Detecting Production Phases Based on Sensor Values using 1D-CNNs [original]

Detecting Production Phases Based on Sensor

Values using 1D-CNNs

Burkhard Hoppenstedt1,, Manfred Reichert1, Ghada El-Khawaga1, Klaus Kammerer1,

Karl-Michael Winter2, and R¨

udiger Pryss3

1Institute of Databases and Information Systems, Ulm University, Ulm, Germany

2Research and Development, Nitrex Metal Inc., St. Laurent, Canada

3Institute of Clinical Epidemiology and Biometry, University of W¨

urzburg, W¨

urzburg, Germany

Abstract—In the context of Industry 4.0, the knowledge ex-

traction from sensor information plays an important role. Often,

information gathered from sensor values reveals meaningful

insights for production levels, such as anomalies or machine

states. In our use case, we identify production phases through

the inspection of sensor values with the help of convolutional

neural networks. The data set stems from a tempering furnace

used for metal heat treating. Our supervised learning approach

unveils a promising accuracy for the chosen neural network that

was used for the detection of production phases. We consider

solutions like shown in this work as salient pillars in the field of

predictive maintenance.

Index Terms—Deep Learning, Industry 4.0, Sensor Processing,

Predictive Maintenance

I. INTRODUCTION

In comparison to complex deep architectures, 1D convolu-

tional neural networks (1D-CNNs) provide the possibility for

real-time fault detection [1] and sensor analysis. The feature

extraction and selection are done implicitly by the network

itself. In this context, the handling of sensor readings is

essential to build enhanced applications for monitoring and

prediction purposes [2]. In this work, we detect and restore

production phases of a carburizing furnace using the carbon

potential inside the furnace. While automated furnaces offer a

built-in functionality to record start and end of a production

phase, manually operated furnaces do not necessarily provide

such a function. The aim of this approach is to upgrade simple

machines, which only record raw sensor data, with a feature to

identify the production phases based on gathered sensor stream

data. Therefore, a model is trained that classifies a production

Fig. 1. a) Batch/IQ (Chamber with integrated quench) furnace [4]. The load

is put into the hot chamber over the quench with an elevator located at the

bottom position. This approach allows for reloading the hot chamber while

the previous load is still in the hardening bath (quench) b) Typical process

setup for carburizing

phase by recognizing its start and end point. Hereby, the

production workflow is accomplished as follows: A new load

waiting on top of the elevator is moved into the hot chamber

(see Figure 1a). Note that opening the inner furnace door and

pushing the cold load into the hot chamber causes changes

in the gathered sensor values (i.e., a pattern can be observed,

see Figure 2). When the material is heated up and kept on a

specific temperature and potential for a given time, then, at the

end of the cycle, when it is rapidly cooled down using an oil

quench, the transfer from the hot chamber to the quench will

once more cause a change in the gathered sensor values (again,

a pattern can be observed, see Figure 1b). The statistics for the

time [index relative to startpoint)

carbon potential [%]

Fig. 2. Raw sensor values visualized at the process start

three used sensor variables can be seen in Fig. 3a. Note that we

used three months of recorded sensor data, where the carbon

value is recorded at a sample rate of mostly one value per

minute (see Figure 3b). During the recorded three months, 212

production phases were conducted. These phases are denoted

as charges. Notably, the process usually takes up to four hours.

II. THE APPLICATION

Before the actual machine learning can be applied, a pre-

processing is required. First of all, data is stored in a SQL

database. For a simplified workflow using python [3], the

data is transferred into a CSV format. For each charge, we

know the timestamps of the start and end point. Therefore,

we can introduce the tree class labels (start,end,between) and

use them to apply a supervised approach to train a machine

learning model. The task is to assign the correct classification

of the current input sensor values, which can be expressed

arXiv:2004.14475v1 [cs.LG] 24 Apr 2020

frequency

20000 120000

120000

100 200 300

time deltas for carbon level [s]

Measurements

Unit

Maximum

Minimum

Mean

Median

Variance

Temperature Carbon Potential Oil Bath

64 373

°C

904.20

0.00

792.26

859.80

174.30

158 978

0.94

0.00

0.52

0.60

0.20

64 373

°C

134.00

0.00

95.61

100

21.57

a) b)

Fig. 3. a) Data set overview b) Histogram of sampling rate from carbon

potential

via sliding windows. For example, a sliding window of size

1200 values, which corresponds to a record length of about 20

minutes, is shifted over the sensor data stream. To each sliding

window, a label is assigned. The label for the whole sliding

window is denoted as the label of the corresponding label

of the centered value in the window. However, we transferred

the classification problem into a regression problem to express

the proximity of the current sliding window to the next phase.

