Toward Process Improvement in Hull Construction: A Combined Work Monitoring and Simulation Framework for Schedule Deviation Analysis

16th International Symposium on Practical Design of Ships and Other Floating Structures PRADS 2025 Ann Arbor, MI, USA, October 19th-23rd 2025 Toward Process Improvement in Hull Construction: A Combined Work Monitoring and Simulation Framework for Schedule Deviation Analysis Chenwei Gui1,*, Jaeyeon Nam1, Tomoyuki Taniguchi2, Kohei Matsuo2and Kazuhiro Aoyama1 1The University of Tokyo, Tokyo, Japan 2National Institute of Maritime, Port and Aviation Technology, Japan Abstract. Modern shipyards face persistent challenges in achieving efficient, on-time production due to the inherent complexity of hull construction. The high uncertainty introduced by diverse product varieties and labour-intensive operations makes it impractical to fully predetermine and optimize production workflows. In practice, shipyards typically develop only rough daily schedules at the shop level, leaving many detailed tasks to the judgment and expertise of field workers. However, the variability in worker behavior, driven by different levels of experience, can introduce additional non-value-adding activities that may lead to significant deviations from the planned schedule. This paper investigates the impact of worker behavior on schedule deviations in a subassembly line by integrating work monitoring tools and agent-based simulation. We employ a YOLO model to extract work-related features from surveillance video feeds. These monitored data form a time series of work activities, which we then segment to pinpoint critical periods where delays are most likely to occur. By tracking workers’ behavior trajectories within these critical windows, we uncover patterns that contribute to suboptimal performance. In parallel, we utilize an agent-based hull construction simulator to generate near-optimal work trajectories, leveraging insights derived from the behavior patterns of veteran workers. By highlighting gaps between simulation-based optimal scenarios and actual workflows, our findings reveal the role of nuanced human factors in delay occurrence. This integrated analysis framework provides clear visibility into how suboptimal work activities arise, enabling more targeted evaluations that can serve as the basis for process improvements in hull construction. Key words: Work Monitoring, Behavior Modeling, Shipbuilding Process Simulation, Schedule Deviation Analysis 1. Introduction Hull construction in modern shipbuilding presents persistent challenges due to its high product variety, reliance on manual labor, and dynamic shop floor conditions. The complexity of assembling highly customized components under varying spatial and temporal constraints makes it difficult to develop detailed, robust production plans. As a result, shipyards typically rely on coarse daily schedules at the shop level, with the finer granularity of task execution left to the judgment of individual workers. This reliance on human decision-making introduces significant uncertainty into the production process, particularly when differences in experience, skill, and situational awareness lead to non-value-adding activities such as repeated adjustments, inefficient movements, or poorly sequenced actions– all of which can ultimately result in production delays. To address the difficulties in shipyard production planning, various simulation-based approaches have been developed to improve schedule accuracy and system performance. Krause et al. [1] proposed a discrete event simulation (DES) framework as a decision-support tool for evaluating alternative investment and scheduling scenarios across shipyard facilities, emphasizing strategic and operational planning. Lee et al. [2] developed a process-centric simulation platform aimed at improving master scheduling accuracy *Correspondence to: [email protected] 1 through forecasting and scenario evaluation. Okubo and Mitsuyuki [3] proposed a method for automatically generating realistic production plans using structured system models for workflow and team allocation. While these works demonstrate the utility of simulation in optimizing shipyard operations, they predominantly focus on macro-level or system-level planning and typically assume ideal task execution and overlook the variability introduced by worker-level decisions—precisely where many delays originate. This gap highlights the need for integrating detailed human behavior into simulation to capture the root causes of micro-level inefficiencies, which can help refine the overall accuracy of production planning. To complement macro-level planning methods and provide insights into shop floor realities, many studies have focused on extracting micro-level information from unstructured on-site data. In particular, surveillance videos accumulated through past shipbuilding projects have been leveraged to capture detailed worker activities, with recent efforts using computer vision techniques to transform this data into structured representations of work behavior. Kim et al. [4] proposed a vision-based system that segments video streams, identifies block components, and compares them with CAD models to track assembly progress. Shinoda et al. [5] introduced a deep