Hierarchical Residual Attribution with Supervised Channel Shortlisting for Sensor-Local and Process Fault Diagnosis in Multivariate Sensor Time Series
TL;DR
This paper proposes Hierarchical Residual Attribution with supervised channel shortlisting (HRA-SL), a diagnostic framework for multivariate sensor windows under a controlled injected-perturbation protocol, establishing a controlled-injection benchmark for reproducible sensor source diagnosis and affected-channel shortlisting on public base signals.
Abstract
Binary anomaly alarms are often insufficient for sensor systems because maintenance actions depend on whether abnormal behavior is localized to a sensor channel or reflects a process-level event. This paper proposes Hierarchical Residual Attribution with supervised channel shortlisting (HRA-SL), a diagnostic framework for multivariate sensor windows under a controlled injected-perturbation protocol. HRA fits calibration-based cross-channel consistency residuals for source-level diagnosis; HRA + spectral is the main source-diagnosis configuration; and HRA-SL adds supervised channel-candidate shortlisting for affected-channel localization. The evaluation uses synthetic, UCI HAR, and UCI Air Quality base signals with split-before-injection protocols over five seeds and five severity levels. HRA + spectral improves source-level macro-F1 over statistical, spectral, statistical + spectral, tuned TCN-AE, tuned USAD-like AE, and GRU-AE reconstruction-feature baselines, reaching 0.780±0.028, 0.655±0.017, and 0.554±0.029, respectively. Ablations show that HRA-Core carries most source-level signal, while residual-attribution features support channel ranking and diagnostic interpretation. The fixed HGB HRA-SL localizer reaches top-1 affected-channel localization of 0.916, 0.714, and 0.715. Under matched target-channel supervision, HRA-SL is practically tied with AE-only HGB on synthetic data (0.916 versus 0.918) and UCI HAR (0.714 versus 0.716), and is higher on Air Quality (0.715 versus 0.652). It also improves over the strongest unsupervised tuned reconstruction ranking baseline by 0.127, 0.357, and 0.048 top-1 under different supervision assumptions. The study establishes a controlled-injection benchmark for reproducible sensor source diagnosis and affected-channel shortlisting on public base signals, providing pre-deployment evidence that can be extended to field-fault studies with maintenance records.