. Inertial profilers

must be moving to measure the road roughness and require a minimum speed. They

have been proven to produce accurate results even if they cannot collect long road un-

3.5 Road profile estimator

dulations. However, spatial frequencies less than 0.01 cycles/m(wavelengths above

100 m) are negligible for road statistics in sense of durability [36]. Other road profilers

have been developed, for instance, the Longitudinal Profile Analyzer [25], whereby a

single-wheel trailer is towed by a car and the movement of the wheel is transformed

to the profile elevation. A similar road profiler was developed by Barbosa [35], who

transformed the wheel movement with the systems transfer function to obtain the

road profile. In 1986, the World Bank published the International Road Roughness

Experiment [13] to establish a standard for road roughness measurements and eval-

uation methods. The authors proposed the International Roughness Index (IRI) to

evaluate the road roughness on a single scale. The IRI is a statistical value that is

computed with a virtual quarter car travelling over a road profile with a constant

velocity of v= 80 km h−1. The accumulated suspension motion y(x) = zs(x)−zu(x)

is divided by the travelled distance L, see Equation (3.11). The sprung mass dis-

placement is defined by zs(x)and the unsprung mass displacement by zu(x).

IRI = 1000 1

∫0

|˙y(x)|dx. (3.11)

IRI has the unit of slope (m km−1). The specific set of parameters characterising

the quarter car system is called The Golden Car. The IRI is well known and widely

accepted in the automobile industry. One advantage is that it measures the vehicle

response to a given road profile and makes a comparison possible. A more detailed

description can be found in [13]. Furthermore, ASTM-E1926 [38] defines a standard

procedure for computing the IRI. Andrén [39] points out, that the IRI is related to

the comfort experienced by a private car, but it is unsuitable for a mathematical

description of the road profile. Accordingly, another method to evaluate the road

roughness is the power spectral density (PSD) of a road profile. The ISO 8608 [6]

defines a method to report the PSD of a given road profile measurement as illustrated

in Figure 3.3. The original PSD of an artificial road profile is shown in the spatial

frequency domain. According to ISO 8608, the road profile can be classified into

eight different road classes (A–H). Furthermore, the ISO proposes a straight line fit

Gd(n)with the two parameters roughness coefficient Gd(n0)and waviness w, see

Equation (3.12). The roughness coefficient Gd(n0)represents the PSD of the road

profile at the reference spatial frequency n0= 0.1 cycles/m.

3 Road classification for two-wheeled vehicles

Figure 3.3 – PSD and straight line fit of an artificial road profile according to ISO 8608.

Gd(n) = Gd(n0)(n

n0)−w

with n0= 0.1 cycles/m,(3.12)

for 0.011 cycles/m< n < 2.83 cycles/m.

The spatial frequency nhas the unit cycles/mand is the inverse of the wavelength

λ. The ISO fixed w= 2, which defines the slope of the fitted PSD. For an evaluation

of the straight line fit and other PSD approximations with more parameters, see [39,

40]. For the description of a road profile with a PSD, the road profile is assumed

to be a homogeneous and isotropic two-dimensional random process, as Dodds and

Robson [41] describe it. On the contrary, Bogsjö [42,43] showed that short segments

of irregularities cannot be modelled with Gaussian processes. Hence, Bogsjö et al.

[44] proposed a new class of random processes called Laplace models to take the

irregularities into account. Furthermore, Johannesson and Rychlik [45] describe a

non-stationary Laplace model to stochastically reconstruct the road profile from

3.5 Road profile estimator

condensed roughness data in form of IRI. They also formulated a relation between

the IRI and the roughness coefficient Gd(n0).

3.5.2 Road profile generation

Synthetic road profiles are required for the development of algorithms to estimate the

road roughness. Tyan et al. [46] discussed the two most commonly used methods to

create synthetic road profiles: shaping filter and approximation with sinusoids. The

first method applies linear digital filters to white noise. This generates coloured noise

and in this case pink noise, whose PSD is characterised by a linear declining slope

in double logarithmic scale as shown in [30,47]. The second method is described by

Cebon [48] and has been applied by González et al. [14], Ngwangwa et al. [15], and

Sun [49]. This method is also utilised by the present research. For the approximation

of a pseudo-random road profile zR(x), a large number of Nsinusoids with different

amplitudes Aiand random phase angles Φiis superimposed, see Equations (3.13)–

(3.14).

zR(x) =

∑

i=1

Aisin (2πnix+ Φi)with (3.13)

Ai=√Gd(ni)∆nand Φi=U[0,2π).(3.14)

The amplitudes Aiare calculated from the straight line fit Gd(ni), see Equa-

tion (3.12), for the Ndifferent spatial frequencies ni. The relation of the PSD to the

amplitude of the road profile is obtained from Fourier analysis, where ∆nis the spa-

tial frequency increment, see Equation (3.14). The random phase angle Φiis taken

from the uniform distribution U. The spatial variable xis defined along the longit-

udinal direction of the road profile. The reference values Gd(n0)are taken from ISO

8608 [6], which are illustrated in Table 3.1 together with an exemplary description of

the different road classes [36] and the maximum possible velocity vmax, respectively.

Additionally, the IRI is reported for the generated pseudo-random test tracks.

The presented method generates random road profiles according to the different

road classes. For the validation of the developed algorithms, random profiles for all

road classes (A–H) were linked together to obtain a complete virtual test track, see

3 Road classification for two-wheeled vehicles

Table 3.1 – Properties of the road classes according to ISO 8608 and Sayers and Karamihas

[36].

ISO Gd(n0)IRI vmax

Class (10−6m3) (m km−1)Description (m s−1)

A16 1.6Airport runways and superhighways 60

B64 3.1Normal pavements 50

C256 6.3Unpaved roads and damaged pavements 30

D1024 12.5Rough unpaved roads 20

E4094 25.1Enduro tracks 10

F16 384 49.7Off-road tracks 5

G65 536 99.1Rough off-road tracks 5

H262 144 198.5Simulation purpose only 5

012345678

Distance x(km)

-1

Road Profile zR(m)

A B C D E F G H

Road Class

Detail

2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3

Distance x(km)

-5

Road Profile zR(cm)

Figure 3.4 – Pseudo random test track comprising road classes A–H and a detailed extract

of class C.

3.5 Road profile estimator

𝑝

𝑏

𝑍

𝑋

𝑌

&2*

𝑧R(𝑥)

𝑧s

𝑧rr 𝑧ft

𝑚rr 𝑚ft

𝜆

𝐽s, 𝑚s𝛩

𝑣

𝑐ft, 𝑘ft

𝑐T, 𝑘T

𝑐rr, 𝑘rr

𝑐T, 𝑘T

𝑠rr 𝑠ft

Figure 3.5 – Full-vehicle model with four DOFs.

Figure 3.4. The upper plot shows the complete test track while the bottom plot

shows a detailed extract of the class C road profile. The characteristic PSD of the

class C road profile is shown in Figure 3.3. The ISO mentions that road class H is

only for simulation purposes. The particular profiles have a length of l= 1 km and

are multiplied by a window function to ensure a smooth transition between them.

This test track enables the development and validation of specific algorithms in order

to estimate the current road roughness.

3.5.3 Full-vehicle model

A motorcycle in its plane of symmetry can be represented by three rigid bodies with

four independent coordinates. Hence, a full-vehicle model with four degrees of free-

dom (DOF) was utilised to describe the system dynamics, as illustrated in Figure 3.5.

Additionally, the spring deflections are illustrated because they are utilised for the

estimation algorithm. The full-vehicle model has been used in several publications

to simulate in-plane dynamics of motorcycles, see [11,12,14,35,50]. The in-plane

dynamics are generally excited by road undulations zR(x)and inertial effects due to

rider manoeuvres such as accelerating and braking. On the assumption of a constant

vehicle velocity v, the in-plane dynamics are reduced to four DOFs: vertical motion

of the chassis zs, pitch of the chassis Θ, rear unsprung mass motion zrr, and front

unsprung mass motion zft. These four DOFs are related to the vibration modes:

bounce, pitch, front wheel hop, and rear wheel hop [12]. The sprung mass mshas

the moment of inertia Jsand comprises the frame, the engine, the chassis, and the

3 Road classification for two-wheeled vehicles

rider. Additionally, parts of the front and rear suspension system are counted to

the sprung mass. The sprung mass is lumped in the centre of gravity (COG). The

rear unsprung mass mrr comprises the rear wheel, the rear brake, and parts of the

rear suspension. The front unsprung mass mft comprises the front wheel, the front

brake, and parts of the front suspension. Furthermore, the main geometric dimen-

sions are illustrated: wheelbase pand perpendicular distance bof the COG from the

rear-wheel Z-axis. The lumped masses are connected with parallel spring-damper

elements, which are represented by reduced stiffness and damping coefficients for

front suspension (kft, cft) and rear suspension (krr, crr). The tyre stiffness and damp-

ing coefficients are defined by kTand cT. The full-vehicle model properties used for

this study are illustrated in Table 3.2. The reduced stiffness and damping coefficients

are derived with the help of a multi-body simulation. The equations of motion of

the full-vehicle model with road excitation are as follows:

M¨

x+C˙

x+Kx=Fwith (3.15)

x=⎛

⎜

⎝

zft

zrr

⎞

⎟

⎠

,M=⎡

⎢

⎣

ms0 0 0

0Js0 0

0 0 mft 0

0 0 0 mrr

⎤

⎥

⎦

,(3.16)

C=⎡

⎢

⎣

cft +crr cft(p−b)−crrb−cft −crr

cft(p−b)−crrb cft(p−b)2+crrb2−cft(p−b)crrb

−cft −cft(p−b)cft +cT0

−crr crrb0crr +cT

⎤

⎥

⎦

,(3.17)

K=⎡

⎢

⎣

kft +krr kft(p−b)−krrb−kft −krr

kft(p−b)−krrb kft(p−b)2+krrb2−kft(p−b)krrb

−kft −kft(p−b)kft +kT0

−krr krrb0krr +kT

⎤

⎥

⎦

,(3.18)

F=⎛

⎜

⎝

kTzR(t) + cT˙zR(t)

kTzR(t−τ) + cT˙zR(t−τ)

⎞

⎟

⎠

.(3.19)

3.5 Road profile estimator

Table 3.2 – Full-vehicle model properties.

Description Symbol Value Unit

Sprung mass ms283 kg

Front unsprung mass mft 26 kg

Rear unsprung mass mrr 32 kg

Rotational inertia sprung mass Js55 kg m2

Reduced stiffness coefficient front suspension kft 17 000 N m−1

Reduced stiffness coefficient rear suspension krr 16 000 N m−1

Reduced damping coefficient front suspension cft 500 N s m−1

Reduced damping coefficient rear suspension crr 1000 N s m−1

Tyre stiffness coefficient kT170 000 N m−1

Tyre damping coefficient cT500 N s m−1

Wheelbase p1.5 m

Perpendicular distance of COG from rear-wheel Z-axis b0.7 m

On the assumption that the rear wheel follows the same road profile as the front

wheel, the road excitation of the rear wheel is the same function as for the front

wheel with a time delay τ:

τ=p

v.(3.20)

This implies that the excitation of the system depends on the vehicle velocity

v. Moreover, this formulation makes a single-input multiple-output (SIMO) system

definition possible to analyse the dynamic behaviour of the system as a function of

the road roughness zR(t). Linear time-invariant (LTI) systems are suitable to analyse

the response of a system to an arbitrary input in the time and frequency domain. The

aforementioned full-vehicle model is therefore formulated as an LTI system, since the

matrices are linear and the solution is linear shift-invariant. The relation between

the excitation and the response of the system is described with transfer functions

H(s), which relate the output Y(s)to the input X(s)in the frequency domain:

H(s) = Y(s)

X(s)=L{y(t)}

L{x(t)}.(3.21)

The Laplace transformation of the system equations needs to be derived to obtain

the transfer functions. For the formulation with just one input variable for the road

roughness ZR(s), the shift theorem as well as the derivation theorem from the Laplace

transformation rules are used as follows:

3 Road classification for two-wheeled vehicles

L{zR(t)}=ZR(s),(3.22)

L{zR(t−τ)}=ZR(s)e−τs,(3.23)

L{˙zR(t)}=sZR(s),(3.24)

L{˙zR(t−τ)}=sZR(s)e−τs.(3.25)

The Laplace transformation of the system equations assuming zero initial condi-

tions is given by:

s2MX(s) + sCX(s) + KX(s) = ⎛

⎜

⎝

kTZR(s) + scTZR(s)

kTZR(s)e−τs +scTZR(s)e−τs

⎞

⎟

⎠

.(3.26)

X(s)is the Laplace transformation of the input x(t). The equations of the spring

deflections sft, srr are derived and subsequently transformed to the frequency domain,

see Equations (3.27)–(3.28).

Sft(s) = Zs(s)+(p−b)Θ(s)−Zft(s),(3.27)

Srr(s) = Zs(s)−bΘ(s)−Zrr(s).(3.28)

The onboard signals comprise the front spring deflection sft and the rear spring

deflection srr. For this reason, the following two transfer functions are formulated:

Hft(s, v) = Sft(s)

ZR(s), Hrr(s, v) = Srr(s)

ZR(s).(3.29)

The input variable is the road roughness ZR(s), respectively. The transfer func-

tions can be derived by solving Equations (3.26)–(3.28) numerically. Since the road

roughness excites the rear wheel depending on the time delay τ, the transfer func-

tions are functions of the vehicle velocity v. Figure 3.6 shows the magnitude part

of the bode plot of the two transfer functions for a vehicle velocity of v= 15 m s−1.

These transfer functions describe the frequency response of the respective output

3.5 Road profile estimator

10−1100101102

Frequency f(Hz)

10−3

10−2

10−1

100

101

Magnitude (abs)

Hft(f)

Hrr(f)

Figure 3.6 – Bodeplot of transfer functions for v= 15 m s−1.

signals to the road excitation. In the following, they are used to estimate the road

profile with the onboard signals.

3.5.4 Estimation algorithm

The relation between the response and the road profile in the time frequency domain

fcan be expressed with transfer functions, as shown in Equation (3.29). To apply

the transfer functions to the PSD signals, the magnitude part of the transfer function

needs to be squared. This relates the PSD of the response signal PSDRes(f)to the

PSD of the road profile PSDRoad(f), as shown in Equation (3.30).

|H(f)|2=PSDRes(f)

PSDRoad(f)⇒PSDRoad(f) = PSDRes(f)

|H(f)|2.(3.30)

This means that the PSD of the road profile can be determined with the PSD of the

response signal and the associated transfer function. The PSD must be transformed

to the spatial frequency domain nto determine the road class according to ISO 8608

3 Road classification for two-wheeled vehicles

with the following relation:

n=f

vand PSDRoad(n) = vPSDRoad(f).(3.31)

The formulation PSDRoad(n) = vPSDRoad(f)follows from the definition of the

PSD as the squared amplitude per frequency increment. In summary, the character-

istic PSD of the road profile can be determined with the following expression:

PSDRoad(n) = vPSDRes(f)

|H(f)|2.(3.32)

This approach is based on the assumption of a constant velocity v. For the devel-

opment of a method that is feasible to estimate the current road roughness under

realistic operating conditions, the algorithm must be fast and independent of a con-

stant vehicle velocity. Hence, the approach is modified as follows: The transfer

functions are calculated in advance for a range of different velocities. As a result, the

function stack H(f, v)provides the correct transfer function Hv(f)depending on the

velocity. For a continuous estimation of the PSD of the road profile, an algorithm was

developed, as illustrated in Figure 3.7. Starting from the onboard signals, a circular

ring buffer records the required signals for a given time span ∆tbuf. Subsequently,

the mean value of the velocity ¯vis calculated for the extracted time frame. In the

meantime, the fast Fourier transform (FFT) and PSD of the signal PSDRes(f)are

computed in the time domain. The correct transfer function is interpolated from the

transfer function stack for the given mean velocity ¯vand the calculated frequencies

fresulting from the FFT. Afterwards, the PSD of the road profile PSDRoad(n)can

be calculated according to Equation (3.32), followed by the classification algorithm.

The choice of the time span ∆tbuf has an influence on the classification quality and

is evaluated in Section 3.6.3.

3.5.5 Road profile classification

As introduced in Section 3.5.1, the PSD of the road profile can be evaluated by ISO

8608 [6]. According to this standard, the PSD of the road profile can be classified

either at the spatial reference frequency n0= 0.1 cycles/mor multiple times in every

single octave band. The first method has the disadvantage that the classification res-

ult depends only on the PSD value at the spatial reference frequency. Furthermore,

3.5 Road profile estimator

Transfer FunctionOnboard Signals PSD Road Profile

FFT -> PSD ClassificationCircular Buffer

Δ𝑡buf

Response

Extract

𝑣 𝑓 PSDRes(𝑓)

𝐻

𝑣(𝑓)

PSDRoad(𝑛)

𝐻(𝑓, 𝑣)𝐻(𝑓, 𝑣)

Figure 3.7 – Flow chart of the road profile estimation algorithm.

it assumes that a straight-line fit can approximate the PSD. As Andrén [39] showed,

PSDs from real road measurements deviate from this assumption. The classification

in every single octave band is also infeasible for the present research, since different

classification results in the single octave bands must be stored for every time seg-

ment. Moreover, this method gives indistinct classification results. Hence, the main

objective is the classification of a road segment by its PSD into a single category

with the maximum of information provided.

For this reason, the authors propose a novel method to classify the PSD of a road

profile for a given time- or distance segment. After the PSD of the road profile has

been calculated according to Equation (3.32), it is smoothed in 10 octave bands,

which are illustrated in Table 3.3 and which are proposed by ISO 8608 [6]. The

centre frequency in each octave band is calculated by nc= 2EXP. The authors

propose the following novel smoothing algorithm. Weighted average values of the

PSD in the respective octave bands are calculated to smooth the PSD. The weighted

average is performed by a normalised Gaussian window function for each octave

band. This smoothing algorithm ensures a correct value of the PSD at the respective

centre frequency nc, even if only a few values are available within each octave band.

This is necessary since different velocities result in different spatial frequency at

3 Road classification for two-wheeled vehicles

Table 3.3 – Octave bands and geometric mean values for road classification.

Geometric mean values for road classes according to ISO 8608

ncGd(nc)(10−6m3)

EXP (cycles/m) A B C D E F G H

−7 0.007 86 2621 10 486 41 943 167 772 671 089 2 684 354 10 737 417 42 949 668

−6 0.0156 655 2621 10 486 41 943 167 772 671 089 2 684 354 10 737 417

−5 0.0313 164 655 2621 10 486 41 943 167 772 671 089 2 684 354

−4 0.0625 41 164 655 2621 10 486 41 943 167 772 671 089

−3 0.125 10 41 164 655 2621 10 486 41 943 167 772

−2 0.25 2.56 10 41 164 655 2621 10 486 41 943

−1 0.5 0.64 2.56 10 41 164 655 2621 10 486

0 1 0.16 0.64 2.56 10 41 164 655 2621

1 2 0.04 0.16 0.64 2.56 10 41 164 655

2 4 0.01 0.04 0.16 0.64 2.56 10 41 164

which the PSD is calculated. In addition, the smoothing in octave bands provides a

uniform distribution of PSD values over the spatial frequencies. This is also necessary

because the classification is treated in the logarithmic domain. The output of the

smoothing algorithm is PSDSmoothed(nc). At this point, the algorithm can handle

multiple estimates of the road profile from different onboard sensors. Subsequently,

one smoothed spectrum is calculated.

The classifier is formulated as a minimum distance classifier, which calculates the

distances of PSDSmoothed(nc)from the geometric mean values of the different road

classes in every single octave band. The matrix M(nc,class)of geometric mean

values of the different road classes in the respective octave bands is illustrated in

Table 3.3. The road profile is classified as the road class with the minimum sum of

distances, according to Equation (3.33).

Road class = min {10

∑

i=1

|log10[PSDSmoothed(nc,i)] −log10[M(nc,i,class)]|}.(3.33)

Figure 3.8 illustrates three different examples of road profile classification for dif-

ferent segments. The classified road classes are denoted beside the graphs. At first,

the original PSDs of the road profile estimation algorithm are smoothed in the spe-

cified octave bands, which are highlighted with dashed lines. For these examples,

the front and rear spring deflection signals have been used to estimate the road pro-

file. It can be seen that the smoothed PSD is the mean of the two provided PSD

3.5 Road profile estimator

Figure 3.8 – Examples of road profile classification.

estimates PSDft(n)and PSDrr(n). Finally, the minimum sum of distances classifies

each road segment. The PSDs are located at different spatial frequencies, which is a

result of different velocities at each segment. For example, road segment categorised

to class A was driven at a velocity of vA= 60 m s−1. Road segment class D was

driven at a velocity of vD= 15 m s−1. Road segment class G was driven at a velocity

of vG= 5 m s−1. Further influences on the frequencies resulting from the FFT are

the time span ∆tbuf and the sample frequency, which is set to f= 100 Hz.

