Unsupervised Fault Event Detection in Historical Industrial Telemetry With Temporal Event Structuring

In modern industrial process monitoring of multivariate time series, the effectiveness of data-driven event detection is often constrained by scarce labels and the dynamic complexity of real-world telemetry. While unsupervised reconstruction models provide a viable alternative to supervised learning, their direct application to continuous industrial streams frequently yields noisy, fragmented alarms and limited interpretability. This paper presents an unsupervised anomalous event detection framework designed to extract structured anomalous events and interpretable evidence from historical data. Beyond standard residual thresholding, the proposed temporal event-structuring stage integrates a dynamics-aware fault score built from reconstruction-error magnitude and derivatives, together with dual-threshold hysteresis and variable-set continuity rules to stabilize event boundaries and reduce fragmentation. The framework operates without requiring labels for training or event structuring, and is evaluated on DAMADICS Phase III, a representative case study of coupled actuator telemetry with partial annotations, treating detections outside documented windows as unverified rather than false. Experimental results show a mean temporal Coverage of 92.97% across documented windows and a Variable Recall (VarRecall) of 87.04%, with detected events overlapping all 18 documented fault windows. Comparative ablations further confirm the distinct role of derivative terms, hysteresis, and continuity in supporting temporally consistent event structuring, achieving a median fragmentation of 1 segment per window. This event-level consistency enables the retrieval of granular interpretive evidence, including activation chronology and activation-cause attribution, without relying on extensive fault labels.