Causal Wavelet Residual Learning with Observation-Support Gating for Industrial Sensor Anomaly Detection

Abstract

Industrial sensor anomalies must be detected under missingness, noise, drift, and finite operator attention. Point-wise scores alone can obscure the resulting alarm burden. We present RC-WMRAD, a causal wavelet residual detector that combines raw temporal prediction with a left-padded multi-scale Haar path and an observation-support gate. Training uses normal data, thresholds are calibrated independently, score bundles are locked before labels are read, and evaluation uses no point adjustment. Six public benchmarks reveal conditional effects. Against Raw-TCN, RC-WMRAD reduces false alarms on SKAB, SMD, MSL, SMAP, and the railway MetroPT-3 benchmark, but the accompanying ranking and event-coverage effects differ by dataset. Against TranAD-style, SMD shows lower AUPRC and higher event recall, whereas PSM shows lower event recall and uncertain AUPRC. On MetroPT-3, TranAD-style ranks failures more strongly, covers a larger fraction of reported failure intervals, and responds earlier, while RC-WMRAD produces fewer false-alarm segments. Corruption, ablation, and calibration analyses preserve these metric-specific boundaries rather than establishing universal robustness. The study therefore contributes an auditable sensor-monitoring protocol and a scoped account of ranking, partial event coverage, and alarm burden, with no pooled global effect.