One advantage of this classification method is its modular approach. The more

onboard sensors are available, the more robust becomes the classification result. The

smoothing algorithm can easily be extended to calculate the mean values of more

than just one PSD estimation. Additional transfer functions can be derived from the

full-vehicle model depending on the available onboard sensors. After the road class

has been determined, the travelled distance can be calculated from the mean velocity.

The distances in the road classes are incremented by the respective segment distance.

Thus, a distribution of travelled road classes can be recorded for the customer usage

3 Road classification for two-wheeled vehicles

800 900 1000 1100 1200 1300 1400 1500

X(m)

100

200

300

400

Y(m)

01234

Curviness c(rad)

Figure 3.9 – Results of the road curve estimator.

profiles.

3.6 Results and Validation

3.6.1 Validation of the road curve estimator

A test ride with the reference motorcycle as described in Section 3.2 was performed

on a curvy road to validate the road curve estimator. Figure 3.9 shows an extract

of the driven test road and the classified curves, respectively. The coloured sections

highlight the identification of a curve while the colour itself represents the curviness c

of the classified curves. It can be seen for example that curve No. 11 has the highest

curviness score, while curve No. 9 has the lowest curviness score. All curves were

3.6 Results and Validation

Table 3.4 – Properties of the classified road curves.

Road curve No.

Property 1 2 3 4 5 6 7 8 9 10 11 12 13 14

¯rc(m) 43 49 35 98 38 77 35 42 70 35 61 33 28 30

¯γ(◦) 113 85 254 119 218 103 208 78 74 112 182 184 237 130

c(rad) ×10−2169 123 388 226 358 181 319 117 114 166 408 319 396 207

1060 1080 1100 1120 1140 1160

X(m)

140

160

180

200

220

240

Y(m)

¯rc= 35m

¯γ= 208◦

c= 3.19rad

GPS

Curve Detection

Classification Result

Figure 3.10 – Properties of road curve No. 7.

identified by the curve estimator, which indicates the robustness of the developed

algorithm. The particular curve properties mean curve radius ¯rc, mean curve angle

¯γ, and curviness care illustrated in Table 3.4. It can also be seen that the distorted

curve No. 11 was scored with a high curviness even if the mean curve angle ¯γwas

scored not that high in comparison to the other curves. This manifests the proposed

index curviness cas an curve-evaluation index.

Figure 3.10 shows the classification results of curve No. 7 in more detail. The curve

estimator detected the correct beginning and end of the curvature of the road, which

is represented by the solid line (Curve Detection). In addition, the estimated curve

properties are highlighted with a circular arc with geometric dimensions according

to the estimation results (Classification Result). It can be seen that the estimated

circular arc roughly fits to the real road curvature. The overestimation is a result

3 Road classification for two-wheeled vehicles

0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2

X(km)

0.2

0.4

Y(km)

0 2 4 6 8 9 11 13 15 17

Road Curviness C(rad)

Figure 3.11 – Results of the road curviness classification.

of the curve construction with clothoids, as can be seen at the beginning and end of

the curve in Figure 3.10. These parts also contribute to the estimation algorithm.

The calculation of the running mean of the roll angle results in curve properties

that are a compromise between the smallest curve radius and the clothoids of the

curve. The proposed method is well suited to detect and classify curves in order

to collect customer usage profiles and to evaluate the driven curves. In addition,

the road curviness Ccontinuously scores road segments of l= 1 km, as illustrated in

Figure 3.11. It can be seen that the road curviness scores the respective road segment

depending on the amount and curviness of curves within the segment. The road

curviness Cwas scored to 9,11,15,17, and 6(rounded) for the given road segments

No. 1–5. Road segment No. 4 was scored with the highest road curviness C. This

is reasonable due to the amount of sharp curves within the segment. In contrast to

the single curve classification, the continuous classification of road segments makes

a characterisation of the driven roads possible.

The lumped-mass model achieved sufficient results for the scope of customer usage

profiles. The classified curve properties can be counted online, whereas the number

of gradations is chosen by the user and the memory capacities. Since the algorithm

is based on the response of the vehicle, the estimated curve properties are based on

the driving line of the motorcycle. Thus, different curve driving techniques can lead

to different results for the same curve. The differences are assumed to be negligible

3.6 Results and Validation

600

700

800

900

Altitude (m)

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14

Distance (km)

-50

H(m/km)

Figure 3.12 – Results of the road slope classification.

in terms of customer usage profiles.

3.6.2 Road slope classification

The road slope estimator developed by Gorges et al. was validated with the help of

a mountain road in a previous publication [1]. In the present paper, this mountain

road has been utilised to show the results of the road slope classification method.

The mountain road was driven uphill and downhill with the reference motorcycle,

as shown in the upper plot of Figure 3.12. The counting results of the road hilliness

Hare illustrated in the bottom plot for positive and negative values, respectively.

The road hilliness Hcounts the positive and negative elevation gain per kilometre.

The overall elevation gain for this ride was counted to hp= 342 m and hn=−343 m.

Once the road slope estimator is implemented in the vehicle, it is convenient to count

the elevation gain with the presented method. The distribution of an overall elevation

gain as part of the customer usage profiles is favourable for the vehicle development

process, since it affects vehicle design targets and improves the understanding of

3 Road classification for two-wheeled vehicles

the customer behaviour. The travelled elevation gain has a direct influence on the

powertrain design and on the brake design. In addition, the classification of the

particular road segments improves the choice of real test road or for the design of

virtual test tracks.

3.6.3 Validation of the road profile estimator

Real roads are not characterised by a homogeneous road class, as highlighted by

Andrén [39] and Bogsjö [42,43]. Furthermore, arbitrary roads are in general not

surveyed by a road profiler, which makes a validation infeasible. For this reasons,

the validation of the road profile estimator was achieved by numerical simulation,

for which the full-vehicle model was excited by the pseudo-random test track as

presented in Section 3.5.2. The simulation was performed by the ode45-solver of

MATLAB R

, which is based on the Runge–Kutta method. Each of the eight road

class segments (A–H) was travelled with a constant acceleration of the motorcycle.

This guarantees that the estimator was tested under variable velocities and that

all possible frequencies in the range of use have been excited. The maximum velo-

cities vmax for the respective road classes were chosen with respect to the physical

limitations of a motorcycle travelling over the tracks. The minimum velocity is

vmin = 3 m s−1, which is the minimum required velocity for the road profile estim-

ator. Each road class was driven for a time period of t= 100 s. Since the full-vehicle

model has no degree of freedom in the longitudinal direction, the variable velocity

was realised through the transformation of the road profile from the spatial domain

to the time domain. Figure 3.13 shows the results of the simulation. The upper plot

shows the linear slope of the velocity vin every road class segment together with

the maximum velocities vmax, respectively. The last three road classes (F–G) were

driven with a maximum velocity of vmax = 5 m s−1because they represent heavy

off-road tracks which are difficult to ride even for an enduro motorcycle like the

reference vehicle. The middle plot shows the respective road profile zR(t). It gets

rougher with an increase in the road class. The bottom plot shows the classification

results of the road profile estimator. It can be seen that almost all predicted road

classes are classified correctly. The time span was set to ∆tbuf = 1 s. This means

that 100 time segments have been classified per road class. A total of four estimates

have been classified false, but only by a one road-class difference. This indicates that

3.6 Results and Validation

Velocity v(m/s)

Actual Class

A B C D E F G H

-1

zR(m)

0 100 200 300 400 500 600 700 800

Time t(s)

Predicted Class

Figure 3.13 – Validation of the road profile estimator.

the proposed method is robust and highly accurate. As is common in classification

analysis, the results are reported in a confusion matrix, see Table 3.5.

The entries contain the amount of respective classifications. The last row and

column illustrate the percentage of correct classified values in each class. An overall

classification result of 99.5 % was achieved. A higher time span ∆tbuf leads to a more

robust result, since the signal length and thus the frequency content gets higher. On

the other hand, under the assumption of a variable velocity, the transfer function

gets ambiguous and therefore the classification quality gets worse. Additionally, the

road quality can change very fast, so that a longer time period results in an indistinct

classification result. In the end, the choice of the time span ∆tbuf is a compromise

between reaction speed and quality. The underlying method of the PSD calculation

has also an influence on the classification result. Since the smoothing algorithm

3 Road classification for two-wheeled vehicles

Table 3.5 – Confusion matrix of the road profile estimator.

Actual class

A B C D E F G H ∑(%)

Predicted class

A100 0 0 0 0 0 0 0 100

B0 100 0 0 0 0 0 0 100

C0 0 100 0 0 0 0 0 100

D0 0 0 100 0 0 0 0 100

E0 0 0 0 100 1 0 0 99

F0 0 0 0 0 98 1 0 99

G0 0 0 0 0 1 99 1 98

H0 0 0 0 0 0 0 99 100

∑(%) 100 100 100 100 100 98 99 99 99.5

is applied after the PSD calculation, an overlapped PSD calculation method is not

necessary to achieve a robust result. In addition, this would lead to a loss of frequency

resolution, which is essential for the road classification algorithm.

The results show that the frequency approach is successful and highly accurate even

under variable velocity, which had been addressed as a disadvantage of this method

in the past [20,21,23]. The reaction time is fast enough for collecting customer

usage profiles and implementing it into real-time control systems. Furthermore, the

modular approach makes the presented method easily extensible depending on the

available onboard sensors. In addition, the computational effort is less compared to

the alternative methods. An online application is therefore feasible. The developed

road profile estimator requires a full-vehicle model of the motorcycle, the velocity

v, and at least one suspension deflection signal as input. The derivation of transfer

functions requires an LTI system formulation. Thus, the model has to be reduced to

a linear system.

3.7 Summary and Conclusion

The objective of this research was to develop a road classification system with on-

board signals of two-wheeled vehicles. First, a curve estimation algorithm was de-

veloped, which identifies and classifies curves on the driven road. The authors pro-

pose a method to evaluate every single curve by its mean curve radius, mean curve

angle, and its curviness – a quantity that expresses how intense a rider would exper-

ience the curve. The algorithm takes the velocity and the roll angle as input signals.

Beside the event detection, the road curviness continuously evaluates the underlying

3.7 Summary and Conclusion

road to gather information about the driven road segments. The results show that

the curve estimator works as expected and that curves are classified in accordance

with their geometric dimensions. Second, the road slope was utilised to classify the

hilliness of the driven road. An absolute value of the elevation gain can be obtained

through the continuous integration of the road slope over the distance travelled.

The road hilliness continuously evaluates the elevation gain per road segment. The

proposed method takes the velocity and the road slope as input signals.

Third, a road profile estimator was developed. Different evaluation methods are

presented for a scientific classification of the road profile. The authors decided to

utilise a frequency approach, which is fast and easy to implement, but was supposed

to have some disadvantages in the past. The transfer functions of a two-wheeled

vehicle were derived with the help of a full-vehicle model and the Laplace transform-

ation. In contrast to previous publications which utilised a simple quarter-car model,

the presented model excites the front and rear wheel and is formulated with just one

input variable, the road roughness. This enables the correct application of the trans-

fer functions, which describe the relationship between the suspension deflections and

the road profile. The presented estimation algorithm is formulated in the frequency

domain and does not rely on a constant velocity assumption, which was addressed

as a disadvantage so far. The results show that the classification algorithm detects

the correct road class within one second. The road profile estimator is highly ac-

curate and robust. The modular approach makes it easily adaptable and extendable

to the available onboard signals. The classification algorithm is based on the ISO

8608 standard. It has successfully been proven on a virtual test track. It has to be

mentioned that the transfer functions as well as the validation have been achieved by

the same numerical model of the full vehicle model. For this reason, the results are

quite good. The proposed method will be verified with a real motorcycle on existing

roads in further research. The nonlinear behaviour of the tyres, the spring-damper

system and the influence of the vehicle dynamics needs to be investigated. The road

profile estimator works fast and requires no excessive computational effort compared

to other methods. The presented method requires a linearised full-vehicle model of

the motorcycle, the velocity, and at least one suspension deflection signal as inputs.

All of the developed algorithms are feasible to work in real time. For this reason,

an implementation into existing electronic control units is feasible. Vehicle design

3 Road classification for two-wheeled vehicles

targets can incrementally be improved with the knowledge of the road class distri-

bution. It helps understanding the customer usage and it enables to develop further

customer-specific applications. Another benefit is the evaluation of measurement

campaigns compared to the customer profiles. In addition, the distribution of dif-

ferent road classes makes a virtual load acquisition possible, where a virtual vehicle

collects load data on a virtual test track. The vehicle-independent road classifica-

tion enables a comparison between different product segments and different markets.

In the future, detailed knowledge about customer usage profiles offers a variety of

applications throughout the automotive business.

Since no time and location stamps are collected, an implementation of the system

is assumed to be uncritical in terms of data privacy. The data is collected an-

onymously without any customer assignment. The resultant customer distribution

is treated statistically to derive values such as quantiles, which do not correspond

to individual customers. Further research is planned to validate the developed road

profile estimator on real roads of different road classes. In addition, the detection

and classification of irregularities, that is, potholes, within the road segments will be

analysed.

References

[1] C. Gorges, K. Öztürk and R. Liebich. ‘Customer loads of two-wheeled vehicles’.

In: Vehicle System Dynamics 55.12 (2017), pp. 1842–1864.

[2] P. Johannesson and M. Speckert, eds. Guide to load analysis for durability in

vehicle engineering. Chichester: Wiley, 2013, p. 434.

[3] L. Müller. ‘Mehrkörpermodell-basiertes Online Monitoring der Betriebsbeans-

pruchung am Beispiel eines Nutzfahrzeug-Demonstrators’. PhD thesis. Kaiser-

slautern: TU Kaiserslautern, 2011.

[4] C. Matz. ‘Online Berechnung von Fahrwerkskräften auf Basis von Onboard-

Sensorik’. PhD thesis. Clausthal: TU Clausthal, 2015.

[5] M. Karlsson. ‘Load modelling for fatigue assessment of vehicles – a statist-

ical approach’. PhD thesis. Göteborg, Sweden: Chalmers, Göteborg University,

2007.

[6] ISO. Mechanical vibration - Road surface profiles - Reporting of measured data.

Standard No. 8608:2016. Geneva, Switzerland: ISO, 2016.

References

[7] M. Speckert, K. Dressler, N. Ruf et al. ‘The virtual measurement campaign

(VMC) concept’. In: Commercial Vehicle Technology 2014. Kaiserslautern,

Germany: Shaker Verlag, Mar. 2014, pp. 88–98.

[8] F. Biral, M. Da Lio, R. Lot et al. ‘An intelligent curve warning system for

powered two wheel vehicles’. In: Eur. Transp. Res. Rev. 2.3 (2010), pp. 147–

156.

[9] V. Cossalter, A. Doria and R. Lot. ‘Steady turning of two-wheeled vehicles’.

In: Vehicle System Dynamics 31.3 (1999), pp. 157–181.

[10] V. Cossalter and R. Lot. ‘A motorcycle multi-body model for real time simula-

tions based on the natural coordinates approach’. In: Vehicle System Dynamics

37.6 (2002), pp. 423–447.

[11] V. Cossalter. Motorcycle Dynamics. 2nd ed. [place unknown]: LULU, 2006,

p. 372.

[12] M. Tanelli, M. Corno and S. M. Savaresi, eds. Modelling, simulation and control

of two-wheeled vehicles. Chichester: Wiley, 2014, pp. 1–348.

[13] M. W. Sayers, T. D. Gillespie and C. A. V. Queiroz. The international road

roughness experiment. Washington, D.C.: The World Bank, 1986.

[14] A. González, E. J. O’brien, Y.-Y. Li et al. ‘The use of vehicle acceleration

measurements to estimate road roughness’. In: Vehicle System Dynamics 46.6

(2008), pp. 483–499.

[15] H. M. Ngwangwa, P. S. Heyns, F. J. J. Labuschagne et al. ‘Reconstruction of

road defects and road roughness classification using vehicle responses with ar-

tificial neural networks simulation’. In: J. Terramechanics 47.2 (2010), pp. 97–

111.

[16] M. Doumiati, A. Victorino, A. Charara et al. ‘Estimation of road profile for

vehicle dynamics motion: experimental validation’. In: American Control Con-

ference. San Francisco: IEEE, June 2011, pp. 5237–5242.

[17] W. Yu, X. Zhang, K. Guo et al. ‘Adaptive real-time estimation on road dis-

turbances properties considering load variation via vehicle vertical dynamics’.

In: Math. Probl. Eng. 2013 (2013), pp. 1–9.

[18] M. Doumiati, S. Erhart, J. Martinez et al. ‘Adaptive control scheme for road

profile estimation: application to vehicle dynamics’. In: The International Fed-

eration of Automatic Control. Cape Town: IFAC, Aug. 2014, pp. 8445–8450.

[19] M. Doumiati et al. ‘Road profile estimation using an adaptive Youla–Kučera

parametric observer: Comparison to real profilers’. In: Control Engineering

Practice 61.2002 (Apr. 2017), pp. 270–278.

3 Road classification for two-wheeled vehicles

[20] J. C. Tudón-Martínez, S. Fergani, O. Sename et al. ‘Online road profile estim-

ation in automotive vehicles’. In: European Control Conference (ECC). Stras-

bourg: ECC, June 2014, pp. 2370–2375.

[21] J. C. Tudón-Martínez, S. Fergani, O. Sename et al. ‘Adaptive road profile

estimation in semiactive car suspensions’. In: IEEE Transactions on Control

Systems Technology 23.6 (2015), pp. 2293–2305.

[22] M. Burger. ‘Calculating road input data for vehicle simulation’. In: Multibody

Syst. Dyn. 31.1 (2013), pp. 93–110.

[23] W. Fauriat, C. Mattrand, N. Gayton et al. ‘Estimation of road profile variability

from measured vehicle responses’. In: Vehicle System Dynamics 54.5 (2016),

pp. 585–605.

[24] M. Yousefzadeh, S. Azadi and A. Soltani. ‘Road profile estimation using neural

network algorithm’. In: Journal of Mechanical Science and Technology 24.3

(2010), pp. 743–754.

[25] H. Imine, Y. Delanne and N. K. M’Sirdi. ‘Road profile input estimation in

vehicle dynamics simulation’. In: Vehicle System Dynamics 44.4 (2006), pp. 285–

303.

[26] J. J. Rath, K. C. Veluvolu and M. Defoort. ‘Simultaneous estimation of road

profile and tire road friction for automotive vehicle’. In: IEEE Transactions on

Vehicular Technology 64.10 (2015), pp. 4461–4471.

[27] W. Jeong, K. Yoshida, H. Kobayashi et al. ‘State estimation of road surface and

vehicle system using a Kalman filter’. In: JSME Int. J. 33.4 (1990), pp. 528–

534.

[28] N. K. Harris, A. González, E. J. O’Brien et al. ‘Characterisation of pavement

profile heights using accelerometer readings and a combinatorial optimisation

technique’. In: Journal of Sound and Vibration 329.5 (2010), pp. 497–508.

[29] T. P. Nordberg. ‘An iterative approach to road/profile identification utilizing

wavelet parameterization’. In: Vehicle System Dynamics 42.6 (2004), pp. 413–

432.

[30] Y. Qin, M. Dong, F. Zhao et al. ‘Road profile classification for vehicle semi-

active suspension system based on adaptive neuro-fuzzy inference system’. In:

Conference on Decision and Control (CDC). Osaka: IEEE, Dec. 2015, pp. 1533–

1538.

[31] Y. Qin, M. Dong, R. Langari et al. ‘Adaptive hybrid control of vehicle semiact-

ive suspension based on road profile estimation’. In: Shock Vib. 2015.1 (2015),

pp. 1–13.

References

[32] A. Solhmirzaei, S. Azadi and R. Kazemi. ‘Road profile estimation using wavelet

neural network and 7-DOF vehicle dynamic systems’. In: Journal of Mechanical

Science and Technology 26.10 (2012), pp. 3029–3036.

[33] R. S. Barbosa. ‘Vehicle dynamic response due to pavement roughness’. In: J.

Brazilian Soc. Mech. Sci. Eng. XXXIII.3 (2011), pp. 302–307.

[34] R. S. Barbosa. ‘Vehicle dynamic safety in measured rough pavement’. In:

Journal of Transportation Engineering 137.5 (2011), pp. 305–310.