neural network framework that classifies work items from head-mounted camera footage, enabling automated observation of welding operations based on defined task categories. Gui et al. [6] combined YOLO-based object detection with time-series analysis to identify complex work activities by recognizing sequences of basic behaviors in video. While these efforts successfully extract valuable micro-level work information, they rarely close the loop from monitoring to understanding the causes of inefficiencies, nor do they integrate behavioral insights into planning or simulation models for process improvement. To help bridge this gap, this study proposes an integrated framework that combines work monitoring and simulation to systematically analyze schedule deviations in hull construction. The core idea is to capture and model worker behavior during production using computer vision-based monitoring, and then compare these observed patterns with standard workflows generated through simulation. This comparison allows us to quantify deviations in task execution and identify behavior-driven inefficiencies that contribute to delays. The framework consists of two main components. First, a work monitoring system extracts time-series data of worker activities from surveillance footage, allowing us to detect critical periods where schedule deviations are likely to occur. The captured behavior is then represented in a structured format using a Behavior Pattern Matrix (BPM), which encodes the duration of tasks at specific locations as well as transitions between them. Second, an agent-based simulation generates near-optimal workflows based on predefined process rules and physical constraints, serving as a reference for comparison. By analyzing the differences between the monitored and simulated behavior patterns, the framework enables a detailed evaluation of how actual worker actions contribute to deviations from planned operations. This study makes several key contributions to the field of shipbuilding production planning and control. First, it proposes a systematic method that combines computer vision techniques with time-series analysis to detect task execution periods where schedule delays are likely to originate. Second, it introduces a matrixbased representation of micro-level work behavior, enabling quantitative evaluation of worker performance and establishing a structured link between observed actions and simulation-based planning—serving as a proof of concept for integrating human behavior into digital twins. Third, the framework offers a workcentric perspective on schedule deviation analysis, shifting focus from system-level processes to individual decision-making and its impact on production outcomes. Finally, the approach is validated through a case study in a real shipyard subassembly process, demonstrating its feasibility and practical relevance. The remainder of this paper is organized as follows. Section 2 describes the shipyard context and the monitoring framework used to detect deviation-prone periods based on visual data. Section 3 introduces the agent-based simulation environment and explains how standard workflows are generated for comparison. Section 4 presents the results of the case study, including the extraction of behavior patterns from monitoring data, the generation of simulated workflows, and a quantitative comparison between the two to analyze behavioral deviations. Finally, Section 5 summarizes the findings and discusses directions for future research and practical implementation. 2 2. Work Monitoring for Schedule Deviation Detection and Behavioral Modeling 2.1. Overview of the Subassembly Line and Fitting Tasks In hull construction, the subassembly process serves as a foundational step in forming the structural blocks that make up the vessel’s hull. It involves the fabrication and assembly of smaller components —such as straight or curved plates and their corresponding strengthening stiffeners— into larger sub-blocks, called sub-assemblies, which are later joined during block assembly stages [7]. The accuracy and efficiency of the subassembly process directly impact the quality and timeliness of downstream operations, making it a critical phase in the shipbuilding workflow. F1 F2 RB W1 W2 Subassembly Line Figure 1.: Layout of the sequential workstations in the subassembly line. The subassembly line examined in this study comprises a series of sequential workstations, as illustrated in Figure 1.. Plates are transported along a conveyor through the following stations: •F1: Part feeding and alignment - Base plates are carried into the workstation, and stiffeners are then delivered and aligned near the base plate for initial positioning. •F2: Fitting - Workers adjust and temporarily fix stiffeners onto the base plate. •RB: Robotic Main Welding - Automated welding for most of the weldable sections. •W1: Manual Welding - Workers perform manual welding on areas inaccessible to robots, completing the structural connections. •W2: Finishing and Inspection Final quality checks and adjustments are made before the subassemblies progress to the next stage. Among the various tasks performed on the subassembly line, the fitting operation at the F2 workstation is of particular interest due to its manual and decision-intensive