SHAP-Based Feature Selection with a Hybrid Convolutional and Recurrent Deep Learning Framework for Remaining Useful Life Prediction of Aero-Engines

Abstract

Prognostics and health management for aeroengines is crucial, especially for reducing catastrophic loss of life and minimizing maintenance costs. Accurate Remaining Useful Life (RUL) estimation optimizes schedules, lowers costs, and enhances safety. Physics-based modeling for RUL assessment is challenging due to its inability to fully account for system and environment-related dynamics. Therefore, machine and deep learning approaches are highly recommended. This paper proposes a new method for RUL prediction of turbofan engines using the well-known (CMAPSS) dataset. Generally, data from acquired engines may be corrupted by anomalies due to sensor failures or environmental disturbances, which can affect the accuracy of prediction models. Therefore, this paper combines advanced techniques for reducing irrelevant and redundant sensor signals by applying Shapley Additive Explanations (SHAP) to refine feature selection by quantifying each sensor’s contribution to RUL prediction, ensuring both interpretability and efficiency. Then an anomaly detection and removal followed by RUL prediction with deep learning are used for such complex tasks. Specifically, K-means clustering, autoencoders, Temporal Convolutional Networks (TCN) and Long Short-Term Memory (LSTM) networks. This approach enables effective data preprocessing by detecting and removing anomalies, thus enhancing the quality of the training data. Moreover, uncertainty analysis is performed to assess the reliability of the predictions, including confidence intervals for the results. The experimental results show substantial improvements in predictive accuracy, confirmed by better evaluation metrics, numerical evaluations, and the coefficient of determination, compared to traditional approaches. The TCN-LSTM model achieved an average performance improvement of 24.05% on FD001, 12.44% on FD003, and 4.63% on FD002 (using Monte-Carlo uncertainty estimation). Furthermore, for the FD004, a performance decrease of -4.68