[35] R. S. Barbosa. ‘Vehicle vibration response subjected to longwave measured

pavement irregularity’. In: J. Mech. Eng. Autom. 2.2 (2012), pp. 17–24.

[36] M. W. Sayers and S. M. Karamihas. The little book of profiling. Michigan:

University of Michigan, 1998.

[37] E. B. Spangler and W. J. Kelly. GMR road profilometer: a method for meas-

uring road profile. Tech. rep. GMR-452. Warren, Michigan: General Motors

Corporation Research Laboratories, 1964.

[38] ASTM International. Standard practice for computing international roughness

index of roads from longitudinal profile measurements. Standard No. E1926-08.

West Conshohocken (PA): ASTM International, 1999.

[39] P. Andrén. ‘Power spectral density approximations of longitudinal road pro-

files’. In: International Journal of Vehicle Design 40.1/2/3 (2006), pp. 2–14.

[40] B. R. Davis and A. G. Thompson. ‘Power spectral density of road profiles’. In:

Vehicle System Dynamics 35.6 (2001), pp. 409–415.

[41] C. J. Dodds and J. D. Robson. ‘The description of road surface roughness’. In:

Journal of Sound and Vibration 31.2 (1973), pp. 175–183.

[42] K. Bogsjö. ‘Development of analysis tools and stochastic models of road profiles

regarding their influence on heavy vehicle fatigue’. In: Vehicle System Dynamics

44.sup1 (2006), pp. 780–790.

[43] K. Bogsjö. ‘Road profile statistics relevant for vehicle fatigue’. PhD thesis.

Lund, Sweden: Lund University, 2007.

[44] K. Bogsjö, K. Podgórski and I. Rychlik. ‘Models for road surface roughness’.

In: Vehicle System Dynamics 50.5 (2012), pp. 725–747.

[45] P. Johannesson and I. Rychlik. ‘Modelling of road profiles using roughness

indicators’. In: International Journal of Vehicle Design 66.4 (2014), p. 317.

[46] F. Tyan, Y.-F. Hong, S.-H. Tu et al. ‘Generation of random road profiles’. In:

J. Adv. Eng. 4.2 (2009), pp. 151–156.

[47] F. Schube. ‘Beitrag zur numerischen Simulation des Wirbelsäulenverhaltens

eines Kraftfahrers infolge durch Straßenunebenheiten induzierter Ganzkörper-

schwingungen’. PhD thesis. Aachen, Germany: RWTH Aachen, 2002.

3 Road classification for two-wheeled vehicles

[48] D. Cebon. Handbook of vehicle-road interaction. Advances in engineering. Boca

Raton, Florida: CRC Press, 2000.

[49] L. Sun. ‘Simulation of pavement roughness and IRI based on power spectral

density’. In: Math. Comput. Simul. 61.2 (2003), pp. 77–88.

[50] V. Cossalter et al. ‘Frequency-domain method for evaluating the ride comfort

of a motorcycle’. In: Vehicle System Dynamics 44.4 (2006), pp. 339–355.

4 Impact detection and experimental

road roughness classification

Impact detection using a machine learning approach and

experimental road roughness classification

Christian Gorgesa, Kemal Öztürka, Robert Liebichb

aBMW AG, Petuelring 130, 80788 Munich, Germany;bChair of Engineering Design

and Product Reliability, Berlin Institute of Technology, Straße des 17. Juni 135,

10623 Berlin, Germany

(This is an Accepted Manuscript of an article published by Elsevier in Mechanical Systems and Signal

Processing on 15/02/2019, available online: https: // doi. org/ 10. 1016/ j. ymssp. 2018. 07. 043 .)

First, this publication presents the experimental validation of a road roughness classific-

ation method. Second, an impact detection strategy for two-wheeled vehicles is proposed

including a classification of service loads, mild special events, and severe special events. The

methods presented utilise the vehicle’s onboard signals to gather field data. The modu-

lar road roughness classification system operates with the vehicle’s transfer functions, and

continuously classifies the road profile, according to ISO 8608. The method was success-

fully validated on test tracks with known road profiles. The impact detection strategy was

developed using a supervised machine learning technique. Six road obstacles were ridden

over using different velocities to invoke mild and severe special events. The most popular

classifiers were trained for comparison and prediction of future observations. The developed

impact detection strategy shows a high accuracy and was successfully validated using a k-

fold cross-validation. The combination of the road roughness classification system and the

impact detection strategy, enables a holistic field data acquisition of customer usage profiles,

in the context of durability engineering. The collection of customer usage profiles improves

vehicle design targets and enables a virtual load acquisition.

Keywords: Road roughness classification, ISO 8608, impact detection, supervised machine

learning, customer usage profiles, two-wheeled vehicles

4 Impact detection and experimental road roughness classification

4.1 Introduction

The present publication combines a road roughness classification method with an im-

pact detection strategy, in the context of durability engineering. Both methods are

presented as an application for two-wheeled vehicles and were validated by experi-

ments. The estimation and evaluation of the actual road roughness with the vehicle’s

onboard signals, has been the focus of many publications. The information about

road roughness is often applied for active suspension systems or road surface mainten-

ance. In the present research, the current road roughness is revealed and classified,

to obtain a distribution of driven road classes, in the sense of field data collection.

The underlying method of estimating the road roughness was developed by Gorges et

al. [1] and is experimentally validated in the present publication. The modular road

roughness classification system utilises the vehicle’s transfer functions and classifies

the road profile according to ISO 8608 [2]. A measurement campaign was carried

out, which included rough, unpaved roads and obstacles for the experimental valid-

ation. The methods have been tested and validated with a real motorcycle ridden

on test tracks, which were surveyed by a 3D roughness measurement system. The

collection of driven road classes is part of gathering customer usage profiles, which

improve vehicle design targets, especially in terms of lightweight design. Further-

more, customer usage profiles enable a virtual load acquisition on virtual test tracks.

Thus, the knowledge about driven road classes is an improvement in the product

development.

In the context of durability, the combination of customer behaviour and road

roughness indicates whether the loads that occur are considered as service loads,

special events, or even misuse events, as Figure 4.1 shows. According to the authors,

an increase in the vehicle velocity induces higher loads for a given road roughness,

and vice versa, see Figure 4.1a. Especially in the automotive industry, and referring

to Matz [3] and Pötter [4], customer loads are divided in three categories: service

loads, special events, and misuse events, as shown in Figure 4.1b. On the one hand,

this separation is characterised by statistical considerations. On the other hand, this

distinction is necessary for product liability and warranty. For example, an off-road

motorcycle is designed for unpaved roads and even small jumps, which means, up to a

specific threshold, the impacts of these events are regarded as service loads. While for

a superbike, the manoeuvres described would be regarded as special events or misuse

4.1 Introduction

Road roughness

Velocity

Misuse

events

Service

loads

(a) as a combination of road

roughness and vehicle velocity

(b) as a probability distribution

Figure 4.1 – Classification of customer loads.

events. Johannesson and Speckert [5] described the customer load distribution “in

terms of vehicle-independent load environment together with the vehicle usage and

the vehicle dynamics”. This description coincides with the presented derivation of

customer loads.

Service loads occur during the normal use of the vehicle, which is often called

the intended purpose. They can be described by a continuously distributed load

spectra during the life of the vehicle. In the case of a motorcycle, service loads com-

prise acceleration and brake manoeuvres, cornering, and loads that occur due to the

roughness of the road surface. In addition to the service loads, the intended pur-

pose also includes the occurrence of special events. Special events are rare compared

to service loads, and they induce a higher load on the vehicle components. They

are often characterised by impacts from sudden events, for example, driving over a

pothole. As special events are part of the intended use of the vehicle, the components

must be designed to sustain the loads. After the occurrence of a special event, the

vehicle still has to be fully operational.

Misuse events are per definition not part of the intended purpose, but they are also

considered during the vehicle design process. In engineering, the fail-safe principle is

applied, which means that the components should deform plastically along the load

path. This is called the damage chain. Figure 4.1b shows that in misuse events, the

load severity typically coincides with the structural strength of the components. The

customer should be able to clearly identify the damaged structure, and recognise that

the components were over-exposed in consequence of the misuse event. Misuse events

4 Impact detection and experimental road roughness classification

are characterised by a higher load level that exceeds a defined threshold, and are also

often the consequence of impacts, for example riding against, or over a significant

obstacle. The load threshold between special events and misuse events, is often

determined by numerical simulation and validated with experiments. Further aspects

of misuse, in the context of structural durability, have been discussed by Köhler et

al. [6], Hauke [7], and Berger et al. [8]. The methods presented can be categorised as

condition monitoring systems. For two-wheeled vehicles, the publication of Gorges et

al. [9] shows a real-time, wheel force calculation, with subsequent rainflow counting,

to derive customer loads. An example for passenger cars can be found in Matz [3].

The road roughness classification method presented, is designed for a continuous

evaluation of the driven road classes. Due to the restrictions of the underlying full-

vehicle model, the method works under normal operation conditions, which means

service loads. The system is not intended for the detection of single events, for ex-

ample passing over obstacles, nor for the evaluation of the loads that occur. For

this reason, the second part of the present study investigates the development of an

impact detection strategy. Since the threshold between service loads, special events,

and misuse events is, not defined by specific loads and strains, a machine learning

approach was evaluated. Different road obstacles were ridden over, at different velo-

cities, to gather an adequate data set of labelled observations. Subsequently, various

supervised machine learning techniques, in form of classification, were evaluated. It

is called supervised machine learning, because the data set was labelled before the

classifier was trained. Knowledge about the distribution of special events during the

product’s life, is highly valuable for improving vehicle design targets.

This paper is organised as follows: Section 4.2 describes the measurement and

evaluation of longitudinal road profiles. Section 4.3 presents the experimental set-

up and the test tracks for the measurement campaign. Section 4.4 briefly explains

the methodology for the road roughness classification, and shows the adjustments

for real operation conditions. The impact detection strategy is presented in Sec-

tion 4.5. Finally, Section 4.6 shows the results of the study. Section 4.7 provides the

publication’s findings.

4.1 Introduction

4.1.1 Literature review for road roughness classification

As mentioned above, information about the road roughness has various applications.

González et al. [10], Harris et al. [11], and Ngwangwa et al. [12] presented methods

for estimating road roughness in the context of road maintenance. Furthermore, the

evaluation of the current road roughness makes active suspension systems possible, as

shown in [13–24]. Burger [25] and Fauriat et al. [26] developed methods of deriving

customer usage profiles in terms of durability, as was also the focus of Gorges et

al. [1]. Different techniques were utilised to estimate the road roughness. For

example, Ngwangwa et al. [12] and Yousefzadeh et al. [27] applied an Artificial

Neuronal Network (ANN) to reconstruct and classify the road profile, depending on

the measured vehicle responses. ANNs usually require high computational efforts for

an online application and a large set of training data. Recently, Qin et al. [21] utilised

deep learning techniques for classifying the road profile. Sliding mode observers were

developed by Imine et al. [28] and Rath et al. [29]. More examples of the application

of control theory can be found in Doumiati et al. [15,16] and Tudón-Martínez et al.

[18]. The control theory methods cited require more signals than the present set-up

can provide. The application of Kalman filters was investigated by Doumiati et al.

[13], Yu et al. [14], Jeong et al. [30], Fauriat et al. [26], Wang et al. [19], and Qin et

al. [22]. Tudón-Martínez et al. [17] also examined an H∞observer to estimate the

road roughness. An inverse control problem was formulated, and solved, by Burger

[25] to estimate the road profile with the help of the control-constraints method,

which requires the solution of differential-algebraic equations. Other mathematical

optimisation techniques were applied by Harris et al. [11] and Nordberg [31]. An

application of wavelet transformation was developed by Qin et al. [23,32,33] and

Solhmirzaei et al. [34]. Ben Hassen et al. [24] utilised the Independent Component

Analysis (ICA) for estimating the road profile with the responses of a full-vehicle

model. The rather simple, but fast, approach of estimating the road profile in the

frequency domain with the help of transfer functions, has been published by González

et al. [10] and Barbosa [35–37]. This approach was also utilised by Gorges et al.

[1] and extended to a full-vehicle model with a delayed real-wheel excitation. The

novel contributions of this research were a sliding window with a small time span

and velocity-dependant transfer functions. The proposed method of road roughness

classification was validated with a full-vehicle model and a numerical simulation.

4 Impact detection and experimental road roughness classification

Since the study showed quite promising results, the method has now been extended

to real working conditions, and validated with the help of experiments from the

measurement campaign.

4.1.2 Literature review for the detection of road irregularities

A lot of road condition monitoring systems already exist. For example, De Zoysa

et al. [38] developed a system for public transport, to monitor road deterioration in

third world countries. They used a direct correlation of acceleration signals to the

road surface condition. Another example of crowd-based monitoring systems, is the

Pothole Patrol (P2)[39], which makes use of the GPS and vibration sensors mounted

on taxis to monitor the civil infrastructure in Boston. They use a threshold detection

z_peak to detect potholes, and a speed vs. z_ratio filter, to deal with the velocity

dependency of the signals. The system was successfully validated to detect potholes

and other severe road anomalies. The same technique is used by the Nericell and

TrafficSense project, published by Mohan et al. [40]. Perttunen et al. [41] used a FFT

transformation of the acceleration signals to extract frequency domain features and a

method of linear regression to remove the speed dependency. Tai et al. [42] developed

a smartphone-based road anomaly detector, especially for motorcycles. They used

machine learning techniques to train a classifier. Mednis et al. [43] developed an

Android based smartphone application as a layer, for the existing navigation system

Waze. They also utilised the acceleration signals collected from the smartphones, to

detect irregularities. Further examples of crowd-sourced pothole detection systems

that utilise smartphone sensors, are the Streetbump project from Carrera et al. [44],

bump detection from Hoffmann et al. [45], S-Road Assist from Sharma et al. [46],

Pothole detection from Wang et al. [47], RoADS from Seraj et al. [48], or the work

of P.M. and Gopi [49], who used Gaussian model-based mining to detect abnormal

events. A recent work from Fox et al. [50] shows the development of a multi-lane

pothole detector, using accelerometer data from embedded vehicle sensors. They

used Support Vector Machines (SVM) as machine learning technique. The research

of Cong et al. [51] shows the application of wavelet packet decomposition for feature

extraction of acceleration signals to detect road anomalies. A one-class SVM was

used as a classifier. A more sophisticated approach was carried out by Li et al. [52].

The authors developed a model-based pothole detection application, which exploits

4.2 Longitudinal road profiles

a multi-phase dynamic model of the tyre. A Bayesian estimation and an Unscented

Kalman Filter (UKF) estimate the current mode. The angular wheel velocity, vehicle

velocity, and vertical acceleration are provided as inputs.

All of the projects mentioned can be categorised as response-type detection sys-

tems, since they utilise the acceleration signal from either the vehicle itself, or from

an additionally-mounted embedded system, or smartphone. As well as these, vision-

based systems have evolved due to the increase in advanced driver assistance systems,

which include stereo cameras and Radar or Lidar sensors. Some examples of ultra-

sonic applications, are the pothole detection systems from Hedge et al. [53] and

Madli et al. [54]. Examples of stereo vision based applications for pothole detection,

can be found in [55–59]. A Lidar application can be found in [60]. The related work

makes significant efforts to detect potholes and other severe road anomalies, in the

sense of road surface condition monitoring. Therefore, the velocity dependency in

the response signals has been removed to identify the road irregularities, exactly as

they occur on the road. From a durability point of view, whether a vehicle drives

over a pothole with a low or high velocity, is of major importance. Hence, the impact

generated is dependant on both the obstacle itself and the velocity. In most cases,

the velocity of the vehicle decides whether the manoeuvre is a service load, special

event or even, misuse event. For this reason, the authors decided not to develop

a pothole detection system, but an impact detection system instead. This system

should be able to detect and classify special events, regardless of whether it was

a pothole, manhole, speed bump, or kerb. Due to the stochastic nature of special

events, the detection strategy has been chosen to be an event detection in the time

domain. With the help of measurements of such special events, an impact detection

strategy was developed.

4.2 Longitudinal road profiles

A definition and illustration of longitudinal road profiles can be found in the previous

research by Gorges et al. [1]. Longitudinal road profiles are defined as slices of the

road surface in the direction of the road. In contrast, lateral road profiles describe the

roughness and texture of the road, as a function of the track width. This is important

for two-track vehicles but is ignored in the case of single-track vehicles. The influence

of driving manoeuvres is neglected in the present research, and the longitudinal road

4 Impact detection and experimental road roughness classification

profile is assumed to be the driven road profile. For the majority of the distance

driven, this assumption appears to be justified. Furthermore, this simplification

is necessary for the validation of the road roughness classification system with the

measured test tracks.

4.2.1 Road profilers

Road profilers or profilometers, measure the texture of road surfaces. They exist in

several variations, but all require three key elements: a reference elevation, a height

relative to the elevation, and a longitudinal distance [61]. General Motors Research

Laboratories [62] developed the first inertial profilers in 1960, to measure large road

networks at high speed. The inertial reference is measured by an accelerometer and

the relative height is measured by a non-contacting sensor, e.g. a laser transducer.

The measurement of road roughness with inertial profilers requires a minimum speed.

These methods have been proven to produce accurate results, even if they cannot

collect long road undulations. However, in terms of durability, spatial frequencies

less than n= 0.01 cycles/m(wavelengths above 100 m) are negligible [61]. Further

road profilers have been developed, for instance, the Motor Industry Research As-

sociation (MIRA) device [63] or the Longitudinal Profile Analyzer [28], where a car

tows a single-wheel trailer and the movement of the wheel is transformed into the

profile elevation. Barbosa [37] developed a similar road profiler, whereby the wheel

movement is transformed using the systems transfer function to calculate the road

profile. Once the road profiles are collected as a function of the distance, they can

be evaluated with different techniques.

In this study, a mobile, multi-sensor, measuring system from 3D Mapping Solutions

GmbH1has been utilised to survey the test tracks. High-performance laser scanners

sample the road surface in three dimensions with a precision <1 mm. The scanned

surface is provided in the OpenCRG R

2-format. It is an open standard file format for

the evaluation of high-precision microscopic road surface data, for analysing handling,

ride comfort, and durability. Further information about the file format can be found

in [64,65]. After the data had been processed and saved in the described format,

a longitudinal road profile was extracted for a given vehicle trajectory. For the

13D Mapping Solutions GmbH, Raiffeisenstrasse 16, D-83607 Holzkirchen, http://www.

3d-mapping.de

2Curved Regular Grid, http://www.opencrg.org

4.3 Measurement campaign

extraction of the road profile, a coarser sampling rate of n= 20 cycles/mwas chosen,

because the evaluation of the road profiles does not require a fine resolution, due to

the enveloping effect of the tyres.

4.2.2 Evaluation of longitudinal road profiles

The evaluation of longitudinal road profiles was discussed in the previous research

by Gorges et al. [1]. In summary, there are two common methods for evaluating

the road roughness; the International Roughness Index (IRI) and the Power Spec-

tral Density (PSD) approach. The IRI represents the road roughness on a single

scale, whereas the PSD approach allows a classification of the road roughness into

eight different road classes (A–H). In addition, the classification can be performed at

different spatial frequencies. In general, the PSDs of road profiles are evaluated for

standard road segment lengths; for example, 1 km. The previous research by Gorges

et al. [1] introduced a novel smoothing algorithm to handle PSDs even for small

time spans, which is necessary for a fast detection of the current road roughness.

The present research experimentally validates this methodology, which is reviewed

briefly in Section 4.4. Examples of smoothed PSDs can be found in Section 4.3.2, at

the evaluation of the test tracks.

4.3 Measurement campaign

4.3.1 Experimental set-up

A motorcycle (BMW R1200GS) was mounted with data-logging devices for exper-

imental tests and validation of the algorithms. The following onboard signals were

logged during the measurement campaign, for an offline simulation of the developed

algorithms:

•vehicle velocity v,

•front and rear suspension travel sft,srr, and

•model-based signals (e.g. roll angle φ).

These signals were logged through the Controller Area Network (CAN) bus. Ad-

ditionally, the vehicle was equipped with a Global Positioning System (GPS) logging

device, which provided information about its position, for subsequent validation.

4 Impact detection and experimental road roughness classification

The logged signals were imported into a MATLAB R

environment to develop and

validate the presented algorithms. The discrete model uses the same time step size

as the vehicle’s onboard system, which is set to ts= 0.01 s. This, in principle, en-

ables an online application of the developed algorithms. The evaluation of specific

hardware requirements for an implementation of the methods into existing or new,

production vehicles, is not part of the present study.

