Interpretable RUL Prediction via Hybrid Deep Evidential clustering for Aero-Engine Health Monitoring

Abstract

Remaining Useful Life (RUL) prediction is critical for predictive maintenance in safety-critical systems such as aerospace engines. While deep learning models achieve high predictive accuracy, they often lack interpretability and reliable uncertainty estimation. This paper proposes the Hybrid Deep Evidential Clustering (HDEC) framework to address both challenges. A CNN–LSTM–GRU backbone first extracts degradation features from multivariate time-series data. These features are then clustered using NN-EVCLUS, an evidential clustering approach based on Dempster–Shafer theory, which groups engines according to their degradation stage. A dedicated RUL predictor is trained for each cluster to enable specialized and interpretable predictions. Engines with uncertain cluster membership are handled through soft memberships, allowing RUL estimation as a weighted combination of cluster-specific predictors instead of hard assignment. Experiments on the NASA C-MAPSS dataset demonstrate that HDEC improves predictive performance while providing well-calibrated uncertainty estimates and interpretable degradation-regime assignments.