nature. In this task, workers are responsible for accurately positioning stiffeners on the base plate, making real-time adjustments, and determining the sequence and location of initial tack welds to temporarily secure components. Although each fitting job is guided by a shop-level production plan that defines the spatial layout and expected completion timeline, the detailed execution—especially the assembly sequence—is left to the discretion of individual workers. This reliance on human judgment introduces variability into the process. Micro-level decisions, influenced by experience, local constraints, or working conditions, can lead to suboptimal workflows that include repeated adjustments, unnecessary movement, or rework. Such deviations from the ideal sequence not only reduce efficiency at the task level but can also contribute to broader delays in the production schedule. In this study, we focus on the fitting task as a representative example to investigate how individual behavior and on-site decision-making impact schedule adherence in hull construction. 3 2.2. Visual Monitoring Framework for Deviation-Prone Task Detection To identify inefficiencies that may contribute to schedule deviations, a visual monitoring framework was developed to capture and analyze worker activities and workstation status throughout the production process. As shown in Figure 2., the framework combines object detection with time series analysis to provide a structured view of production behavior. (a) Time Series Analysis for Job Segmentation (b) Time Series Analysis for Local Feature Visualization Worker Postures Stand Squat Weld Video Frames YOLO11 Image Classification Object Detection Workstation Status Ground Base Plate Delivered WIP Subassembly Figure 2.: Visual monitoring framework for the identification of potential sources of schedule deviation. (a) Segmented workstation status timeline showing detected job boundaries alongside the planned job schedule for comparison. (b) Worker posture time series with corresponding local work feature curves, illustrating variations in work patterns across different jobs. The monitoring begins with object detection using a fine-tuned YOLO11n model [8]. For worker posture detection, the model was trained on 975 annotated images collected from three different shipyards, augmented to over 3,700 images. It detects three postures: standing, squatting, and welding, reflecting different stages of task execution, including incidental actions such as positioning or adjustments. Workstation status is identified using YOLO11n with a classification head YOLO11n-cls, fine-tuned on 475 images to classify five overall stages: ground, base plate, delivered, WIP, and subassembly. Due to the complex borders and overlapping elements in shipyard environments, conventional object detection struggled with accurately detecting the status of individual components. Image classification was therefore used to assess the overall status of the workstation, where ground indicates no plate present, base plate signals plate placement, delivered marks stiffener delivery, WIP shows ongoing fitting work, and subassembly denotes completion. These detection results are processed through time series analysis. Workstation status changes are used to segment the job into distinct phases, while sequences of worker postures help visualize local work patterns. Together, these elements support the identification of variations in task execution that may signal inefficiencies. As shown in Figure 2.a, job segmentation is performed based on the time series of workstation status. The raw status data is first smoothed using a moving average, after which Change Point Detection (CPD) is applied to identify potential transitions between jobs. The CPD algorithm formulates the segmentation task as an optimization problem, seeking breakpoints that maximize the selected goodness of fit measure for the segments [9]. To capture non-linear changes typical in production processes, we use kernel CPD with a radial basis function (RBF) kernel. This allows detection of various changes in the distribution, not limited to shifts in mean or variance, and is particularly effective for complex structured 4 data [10]. A low penalty setting in CPD algorithm enables the detection of multiple candidate breakpoints, as shown in the second row of Figure 2.a. These candidates are further refined using rule-based filtering, where patterns aligned with conveyor movements serve as indicators of true job boundaries (third row). Comparing the filtered segmentation (third row) with the planned job timeline (fourth row), it is evident that the third and fourth jobs took longer than expected, with the third job showing a particularly significant delay. To provide more specific insights for schedule deviation analysis, local work features were analyzed within each segmented job, as shown in Figure 2.b. The first three rows display the YOLO-based posture detections over time (count of standing, squatting, and welding), followed by CPD results applied within each job segment, and the computed local feature values for these segments. Two key features were used: local main work rate, defined as welding time divided by total work time (in man-hours), and progress increment, calculated as the welding time relative to the total welding time of the job. Abnormal patterns in these features, such as low main work rates or stagnation in progress, suggest inefficiencies. In Job 3, two distinct plate periods were observed, along with notably low main work rates, indicating this job may be a major contributor to the overall delay. This provides supplementary evidence linking specific work behavior to schedule deviations. 