4.3.2 Test tracks for road roughness classification

Nine different test tracks were ridden over to validate the road roughness classification

algorithm. Table 4.1 shows the properties of the selected test tracks. Six of them were

surveyed by the 3D roughness measurement system in a previous study by BMW,

which are denoted by the Curved Regular Grid (CRG) column (y/n). The ISO

classification result was achieved at the reference spatial frequency n0= 0.1 cycles/m,

see Section 4.2.2 for the evaluation method. The IRI was calculated directly from

the longitudinal road profile, whereas the ISO classification results from the profile’s

PSD. Test track No. 1, is a high-speed test track at BMW’s test and performance

centre, see Figure 4.2. The track is characterised by a very smooth surface, made

for high velocity manoeuvres. The classification results are ISO class A and a low

IRI, which means that very little suspension motion occurs while driving on the

test track. Test track No. 2 is an artificial, bumpy, country road, also located at

BMW’s test and performance centre. It was constructed to test vehicles in terms of

comfort and long-wave excitation. The profile was classified as ISO class B. Another

example of an ISO class B road, is test track No. 3. This dilapidated country road

is located in the Munich countryside. It is characterised by a long-wave excitation

and a crumbled surface with some defects.

Test track No. 4 represents fine cobble stone. This road surface is characteristic

for the historic centres of European cities, for example Milan or Rome, and excites

the vehicle at very short wavelengths. It was also classified as ISO class B. An extract

of test track No. 4’s longitudinal road profile, is illustrated in Figure 4.3. It also

shows the histogram of the road profile, together with a Gaussian distribution. The

similarity between the probability distribution and the normal distribution, shows

the stochastic nature of road profiles, as discussed in Section 4.2.2. An example of

an ISO class C road, is test track No. 5. This test track is also located at the test

4.3 Measurement campaign

Table 4.1 – Properties of test tracks.

ISO L IRI CRG

Test track No. Class Description (m)(m km−1)(y/n)

#1 A High-speed test track 7428 1.13 y

#2 B Bumpy country road 916 3.94 y

#3 B Dilapidated country road 720 4.21 y

#4 B Fine cobblestone 266 5.40 y

#5 C Dilapidated concrete panels 209 6.24 y

#6 C Coarse cobblestone 495 7.97 y

#7 - Gravel road 836 - n

#8 - Rough unpaved roads 6742 - n

#9 - Enduro fun park 12 296 - n

and performance centre, and represents a bad highway made of dilapidated concrete

panels. The connections between the concrete panels are rough, and this results in

mid-wave excitation of the vehicle. This pavement is specific for older, dilapidated

highways. Test track No. 6, which is part of the castle square at the Nymphenburg

Palace in Munich, is made of coarse cobblestone. The road surface induces short-

wave excitation. It was also classified as ISO class C.

Test track No. 7 consists of a gravel road and is located in the Munich countryside.

It is a public road with some small defects. The gravel induces high-frequency excit-

ation. Rough, unpaved roads are represented by test track No. 8, which is part of the

test and performance centre. This test track consists of several natural and artificial

dirt tracks, which were constructed for off-road motorcycle tests. These trails rep-

resent the tracks which are used during the increasingly popular enduro motorcycle

adventure tours. The roughest test track within the measurement campaign, is test

track No. 9. It is part of the enduro fun park in Hechlingen, Germany, where special

manoeuvres for off-road training are practised. These tracks are characterised by

rough unpaved roads with large defects, potholes, and obstacles, for example, fallen

tree trunks, roots, and stones. Test tracks No. 7–9 are ordered according to their

subjectively perceived, roughness.

ISO classes of C and higher are characterised by unpaved roads, as the road sur-

face gets more and more dilapidated with a higher roughness classification. As a

consequence, these unpaved roads cannot be surveyed correctly. This is not a re-

striction of the measurement system but rather of the topology itself. On the one

hand, the road surface of unpaved roads can change due to weather and wind. On

4 Impact detection and experimental road roughness classification

#1: High-speed test track #2: Bumpy country road #3: Dilapidated country road

#4: Fine cobblestone #5: Dilapidated concrete panels #6: Coarse cobblestone

#7: Gravel road #8: Rough unpaved roads #9: Enduro fun park

Figure 4.2 – Test tracks for the road roughness classification.

the other hand, the vehicle itself displaces the loose underground as it rolls over

it. This implies that the validation of the road roughness classification system for

unpaved roads cannot be achieved by 3D roughness measurement. However, the au-

thors propose to evaluate the unpaved roads presented, which have no measurement

data available. The validation of these roads was achieved by a relative comparison

between them, and also by comparing them to the surveyed roads.

Figure 4.4 shows the smoothed PSD of the test tracks No. 1, 2, 5, and 6, together

with the ISO classification thresholds of the classes A–E. After the PSDs of the road

profiles had been calculated, the smoothing algorithm computed one data point for

each of the ten octave centre frequencies. The octave bands are highlighted with

4.3 Measurement campaign

0 20 40 60 80 100

-20

-10

0 0.1 0.2

Figure 4.3 – Road profile and histogram of test track No. 4.

dashed lines. The reference spatial frequency n0= 0.1 cycles/mis also highlighted

with a solid line, at which the simple ISO classification is performed. According

to this simplified classification procedure, the PSD at the other spatial frequencies,

does not contribute to the classification result. This follows from the straight-line

approximation of the PSD. On the contrary, the smoothed PSDs show that only

test track No. 1 could be approximated by a straight line fit. Test track No. 5

also shows a more or less linear behaviour, excluding the first and last octave. The

bumpy country road has a high degree of roughness at low spatial frequencies, with

a decrease in roughness at the higher spatial frequencies. In contrast, test track No.

6 has a lower roughness at the mid spatial frequencies, compared to the roughness

at the lower and higher spatial frequencies. As Andrén [66] also discovered, the

straight-line approximation is not suitable for arbitrary test tracks.

In summary, a profile of a real, existing, road can be classified either by the sim-

plified straight-line approximation, or in multiple octave bands at different spatial

frequencies. In the case of the straight-line approximation, the PSD is evaluated

solely at the reference spatial frequency. Other spatial frequencies are ignored and

this results in an inaccurate classification result. On the other hand, a classification

at multiple octave bands is also impractical, since this results in multiple classifica-

tion results. Therefore, a more convenient classification method is proposed by the

authors in the previous research [1]. Thus, a short road segment is classified by a

4 Impact detection and experimental road roughness classification

Figure 4.4 – Smoothed PSDs of selected test tracks.

minimum distance classifier into one of the ISO classes. The treatment of short road

segments is one main feature of the presented method, since it calculates the PSD for

every ∆tbuf = 1 s. The segment lengths are accumulated simultaneously depending

on the road class. This results in a distribution of distances per ISO class for a given

road which enables a more detailed classification result for the individual test tracks.

Examples are given in Section 4.6.

4.3.3 Road obstacles for impact detection

For the development and validation of the impact detection strategy, several road

obstacles were ridden over with different velocities. The onboard signals were logged

for posterior development and validation of the algorithms. Figure 4.5 illustrates the

road obstacles. At first, two railway crossings were tested. Railway crossings are

common in rural areas. Depending on the velocity, the discontinuous road surface

induces an impact. Next, a relatively large speed table, which can be found in

south Europe, south America, and Asia, was tested. The speed table has a short

4.3 Measurement campaign

#1: Railroad crossing lateral #2: Railroad crossing angular #3: Speed table

#4: Pothole small #5: Pothole large #6: Kerb

Figure 4.5 – Road obstacles.

slope up to a height of h= 20 cm on both sides. This induces an impact at the

beginning and at the end of the speed table. Two different sizes of potholes were

ridden over to evaluate small road irregularities (#4) and larger road defects (#5).

Both potholes had a depth of d= 5 cm, while the small potholes had a maximum

length of l= 30 cm and the large pothole had a length of l= 1 m. Finally, a kerb

with a height of h= 14 cm, was tested to demonstrate the threshold between special

events and misuse events. The motorcycle was ridden at right angles to the kerb to

avoid an accident. Table 4.2 shows the properties of the test manoeuvres in detail.

The number of experiments and the ridden velocity range are illustrated.

Since the set-up allowed no evaluation of the strength of the impact, the test rider

was asked to allocate the special events to two groups, mild or severe. A mild special

event was defined as an impact that was perceived by the rider, but did not lead to

a loss of control or a high feeling of discomfort. A severe special event was defined

as the perception of a high degree of discomfort, in combination with a near miss.

It was defined as the maximum possible special event, with a transition to misuse

events. In total, 52 special events were measured; 46 of them were labelled as mild

special events, whereas 6 were labelled as severe special events. Please note that only

4 Impact detection and experimental road roughness classification

Table 4.2 – Properties of measured special events.

Obstacle Velocity Label

No. Description Special events (m s−1) Mild Severe

#1 Railway crossing lateral 6 13.9–16.66 0

#2 Railway crossing angular 6 13.9–16.66 0

#3 Speed table 6 3.1–7.26 0

#4 Pothole small 15 5.5–8.315 0

#5 Pothole large 12 4.7–17.512 0

#6 Kerb 7 3.3–7.21 6

∑52 46 6

the kerb manoeuvres were labelled as severe special events.

4.4 Road roughness classification

The present research is based on the theoretical studies about road roughness classi-

fication by Gorges et al. [1], to reveal the vehicle-independent distribution of driven

road classes. In the previous research, the proposed method of road roughness clas-

sification was validated with a full-vehicle model and a numerical simulation. Since

the study showed quite promising results, a measurement campaign was carried out

and the method has now been tested and validated with a real motorcycle ridden

on real test tracks. The main concept of the underlying method will be explained

briefly, and the extensions for real operating conditions will then be presented.

4.4.1 Methodology

The modular road profile classification system utilises the vehicle’s transfer functions

to compute the excitation from the resultant responses. The velocity-dependant

transfer functions had already been calculated from a numerical, full-vehicle model

of the motorcycle. The excitation is the longitudinal road profile and the resultant

responses are measured by the vehicle’s onboard sensors. In the present research, the

suspension travel of the front and rear suspension system have been applied as the

response signals. The response signals are recorded in the time domain with the help

of a sliding window that has a time span of ∆tbuf = 1 s. Subsequently, the signal

extracts are transformed into the time frequency domain with the help of the Fourier

transformation. The PSD of the response signal is also computed, which results in

4.4 Road roughness classification

PSDRes(f). Meanwhile, the transfer functions H(f)are interpolated according to the

current velocity v. Finally, the PSD of the road profile PSDRoad(n)can be computed

according to Equation (4.1).

PSDRoad(n) = vPSDRes(f)

|H(f)|2.(4.1)

After the PSD for a given time or distance segment has been computed, the

novel smoothing algorithm, introduced by Gorges et al. [1], is applied. The PSD is

smoothed in 10 octave bands, which are proposed by ISO 8608 [2]. The centre fre-

quency in each octave band is calculated by nc= 2EXP. The weighted average values

are computed for each octave band using a normalised Gaussian window function.

This ensures a good approximation of the PSD at the respective centre frequency,

even if only a few values are available within each octave band. This is especially

necessary for the evaluation of short test tracks and for the implementation of a real-

time road roughness classification. In addition, the PSDs from both the front and

the rear suspension sensor can be treated together to achieve a more robust classi-

fication result, since more information is evaluated. The result is the smoothed PSD

at the ten octave centre frequencies. Subsequently, a minimum distance classifier is

applied, which computes the class distance from the ISO [2] classes in every octave

band. The class with the lowest error is chosen to be the classification result, as can

be found in Gorges et al. [1].

4.4.2 Extension for experimental validation

In the present publication, the validation of the road roughness classification was

achieved by riding a real motorcycle on the test tracks. Therefore, the numerical

full-vehicle model providing the transfer functions, must describe the real vehicle

as well as possible. Since the transfer functions can only be derived from a linear

time-invariant model, the nonlinear behaviour of the spring-damper systems must

be linearised. In addition, the reduced stiffness and damping coefficients need to be

derived, as described by Cossalter [67] and Tanelli et al. [68]. The test motorcycle

is designed with a BMW-Telelever front suspension system. It is characterised by

one single suspension strut in the X-Z-plane. This constructions allows a separation

of the suspension functions. The telescopic sliders contribute to the wheel control,

4 Impact detection and experimental road roughness classification

𝑍

𝑋

𝑌𝑧R(𝑥)

𝑘s

𝑚ft

𝑐T, 𝑘T

𝑣

𝜆

𝑧ft

𝑙s

(a) BMW-Telelever suspension system

𝑍

𝑋

𝑌𝑧R(𝑥)

𝜆

𝑚ft

𝑐ft, 𝑘ft

𝑐T, 𝑘T

𝑣

𝑧ft

𝑚s

(b) Front part of reduced full-vehicle model [1]

Figure 4.6 – Derivation of reduced stiffness and damping coefficients.

whereas the single suspension strut is responsible for the vehicle’s comfort and road-

holding. In addition, the system is known for its anti-dive behaviour in braking

situations. Figure 4.6 illustrates the derivation of the reduced stiffness and damping

coefficients kft and cft from the BMW-Telelever front suspension system. The velocity

ratio βis defined as the ratio between the suspension travel lsand the vertical wheel

movement zft, see Equation (4.2).

β=∂ls

∂zft

.(4.2)

The velocity ratio βwas determined with the help of a multi-body simulation and,

for the test motorcycle, is a constant value. For the derivation of the reduced stiffness

and damping coefficients kft and cft, the power balance between the wheel force Fz|ft

and the spring force Fsis applied, see Equation (4.3).

Fz|ft ˙zft =Fs˙

ls=Fs

∂ls

∂zft

˙zft with Fs=F0+ks(l0−ls).(4.3)

The spring force Fsis calculated with the spring coefficient ksand the variation

in suspension travel ls, where F0is the spring force at the initial suspension travel

l0. The power balance reduces to

Fz|ft =Fs

∂ls

∂zft

=Fsβ. (4.4)

100

4.4 Road roughness classification

The reduced stiffness coefficient kft is defined as

kft =∂Fz|ft

∂zft

.(4.5)

Under the assumption of a constant velocity ratio βit reduces to

kft =ksβ2.(4.6)

The same approach is used for the derivation of the reduced damping coefficient

cft, and for the derivation of the coefficients for the rear suspension system. In a

nutshell, the reduced stiffness and damping coefficients can be derived with the help

of the constant velocity ratio βand the linearised spring and damper coefficients.

In order to complete the derivation of the full-vehicle model, the real motorcycle

must be reduced to the sprung mass ms, the front unsprung mass mft, and the rear

unsprung mass mrr.

In addition, to work under real operating conditions, the road roughness classific-

ation algorithm also needed some adjustments. First, the front and rear suspension

travel response signals are transformed to the vertical wheel movement. This can be

achieved with the help of the velocity ratio β. Moreover, as a preparation for the

Fourier transformation, the signal extracts are detrended and windowed. The win-

dow function was set empirically to a Tukey window with a taper coefficient r= 0.05.

In addition to the signal preparation for the classification algorithm, the model as-

sumptions must be challenged. In the theoretical study by Gorges et al. [1], the

degrees of freedom of the numerical full-vehicle model are the same as for the model

from which the transfer functions are derived. Eventually, this led to quite good

results. The real motorcycle has roughly the same degrees of freedom in the X–Z-

plane as the numerical model, but not only the road profile but also the longitudinal

dynamics can excite the sprung mass. In the case of significant longitudinal vehicle

dynamics, such as accelerating and braking, the sprung mass movement would lead

to a false interpretation of the road roughness. Therefore, an on/off logic was imple-

mented to restrict the pitch influence on the road roughness classification. Since no

pitch signal was available on the test motorcycle, the longitudinal acceleration was

utilised to detect acceleration and brake events. Thus, the classification algorithm

pauses when a certain longitudinal acceleration value is exceeded. Furthermore, the

101

4 Impact detection and experimental road roughness classification

Figure 4.7 – Lower and upper bound of spatial frequency ndepending on the velocity v.

model assumptions are only valid for in-plane dynamics. The algorithm also pauses

when the measured angular yaw rate exceeds a certain value, for example during

a cornering manoeuvre. Since the classification algorithm is based on the vehicle

response, a minimum velocity of v≥3 m s−1must be exceeded so that the roughness

classification achieves correct results. When at least one condition is violated, the

classification algorithm discards the current signal extract. In this case, the rejected

signal extract is classified as unknown roughness.

The sampling rate of the vehicle’s bus system is ts= 0.01 s and the sliding window

for the signal extracts has a time span of ∆tbuf = 1 s. This combination results

in a minimum time frequency of fmin = 1 Hz and a maximum time frequency of

fmax = 50 Hz after the Fourier transformation. The transformation of the PSD from

the time frequency fto the spatial frequency n, depends on the velocity v. As a

consequence, the PSD is defined by a restricted spatial frequency domain depending

on the velocity v. Figure 4.7 illustrates the lower and upper bounds of the restricted

spatial frequency domain of the transformed PSD as functions of the velocity. Be-

cause there is a moving spatial frequency domain, the classification algorithm was

designed to handle varying spatial frequencies. It can be seen that independent of the

velocity v, a minimum of five octave bands can be evaluated, which are highlighted

with the dashed lines.

102

4.5 Impact detection

𝑍

𝑋

𝑌

𝜔, 𝛼

𝐽𝑟𝐹

𝐹

𝑙s𝑣

𝑠ft = 𝑙0− 𝑙s

Figure 4.8 – Schematic view of an impact.

4.5 Impact detection

The road roughness classification system is designed to classify the road roughness

independently of the velocity. However, from the viewpoint of durability, the combin-

ation of the vehicle velocity and the road surface roughness induces varying loads that

need further investigation. Therefore, an impact detection strategy was developed

to detect and classify special events, for example passing over a kerb.

4.5.1 Model of an impact event

Figure 4.8 shows a schematic view of an impact. The angular front wheel velocity

ωand the variation in suspension travel sft are available for the detection of special

events. The variation in suspension travel is given by sft =l0−ls, where l0is the

length of the uncompressed spring. The impact induces a contact force Fc, which is

decomposed into its components Fxand Fz. The contact force induces a change in

the angular momentum Lof the front wheel, see Equation (4.7).

∆L=Jα, α =dw

dt, ω =v

r.(4.7)

4.5.2 Signal processing and feature extraction

Figure 4.9 shows the onboard signals for the angular front wheel velocity ωand the

variation in suspension travel sft during a kerb crossing, together with their time

derivatives, angular wheel acceleration αand suspension travel velocity ˙sft. The

103

4 Impact detection and experimental road roughness classification

Figure 4.9 – Onboard signals during kerb crossing.

change in the angular momentum can be seen as a peak in the angular wheel velocity

ω, or in the angular acceleration αof the front wheel. In addition, the z-component of

the contact force lifts the wheel over the kerb and compresses the suspension spring.

Hence, the impact can also be seen in the suspension signal sft and its derivative,

the suspension travel velocity ˙sft. The kerb crossing manoeuvre is representative

for all the other special events, because all the events have in common that the

impact induces an x- and z-component of the contact force. For the application of

a classification method, feature extraction from the input signals is necessary. The

authors tested some of the statistical key numbers, for example standard deviation,

variance, root mean squared error, and peak-to-peak on sliding windows with a time

span of ∆tbuf = 1 s. Studies have shown that these key numbers are highly correlated

and in terms of feature selection, the authors decided to pick the peak-to-peak values

of the angular wheel acceleration αand the suspension velocity ˙sft as features.

The choice of the time derivatives of the original measured onboard signals angular

velocity ωand variation in suspension travel sft, has two major benefits: First, the

104

4.5 Impact detection

Figure 4.10 – Scatter plot of the training set.

time derivatives act as a numerical high-pass filter. Thus, the influence of sensor

drifts and offsets is prevented. Second, particularly for the suspension travel signal,

only the time derivative distinguishes between a spring compression as a result of a

special event, and a spring compression as a result of vehicle dynamics, for example,

braking and acceleration events. As a consequence, the suspension travel velocity ˙sft

is more practical for the robust detection of special events. Since the units and the

dimensions of the selected features are different, it is common in machine learning

techniques to standardise the input variables. Accordingly, the features have been

standardised by the z-score, and are henceforth referred as x1and x2as given in

Equation (4.8).

x1=(α−µα)

σα

and x2=( ˙sft −µ˙sft )

σ˙sft

.(4.8)

4.5.3 Training set

Figure 4.10 shows a scatter plot of the observations in the x1–x2plane. For this,

the measured onboard signals from the 52 special events were processed using the

sliding window, feature extraction, and standardisation. For the classification and

detection of special events, it is necessary to also include observations that represent

the normal usage of the motorcycle. For this reason, the measurements for test

tracks No. 1–6 from the road roughness classification, were also processed and are

105

4 Impact detection and experimental road roughness classification

labelled as service loads. In addition, several measurements for dynamic manoeuvres

have been provided, so that the classifier can distinguish between normal usage and

special events. For a robust classification algorithm, it is essential to provide as

many observations as possible in the training set. On the one hand, the scatter plot

shows a clear distinction between mild and severe special events, which confirms

the correct labelling of the observations. On the other hand, the transition between

service loads and mild special events is ambiguous. Some normal usage manoeuvres

also lead to a spring compression and/or an angular wheel acceleration, for example

intense braking events. In total, the training set consists of 2459 service loads, 46

mild special events, and 6 severe special events.

