Process mining-driven modeling and simulation to enhance fault diagnosis in cyber-physical systems

A new process-mining pipeline turns raw robotic-arm sensor data into interpretable Petri-net models that both explain and simulate faults—giving manufacturing leaders faster root-cause insights and more accurate digital twins.

1. What the method is

The paper introduces an unsupervised fault-diagnosis workflow that detects collective anomalies in multivariate time-series data, converts them into event logs, and then applies process-mining to learn timed Petri-net models. Those models can be simulated to reproduce abnormal behaviour, yielding both interpretable explanations and synthetic fault data for downstream analytics.

2. Why the method was developed

Traditional fault-diagnosis strategies in cyber-physical systems rely on expert-crafted rules, opaque deep-learning detectors, or labour-intensive physical models that seldom support realistic simulation. The authors aim to bridge this gap by combining process-mining transparency with stochastic simulation, so that engineers can automatically build fault dictionaries, accelerate root-cause analysis, and feed more faithful digital-twin loops without sacrificing explainability or accuracy.

3. Who should care

• Heads of Manufacturing & Operations
• Digital-Twin and Industry 4.0 Program Leads
• Reliability & Maintenance Engineers

4. How the method works

First, multivariate sensor streams are segmented into windows and clustered to spot collective anomalies. Each anomalous window is translated into an event log with state transitions. Integer Linear Programming and the Inductive Miner with filtering are then applied to discover Petri-net structures capturing the faulty workflow. Timing distributions extracted from logs are embedded in the nets, enabling Monte-Carlo simulation of the abnormal sequence. The resulting models provide both a visual explanation of causal chains and a generator of synthetic traces that augment scarce fault data for classification or digital-twin what-if studies.

5. How it was evaluated

Experiments used the open Robotic Arm Dataset (RoAD), a benchmark covering velocity- and weight-related anomalies in smart-manufacturing lines. Fitness, precision, root-mean-square-error (RMSE), and model-complexity metrics assessed how well simulated traces matched real faults, while a downstream classifier measured fault-type recognition accuracy.

6. How it performed

The best Petri-net achieved a fitness of 0.79 on velocity anomalies and delivered classification accuracy above 91 % with only ±7.8 % variance, outperforming models built from raw data alone. Simulation RMSE stayed below 0.06, and the velocity-to-weight fitness ratio peaked at 1.50, indicating robust generalisation across fault types:contentReference[oaicite:7]{index=7}. (Source: arXiv 2506.21502, 2025)