2.3. Behavior Pattern Matrix Construction from Monitoring Data To enable detailed analysis of worker behavior during the fitting task, a quantitative representation was developed in the form of a Behavior Pattern Matrix (BPM), which captures both the duration of work at each workplace (typically a stiffener) and the transitions between workplaces. The BPM is defined as a square matrix, where diagonal entries indicate the total time spent at each workplace, and off-diagonal entries represent the number of transitions from one workplace to another. This structure allows for a compact yet expressive view of both task focus and movement patterns, making it suitable for comparison, visualization, and further analysis. To construct the BPM from monitoring data, DBSCAN clustering [11] was applied to the spatialtemporal distribution of detected welding points, enabling the identification of discrete welding workplaces. These clusters were then aligned with spatial arrangement planning data to map them to specific stiffener locations on the plate. Once workplaces were established, the durations of welding and incidental work were accumulated to compute the diagonal elements. For the off-diagonal elements, the welding time series was compressed into a one-dimensional sequence of workplace visits, from which transitions were counted. Short-duration, noisy movements were filtered out to avoid overestimating transitions caused by minor hesitations or detection errors. This matrix format not only enables a structured comparison of actual and simulated workflows but also provides a basis for downstream feature extraction and network-based analysis. Metrics such as modularity, transition entropy, or centrality can be derived from the BPM to assess behavioral patterns and efficiency. By preserving both location-specific engagement and task flow, the BPM serves as a foundational representation for identifying deviations, evaluating performance, and understanding how execution unfolds across space and time. 3. Simulation-Based Baseline Modeling and Comparison Method 3.1. Simulation Mechanism for Standard Workflow Generation To better support precise production planning, especially in estimating workloads and defining benchmark workflows, a multi-agent-based simulator was used to realistically model the shipbuilding process, including incidental work. This simulator was developed by the National Maritime Research Institute (NMRI) and is implemented on the Unity platform [12]. It captures both the behavior of individual workers and the physical layout of the shipyard in a virtual environment. As shown in Figure 3., the simulator is composed of three main elements: A virtual shipyard, which models both the facilities and the products being assembled. Worker agents, which act autonomously, 5 Facilities Product Work Task 1 Task 2 Basic Task 1 Basic Task 2 Basic Task 3 Virtual Shipyard (Environment) Worker Agent Observation Action (Task Selection & Execution) Pattern A Decision-Making Model Pattern B Pattern C Workflow A Workflow B Workflow C Figure 3.: Overview of the agent-based shipyard simulator and its mechanism. making decisions based on their surroundings. A hierarchical work model where tasks are structured into layers, breaking down complex activities into simpler, more manageable basic tasks. Before running the simulation, a Bill of Process (BOP) needs to be defined for each product, outlining the required steps for assembly. Once the simulation starts, each worker agent evaluates available tasks and selects what to do next based on their perception of the environment and predefined decision-making rules. This selection process is governed by a priority function 1: ftask =1 k(Wp+b)×Y pc∈p sgn(pc)(1) where: •prepresents the feature vector of the environment, including the current task and its context. •Wis a weight vector that determines the importance of each feature. •bis a bias term related to the type of work. •kis the number of features in p, used for normalization. •pcrefers to specific constraint-related features, such as task precedence, and sgn(pc)adjusts priority based on whether constraints are satisfied. After selecting a task, it is broken down into basic tasks, whose durations are calculated using simulation parameters. For example, the time for a fitting task is determined by the length of the welding line divided by the worker’s fitting speed, plus additional time for adjusting positions based on the stiffener’s condition. This setup allows the simulator to account for both main and incidental work, producing realistic workflows that can later be compared to actual observed behavior. While it is difficult to define a single optimal workflow in practice, the simulator generates baseline workflows by implementing three different task selection strategies. These strategies represent different approaches to sequencing and decision-making, providing varied behavior patterns for analysis: 6 1. Autonomous Strategy: Agents prioritize minimizing movement distance, selecting the nearest available workplace at each decision point. This models an efficiency-driven approach without adherence to a predefined sequence. 