4.6 Results and Discussion

This section contains the evaluation of the methods presented. The results of the

road roughness classification algorithm were compared to the classification results of

the laser-scanned road profiles. For the impact detection strategy, different machine

learning techniques were evaluated and validated with the k-fold cross-validation

method.

4.6.1 Validation of the road roughness classification

Since the road roughness classification algorithm determines a road class for every

∆tbuf = 1 s, the resultant distances per road class are accumulated. As discussed in

Section 4.3.2, a single classification for the complete measured road profile leads to

inaccurate results. Therefore, the measured road profile was segmented according

to the distance for which the motorcycle stayed within the sliding window. Only

this method enables a comparison of the results. For example, at test track No. 1,

the motorcycle had a mean velocity of vmean = 26 m s−1and therefore the segment

length was set at lseg = 26 m. On the one hand, this method guarantees that the same

number of classification results are compared. On the other hand, it ensures that the

measured road profile is evaluated at the same spatial frequencies as the classification

algorithm that uses the response signals. Figure 4.11 shows the validation results

for test tracks No. 1–6, while CRG data represents the measured road profile. The

percentage of correctly classified segments is shown in Table 4.3.

106

4.6 Results and Discussion

#1: High-speed test track #2: Bumpy country road #3: Dilapidated country road

#4: Fine cobblestone #5: Dilapidated concrete panels #6: Coarse cobblestone

Figure 4.11 – Results of the road roughness classification for test tracks No. 1–6.

Table 4.3 – Results of the road roughness classification.

Test track No. #1 #2 #3 #4 #5 #6

Correct classified (%)99.8 87.5 88.2 91.8 95.7 95.4

Test track No. 1 is characterised by the smoothest surface, and was almost com-

pletely classified as an ISO class A road by both the classification algorithm and the

measured data. A classification result of 99.8 % was achieved. The main classific-

ation result for test track No. 2 is ISO class B. However, a third of the test track

was classified to ISO class A and C. A coincidence of 87.5 % was achieved. A similar

result shows the comparison of test track No. 3, where a coincidence of 88.2 % was

obtained. Almost 80 % of test track No. 4 was classified as ISO class B and roughly

20 % as ISO class C. A classification result of 91.8 % was achieved. In contrast, test

track No. 5 was mainly classified as ISO class C (≈60 %). A smaller part was classi-

fied as ISO class B (≈40 %). An overall classification result of 95.7 % was obtained

for this test rack. The roughest test track in the present measurement campaign is

107

4 Impact detection and experimental road roughness classification

#7: Gravel road #8: Rough unpaved roads #9: Enduro fun park

Figure 4.12 – Results of the road roughness classification for test tracks No. 7–9.

test track No. 6, which was classified as almost 80 % ISO class C, and some parts as

class B and D. An overall classification result of 95.4 % was achieved.

Figure 4.12 shows the validation results for the test tracks No. 7–9. Due to their

loose surfaces, these test tracks have no measurement data available, as discussed in

Section 4.3.2. Test track No. 7 was classified as a mixture of road classes, from ISO

class A to C. Test track No. 8 was subjectively perceived as being much rougher and

harsher than test track No. 7, as the relative comparison of the classification results

also shows. This test track was mainly classified as ISO class C–E. The highest

classification results were obtained at test track No. 9, which was mainly classified

as ISO class D–E. On the one hand, the classification results obtained reflect the

subjective feeling of roughness quite well in comparison to the roads that have been

validated with measurements. On the other hand, the relative comparison of the

classification results of the test tracks No. 7–9, represent the rider’s perception

of increasing roughness. In summary, the road roughness classification algorithm

developed, achieves classification results of more than 87 %, which is sufficient for

the purpose of road classification in terms of customer usage profiles. In contrast to

the simplified classification method discussed in Section 4.3.2, the method presented

of dividing road segments, leads to a more detailed classification result. Compared

to Table 4.1, the classification result is more detailed. This was also expected due to

the non-linear shape of the PSDs of the measured road profiles, compare Figure 4.4.

Accordingly, a single road can be characterised by different road classes at different

spatial frequencies. For this reason, the method presented of dividing road segments,

is proposed to be preferable for classifying roads and test tracks. The results from

108

4.6 Results and Discussion

the divided road segments need to be summarised to achieve a distribution of driven

distances per road class.

The differences between the road roughness classification results and the measured

road profiles, have two reasons: First, the full-vehicle model is a simplification of the

real motorcycle. This implies that nonlinearities are ignored, and reduced stiffness

and damping coefficients have been applied. Moreover, in the event of small height

differences, for example, gravel, the tyre smooths the road profile. On the contrary,

small gaps cannot be measured correctly due to the geometric dimensions of the tyre.

Furthermore, the model does not consider the influence of longitudinal dynamics on

the response signals, as discussed in Section 4.4. Second, the method of validation has

some potential error sources. The segmentation of the measured and the driven road

is not equally accurate. This means that the start and end points of the test tracks

have been determined by the GPS position of both the measurement vehicle and the

test motorcycle. This might be a source of error in the evaluation of the individual

road segments. In addition, the trajectory of the motorcycle and especially the lateral

position on the test track, is ambiguous and therefore, the correct longitudinal road

profile cannot be derived from the CRG-data. During the measurement campaign,

the rider attempted to ride in the centre of the lane as much as possible. Hence, the

longitudinal profile of the three-dimensional measurement data has also been derived

from the middle of the lane. In summary, model and systematic failures affect the

validation of the road roughness classification.

4.6.2 Validation of the impact detection

After the training set had been created, different machine learning techniques were

applied. In the terminology of machine learning, the current problem is described as

supervised machine learning. According to Suthaharan [69], in supervised learning,

all classes are known and class boundaries are well defined in the given training

set. The objective is to generate rules and to train a classifier, using the training

set, so that new, unlabelled data can be classified. This is also called classification.

Although this classification problem is trivial, it is well suited for a study on the

principles of classification algorithms. Especially because the feature space can be

plotted because the problem is two-dimensional. Hence, nine popular classification

methods were tested on the training set, as can be seen in Table 4.4. The accuracy of

109

4 Impact detection and experimental road roughness classification

Table 4.4 – Properties of classification methods.

No. Classifier Prediction speed Overfitting Accuracy

#1 Logistic Regression 1n99.60

#2 Decision Tree 1.5n99.64

#3 Discriminant Analysis (linear) 3.5n99.57

#4 Discriminant Analysis (quadratic) 4.5y98.93

#5 Naive Bayes 4.5n99.01

#6 SVM (linear) 5n99.41

#7 KNN, (k=1) 7n99.53

#8 SVM (quadratic) 8n99.76

#9 SVM (Gaussian) 15 y99.37

the classifiers was tested by the k-fold cross-validation, where the number of subsets

was set to k= 5. In machine learning, the problem with validation is the division of

the data set into a training set and a testing set. There is a target conflict between

the maximum amount of information in the training set to provide the best possible

training data, and the maximum of testing data to evaluate the performance of the

trained classifier. Cross-validation partitions the data set into kfolds, and for each

fold, a model is trained using the out-of-fold observations. The performance is tested

using the in-fold data. This procedure is repeated ktimes, and finally, the average

performance over all folds is calculated. This method has the advantage that all

observations are used for training and testing. Cross-validation is only used for the

evaluation of the classifier, in terms of accuracy. The final classifier was trained using

the complete training set. The classifiers are ordered depending on the normalised

prediction speed based on the fastest algorithm. Hence, the fastest classifier is the

logistic regression.

Overfitting is a phenomenon in which the classifier models the training data too

well, so that the performance on new data decreases. In the present study, the

problem of overfitting is handled by the so-called decision surfaces, where the solu-

tion space is evaluated by the expectation and experience of the user. Hence, the

benchmark shape was plotted in the form of decision surfaces, as illustrated in Fig-

ure 4.13. According to the authors’ expectation, a threshold on both features x1

and x2must be exceeded to distinguish between service loads and special events. In

addition, severe special events are also distinguished from mild special events by the

exceeding of thresholds. The boundaries need not necessarily be in linear form, they

could also be in radial form. This benchmark shape helps to evaluate the problem

110

4.6 Results and Discussion

#0: Benchmark shape

#1: Logistic Regression #2: Decision Tree #3: Linear Discriminant

#4: Quadratic Discriminant #5: Naive Bayes #6: SVM (linear)

#7: KNN, (k=1) #8: SVM (quadratic) #9: SVM (Gaussian)

Figure 4.13 – Decision surfaces for different classifiers.

of overfitting and the treatment of future data. It is more an orientation for the user

than a nominal target. Since all the classifiers achieved an accuracy of almost 100 %,

further investigations have been carried out. Due to the two-dimensional problem,

it is possible to plot the decision surfaces for all the classifiers, see Figure 4.13. This

111

4 Impact detection and experimental road roughness classification

opportunity offers a benefit, because the decision surfaces show graphically how the

classifiers separate the observations. In addition, the treatment of future data can

be estimated. First, a multinomial logistic regression was tested, which is fast and

easy to implement. The algorithm calculates probabilities for each observation. The

linear behaviour of the classifier is readily identifiable. The decision surfaces show a

good separation of the classes. Second, a coarse binary decision tree with a maximum

number of four leafs, was trained. Decision trees are easy to interpret and make fast

predictions. The boundaries of the decision surfaces are parallel to the coordinate

system, which is characteristic for decision trees. This classifier comes closest to the

benchmark shape.

Third, a linear and a quadratic discriminant analysis were tested. The algorithms

assume that different classes generate data based on different Gaussian distributions.

Although these algorithms make fast predictions, the decision surfaces are not sat-

isfactory. For the linear discriminant analysis, the boundary of severe special events

is too close to the mild special events. Overfitting occurs for the quadratic discrim-

inant analysis, which means that the classifier is trained specifically for the training

set, but unknown data could be incorrectly classified. Next, a naive Bayes classifier

was tested, which is based on the Bayes’ theorem. It assumes independence between

features. In the present study, the features are highly correlated, but nevertheless

the classifier shows a good performance. The shape of the decision surfaces are sat-

isfactory. Subsequently, the family of support vector machines (SVM) was evaluated

with different kernel functions. SVMs classify data by finding the hyperplane that

separates the data with the largest margin between the classes. The linear imple-

mentation shows a good separation between the classes, while the space for special

events is too large at the quadratic implementation. The implementation with the

Gaussian kernel function is characterised by strong overfitting, which can be seen

on the small islands to detect severe special events. However, SVMs show a me-

dium to slow prediction performance, especially for multi-class problems. Finally,

the k-Nearest-Neighbour classifier was evaluated with k= 1. This algorithm tries to

classify observations depending on the kclosest points. This technique has a medium

prediction speed compared to the others. Kotsiantis et al. [70] provide a review of

classification methods.

The authors decided to implement the binary decision tree for the prediction of

112

4.6 Results and Discussion

Table 4.5 – Confusion matrix of binary decision tree.

Actual Class

Service loads Mild special events Severe special events TPR (%)

Prediction

Service loads 2456 6 0 99.8

Mild special events 3 40 0 93.0

Severe special events 0 0 6 100.0

PPV (%) 99.9 87.0 100 99.6

future observations. The classifier shows a high accuracy after k-fold cross-validation

and the decision surfaces clearly separate the classes. In addition, the classifier rep-

resents the benchmark shape and is fast on the prediction of future data. After the

5-fold cross-validation, the algorithm shows an accuracy of 99.6 %, as can be seen in

Table 4.5. The numbers within table represent the number of observations. The last

column shows the True Positive Rate (TPR) and the last row shows the Positive Pre-

dictive Value (PPV). For example, 87 % of the actual mild special events have been

predicted correctly and 93 % of the predicted mild special events are actually true. It

is important, that severe special events are predicted and classified 100 % correctly,

since they have the highest impact on the vehicle. The losses between service loads

and mild special events are acceptable, since the scatter plot also shows an ambigu-

ous assignment. However, this behaviour was expected. Figure 4.14 illustrates the

final binary decision tree with four leaves.

𝑥2< 2.21 𝑥2≥ 2.21

𝑥1< 3.11 𝑥1≥ 3.11 𝑥1≥ 9.32𝑥1< 9.32

Service

loads

Mild

special event

Mild

special event

Severe

special event

Figure 4.14 – Binary decision tree.

In summary, a sliding window buffers the angular front wheel velocity ωand the

113

4 Impact detection and experimental road roughness classification

variation in suspension travel sft for a time span of ∆tbuf = 1 s. Subsequently, the

time derivatives of these two signals are calculated. The peak-to-peak values are

extracted and normalised by the z-score, to provide the input variables x1and x2.

The predictor is a trained classifier of decision tree type, and continuously classifies

the observations into service loads, mild special events, and severe special events.

The resultant classifier is based on the training set achieved from the test mo-

torcycle riding with different velocities over the road obstacles. The application of

the classifier is dedicated to this experimental set-up. Varying velocities are covered

by the generation of the training set, whereas a different type of motorcycle would

probably lead to incorrect classification results. Thus, it is recommended to gen-

erate individual training sets for other types of motorcycles. Both measuring the

front suspension signal as well as riding over the presented obstacles is part of the

endurance tests at the product validation. Thus, there is no additional effort except

of labelling the special events.

4.7 Conclusion

The objective of this research was the experimental validation of a road roughness

classification method and the development of an impact detection strategy, in the

context of durability engineering. Both methods were validated with the help of a

measurement campaign. Nine test tracks and six obstacles were evaluated, while the

test tracks had been surveyed by a laser scanner in a previous BMW study. The

methods developed utilise the vehicle’s onboard signals and are therefore suitable to

collect customer usage profiles in terms of field data.

The road roughness classification algorithm was experimentally validated, after the

proof of concept had been achieved theoretically in a previous publication by Gorges

et al. [1]. The method was extended to work under real operating conditions. The

results show an accuracy of more than 87 % on the test tracks. Furthermore, the

validation method shows that the evaluation of divided road segments leads to a

more detailed classification result, compared to the classification results achieved

by processing the complete road profile. This fact confirms the application of the

developed road roughness classification method, because in the presented real-time

implementation, the road roughness is also processed in small road segments de-

pending on the velocity. In summary, the road roughness classification algorithm

114

4.7 Conclusion

was proved to achieve adequate results on real test tracks. The methodology was

experimentally validated, even on rough, unpaved roads, and can be implemented

into new products to collect customer usage profiles. The presented algorithm re-

quires a linearised full-vehicle model of the motorcycle, the velocity, and at least one

suspension deflection signal, as inputs.

The road roughness classification system continuously classifies the road rough-

ness, independently of the velocity and the vehicle. From the viewpoint of durabil-

ity, only the combination of road roughness and vehicle velocity determines whether

the induced loads are service loads, special events, or misuse events. Therefore, the

impact detection strategy was developed. Based on the same input signals, velocity

and spring deflection, a classifier was trained for detecting special events. Different

road obstacles were ridden over with different velocities to generate a broadly dis-

tributed data set. The test rider divided the events into two classes: mild special

events and severe special events. The peak-to-peak values of the time derivatives of

the angular front wheel velocity and the suspension travel, were chosen as features.

After normalisation of the features, different machine learning techniques were eval-

uated on the given training set, in terms of prediction speed, accuracy, overfitting,

and shape of the decision surfaces. The decision tree was chosen, with an accuracy

of 99.6 %. The validation was performed by a k-fold cross-validation. The results

show that a distinction between service loads, mild special events, and severe special

events is possible. The impact detection strategy can easily be appended to the ex-

isting road roughness classification system, since the algorithm works with the same

onboard signals and the same sliding window, as the road roughness classification.

A schematic flowchart of the impact detection strategy is provided in Figure 4.15.

Further research should investigate the robustness of the classifier against changes of

the vehicle attributes; for example, changes of the vehicle weight or manipulations

of the suspension and tyre characteristics.

All of the algorithms developed work in real time. Thus, an implementation into

existing electronic control units is possible. The distribution of driven road classes

improves vehicle design targets. Furthermore, it enables a virtual load acquisition

and makes the generation of synthetic road profiles possible. In addition, the amount

and strength of special events helps to define maximum loads. The combination of

distributed road classes and special events, completes the description of customer

115

4 Impact detection and experimental road roughness classification

Onboard Signals

𝑥1, 𝑥2

Feature Extraction Prediction

𝑡buf

Figure 4.15 – Flowchart of the impact detection strategy.

usage profiles. Further research should investigate the possibilities of a real-time ap-

plication for the information gathered. For example, the knowledge of severe special

events could enable applications for predicted maintenance. Moreover, the statist-

ical treatment of the gathered field data and the derivation of design loads, needs

investigation. Finally, the interaction between field data collection and a virtual load

acquisition, must be examined from the viewpoint of the vehicle design process.

References

[1] C. Gorges, K. Öztürk and R. Liebich. ‘Road classification for two-wheeled

vehicles’. In: Vehicle System Dynamics 56.8 (2018), pp. 1289–1314.

[2] ISO. Mechanical vibration - Road surface profiles - Reporting of measured data.

Standard No. 8608:2016. Geneva, Switzerland: ISO, 2016.

[3] C. Matz. ‘Online Berechnung von Fahrwerkskräften auf Basis von Onboard-

Sensorik’. PhD thesis. Clausthal: TU Clausthal, 2015.

[4] K. Pötter. ‘Structural durability and integrity in vehicle design’. In: 18th European

Conference on Fracture. Dresden, Germany: DVM, Aug. 2010.

[5] P. Johannesson and M. Speckert, eds. Guide to load analysis for durability in

vehicle engineering. Chichester: Wiley, 2013, p. 434.

[6] M. Köhler, S. Jenne, K. Pötter et al. Load Assumption for Fatigue Design of

Structures and Components. Berlin, Heidelberg: Springer, 2017, pp. 1–226.

[7] M. Hauke. ‘Simulation des Missbrauchverhaltens von Gesamtfahrzeugen’. PhD

thesis. Munich, Germany: TU München, 2004.

[8] C. Berger, K.-G. Eulitz, P. Heuler et al. ‘Betriebsfestigkeit in Germany – an

overview’. In: International Journal of Fatigue 24.6 (June 2002), pp. 603–625.

116

References

[9] C. Gorges, K. Öztürk and R. Liebich. ‘Customer loads of two-wheeled vehicles’.

In: Vehicle System Dynamics 55.12 (2017), pp. 1842–1864.

[10] A. González, E. J. O’brien, Y.-Y. Li et al. ‘The use of vehicle acceleration

measurements to estimate road roughness’. In: Vehicle System Dynamics 46.6

(2008), pp. 483–499.

[11] N. K. Harris, A. González, E. J. O’Brien et al. ‘Characterisation of pavement

profile heights using accelerometer readings and a combinatorial optimisation

technique’. In: Journal of Sound and Vibration 329.5 (2010), pp. 497–508.

[12] H. M. Ngwangwa, P. S. Heyns, F. J. J. Labuschagne et al. ‘Reconstruction of

road defects and road roughness classification using vehicle responses with ar-

tificial neural networks simulation’. In: J. Terramechanics 47.2 (2010), pp. 97–

111.

[13] M. Doumiati, A. Victorino, A. Charara et al. ‘Estimation of road profile for

vehicle dynamics motion: experimental validation’. In: American Control Con-

ference. San Francisco: IEEE, June 2011, pp. 5237–5242.

[14] W. Yu, X. Zhang, K. Guo et al. ‘Adaptive real-time estimation on road dis-

turbances properties considering load variation via vehicle vertical dynamics’.

In: Math. Probl. Eng. 2013 (2013), pp. 1–9.

[15] M. Doumiati, S. Erhart, J. Martinez et al. ‘Adaptive control scheme for road

profile estimation: application to vehicle dynamics’. In: The International Fed-

eration of Automatic Control. Cape Town: IFAC, Aug. 2014, pp. 8445–8450.

