Speakers
Description
The development of data-driven fault-diagnosis systems is often constrained by the limited availability of representative faulty data. Industrial faults may occur rarely, require long observation periods to capture, or be unsafe and costly to reproduce under actual operating conditions. Consequently, operational datasets frequently contain abundant normal or healthy data but insufficient examples covering different fault types, locations, and severity levels. Anomaly-detection methods address part of this limitation by learning normal system behaviour and identifying deviations as abnormal. However, such methods may only indicate that abnormal behaviour has occurred and may be unable to reliably determine the fault type, location, or severity. This paper presents fault injection methodology as a complementary technique for generating controlled and labelled faulty data when field data are unavailable, incomplete, or impractical to collect. Faults are systematically introduced into a physical system, simulation model, digital twin, or historical healthy dataset by modifying selected parameters, signals, components, or operating conditions. The resulting data can represent predefined fault categories, locations, progression patterns, and severity levels. These labelled datasets can then be used to train and evaluate diagnostic algorithms for fault detection, isolation, classification, localization, and severity estimation. The methodology includes selecting credible fault modes, defining injection mechanisms and severity levels, generating data across representative operating conditions, validating the realism of the injected responses, and assessing algorithm robustness. Fault injection does not replace real faulty data; rather, it provides a structured means of supplementing limited observations and improving diagnostic coverage.
| Topic Area | Smart Operations and Maintenance |
|---|