Remaining Useful Life Prediction of Turbofan Engines Using Metaheuristic-Optimised LSTM with SHAP-Based Explainability

Abstract

This study presents a deployment-aware framework for explainable remaining useful life prediction of turbofan engines, integrating metaheuristic hyperparameter optimisation with deep learning and SHAP-based explainability. Long Short-Term Memory (LSTM) models are optimised using a genetic algorithm (GA) and particle swarm optimisation (PSO), while an untuned bidirectional LSTM (BiLSTM) serves as the architecturally stronger reference. This asymmetric design quantifies the regime-dependent conditions under which automated tuning of a single-direction LSTM substitutes for the additional cost of a bidirectional architecture. The framework is evaluated on the NASA C-MAPSS benchmark across all four sub-datasets (FD001–FD004) under an engine-wise validation split that eliminates window-level data leakage and across N=10 independent runs analysed with Kruskal-Wallis and Dunn post-hoc tests. Results demonstrate that the substitution is regime-dependent. Under single-condition operation (FD001), the mean RMSE values converge within 14.74–14.86 cycles with no significant differences (p=0.545); PSO-LSTM attains the numerically lowest mean (14.736±0.489 cycles) and BiLSTM has the lowest variance (CV = 1.51%). GA-LSTM achieves the lowest mean RMSE under dual-fault operation (14.015±0.636 cycles on FD003). Under multi-condition operation (FD002, FD004), BiLSTM delivers the lowest mean RMSE with minimal variability (30.83 and 35.28 cycles), while metaheuristic-tuned LSTMs collapse to the mean-prediction baseline in ten of ten runs on at least one subset. SHAP attributions are reshaped by engine complexity: HPC outlet pressure (Ps30, phi), exhaust gas temperature (T50) and core rotational speed (Nc) dominate under single-condition operation, while operating-condition descriptors emerge as primary contributors under multi-condition dynamics. The framework provides an automated, interpretable, and statistically validated approach for turbofan engine prognostics.