[16] M. Doumiati et al. ‘Road profile estimation using an adaptive Youla–Kučera

parametric observer: Comparison to real profilers’. In: Control Engineering

Practice 61.2002 (Apr. 2017), pp. 270–278.

[17] J. C. Tudón-Martínez, S. Fergani, O. Sename et al. ‘Online road profile estim-

ation in automotive vehicles’. In: European Control Conference (ECC). Stras-

bourg: ECC, June 2014, pp. 2370–2375.

[18] J. C. Tudón-Martínez, S. Fergani, O. Sename et al. ‘Adaptive road profile

estimation in semiactive car suspensions’. In: IEEE Transactions on Control

Systems Technology 23.6 (2015), pp. 2293–2305.

[19] Z. Wang et al. ‘Road profile estimation for suspension system based on the

minimum model error criterion combined with a Kalman filter’. In: J. Vibroen-

gineering 19.6 (Sept. 2017), pp. 4550–4572.

[20] Z. Wang et al. ‘Suspension system state estimation using adaptive Kalman

filtering based on road classification’. In: Vehicle System Dynamics 55.3 (Mar.

2017), pp. 371–398.

117

4 Impact detection and experimental road roughness classification

[21] Y. Qin et al. ‘Road excitation classification for semi-active suspension system

with deep neural networks’. In: J. Intell. Fuzzy Syst. 33.3 (Aug. 2017), pp. 1907–

1918.

[22] Y. Qin et al. ‘Road profile estimation for semi-active suspension using an ad-

aptive Kalman filter and an adaptive super-twisting observer’. In: 2017 Am.

Control Conf. Seattle, USA: IEEE, May 2017, pp. 973–978.

[23] Y. Qin et al. ‘Road excitation classification for semi-active suspension system

based on system response’. In: Journal of Vibration and Control (2017), pp. 1–

17.

[24] D. Ben Hassen et al. ‘Road profile estimation using the dynamic responses of

the full vehicle model’. In: Applied Accoustics May (Dec. 2017), pp. 0–1.

[25] M. Burger. ‘Calculating road input data for vehicle simulation’. In: Multibody

Syst. Dyn. 31.1 (2013), pp. 93–110.

[26] W. Fauriat, C. Mattrand, N. Gayton et al. ‘Estimation of road profile variability

from measured vehicle responses’. In: Vehicle System Dynamics 54.5 (2016),

pp. 585–605.

[27] M. Yousefzadeh, S. Azadi and A. Soltani. ‘Road profile estimation using neural

network algorithm’. In: Journal of Mechanical Science and Technology 24.3

(2010), pp. 743–754.

[28] H. Imine, Y. Delanne and N. K. M’Sirdi. ‘Road profile input estimation in

vehicle dynamics simulation’. In: Vehicle System Dynamics 44.4 (2006), pp. 285–

303.

[29] J. J. Rath, K. C. Veluvolu and M. Defoort. ‘Simultaneous estimation of road

profile and tire road friction for automotive vehicle’. In: IEEE Transactions on

Vehicular Technology 64.10 (2015), pp. 4461–4471.

[30] W. Jeong, K. Yoshida, H. Kobayashi et al. ‘State estimation of road surface and

vehicle system using a Kalman filter’. In: JSME Int. J. 33.4 (1990), pp. 528–

534.

[31] T. P. Nordberg. ‘An iterative approach to road/profile identification utilizing

wavelet parameterization’. In: Vehicle System Dynamics 42.6 (2004), pp. 413–

432.

[32] Y. Qin, M. Dong, F. Zhao et al. ‘Road profile classification for vehicle semi-

active suspension system based on adaptive neuro-fuzzy inference system’. In:

Conference on Decision and Control (CDC). Osaka: IEEE, Dec. 2015, pp. 1533–

1538.

118

References

[33] Y. Qin, M. Dong, R. Langari et al. ‘Adaptive hybrid control of vehicle semiact-

ive suspension based on road profile estimation’. In: Shock Vib. 2015.1 (2015),

pp. 1–13.

[34] A. Solhmirzaei, S. Azadi and R. Kazemi. ‘Road profile estimation using wavelet

neural network and 7-DOF vehicle dynamic systems’. In: Journal of Mechanical

Science and Technology 26.10 (2012), pp. 3029–3036.

[35] R. S. Barbosa. ‘Vehicle dynamic response due to pavement roughness’. In: J.

Brazilian Soc. Mech. Sci. Eng. XXXIII.3 (2011), pp. 302–307.

[36] R. S. Barbosa. ‘Vehicle dynamic safety in measured rough pavement’. In:

Journal of Transportation Engineering 137.5 (2011), pp. 305–310.

[37] R. S. Barbosa. ‘Vehicle vibration response subjected to longwave measured

pavement irregularity’. In: J. Mech. Eng. Autom. 2.2 (2012), pp. 17–24.

[38] K. De Zoysa, C. Keppitiyagama, G. P. Seneviratne et al. ‘A public transport

system based sensor network for road surface condition monitoring’. In: Proc.

2007 Work. Networked Syst. Dev. Reg. - NSDR ’07. New York, USA: ACM

Press, 2007.

[39] J. Eriksson, L. Girod, B. Hull et al. ‘The pothole patrol: Using a Mobile Sensor

Network for Road Surface Monitoring’. In: Proceeding 6th Int. Conf. Mob. Syst.

Appl. Serv. - MobiSys ’08. New York, USA: ACM Press, 2008.

[40] P. Mohan, V. N. Padmanabhan and R. Ramjee. ‘Nericell: Rich Monitoring of

Road and Traffic Conditions using Mobile Smartphones’. In: Proc. 6th ACM

Conf. Embed. Netw. Sens. Syst. - SenSys ’08. New York, USA: ACM Press,

2008.

[41] M. Perttunen et al. ‘Distributed Road Surface Condition Monitoring Using

Mobile Phones’. In: Lect. Notes Comput. Sci. Ed. by C.-H. Hsu et al. Springer-

Verlag Berlin Heidelberg, 2011, pp. 64–78.

[42] Y. Tai, C. Chan and J. Hsu. ‘Automatic road anomaly detection using smart

mobile device’. In: Conference on Artificial Intelligence and Applications (TAAI).

Hsinchu, Taiwan, Nov. 2010.

[43] A. Mednis, G. Strazdins, R. Zviedris et al. ‘Real time pothole detection using

Android smartphones with accelerometers’. In: International Conference on

Distributed Computing in Sensor Systems and Workshops (DCOSS). IEEE,

June 2011, pp. 1–6.

[44] F. Carrera, S. Guerin and J. B. Thorp. ‘BY THE PEOPLE, FOR THE PEOPLE:

THE CROWDSOURCING OF "STREETBUMP": AN AUTOMATIC POTHOLE

MAPPING APP’. In: ISPRS - Int. Arch. Photogramm. Remote Sens. Spat. Inf.

Sci. XL-4/W1.4W1 (May 2013), pp. 19–23.

119

4 Impact detection and experimental road roughness classification

[45] M. Hoffmann, M. Mock and M. May. ‘Road-quality classification and bump

detection with bicycle-mounted smartphones’. In: UDM’13 Proc. 3rd Int. Conf.

Ubiquitous Data Min. Vol. 1088. Beijing, China, 2013, pp. 39–43.

[46] H. Sharma, S. Naik, A. Jain et al. ‘S-road assist: Road surface conditions and

driving behavior analysis using smartphones’. In: 2015 International Confer-

ence on Connected Vehicles and Expo (ICCVE). IEEE, Oct. 2015, pp. 291–

296.

[47] H.-w. Wang, C.-h. Chen, D.-y. Cheng et al. ‘A Real-Time Pothole Detection

Approach for Intelligent Transportation System’. In: Mathematical Problems

in Engineering 2015 (2015), pp. 1–7.

[48] F. Seraj, B. J. van der Zwaag, A. Dilo et al. ‘RoADS: A Road Pavement Mon-

itoring System for Anomaly Detection Using Smart Phones’. In: Lect. Notes

Comput. Sci. Ed. by M. Atzmueller et al. Vol. 9546. Lecture Notes in Computer

Science. Springer International Publishing, 2016, pp. 128–146.

[49] H. P. M. and V. P. Gopi. ‘Vehicle Vibration Signal Processing for Road Surface

Monitoring’. In: IEEE Sens. J. 17.16 (Aug. 2017), pp. 5192–5197.

[50] A. Fox, B. V. Kumar, J. Chen et al. ‘Multi-Lane Pothole Detection from Crowd-

sourced Undersampled Vehicle Sensor Data’. In: IEEE Trans. Mob. Comput.

16.12 (Dec. 2017), pp. 3417–3430.

[51] F. Cong, H. Hautakangas, J. Nieminen et al. ‘Applying wavelet packet decom-

position and one-class support vector machine on vehicle acceleration traces

for road anomaly detection’. In: Lecture Notes in Computer Science 7951

LNCS.PART 1 (2013), pp. 291–299.

[52] Z. Li, I. Kolmanovsky, E. Atkins et al. ‘Road anomaly estimation: Model based

pothole detection’. In: 2015 Am. Control Conf. Vol. 2015-July. IEEE, July

2015, pp. 1315–1320.

[53] S. Hegde, H. Mekali and G. Varaprasad. ‘Pothole detection and inter vehicular

communication’. In: 2014 IEEE Int. Conf. Veh. Electron. Saf. IEEE, Dec. 2014,

pp. 84–87.

[54] R. Madli, S. Hebbar, P. Pattar et al. ‘Automatic Detection and Notification of

Potholes and Humps on Roads to Aid Drivers’. In: IEEE Sens. J. 15.8 (Aug.

2015), pp. 4313–4318.

[55] Z. Zhang, X. Ai, C. K. Chan et al. ‘An efficient algorithm for pothole detection

using stereo vision’. In: 2014 IEEE Int. Conf. Acoust. Speech Signal Process.

IEEE, May 2014, pp. 564–568.

[56] Y. Jo and S. Ryu. ‘Pothole Detection System Using a Black-box Camera’. In:

Sensors 15.12 (Nov. 2015), pp. 29316–29331.

120

References

[57] D.-W. Jang and R.-H. Park. ‘Pothole detection using spatio-temporal saliency’.

In: IET Intell. Transp. Syst. 10.9 (Nov. 2016), pp. 605–612.

[58] P. Wang, Y. Hu, Y. Dai et al. ‘Asphalt Pavement Pothole Detection and Seg-

mentation Based on Wavelet Energy Field’. In: Math. Probl. Eng. 2017 (2017),

pp. 1–13.

[59] A. Tedeschi and F. Benedetto. ‘A real-time automatic pavement crack and

pothole recognition system for mobile Android-based devices’. In: Adv. Eng.

Informatics 32 (Apr. 2017), pp. 11–25.

[60] B.-h. Kang and S.-i. Choi. ‘Pothole detection system using 2D LiDAR and

camera’. In: 2017 Ninth Int. Conf. Ubiquitous Futur. Networks. IEEE, July

2017, pp. 744–746.

[61] M. W. Sayers and S. M. Karamihas. The little book of profiling. Michigan:

University of Michigan, 1998.

[62] E. B. Spangler and W. J. Kelly. GMR road profilometer: a method for meas-

uring road profile. Tech. rep. GMR-452. Warren, Michigan: General Motors

Corporation Research Laboratories, 1964.

[63] R. La Barre, R. Forbes and S. Andrew. The Measurement and Analysis of Road

Surface Roughness. Tech. rep. 1970/5. Nuneaton, Warwickshire (UK): MIRA

(The Motor Industry Research Association), 1969.

[64] H. Gimmler, D. Ammon and J. Rauh. Road profiles: mobile measurement, data

processing for efficient simulation and assessment of road properties. Tech. rep.

VDI-Report No. 1912. Dusseldorf, Germany: VDI, 2005.

[65] J. Rauh and M. Mössner-Beigel. ‘Tyre simulation challenges’. In: Vehicle Sys-

tem Dynamics 46.sup1 (2008), pp. 49–62.

[66] P. Andrén. ‘Power spectral density approximations of longitudinal road pro-

files’. In: International Journal of Vehicle Design 40.1/2/3 (2006), pp. 2–14.

[67] V. Cossalter. Motorcycle Dynamics. 2nd ed. [place unknown]: LULU, 2006,

p. 372.

[68] M. Tanelli, M. Corno and S. M. Savaresi, eds. Modelling, simulation and control

of two-wheeled vehicles. Chichester: Wiley, 2014, pp. 1–348.

[69] S. Suthaharan. Machine Learning Models and Algorithms for Big Data Clas-

sification. Vol. 36. Integrated Series in Information Systems. Boston, MA:

Springer US, 2016, p. 364.

[70] S. B. Kotsiantis, I. D. Zaharakis and P. E. Pintelas. ‘Machine learning: a review

of classification and combining techniques’. In: Artif. Intell. Rev. 26.3 (Nov.

2006), pp. 159–190.

121

5 Discussion

5.1 Summary of the developed methods

This cumulative PhD thesis shows the development and validation of methods for

identifying customer usage profiles of two-wheeled vehicles with the onboard sensors.

The main idea of collecting field-data for the purpose of improving the vehicle de-

velopment was proven to be feasible and produces reliable results. The methods de-

veloped can now be implemented into mass production vehicles for collecting crowd-

sourced data. The detailed information about customer usage profiles will improve

the entire vehicle design process, beginning at the definition of vehicle requirements,

and ending at the validation of the complete vehicle, as introduced in Section 1.1.

The first publication [1] shows the calculation and subsequent counting of the

wheel forces with the onboard signals. Moreover, the estimation of the road slope

and the total vehicle mass was achieved. The vehicle loading is of major importance

and has not yet been estimated on two-wheeled vehicles. On the small scale, the

knowledge about vehicle loading enables further online applications; for example, a

pre-load of the suspension spring. On a large scale, the distribution of the vehicle

loading reveals the amount of rides achieved with an additional passenger, which is

not simply a question of occurring loads, but also comfort and safety. The same

benefits follow from the knowledge of the current road slope and the distribution of

elevation gains considering all observed vehicles. A road slope estimation has also not

yet been implemented into two-wheeled vehicles. Finally, the online calculation and

subsequent counting of occurring wheel forces is a novelty in durability engineering of

two-wheeled vehicles. The comparison between the simulation results and the wheel

forces measured with wheel-load transducers shows a high coincidence. As a result,

the collected rainflow matrices can directly be used to derive verification loads on

full-vehicle test beds. The methodology of calculating the wheel forces directly from

the resultant motions is fast in terms of computation time, since the operations are

123

5 Discussion

basic math.

The second publication [2] presents methods for the vehicle-independent road clas-

sification with onboard signals. The road can be classified in terms of curviness,

hilliness and - most importantly - road roughness according to ISO 8608 [24]. These

road characteristics can be revealed independent of the vehicle and are of general

interest throughout the vehicle development process. The publication shows how

road curves can be detected and subsequently classified. Beside online applications,

the distribution of driven road curves helps manufactures to chose appropriate test

tracks for measurements or endurance testing. The travelled curviness can also be

used to reconsider product placement and vehicle requirements. In addition, the pub-

lication shows the classification of driven hilliness and elevation gains, which serve

the same purpose. Finally, the information about road roughness could be used to

provide driving assistance systems with the information about the current road un-

evenness; for example, active damping control systems. In the context of durability,

road roughness causes vibrations and additional dynamic wheel loads. The distribu-

tion of different road classes helps engineers choosing the appropriate quality of test

tracks. The VMC project from Speckert et al. [25] aims at the planning of measure-

ment campaigns with the same three road characteristics: curviness, hilliness and

road roughness. In addition, product requirements can be improved with a detailed

knowledge of the driven road classes. A comparison of different regions and seg-

ments could also be used to derive market-specific vehicle requirements. Moreover,

knowledge about the road characteristics enables a virtual load acquisition, where

customer loads are simulated on virtual test tracks. Thus, a detailed multi-body

system of a motorcycle can be simulated on a given mix of different road classes to

obtain information about the occurring loads before any expensive prototype must be

builded up. The implementation of the curviness and hilliness classification is trivial

once an inertial measurement unit (IMU) is part of the onboard signals, whereas

the estimation of the current road roughness is a novelty on two-wheeled vehicles.

Especially the use of velocity-dependent transfer functions in combination with a

relatively short sliding window has not yet been achieved. Furthermore, the method

presented is designed as a modular road roughness estimator, which can be adapted

to the available onboard signals; for example, spring deflection or any other response

signal of the full-vehicle model. Therefore, it can also be extended to the use on pas-

124

5.1 Summary of the developed methods

senger cars. Put simply, the second publication shows the theoretical derivation of

the road roughness classification method and a validation approach using a numerical

full-vehicle model.

The third publication [3] shows the experimental validation of the road rough-

ness classification method. Thus, a measurement campaign was achieved, where

the test motorcycle rode the same road segments as a laser-scanning road profilo-

meter, ranging from smooth high-speed test tracks to rough off-road tracks. The

validation shows that the road roughness estimator provides reliable results with an

adequate accuracy. The algorithm detects a different road class within the sliding

window of ∆tbuf = 1 s, which in principle enables further online applications. The

methodology of using transfer functions and estimating the road roughness directly

in the frequency domain is straightforward, since the classification by ISO 8608 is

also described in the frequency domain. The approach presented is thus considered

as pragmatic and target-oriented. As the road surface becomes rougher and from

the viewpoint of durability, a combination of road obstacles and vehicle velocity

leads to the occurrence of special events, as introduced in Section 1.2. Thus, the

measurement campaign also included the passing of obstacles with different vehicle

velocities. A machine learning (ML) approach was used on the labelled test set of

special events to construct a classifier that can distinguish between service loads,

mild- and severe special events. The suspension travel and the front wheel velocity

were used as input. The procedure of data labelling, feature extraction, generation

of a test set, and subsequently exploring well-known classifiers, is state of the art in

artificial intelligence. The novelty of this publication is the generation of a new test

set for ML algorithms in the context of durability. The choice of a ML approach

seams reasonable, since a direct modelling of the dynamic physical processes during

the impact would require a non-linear model of the front suspension system and -

most importantly - the tire. In addition, the labelling of the events to service loads

and two kinds of special events could only be undertaken by the rider, due to the

statistical nature and the ambiguous separation between service loads and special

events. The result is thus also dependent on the test rider, who should be aware of

product liability and durability. Moreover, he should anticipate the customer percep-

tion of mild and severe special events. Since every product has different suspension

and tire specifications, the approach of ML seems fast to achieve and can easily be

125

5 Discussion

implemented into the vehicle development process. However, the procedure requires

a well labelled test set, as is usual at the application of supervised ML techniques.

To sum up, the methods presented enable a holistic approach of identifying cus-

tomer usage profiles with the onboard signals. Complete customer usage profiles

can be collected with these methods, once the algorithms have been implemented in

mass production vehicles. The main research question of the present research project

was the investigation of methods to acquire the customer usage profiles, which has

successfully been answered. Once this crowd-sourced data has been collected, it is

of major importance to implement a reasonable usage of this data into the vehicle

development process. This is discussed by the next Section. Within this industry-

sponsored PhD program, the following patent applications have been submitted and

are the property of BMW:

•Road curve characteristics PA 2016211739 DE

•Mass estimation PA 2017209746 DE

•Road slope estimation PA 2017209747 DE

•Road roughness classification PA 2017219767 DE

5.2 Implementation into the vehicle development process

The process of capturing value out of data can be described by the information

value loop, see Figure 5.1. It is often referred by the description of the Internet of

Things (IoT); for example, by Deloitte [26]. In the present context, the information

value loop is used to describe how the methods developed for identifying customer

usage profiles can be implemented into the vehicle development process. According

to Raynor and Cotteleer [26], create means the use of sensors to create information

about a physical system or state. communicate represents the transmission of data

from one place to another. aggregate stands for the gathering of information created

at different times or from different sources. It also includes the transformation of

raw data to high-quality data. analyse describes the process of data analysis with

the target of data insight that leads to descriptions, predictions or prescriptions for

action. Finally, act means initiating, maintaining, or changing a physical event or

state, and in the present context, making data-driven decisions. It is important

126

5.2 Implementation into the vehicle development process

ACT

CREATE

COMMUNICATEAGGREGATE

ANALYSE

Figure 5.1 – Information value loop.

to consider these steps not as sequential tasks, but more as a closed loop. The

loosely process of arbitrary data generation and later data usage has shown major

disadvantages in the past. Therefore, it is convenient to start with the act step of

asking the right questions for solving the underlying problem. Once the purpose of

the data is identified, the correct acquisition and subsequent analysis of this data

can be achieved. The following sub-sections describe the different steps of gathering

customer usage profiles.