2. Standard Strategy: Agents follow a recommended sequence based on standard operating procedures, representing strict compliance with predefined work instructions. 3. Heuristic Strategy: Agents assign higher priority to horizontal stiffeners while maintaining flexibility in choosing subsequent workplaces. This hybrid strategy balances efficiency with structural considerations, modeling expert-like behavior. In actual production, some precedence constraints govern the sequence of assembly. If these constraints are not followed, additional time and rework risk can arise due to difficulties such as poor welding posture or limited access. In the simulator, when a generated workflow violates such constraints, an additional time penalty is applied based on the specific condition of the product. 3.2. Behavior Pattern Comparison Method for Schedule Deviation Analysis To enable structured comparison between actual and simulated workflows, both were transformed into Behavior Pattern Matrices (BPMs) with a unified format. Each BPM was interpreted as a directed weighted graph (DWG), where nodes corresponded to workplaces and edge weights represented the frequency of transitions between them. Based on this representation, clustering was performed using greedy modularity maximization [13] to identify cohesive substructures within the workflow, reflecting the underlying structure of task execution. In addition to visual comparisons, a set of quantitative features was extracted. From the diagonal entries of the BPM, time-based statistics—including average, standard deviation, and skewness of workplace durations— were computed to describe the typical workload distribution, variability, and effort balance. For structural analysis, the off-diagonal entries were treated as an adjacency matrix, and several graphtheoretic features were derived. Modularity was used to quantify the strength of community structure in the transition network, indicating how well the workflow divided into distinct phases. The average degree was calculated to represent the number of transitions per workplace, reflecting overall workflow complexity. The clustering coefficient [14] calculated using the weighted directed formulation based on geometric means, indicating the presence of tight transition loops possibly associated with rework or localized loops. Eigenvector centrality [15] was also calculated to capture the quality of connections based on the idea that a node is important if it is connected to other important nodes. Lastly, graph density was measured to assess the overall level of interconnectivity across the network. To complement these structural features, the BPM was also normalized into a transition probability matrix, allowing the entropy of movements to be calculated. This value reflected the degree of randomness in a worker’s transitions, with higher entropy suggesting less predictable or less structured workflows. These timeand graph-based metrics together formed a feature vector that enabled quantitative similarity analysis between actual and simulated behavior, supporting the identification of behavior-driven deviations and guiding improvement efforts. 4. Results and Discussion 4.1. Behavior Pattern Extraction and Analysis from Monitoring Data As illustrated in Figure 2., Job 3 was identified as a potential source of schedule deviation. This job involved two plates occupying the entire workshop area, placed symmetrically on the left and right sides, as shown in Figure 4.a. Both plates had identical designs but were assigned to two different workers— Worker A on the left and Worker B on the right—who performed the tasks independently. According to expert evaluation based on repeated video reviews, Worker B appeared to exhibit higher efficiency and skill in task execution, although his sequence did not always align with the predefined standard workflow. In contrast, Worker A generally adhered to the suggested sequence but exhibited frequent local movements, particularly within confined areas, suggesting possible hesitation or repeated adjustments. 7 0 1 2 3 45 6 7 8 9 10 11 0 1 2 3 4 5 6 7 8 9 10 11 (a) Layout of workplaces completed in Job3 at the F2 station. 