Hereby, the label value lcan be interpreted as follows.

l=









start if l≈1

between if l≈0

end if l≈ −1

(1)

Before preprocessing was applied, the labels start and end

were assigned only at the exact position of a phase change.

After processing, the values in the area of a phase change

can be marked as different, as they differ from value 0, which

means a change was detected. In contrast, in the case of a start,

the generated label corresponds exactly to value 1at the exact

timestamp tex , for which the charge started and is reduced

linearly along the sliding window sw of duration tsw . Hereby,

the new label lnew , for a sensor measurement with value vsm

and time tsm , is assigned as follows:

distance =abs(tex

−tsm ), lnew = 1 −(distance/(tsw /2) (2)

The label workflow can be seen in Fig. 4. First, the real labels

are generated on a scale from −1to +1. Based on these labels,

a 1D-CNN is trained. The results can be seen in the curve

predicted label. To transfer the results of the predicted curve

back to a unique timestamp, local maxima and minima are

calculated. Finally, a threshold of 0.5ensures that only strong

peaks are accepted. It is evident that the predicted label fits

well to the real label. Therefore, the procedure seems suitable

for the detection of production phases. The machine learning

model is trained using the framework Keras [5]. We used the

network architecture presented in Fig. 5. Hereby, the input is

provided by the current sliding window. Next, a convolution

layer with 64 filters of kernel size 3 is applied and summarized

in another layer using max pooling. The generated matrix from

the filter results is then flattened and serves as the input for a

dense network. By using the filter-based approach, the network

automatically weights more important features. To validate the

approach, the data set is split up with a 80/20 ratio of train

and test set. With the used network, we are able to restore the

timestamps with a mean distance to the actual starting point

of 93.12 seconds without any false positive detection.

Real Label

Predicted Label

Local Max ‚Start‘

Local Min ‚End‘

time

label value

Fig. 4. Generated label values with prediction and filtering

Input

(1200,1)

Conv1D

(1200,64)

MaxPooling

(600,64)

Flatten

(38400)

Dense

(50)

Dense

(1)

Fig. 5. Architecture of the 1D-CNN

III. SUMMARY AND OUTLOOK

In this paper, we showed a workflow how to use 1D convo-

lutional neural networks (1D-CNNs) to recognize production

phases based on raw sensor value input. Based on simple

timestamps for start and end points, this classification is then

transformed into a regression problem. Finally, the results are

filtered to receive the best fitting prediction. The approach

showed a good precision and can be used for further analysis.

In future approaches, similar procedures could be applied to

detect more complex states and use these states as input for

advanced analytics, such as predictive maintenance [6].

REFERENCES

[1] L. Eren et al., ”A Generic Intelligent Bearing Fault Diagnosis System

Using Compact Adaptive 1D CNN Classifier”, Journal of Signal Pro-

cessing Systems, vol. 91, no. 2, pp. 179-189, 2018.

[2] K. Kammerer et al., ”Anomaly Detections for Manufacturing Systems

Based on Sensor Data-Insights into Two Challenging Real-World Pro-

duction Settings”, Sensors, vol. 19, no. 24, p. 5370, 2019.

[3] J. Brownlee, ”How to Develop Convolutional Neural Network Models

for Time Series Forecasting”, Machine Learning Mastery, 2018. [On-

line]. Available: https://machinelearningmastery.com/how-to-develop-

convolutional-neural-network-models-for-time-series-forecasting/. [Ac-

cessed: 04- Jan- 2020].

[4] K.-M. Winter et al., ”Principles of process technology of thermochemical

surface engineering of steels”. In Thermochemical Surface Engineering

of Steels. Woodhead Publishing Series in Metals and Surface Engineer-

ing, no. 62, pp. 141-206, 2015.

[5] F. Chollet et al., ”Keras”, Keras.io, 2015. [Online]. Available:

https://keras.io/. [Accessed: 04- Jan- 2020].

[6] B. Hoppenstedt et al., ”Techniques and Emerging Trends for State of the

Art Equipment Maintenance SystemsA Bibliometric Analysis”, Applied

Sciences, vol. 8, no. 6, p. 916, 2018.