Trustworthy Equipment Monitoring via Cascaded Anomaly Detection and Thermal Localization

TL;DR

The paper tackles trustworthy equipment monitoring by challenging the default of multimodal fusion when modality informativeness is asymmetric. It proposes a cascaded architecture in which a Random Forest detector operates on engineered statistical features from sensor time-series, followed by a CNN with spatial attention for post-detection thermal localization. An explainability pipeline combining TreeSHAP, temporal and spatial attention visualizations, and gate-weight analysis reveals a modality bias in fusion models and provides actionable auditing tools. Across 13,121 real-world samples, the cascaded approach with stage-specific roles significantly outperforms LSTM and end-to-end fusion, achieving 94.66% F1 and 99.61% AUROC for detection, while enabling interpretable localization that supports trust and regulatory compliance. These findings advocate for architecture design that respects modality roles and emphasize XAI-driven diagnostics to guide industrial deployment and transferability.

Abstract

Predictive maintenance demands both accurate anomaly detection and interpretable explanations. We demonstrate that naive multimodal fusion of sensor time-series and thermal imagery can degrade performance, and instead propose a cascaded, hybrid architecture. Our approach utilizes Random Forest on statistical sensor features for detection ($94.66\%$ F1), triggering a CNN with spatial attention for thermal fault localization only post-detection. Rigorous analysis reveals that statistical feature-based detection significantly outperforms both LSTM ($89.57\%$ F1) and end-to-end fusion ($84.79\%$ F1) at typical industrial noise levels. However, we identify a critical noise crossover phenomenon: while Random Forest excels at low noise, deep learning approaches demonstrate superior resilience at high noise ($σ> 0.3$). Additionally, we introduce an explainability pipeline integrating TreeSHAP and attention heatmaps to diagnose "modality bias," where fusion models irrationally favor weaker thermal inputs. Validated on 13,121 real-world samples from automated transport systems, this work provides evidence-based guidelines for model selection, proving that traditional machine learning often surpasses complex deep learning for industrial monitoring while offering superior interpretability.

Figures (8)

• Figure 1: Hybrid Cascaded Anomaly Detection Framework. Stage 1 uses Random Forest on statistical sensor features for high-accuracy detection (94.7% F1). Stage 2 activates post-detection to localize faults on thermal images using CNN with spatial attention.
• Figure 2: Ablation comparison across model variants. Statistical RF achieves the highest performance (94.7% F1), while adding thermal features via naive fusion degrades results.
• Figure 3: Learning curves showing F1-score versus training data fraction. Random Forest (blue) consistently outperforms LSTM (orange) and Fusion (green) models across all data fractions, with shaded regions indicating standard deviation across 3 seeds. RF achieves 90.6% F1 with only 10% of data, demonstrating superior data efficiency.
• Figure 4: TreeSHAP feature importance for the Random Forest model. Statistical features from the NTC temperature sensor (mean, std, min, max) collectively contribute most to predictions, aligning with domain knowledge.
• Figure 5: Temporal attention weights for the LSTM baseline across time steps. Later time steps receive higher attention (0.049 $\rightarrow$ 0.051), indicating that recent observations are most predictive of equipment state.