Behavioural Fault Signature Analysis for Automated Manufacturing Systems Using Dynamic Signal Sandboxing and Control-Sequence N-Grams

I. INTRODUCTION

Modern automated production lines exhibit fault behaviour that varies with wear, load and calibration drift in ways static threshold checks cannot anticipate, while the underlying actuator command and control-sequence patterns remain relatively stable across occurrences within the same failure family, motivating dynamic behavioural analysis.

II. METHODOLOGY

A total of 3,400 labelled control-sequence traces, comprising faulty runs from six major failure families and an equal number of healthy runs, were replayed in an isolated test-rig sandbox for 180 seconds each, with control-opcode sequences extracted from logged instruction traces converted to 3-gram frequency vectors and combined with actuator-command frequency features, used to train a random forest classifier with 400 trees.

III. RESULTS AND EVALUATION

The classifier achieved 96.8 percent binary fault-detection accuracy (faulty versus healthy) and, among correctly detected faulty traces, correctly attributed the specific failure family in 91.2 percent of cases based on hierarchical clustering of the same feature set, with the two most confused family pairs sharing a common underlying actuator subsystem.

IV. CONCLUSION

Control-opcode n-gram and actuator-command behavioural features extracted via sandboxed dynamic replay provide robust fault detection resilient to gradual drift, along with useful family attribution for maintenance response. Future work will evaluate detection latency trade-offs for shorter sandbox replay windows.

V. REFERENCES