Process mining-driven modeling and simulation to enhance fault diagnosis in cyber–physical systems
Cyber–Physical Systems (CPSs) tightly interconnect digital and physical operations within production environments, enabling real-time monitoring, control, optimization, and autonomous decision-making that directly enhance manufacturing processes and productivity. The inherent complexity of these systems can lead to faults that require robust and interpretable diagnoses to maintain system dependability and operational efficiency. However, manual modeling of faulty behaviors requires extensive domain expertise and cannot leverage the low-level sensor data of the CPS. Furthermore, although powerful, deep learning-based techniques produce black-box diagnostics that lack interpretability, limiting their practical adoption. To address these challenges, we set forth a method that performs unsupervised characterization of system states and state transitions from low-level sensor data, uses several process mining techniques to model faults through interpretable stochastic Petri nets, simulates such Petri nets for a comprehensive understanding of system behavior under faulty conditions, and performs Petri net-based fault diagnosis. The method begins with detecting collective anomalies involving multiple samples in low-level sensor data. These anomalies are then transformed into structured event logs, enabling the data-driven discovery of interpretable Petri nets through process mining. By enhancing these Petri nets with timing distributions, the approach supports the simulation of faulty behaviors. Finally, faults can be diagnosed online by checking collective anomalies with the Petri nets and the corresponding simulations. The method is applied to the Robotic Arm Dataset (RoAD), a benchmark collected from a robotic arm deployed in a scale-replica smart manufacturing assembly line. The application to RoAD demonstrates the method’s effectiveness in modeling, simulating, and classifying faulty behaviors in CPSs. The modeling results demonstrate that our method achieves a satisfactory interpretability-simulation accuracy trade-off with up to 0.676 arc-degree simplicity, 0.395 R2, and 0.088 RMSE. In addition, the fault identification results show that the method achieves an F1 score of up to 98.925%, while maintaining a low conformance checking time of 0.020 s, which competes with other deep learning-based methods.