An Investigation into Unsupervised Anomaly Detection for Data-Driven Predictive Maintenance

This study explores anomaly detection in predictive maintenance, focusing on unsupervised learning techniques for online condition monitoring. The shift from model-based to data-driven approaches, facilitated by advancements in sensing technologies and computer science, is examined within the predictive maintenance framework. The efficacy of prominent unsupervised anomaly detection methods in identifying abnormal patterns in equipment data streams, particularly streaming time-series data, is assessed. The study evaluates three classes of methods for unsupervised anomaly detection - Local Outlier Factor, Isolation Forests, and Long Short-Term Memory - quantitatively and qualitatively, for different unique contexts in the case of machine vibration and temperature monitoring. Results indicate that although the Long Short-Term Memory model outperforms the other models across various metrics, there is still the challenge with root cause analysis. An improvement on the use of Long Short-Term Memory as a detection technique is presented to address this specific challenge of root cause analysis, particularly in the case of gradual degradation and/or cascading faults. Further recommendations on evaluating methods for unsupervised detection in predictive maintenance are also offered.