Decoding Degradation: An Interpretable Regime-Aware Framework For Sustainable Predictive Maintenance In Cyber-Physical Systems

The integration of Artificial Intelligence into safety-critical industrial systems is hampered by the ”black-box” problem, where high-performance models lack the interpretability required for operational trust. This is particularly acute in predictive maintenance (PdM) for cyber-physical systems, where existing modelsoften treat degradation as a monolithic process, failing to capture the differential impact of varying operational stresses. This paper introduces the Regime-Aware Prognostics (RAP) framework, a novel, two-stage computational methodology that resolves this performance-interpretability trade-off. The RAP framework first uses unsupervised clustering to decode discrete operational regimes from high-frequency sensor data. It then leverages these regimes as explicit, interpretable features in a specialized supervised model to predict Remaining Useful Life (RUL). We validate this framework on the critical problem of proactive fuel filter maintenance in heavy mining equipment, a sector demanding both high reliability and transparent decision-making. Using industrial SCADA data, our validation, centered on Sur vival Analysis (Cox Proportional Hazards), moves beyond mere prediction to pro vide deep mechanistic insight. We quantitatively prove that the operational regime is a primary, statistically significant predictor of equipment failure (Hazard Ratio > 3.2), an insight entirely missed by conventional methods. The RAP model achieves an R2 of 0.88 and reduces prognostic error (RMSE) by over 50% compared to benchmarks, translating to an estimated 15-20% reduction in maintenance costs and a >10%increase in equipment uptime. This work presents a generalizable and fully reproducible (Docker, DVC, Zenodo) blueprint for building trustworthy AI, offering a paradigm shift from time-based to truly condition-and-regime-based maintenance, with significant implications for sustainable and efficient industrial operation