Physics-guided explainable predictive maintenance framework for industrial IoT systems

Machine learning models are increasingly deployed in industrial systems for predictive maintenance, enabling failure detection before breakdowns occur. However, most existing approaches optimise solely for statistical performance and do not explicitly enforce consistency with established physical relationships, which can yield predictions that violate known thermomechanical trends and degrade reliability under varying operating conditions. This paper presents the Physics-Guided Explainable Predictive Maintenance (PG-XPM) framework, a six-module architecture that integrates thermomechanical domain knowledge into a gradient-boosted classification pipeline through engineered features, structural constraints, and a domain-aware evaluation metric. We use the term physics-guided to denote this feature- and constraint-level integration, and explicitly distinguish it from physics-informed approaches that embed governing differential equations directly into the training loss (e.g. physics-informed neural networks); PG-XPM does not include a physics-based loss term or a dynamical degradation model, and this scoping is stated explicitly to avoid overstating the mechanism of integration. The framework augments XGBoost with four closed-form features—mechanical power, thermal gradient, strain index, and thermal load ratio—applies selective monotonic constraints exclusively to these physics-derived inputs, and employs a load-adaptive decision threshold that modulates classification sensitivity according to real-time mechanical load. A domain-aware evaluation metric, the Physics Violation Index (PVI), is introduced to quantify the degree to which predicted failure probabilities respect the monotonicity trend that failure risk increases with mechanical load, formulated as the Spearman rank correlation between a composite load index and predicted failure probabilities (with Kendall’s \(τ \) reported as a robustness check); values approaching \(+1\) indicate stronger physical consistency. Evaluated on the AI4I 2020 Predictive Maintenance Dataset with a real (not merely described) SMOTE-based ablation and 5-fold stratified cross-validation, the proposed framework reduces the false negative rate from a mean of 0.360 to 0.153 (a 57.5% mean relative reduction, paired \(t=-9.19\) , \(p<0.001\) , consistent across all folds) relative to a well-configured, non-resampled XGBoost baseline, alongside a statistically significant improvement in PVI (0.291 to 0.348, \(p=0.017\) ). A sensitivity sweep over the adaptive-threshold coefficients and a latency/footprint profiling study are reported to support the framework’s suitability for resource-constrained IIoT edge gateways. SHAP-based analysis confirms that model decisions are driven predominantly by mechanically and thermally meaningful features; we term this observation physics-validated explainability only as informal shorthand for attribution-feature consistency, not as a claim of causal or domain-expert-validated mechanistic explanation.