400 200 0 200 400 Time (s) WP 0 WP 1 WP 2 WP 3 WP 4 WP 5 WP 6 WP 7 WP 8 WP 9 WP 10 WP 11 0.070.06 0.110.04 0.370.03 0.080.31 0.130.33 0.280.48 0.200.27 0.190.07 0.350.77 0.090.03 0.030.07 0.080.12 Welding (A) Incidental Work (A) Welding (B) Incidental Work (B) (b) Comparison of welding and incidental work times for Worker A and Worker B across all workplaces. Figure 4.: Workplace distribution and worker efficiency comparison for Job 3. Figure 4.a shows the workplaces (stiffeners) completed at the F2 station, while Figure 4.b compares the welding and incidental work times per workplace for both workers. The stacked bars represent the time spent on welding and incidental tasks, and the value at the end of each bar represents the main work rate at that location. Overall, Worker B maintained higher main work rates across most workplaces, supporting the expert judgment regarding superior task execution skills. In contrast, Worker A recorded higher incidental work time at several locations. The significant time spent at WP 0 may indicate an external disturbance or abnormal difficulty at that location. To further investigate the differences in work patterns, Behavior Pattern Matrices (BPMs) were constructed for both workers, as shown in Figure 5.. These matrices capture both the time spent at each workplace (diagonal elements) and the frequency of transitions between workplaces (off-diagonal elements). Figure 5.a and 5.b display the BPMs of Worker A and Worker B, respectively. Worker A’s BPM reveals more scattered transitions and higher variability in work time across workplaces, while Worker B’s BPM shows more focused activity, particularly around WP 3, WP 9, and WP 10. To visualize these dynamics more intuitively, the BPMs were also represented as directed weighted graphs, as shown in Figure 5.c and 5.d. In these graphs, node positions correspond to the actual spatial layout of the workplaces. Node sizes reflect the total time spent at each workplace, and edge widths represent the frequency of transitions. Additionally, node colors indicate clusters obtained via graph-based community detection, which correspond to the modular blocks visible in the matrix representation. Figure 5.c shows that Worker A followed a highly interconnected movement pattern, with dense transitions among WP 4, WP 6, WP 7, and WP 9, indicating complex and potentially inefficient task routing. In contrast, Figure 5.d shows Worker B’s workflow as more modular and spatially cohesive. From an intuitive perspective, B’s graph appears more regular, with clearly defined task phases and fewer cross-region transitions. This visualization reinforces the matrix-based analysis, illustrating how individual behavior differences—especially in movement structure and task allocation—can significantly influence execution efficiency and highlight areas where targeted guidance or intervention may improve performance. 4.2. Standard Workflow Generation via Simulation To establish baseline workflows for comparison, simulations were conducted using the three task selection strategies introduced earlier: autonomous, standard, and heuristic. The resulting Behavior Pattern Matrices (BPMs) and their corresponding DWG are presented in Figure 6.. The autonomous strategy minimizes total movement by allowing the agent to select the nearest available workplace at each step. However, this often leads to violations of precedence constraints— particularly those requiring horizontally placed stiffeners (typically central nodes) to be tack welded before their vertically placed neighbors. When such constraints are ignored, time penalties are imposed to simulate the 8 01234567891011 Workplace 0 1 2 3 4 5 6 7 8 9 10 11 Workplace 484 22000000000 21341000000000 0160100100000 10018701200000 00001560200100 0000058200000 000431 214 50000 000000495 1 1 0 1 0000001046000 00000002020701 00000000011781 00000010002158 0 1 2 3 4 5 Transitions (Off-Diagonal) 0 100 200 300 400 Time (Diagonal) (a) Behavior pattern matrix of Worker A (left plate). 01234567891011 Workplace 0 1 2 3 4 5 6 7 8 9 10 11 Workplace 190 10000000000 1 120 0 200000000 0197100000000 012131 1 3010000 1002960000000 0002 2 65000000 0001018102000 0000000961000 00000030 52 0 0 0 000000000297 12 0000001001208 0 0000000001152 0.0 0.5 1.0 1.5 2.0 2.5 3.0 Transitions (Off-Diagonal) 0 50 100 150 200 250 Time (Diagonal) (b) Behavior pattern matrix of Worker B (right plate). 0 1 2 3 4 5 6 7 8 9 10 11 (c) BPM network of Worker A (left plate). 0 1 2 3 4 5 6 7 8 9 10 11 (d) BPM network of Worker B (right plate). Figure 5.: Comparison of BPM and corresponding network graphs for Worker A and Worker B. 01234567891011 Workplace 0 1 2 3 4 5 6 7 8 9 10 11 Workplace 59 10000000000 0117 1 010000000 1 1 194 000000000 0005800100000 0000122 1 000000 00010124 000000 0000005610000 0000000114 1 000 00000000111 001 0000000005500 0000000001 122 0 00000000001 122 0.0 0.2 0.4 0.6 0.8 1.0 Transitions (Off-Diagonal) 0 25 50 75 100 125 150 175 Time (Diagonal) (a) BPM of simulated autonomous strategy. 01234567891011 Workplace 0 1 2 3 4 5 6 7 8 9 10 11 Workplace 124 00100000000 0106 1 000100000 01 111 000000000 0105600000000 0000104 000010 0 00001 101 000000 0000015800000 0000000102 0010 00000001 101 000 00000000161 0 0 0000000000112 1 00000000000112 0.0 0.2 0.4 0.6 0.8 1.0 Transitions (Off-Diagonal) 0 20 40 60 80 100 120 Time (Diagonal) (b) BPM of simulated standard strategy. 01234567891011 Workplace 0 1 2 3 4 5 6 7 8 9 10 11 Workplace 124 00100000000 0113 0010000000 01 113 000000000 0005600100000 0000102 0010000 0010 0 104 000000 000000560010 0 0100000107 0000 0000010 0 102 000 0000000005801 0000000010102 0 00000000001 101 0.0 0.2 0.4 0.6 0.8 1.0 Transitions (Off-Diagonal) 0 20 40 60 80 100 120 Time (Diagonal) (c) BPM of simulated heuristic strategy. 0 1 2 3 4 5 6 7 8 9 10 11 (d) BPM network graph of automated strategy. 0 1 2 3 4 5 6 7 8 9 10 11 (e) BPM network graph of standard strategy. 0 1 2 3 4 5 6 7 8 9 10 11 (f) BPM network graph of heuristic strategy. Figure 6.: BPM and network graphs of the simulated baseline workflows. 9