Act: Defining the purpose

In the present context, the purpose of the data, which is, the customer usage profiles,

is improving the vehicle development process by validating the load assumptions. As

introduced in Section 1.2, from the viewpoint of durability, the need for customer

usage profiles is given. Concerning the vehicle development process, the impact

of customer usage profiles covers the definition of initial vehicle specifications, the

127

5 Discussion

derivation of vehicle requirements, the virtual load acquisition and the validation of

the vehicle and its components, see Figure 1.1.

Create: Data acquisition

First, the customer usage profiles must be acquired on a small scale, which means

the methods of gathering customer usage profiles must be implemented into new

vehicles. This is the most difficult part, since both the development and the actual

implementation of the methods are time- and cost-intensive. Moreover, it is often

not possible to calculate a profitable business case for an application that needs CPU

capacity, memory and it does not seem to improve the customer experience upon first

glance. However, referring to the judgement of Wixom and Ross (see Section 5.5),

“using data to improve operational processes and boost decision-making quality may

not be the most glamorous path to monetising data, but it is the most immediate”

[27]. This rethinking from old-fashioned business case-triggered projects to forward-

looking data collecting projects is part of the digital transformation and concerning

the automotive industry it has only just begun.

Since the implementation of the methods collecting the customer usage profiles

is part of a traditional function development for electronic control units (ECUs),

it follows the same steps as the vehicle development process. At the beginning,

general requirements concerning the customer usage profiles must be defined. This

comprises the choice of the features and the classification scheme, which can be

multi-dimensional; for example, vehicle velocity versus engine speed. In addition,

the resolution of the classification results, which is the bin width, must be defined. It

is convenient to collect mutual requirements for the data classification together with

other departments and company divisions, since most of the data output is generally

also valuable for others. After the requirements have been defined as detailed as

possible, the algorithms can be implemented.

It is important to validate the proper function of the data collecting methods to

guarantee a high quality of data. Moreover, once the algorithms are implemented into

ECUs, it is cost-intensive to improve them by software updates. In addition, data

analysis is highly dependent on the data basis. In the past, the most time-consuming

part of data analysis was the data preparation and the detection of faulty data sets,

including data filtering, outlier detection and plausibility check. In summary, the

128

5.2 Implementation into the vehicle development process

detailed definition of data requirements, and the proper implementation including

validation and verification, is necessary for an efficient data acquisition.

Communicate: Data transfer

The transmission of the customer usage profiles to the manufacturer can be achieved

by automatic read-out procedures during workshop appointments, or by wireless

communication via mobile connection. Thus, the data is transferred from the vehicles

to a manufacturer’s database.

Aggregate: Data aggregation

Once the raw data of each individual vehicle has been transferred to the database,

it must be aggregated and improved to result in a high-quality data set. The ag-

gregation from the individual data sets to customer usage profiles is regarded as the

transformation from the small to the large scale. This is usually covered by the

following steps: data cleaning, data filtering, data joining, calculation of enhanced

values, data clustering, interpolating of time stamps and adding of meta data. The

result is often referred as a prepared data set which now includes the entirety of

observed vehicles.

Analyse: Generating insight

The analysis of the prepared data set generates insight and value. It comprises

descriptive statistics, unsupervised ML techniques such as clustering, supervised ML

such as classification, pattern recognition, as well as many other statistical methods

depending on the purpose. The next section provides some examples regarding

statistics.

Act: Improving the vehicle development process

In this research, the value of the customer usage profiles is improving the vehicle

development process, ranging from the design stage to the validation procedure.

Each feature of the customer usage profiles addresses a different need, as discussed

by the single publications.

129

5 Discussion

Lateral Y

Longitudinal X

Vertical Z

Brake force

Figure 5.2 – Durability test rig for motorcycles c

BMW AG.

One example of the application to the vehicle validation procedure is the usage of

the rainflow matrices resulting from the wheel force calculation. Figure 5.2 shows

a durability test rig of BMW, where the wheel forces of the rear wheel are applied

to the frame of the vehicle. Since the test rig is limited regarding the amplitudes

and frequency of the wheel forces, an iteration process is achieved, which compares

the resulting rainflow matrices of the wheel forces to the pre-defined load targets.

In the past, these load targets were defined by measurements with predecessors. By

using the presented wheel force calculation method, these load targets can directly

be derived from the rainflow matrices of the customer usage profiles given a specific

quantile; for example, the 99 % quantile. More information about the application

and measurement of wheel forces can be found by Kuchler and Schrupp [28].

130

5.3 Statistical considerations

Figure 5.3 – Annual mileage of a population of vehicles.

5.3 Statistical considerations

The statistical treatment of crowd-sourced data is not trivial. Since the methods

presented have not yet been implemented into new vehicles, some statistical consid-

erations are presented using already observed data. A feature containing just one

value per vehicle can easily be presented by its probability distribution; for example,

mileage or operating hours. Figure 5.3 shows the probability distribution of the an-

nual mileage of a population of vehicles, whereby the annual mileage is representative

for any other single value feature. The distribution can be represented by a bar plot,

whereby the bin width can be chosen at one’s pleasure. The bar plot identifies local

extrema and illustrates the shape of the underlying distribution. In addition, the

cumulative probability is shown with a red line. This representation helps identifying

individual quantiles and helps answering questions such as “80 % of the vehicles have

an annual mileage less then xkilometres per year”.

The most common quantiles are illustrated with the green, yellow and red solid

vertical lines, nominally, the 50 %, the 95 %, and the 99 % quantile. The 50 % quantile

131

5 Discussion

is defined to be the median, which means that half of the vehicles drive less, and

half of the vehicles drive more than this value. The median is often used to obtain

a general assessment for the regarded feature and for comparing different groups;

for example product segments, or regions. The other quantiles are frequently used

for the derivation of design loads and product specifications. Especially the 99 %

quantile is used for safety critical parts and components.

It is common to provide 95 % confidence intervals for the quantiles, which are

highlighted with the dashed lines. The confidence intervals are calculated to evaluate

the reliabilities of the quantiles and they provide a range of values which are likely to

contain the population parameter of interest. Since the underlying distribution is not

assumed to be standard normal distributions, it is recommended to use the bootstrap

method to calculate the confidence intervals. The bootstrap method does not assume

a theoretical distribution of the sample; instead, it uses multiple bootstrap samples

of the actual distribution to calculate a confidence interval.

Suppose a distribution Fwith ndata points x1, x2, ..., xn. An empirical bootstrap

sample F∗is a resample of the same size nwith x∗

1, x∗

2, ..., x∗

n. Given that uis a

statistic from the original distribution F- for example, the 50 % quantile - the boot-

strap principle states that the variation of uis well approximated by the variation of

u∗from the bootstrap samples. The resampling is achieved randomly with replace-

ment. The confidence interval increases with a higher quantile, as shown in Figure

5.3, since the variability for the single statistic also increases. More information

about the bootstrap method can be found in Orloff and Bloom [29].

Every feature of the customer usage profiles can be grouped by different meta

data; for example, by product segment or region. Figure 5.4 shows a boxplot for the

median air temperature per vehicle during the ride of different regions. The boxplot

is well defined in statistics: the line in the middle of the box is the sample median. If

the median is not centred in the box, it is an indicator for sample skewness. The top

and bottom of each box is the 25 % and 75 % quantile of the sample, respectively.

The distance between the top and bottom is the interquartile range. The whiskers

are the lines extended below and above the boxes, which are, in the present research,

defined as the 1 % and 99 % quantile of the sample, respectively. Sample points

outside the whiskers are highlighted as outliers. It is evident that the different

regions are characterised by different air temperatures. This in principle enables

132

5.3 Statistical considerations

Figure 5.4 – Median air temperature of different regions.

market-specific product requirements. However, it is important to highlight that

there are no overall customer usage profiles. Instead, an allocation of vehicles to

groups shows that different regions demand different requirements. This allocation

can also be made with other meta data; for example, product segment, personal

information or even optional vehicle equipment.

The treatment of a single variable is trivial, compared to the treatment of a dis-

tributed variable per vehicle; for example, the velocity. Each vehicle provides a

time-at-level counting for different velocity ranges, which can be defined in a coarse

or fine classification scheme, depending on the requirements. Figure 5.5 shows a

time-at-level counting for seven velocity bins at the top plot. Furthermore, the three

main statistical applications of these distributed variables are illustrated. First, the

median velocity can be evaluated per vehicle. The velocity medians of all vehicles

can then be represented by a probability distribution in the same way as a single

variable. This analysis provides an overview of how the medians are distributed. It

can also be grouped by segments or regions to work out differences.

133

5 Discussion

Q50 Ø 𝑘=5

7𝑡𝑘

Figure 5.5 – Statistical treatment of distributed variables.

Second, all observed distributions can be summed up to obtain an average distri-

bution with the same bins as the original classification result. Before the minutes at

the different velocity ranges can be summed up, each vehicle distribution has to be

extrapolated to a mutual mileage target or a mutual operating hours specification,

so that a comparison is feasible. The average distribution has the benefit that the

original classification scheme is still existent, but extreme customer behaviour cannot

be revealed. Thus, it provides a characteristic average vehicle velocity profile.

Third, specific questions can be evaluated; for example, when single bins are

grouped together. This allows the treatment of extreme usage. In Figure 5.5, the

upper right plot shows the distribution of the bins 5–7 summarised. In the case of

the vehicle velocity, it shows how much time is driven over a given threshold of velo-

city. The cumulative probability is indicated with the solid red line, which provides

information about the different quantiles; for example, the median. Other examples

134

5.4 Application to four-wheeled vehicles

of distributed variables are the engine speed, throttle position, gear position, engine

torque and many more.

In summary, the statistical treatment of customer usage profiles is not straight-

forward and highly depends on the application and the underlying problem. The

example of distributed variables shows that different conclusions can be made de-

pending on the statistical treatment. Therefore, it is recommended to formulate the

question of interest as detailed as possible prior to statistical evaluation. Hence,

this emphasises the importance of the act step at the beginning of the information

value loop. In addition, a fundamental statistical training is recommended for the

applicants, which ensures a successful application and correct treatment of the cus-

tomer usage profiles to derive data-driven decisions. It is worth highlighting that,

the proper meaning of the word, customer usage profile, suggests that the data re-

ceived from a vehicle represents the behaviour of exactly one customer. This is only

valid when the same customer rides the motorcycle all of the time and the vehicle

is not sold to any other person. However, in the context of durability, this is not

significant, since the actual usage of the vehicle is of interest, regardless of who rides

the vehicle. That means that technically the outcome of this type of crowd-sourced

data collection should be named vehicle usage profiles. Since the impact of altern-

ating owners and riders to the overall distribution is presumed to be poor, the term

customer usage profile has established.

5.4 Application to four-wheeled vehicles

In general, the need for customer usage profiles exists in any product development

where the product is exposed to individual and unforeseen loads. The approach of

gathering field-data with the onboard signals is state of the art at the development

of four-wheeled vehicles, since, compared to two-wheeled vehicles, the number of

onboard sensors is even more. The following Section discusses the possibilities of

transferring the methods presented to four-wheeled vehicles.

The estimation of the road slope can easily be transferred to four-wheeled vehicles

and should be even more robust, since the influence of a pitching chassis should be

less, compared to motorcycles. The required sensor is an IMU that measures the

longitudinal acceleration axand an angular wheel velocity sensor, which should be

part of the standard sensors. In modern cars, often a 6-DOF IMU is part of the

135

5 Discussion

onboard sensors, which makes it even more straightforward to estimate the road

slope by integrating the pitch rate. The method of estimating the vehicle mass has

already been implemented for heavy-duty vehicles; for example, by Ritzen et al.

[30], Holm et al. [31], and Lundin et al. [32]. The requirements comprise a detailed

driveline model and the knowledge of all occurring resistance forces. Beside heavy-

duty vehicles, the application of this method is especially interesting for estimating

the mass of a passenger car with a trailer or a caravan. This could enable further

driving assistance systems; for example, a brake preconditioning system. Finally, the

online estimation of the occurring wheel forces has already been realised to passenger

cars by the work of Matz [7], who implemented the methods in the same context

of durability. The equations need to be adjusted to the vehicle dynamics of a four-

wheeler.

Next, the methods for classifying the road profile can also be transferred to cars.

The methodology revealing the curviness needs adjustments, since a four-wheeler has

a different curve driving technique, compared to single-track vehicles. Rather than

utilising the roll angle, the lateral acceleration and the steering angle could be used

to estimate the current curvature κ. It is recommended to pay special attention to

the phenomena of under/over steering and the loss of grip, which is more complex on

four-wheelers. Once the road slope estimation has been integrated, the counting of

elevation gains is also trivial. Transferring the road roughness classification method

requires a full-vehicle model of the four-wheeler, which should be more complex

compared to two-wheeled vehicles. It is recommended to use a multi-body simulation

to derive the correct transfer functions, which relate the vehicle response to the road

excitation. It is not mandatory that every wheel has a suspension sensor, since the

method is designed to be modular. In addition, the transfer functions can be derived

for any other vehicle response signal. Further investigation should be achieved to

the classification scheme, since four-wheelers are exposed to two different tracks on

each axle. It is conceivable to classify the left and right track separately, or build an

average value.

Finally, the impact detection method can easily be transferred, once an angular

wheel velocity sensor and a suspension sensor is part of the onboard signals. The

method should be implemented for both left and right front wheel and further in-

vestigation should be achieved regarding the occurring special events. In general,

136

5.5 Data as a resource

passenger cars are also exposed to lateral forces when driving over a kerb so that

eventually other signals could hold interest; for example, the steering torque. In

summary, the methods presented can be transferred to four-wheeled vehicles, given

an adjusted model of vehicle dynamics.

5.5 Data as a resource

In recent years, data has become an increasingly valuable resource in the industry.

According to Meglena Kuneva, European Consumer Commissioner, and the World

Economic Forum (WEF) in 2011, “Personal data is the new oil of the Internet and

the new currency of the digital world” [33]. It will emerge as a new asset class in

the 21st Century. The WEF defined personal data (and metadata) as digital data

created by and about people. Moreover, they clustered personal data in three groups:

volunteered data, created and explicitly shared by individuals; for example, social

network profiles; observed data, captured by recording the actions of individuals such

as location data when using smartphones; and, inferred data, about individuals based

on the analysis of volunteered or observed data, such as credit scores. According

to the Digital Universe Study from IDC [34], the annual created, replicated and

consumed data, will grow by a factor of 300 from 2005 to 2020.

Several studies investigated the question, how companies can capture value of

their data. According to Wixom and Ross [27], companies can take three approaches

monetising their data. First, improving internal business processes and decisions.

Second, wrapping information around core products and services. Third, selling in-

formation to new and existing markets. From the viewpoint of business, the methods

presented deal with data in terms of improving internal business processes, which

is the vehicle development process, see Figure 1.1. Wixom and Ross describe this

approach as “using data to improve operational processes and boost decision-making

quality may not be the most glamorous path to monetising data, but it is the most

immediate” [27]. Continuing, companies will see positive results when they put

data and analytics in the hands of employees to make data-driven decisions. This

data-based insights will better address customer demands and optimise product de-

velopment. These assessments by Wixom and Ross [27] coincide with the thoughts

of Dinter et al. [35], who evaluated the impact of data as improving the basis of

decision-making, optimising internal processes and customer-oriented product devel-

137

5 Discussion

opment. According to Hartmann et al. [36], the following key activities are required

to generate value of data: data generating/acquisition, processing, aggregation, ana-

lytics, visualisation and distribution. These activities match up with the steps of

the information value loop, see Figure 5.1. The present research deals with the key

activity of data generation in terms of field-data acquisition. The methods presented

enable the collection of customer usage profiles, which are, in the present context,

defined as the data. Thus, this data can be categorised, according to the taxonomy

of Hartmann et al. [36], as internal company data, which is self-generated, and more

specifically, crowd-sourced. As per definition of the WEF, this data is categorised as

observed data.

Data privacy is a hot topic in the public interest and the limiting factor of observing

and utilising customer usage profiles. For the European Union, the Regulation (EU)

2016/679 of the European Parliament and of the Council of 27th April 2016 on

the protection of natural persons regarding the processing of personal data and on

the free movement of such data [37], will apply beginning from the 25th May 2018.

This law applies when dealing with personal data, which describes the principles of

purpose limitation, data minimisation and transparency. In general, the customer

must agree a declaration of consent for the use of personal data, which can be revoked

at any time. In addition, the customer has the right of information about his or her

data. Thus, the manufacturer must answer which data has been collected and for

what purpose.

5.6 Outlook

The present research project has shown the development of methods for gathering

the individual components of customer usage profiles. Next, the methods must be

implemented into the vehicle development process, as shown in Section 5.2. There-

fore, it is recommended to collect the specifications of all business units by starting

with the act step: asking the right questions, which means defining the purpose. This

step will help choosing the appropriate classification resolution and if necessary, the

definition of dependencies and multi-dimensional classifications.

Once the individual datasets are collected, further research should be carried out

to derive customer usage profiles on a large scale. Nowadays, each of the components

of customer usage profiles is treated independent from each other. For example, the

138

5.6 Outlook

Feature 1 Feature 3Feature 2

Persona A

Persona B

Figure 5.6 – Statistical derivation of personas.

99 % quantile (Q99) of the vehicle velocity is used for a validation procedure, as well

as the Q99 for road roughness, and the Q99 for brake pressure, etc. It is unfeasible

that one single customer represents the 99 % quantile of every component of cus-

tomer usage profiles. In addition, the characteristics are often mutually exclusive. A

possible approach would be the definition of stereotypes or so-called personas, which

represent a specific behaviour of a group of customers, see Figure 5.6. For example,

this could be the short-distance rider, the commuter, the tourer, the weekend rider,

the aggressive rider, the race-track rider and many more. These personas would

represent a well-defined combination of the components of customer usage profiles.

This is a statistical problem and could enable further potential of customer-specific

products.

When the collection and derivation of customer usage profiles is established once,

the entire vehicle development process will be much more customer-orientated. On

the one hand, this reduces development costs, since measurements with expensive

prototypes can be avoided. On the other hand, improved vehicle requirements will

reduce warranty claims, and thus quality costs can also be reduced. The virtual

product development will also benefit from the improved vehicle requirements. Fur-

ther research should concentrate on the methodology of virtual load acquisition and

the development of a modular system, consisting of three independent components:

the vehicle concept, the rider and the road characteristics, see Figure 5.7. In the

early stages of product development, the vehicle concept exists in the form of a nu-

merical full-vehicle model, which is state of the art in vehicle development. Next,

139

5 Discussion

Vehicle Concept Road CharacteristicsRider

Persona A

Persona B

Figure 5.7 – Three components of a virtual load acquisition.

the target group of customers can be defined using personas, which describe how the

vehicle will be used. Finally, the road characteristics describe where the vehicle is

used. Thus, the next steps should comprise the derivation of personas and the design

of virtual test tracks with the help of the customer usage profiles.

As Wixom and Ross [27] discussed, these data-based insights will better address

customer demands and optimise product development. Further activities could be

utilising this data for offering customer-specific products. In the motorcycle busi-

ness, this could be offering performance parts - for example, a sport exhaust - to the

group of sporty riders. By contrast, safety parts such as an airbag jacket, could be

advertised to cautious riders. In the near future, it will also be possible to offer digital

services similar to the well-known CE industry. It is conceivable that a customer can

buy extra performance from the in-vehicle app store. Further real-time applications

for customer usage profiles are condition-based maintenance systems. Nowadays, a

predefined workshop interval is given. With the knowledge of customer-specific usage,

every vehicle could be treated individually. Especially the detection of severe special

events could improve in-field quality costs for the manufacturer. Selling the inform-

ation to other manufacturers or third-parties is another possibility of monetising

this data. Officials could be interested in vehicle usage for improving infrastructure.

Insurances could offer individual contributions depending on the driving style.

Collecting crowd-sourced field-data is empowered by the development of autonom-

ous driving vehicles, since the signal processing and data sharing to other vehicles and

the backend is one major facet of this new technology. Data plays a fundamental role

140

5.6 Outlook

in the digital transformation and will gain increasing importance, whereby vehicle

data is simply one facet. Industry 4.0 requires production data for a connected

production and a continuous improvement in the production system. Further, con-

text data such as weather, traffic and mobility will be added to the vehicles as a

digital platform, as well as personal data, such as activities and social media. New

business models will emerge and the connected vehicle will become a high-tech and

high-complex product competing with the well-known CE industry.

141

Bibliography

[1] C. Gorges, K. Öztürk and R. Liebich. ‘Customer loads of two-wheeled vehicles’.

