Abstract Predictive maintenance in safety-critical systems like turbofan engines increasingly relies on machine learning (ML) models to estimate remaining useful life (RUL), but the ‘black box’ nature of these models hinders their adoption and trustworthiness. While traditional ex-ante prognostic metrics (e.g. monotonicity, trendability) are used to pre-screen sensor data, a systematic comparison against the post-hoc explanations of what a model actually learns is lacking. We explore the application of SHapley Additive exPlanations (SHAP) from explainable artificial intelligence (XAI) to investigate feature importance in engine failure prediction using the second dataset of the Commercial Modular Aero-Propulsion System Simulation (CMAPSS). The preprocessing pipeline includes z-score normalisation of sensor data and the calculation of a health index (HI) to quantify system degradation. A power-law fit is applied to the HI to capture the underlying trends of engine wear and failure progression. We use the normalisation data to calculate prognostic feature selection metrics: monotonicity, trendability and prognosability. Then, we train two machine learning models – random forest (RF) regressor and gradient boosting (GB) method – directly from the raw data to predict the RUL based on the actual sensor readings. The SHAP values generated for both models are analysed to identify the features with the most significant impact on RUL predictions. By comparing the SHAP value distributions across models and prognostic predictors, we highlight feature robustness and their relative influence on engine degradation and failure prediction. This work provides insights into the interpretability of machine learning models in prognostics and enhances the understanding of sensor contributions to engine health monitoring. The results demonstrate the effectiveness of SHAP in elucidating feature importance, supporting the development of more transparent and reliable prognostic systems.
Reha Avsar (2026) studied this question.