Degradation mode identification and remaining useful life prediction via an interpretable CNN-BiLSTM framework

Remaining useful life prediction is essential for condition-based maintenance in safety-critical industries, but existing deep learning approaches often exhibit degraded accuracy under variable operating conditions and provide limited interpretability. This paper presents an explainable CNN-BiLSTM framework incorporating regime-specific normalisation and degradation-aware feature engineering. The methodology employs clustering-based regime identification with condition-specific normalisation to mitigate covariate shift, alongside two complementary feature types: delta-from-baseline features capturing cumulative deviation from healthy states, and first-order temporal differences encoding instantaneous degradation rates. Validation on the NASA C-MAPSS benchmark demonstrates competitive performance under single operating conditions and substantial improvements over existing methods under variable conditions, particularly on the most challenging multi-regime, multi-fault subset. Post-hoc explainability analysis reveals that engineered features dominate model predictions, accounting for over 82% of total importance, whilst raw sensor values contribute minimally. The analysis further identifies three degradation mode clusters with characteristic feature utilisation patterns, indicating that the model learns to recognise multiple degradation signatures. Prediction accuracy improves markedly as engines approach failure, demonstrating highest precision when accurate forecasts are most consequential for maintenance decisions.