Trustworthy Equipment Monitoring via Cascaded Anomaly Detection and Thermal Localization

Abstract: Predictive maintenance demands accurate anomaly detection and trustable explanations. Although multimodal fusion of sensor time-series and thermal imagery shows promise, we demonstrate that naive fusion strategies can paradoxically degrade performance. This paper introduces a Cascaded Anomaly Detection framework that decouples detection and localization. Stage 1 employs an LSTM-based sensor encoder with temporal attention for high-accuracy detection, while Stage 2 activates a CNN-based thermal encoder for post-detection fault localization. Our results reveal that sensor-only detection outperforms full fusion by 8.3 percentage points (93.08% vs. 84.79% F1-score), challenging the assumption that additional modalities invariably improve performance. We further contribute an explainability pipeline integrating SHAP, temporal/spatial attention, and gate weight analysis. This analysis uncovers a "modality bias" where fusion models assign 65-87% weight to the weaker thermal modality. Validated on a real-world bearing dataset (78,397 samples), our cascaded approach achieves state-of-the-art accuracy while providing actionable diagnostics for maintenance decision-making.