In: Vehicle System Dynamics 55.12 (2017), pp. 1842–1864.

[2] C. Gorges, K. Öztürk and R. Liebich. ‘Road classification for two-wheeled

vehicles’. In: Vehicle System Dynamics 56.8 (2018), pp. 1289–1314.

[3] C. Gorges, K. Öztürk and R. Liebich. ‘Impact detection using a machine learn-

ing approach and experimental road roughness classification’. In: Mechanical

Systems and Signal Processing 117 (2019), pp. 738–756.

[4] P. Johannesson and M. Speckert, eds. Guide to load analysis for durability in

vehicle engineering. Chichester: Wiley, 2013, p. 434.

[5] J. Ungermann. ‘Zuverlässigkeitsnachweis und Zuverlässigkeitsentwicklung in

der Gesamtfahrzeugerprobung’. PhD thesis. Zurich: ETH Zurich, 2009.

[6] K. Pötter. ‘Structural durability and integrity in vehicle design’. In: 18th European

Conference on Fracture. Dresden, Germany: DVM, Aug. 2010.

[7] C. Matz. ‘Online Berechnung von Fahrwerkskräften auf Basis von Onboard-

Sensorik’. PhD thesis. Clausthal: TU Clausthal, 2015.

[8] M. Köhler, S. Jenne, K. Pötter et al. Load Assumption for Fatigue Design of

Structures and Components. Berlin, Heidelberg: Springer, 2017, pp. 1–226.

[9] D. Schramm, M. Hiller and R. Bardini. Vehicle Dynamics. Vol. 24. 4. Berlin,

Heidelberg: Springer Berlin Heidelberg, 2018.

[10] M. Mitschke and H. Wallentowitz. Dynamik der Kraftfahrzeuge. Wiesbaden:

Springer Fachmedien Wiesbaden, 2014.

[11] B. Heißing, M. Ersoy and S. Gies. Fahrwerkhandbuch. Ed. by B. Heißing, M.

Ersoy and S. Gies. Vol. 1. Wiesbaden: Springer Fachmedien Wiesbaden, 2013.

[12] V. Cossalter. Motorcycle Dynamics. 2nd ed. [place unknown]: LULU, 2006,

p. 372.

[13] G. Charles, R. Goodall and R. Dixon. ‘Model-based condition monitoring at

the wheel–rail interface’. In: Veh Syst Dyn. 46.sup1 (Sept. 2008), pp. 415–430.

[14] T. X. Mei and X. J. Ding. ‘Condition monitoring of rail vehicle suspensions

based on changes in system dynamic interactions’. In: Veh Syst Dyn. 47.9 (Sept.

2009), pp. 1167–1181.

143

Bibliography

[15] N. Tandon, G. Yadava and K. Ramakrishna. ‘A comparison of some condi-

tion monitoring techniques for the detection of defect in induction motor ball

bearings’. In: Mech. Syst. Signal Process. 21.1 (Jan. 2007), pp. 244–256.

[16] C. Aliustaoglu, H. M. Ertunc and H. Ocak. ‘Tool wear condition monitoring

using a sensor fusion model based on fuzzy inference system’. In: Mech. Syst.

Signal Process. 23.2 (2009), pp. 539–546.

[17] M. Schlechtingen. ‘A global condition monitoring system for wind turbines’.

PhD thesis. Technical University of Denmark, 2013. isbn: 9788774753544.

[18] X. Liu, S. Alfi and S. Bruni. ‘An efficient recursive least square-based condition

monitoring approach for a rail vehicle suspension system’. In: Veh Syst Dyn.

54.6 (2016), pp. 814–830.

[19] P. Li et al. ‘Estimation of railway vehicle suspension parameters for condition

monitoring’. In: Control Eng. Pract. 15.1 (Jan. 2007), pp. 43–55.

[20] M. Börner et al. ‘Model based fault detection of vehicle suspension and hy-

draulic brake systems’. In: Mechatronics 12.8 (Oct. 2002), pp. 999–1010.

[21] M. Haddar et al., eds. Applied Condition Monitoring. Springer International

Publishing, 2018.

[22] A. Rupp, A. Masieri and T. Dornbusch. ‘Durability transfer concept for the

monitoring of the load and stress conditions on vehicles’. In: Innov. Automot.

Technol. IAT. 2005.

[23] L. Müller. ‘Mehrkörpermodell-basiertes Online Monitoring der Betriebsbeans-

pruchung am Beispiel eines Nutzfahrzeug-Demonstrators’. PhD thesis. Kaiser-

slautern: TU Kaiserslautern, 2011.

[24] ISO. Mechanical vibration - Road surface profiles - Reporting of measured data.

Standard No. 8608:2016. Geneva, Switzerland: ISO, 2016.

[25] M. Speckert, K. Dressler, N. Ruf et al. ‘The virtual measurement campaign

(VMC) concept’. In: Commercial Vehicle Technology 2014. Kaiserslautern,

Germany: Shaker Verlag, Mar. 2014, pp. 88–98.

[26] M. E. Raynor and M. J. Cotteleer. ‘The more things change. Value creation,

value capture, and the Internet of Things’. In: Deloitte Review 17 (2015),

pp. 50–65.

[27] B. H. Wixom and J. W. Ross. ‘How to monetize your data’. In: MIT Sloan

Manag. Rev. 58.3 (2017), pp. 10–13.

[28] M. Kuchler and R. Schrupp. ‘Mehrkomponenten-Motorradmessnabe’. In: Meas-

uring and testing techniques in the vehicle industry. Nuremberg, Germany:

VDI, 2001 5 10–11 2001, pp. 249–278.

144

Bibliography

[29] J. Orloff and J. Bloom. ‘Bootstrap confidence intervals’. In: Introduction to

Probability and Statistics. Massachusetts Institute of Technology: MIT Open-

CourseWare, License: Creative Commons BY-NC-SA, 2014.

[30] E. Ritzen. ‘Adaptive vehicle weight estimation’. MA thesis. Linköping, Sweden:

Department of Electrical Engineering, Linköping University, 1998.

[31] E. J. Holm. ‘Vehicle mass and road grade estimation using Kalman filter’. MA

thesis. Linköping, Sweden: Department of Electrical Engineering, Linköping

University, 2011.

[32] B. Lundin and A. Olsson. ‘Estimation of vehicle mass using an extended Kal-

man filter’. MA thesis. Gothenburg, Sweden: Department of Signals and Sys-

tems, Chalmers University of Technology, 2012.

[33] World Economic Forum. Personal data : The emergence of a new asset class.

Geneva, Switzerland: World Economic Forum, 2011.

[34] J. Gantz and D. Reinsel. THE DIGITAL UNIVERSE IN 2020: Big Data,

Bigger Digital Shadows, and Biggest Growth in the Far East. Tech. rep. IDC,

2012, pp. 1–16.

[35] B. Dinter et al. Big Data und Geschäftsmodell- Innovationen in der Praxis :

40 + Beispiele. Berlin, Germany: BITKOM, 2015.

[36] P. M. Hartmann et al. ‘Capturing value from big data – a taxonomy of data-

driven business models used by start-up firms’. In: Int. J. Oper. Prod. Manag.

36.10 (2016), pp. 1382–1406.

[37] ‘Regulation (EU) 2016/679 of the European Parliament and of the council of

27th April 2016 on the protection of natural persons with regard to the pro-

cessing of personal data and on the free movement of such data, and repealing

Directive 95/46/EC (General Data Protection Regulation)’. In: OJ L 119/1

(2016).

[38] G. Welch and G. Bishop. An introduction to the Kalman filter. Technical report

TR 95-041. Chapel Hill (NC): Department of Computer Science, University of

North Carolina at Chapel Hill, 2006.

145

A Onboard signals

147

A Onboard signals

Customer usage proﬁles

Sub-models

Sensors

Acceleration X

Acceleration Z

Roll rate

Yaw rate

Inertial Measurement Unit (IMU)

Acceleration X

Vehicle velocity

Road slope

Road slope estimator

Vehicle velocity

Angular rate wheel front

Angular rate wheel rear

Break preassure front

Brake pressure rear

ABS sensors

Suspension travel front

Suspension travel rear

Suspension sensors

Engine torque

Gear

Engine sensor

Roll rate

Yaw rate

Vehicle velocity

Roll angle

Angular rate wheel rear

Engine torque

Gear

Traction force

Driveline model

Vehicle velocity

Road slope

Traction Force

Mass

Mass estimator

Acceleration X

Vehicle velocity

Acceleration Z

Roll rate

Roll angle

Mass

Road slope

Traction force

Angular rate wheel front

Angular rate wheel rear

Break preassure front

Break preassure rear

Suspension travel front

Suspension travel rear

Wheel force calculation

Vehicle velocity

Roll angle

curviness c

Road curviness C

Curviness

Vehicle velocity

Road slope

hilliness h

Road hilliness H

Hilliness

Vehicle velocity

Suspension travel front

Suspension travel rear

ISO Class (A-H)

Road roughness

Vehicle velocity

Suspension travel front

Angular rate wheel front

Impact class

Impact detection

Mass Vehiucle loading

Vehicle loading

Figure A.1 – Block diagrams and signal flow of developed methods.

148

Table A.1 – Overview of sensor signals and model signals.

Sensor / Model Description Symbol Unit

Inertial Measurement Unit

Acceleration COG X axm s−2

Acceleration COG Z azm s−2

Roll rate COG ωxrad s−1

Yaw rate COG ωzrad s−1

ABS sensors

Velocity vm s−1

Angular rate wheel front ωft rad s−1

Angular acceleration wheel front αft rad s−2

Angular rate wheel rear ωrr rad s−1

Brake pressure front pft N m−2

Brake pressure rear prr N m−2

Engine sensor Engine torque TeN m

Gear i-

Suspension sensors

Suspension travel front sft m

Suspension travel rear srr m

Suspension velocity front ˙sft m s−1

Suspension velocity rear ˙srr m s−1

Sub-models

Roll angle φ◦

Mass mkg

Road slope α◦

Traction force FTN

Customer usage profiles

Wheel forces FX, FY, FZN

Mean curve radius ¯rcm

Mean curve angle ¯γ◦

Kappa κm−1

Curviness crad

Road curviness Crad

Elevation gain hm

Road hilliness Hm

Road roughness classification - -

Special event classification - -

149

B Vehicle dynamics

151

B Vehicle dynamics

𝑍

𝑋

𝑌𝑧R

𝑧s

𝑧un

𝑚un

𝑚s

𝑘T

𝑘s

𝑐s

Figure B.1 – Quarter of Vehicle (QoV).

B.1 Quarter of Vehicle

Equations of motion

M¨

x(t) + C˙

x(t) + Kx(t) = F(t)with (B.1)

M=[ms0

0mun],x=(zs

zun),(B.2)

C=[cs−cs

−cscs],K=[ks−ks

−ksks+kT],(B.3)

F(t) = (0

kTzR(t)).(B.4)

152

B.1 Quarter of Vehicle

State space model

q(t) = Aq(t) + Bu(t),(B.5)

y(t) = Cq(t) + Du(t),(B.6)

with

⎡

⎢

⎣

0 0 1 0

0 0 0 1

−ks

−cs

mun

−(ks+kT)

mun

−cs

mun

⎤

⎥

⎦

,q=⎛

⎜

⎝

zun

˙zs

˙zun

⎞

⎟

⎠

,(B.7)

B=⎡

⎢

⎣

mun

⎤

⎥

⎦

,C=⎡

⎢

⎣

1 0 0 0

0 1 0 0

0 0 1 0

0 0 0 1

⎤

⎥

⎦

,D= 0,u(t) = zR(t).(B.8)

Transfer functions

H1(s) = Zs(s)

ZR(s)=

cskT

msmun

s+kskT

msmun

s4+csms+csmun

msmun

s3+ksms+kTms+ksmun

msmun

s2+cskT

msmun

s+kskT

msmun

.(B.9)

H2(s) = Zun(s)

ZR(s)=

mun

s2+cskT

msmun

s+kskT

msmun

s4+csms+csmun

msmun

s3+ksms+kTms+ksmun

msmun

s2+cskT

msmun

s+kskT

msmun

.(B.10)

153

B Vehicle dynamics

𝑝

𝑏

𝑍

𝑋

𝑌

COG

𝑧R(𝑡)

𝑧s

𝑧rr 𝑧ft

𝑚rr 𝑚ft

𝐽s, 𝑚s𝛩

𝑣

𝑐ft, 𝑘ft

𝑘T

𝑐rr, 𝑘rr

𝑘T

𝑧R(𝑡 − 𝜏)

Figure B.2 – Half of Vehicle (HoV).

B.2 Half of Vehicle

Equations of motion

M¨

x(t) + C˙

x(t) + Kx(t) = F(t)with τ=p

v, a =p−b, (B.11)

M=⎡

⎢

⎣

ms0 0 0

0Js0 0

0 0 mft 0

0 0 0 mrr

⎤

⎥

⎦

,x=⎛

⎜

⎝

zft

zrr

⎞

⎟

⎠

,F(t) = ⎛

⎜

⎝

kTzR(t)

kTzR(t−τ)

⎞

⎟

⎠

,(B.12)

C=⎡

⎢

⎣

cft +crr cfta−crrb−cft −crr

cfta−crrb cfta2+crrb2−cfta crrb

−cft −cfta cft 0

−crr crrb0crr

⎤

⎥

⎦

,(B.13)

K=⎡

⎢

⎣

kft +krr kfta−krrb−kft −krr

kfta−krrb kfta2+krrb2−kfta krrb

−kft −kfta kft +kT0

−krr krrb0krr +kT

⎤

⎥

⎦

.(B.14)

154

B.2 Half of Vehicle

State space model with delayed inputs

q(t) = Aq(t) + B1u(t) + B2u(t−τ),(B.15)

y(t) = Cq(t) + D1u(t) + D2u(t−τ),(B.16)

with

⎡

⎢

⎣

00 0 0 1 0 0 0

0 0 0 0 0 1 0 0

0 0 0 0 0 0 1 0

0 0 0 0 0 0 0 1

−(kft +krr)

−kfta+krrb

kft

krr

−(cft +crr)

−cfta+crrb

cft

crr

−kfta+krrb

−(kfta2+krrb2)

kfta

−krrb

−(cfta+crrb)

−(cfta2+crrb2)

cfta

−crrb

kft

mft

kfta

mft

−(kft +kT)

mft

0cft

mft

cfta

mft

−cft

mft

krr

mrr

−krrb

mrr

0−(krr +kT)

mrr

crr

mrr

−crrb

mrr

0−crr

mrr

⎤

⎥

⎦

(B.17)

⎛

⎜

⎝

zft

zrr

˙zs

˙zft

˙zrr

⎞

⎟

⎠

,B1=

⎡

⎢

⎣

mft

⎤

⎥

⎦

,B2=

⎡

⎢

⎣

mrr

⎤

⎥

⎦

,C=

⎡

⎢

⎣

1 0 0 0 0 0 0 0

0 1 0 0 0 0 0 0

0 0 1 0 0 0 0 0

0 0 0 1 0 0 0 0

0 0 0 0 1 0 0 0

0 0 0 0 0 1 0 0

0 0 0 0 0 0 1 0

0 0 0 0 0 0 0 1

⎤

⎥

⎦

(B.18)

D1= 0,D2= 0,u(t) = zR(t).(B.19)

The transfer functions cannot be displayed in the analytical form. Instead, they

can be derived with a numerical computing environment, for example, MATLAB R

,

using the state space representation.

155

C Kalman filter

157

C Kalman filter

C.1 Linear Kalman Filter

The linear Kalman filter addresses the problem of estimating the state x∈Rnof a

discrete-time controlled process given the linear stochastic difference equation

xk=Axk−1+Buk−1+qk−1,(C.1)

with the control input vector u∈Rland the measurement vector z∈Rmthat is

zk=Hxk+rk.(C.2)

The n×nmatrix Arelates the state at the previous time step k−1to the state

at the current time step k. The n×lmatrix Brelates the control input uto the

state x. The m×nmatrix Hrelates the state xto the measurement z. Note

that the matrices A,B, and Hmight change with each time step or measurement.

The process noise qkand the measurement noise rkare assumed to be independent,

white, and with a normal probability distribution

P(q)∼ N(0,Q),(C.3)

P(r)∼ N(0,R),(C.4)

with the process noise covariance matrix Qand the measurement noise covariance

matrix R, which, in practice, might change with each time step or measurement.

The Kalman filter estimates a process by using a form of feedback control. Thus,

the filter estimates the process state at some time and then obtains feedback in the

form of noisy measurements. The time update equations project the current state and

the error covariance estimate forward in time. The measurement update equations

provide the feedback, which means, incorporating a new measurement into the a

priori estimate to obtain an improved a posteriori estimate. Thus, the a priori state

estimate ˆ

kat step krelies on the knowledge of the process prior to step k. The a

posteriori state estimate ˆ

xkis corrected by the measurement zk.

158

C.1 Linear Kalman Filter

Time update (Predictor)

Project state ahead: ˆ

k=Aˆ

xk−1+Buk−1(C.5)

Project error covariance ahead: P-

k=APk−1A⊺+Q(C.6)

with initial estimates for ˆ

x0and P0.

Measurement update (Corrector)

Compute Kalman gain: Kk=P-

kH⊺(HP-

kH⊺+R)−1(C.7)

Update estimate with measurement: ˆ

xk=ˆ

k+Kk(zk−Hˆ

k)(C.8)

Update error covariance: Pk= (I−KkH)P-

k(C.9)

The Kalman gain Kkis the relative weight between the measurement and current

state estimate. A high gain weights the recent measurement more, compared to

a low gain, which follows the model predictions more. The weights are calculated

from the error covariance matrix Pk, which is a measure of the uncertainty of the

model prediction. When the exact system state ˆ

x0is known, P0can be initialised as

zero. After each time and measurement update pair, the process is repeated. This

recursive nature of the Kalman filter is one of its well-known features and therefore

easy to implement and fast in computation time. The measurement noise covariance

matrix Ris usually measured prior to the application of the filter. In general, this

is feasible since the process must be measurable anyway. The determination of the

process noise covariance matrix Qis typically more difficult. The filter performance

can be increased by studying these parameters, which is called filter tuning. More

information can be found in Welch [38].

159

C Kalman filter

C.2 Extended Kalman Filter

The extended Kalman filter (EKF) addresses the problem of estimating the state

x∈Rnof a discrete-time controlled process given the nonlinear stochastic difference

equation

xk=f(xk−1,uk−1,qk−1),(C.10)

with the control input vector u∈Rland the measurement vector z∈Rmthat is

zk=h(xk,rk).(C.11)

The nonlinear function frelates the state at the previous time step k−1to the

state at the current time step k. It takes a control input function uk−1and a zero-

mean process noise qk−1as inputs. The nonlinear function hrelates the state xk

to the measurement zk. The basic operations of the EKF are the same as of the

linear Kalman filter. The EKF linearises around the current mean and covariance

to handle the nonlinear process and measurement equations. It thus requires the

Jacobian matrices A,W,H, and N.

Time update (Predictor)

Project state ahead: ˆ

k=f(xk−1,uk−1,0) (C.12)

Project error covariance ahead: P-

k=AkPk−1A⊺

k+WkQk−1W⊺

k(C.13)

with initial estimates for ˆ

x0and P0.

Measurement update (Corrector)

Compute Kalman gain: Kk=P-

kH⊺

k(HkP-

kH⊺

k+NkRkN⊺

k)−1

(C.14)

Update estimate with measurement: ˆ

xk=ˆ

k+Kk(zk−h(ˆ

k,0)) (C.15)

Update error covariance: Pk= (I−KkHk)P-

k(C.16)

160

C.2 Extended Kalman Filter

Ais the Jacobian matrix of partial derivatives of fwith respect to x:

A[i,j]=∂f[i]

∂x[j]

(ˆ

xk−1,uk−1,0) (C.17)

Wis the Jacobian matrix of partial derivatives of fwith respect to q:

W[i,j]=∂f[i]

∂q[j]

(ˆ

xk−1,uk−1,0) (C.18)

His the Jacobian matrix of partial derivatives of hwith respect to x:

H[i,j]=∂h[i]

∂x[j]

(ˆ

k,0) (C.19)

Nis the Jacobian matrix of partial derivatives of hwith respect to r:

N[i,j]=∂h[i]

∂r[j]

(ˆ

k,0),(C.20)

Note that the Jacobian matrices are updated each time step. According to Welch

and Bishop [38] the most interesting and successful applications have been solved

with EKFs. The most numerical computing tools come along with build-in black-

boxes for the linear and nonlinear Kalman filter.

161