Publikation
Towards Objectively Interpretable Fault Diagnosis forTime-Series Data in Grinding
Tsan Tsai Chan; Kai Lange; Ruoxuan Liu; Alessa Wein; Nina Keßler; Maxx Richard Rahman; Wolfgang Maaß
In: Workshop on Information Technology and Systems. Workshop on Information Technology and Systems (WITS-2024), WITS, 12/2024.
Zusammenfassung
Manually diagnosing mechanical faults is highly laborious, and a lack of strategies to intuitivelyexplain model decisions has precluded effective automated diagnosis by unsupervised systems. Toremedy this, we propose an objectively interpretable probabilistic framework for analysing time-series data in grinding, using a Gaussian mixture model (GMM) trained only on normal processesto assign likelihoods to each 50-ms segment of a grinding signal. A dashboard visualises theselikelihoods in sequence, showing where during a process anomalies crop up and yielding diagramsthat could expedite visual fault diagnosis. Anomalous grinding runs are objectively defined bydeviations from normal training data, facilitating predictive maintenance. We show that ourframework allows various simple GMM-based architectures to outperform a more complexrecurrent one in stability of F2-scores on small training samples with simulated anomaly data.
