Unlocking the Black Box: An Explainable Hybrid Framework for Aero-Engine RUL Prediction
作者
Karthika Natarajan,Vignesh G A,Sarathy S P,Hari Prasath M,K. Aravind
标识
DOI:10.1109/icsscna68616.2026.11546795
摘要
Accurate estimation of Remaining Useful Life (RUL) is essential in the predictive strategies of maintenance of the complex engineering systems, especially aircraft engines. Although deep learning solutions can be effective to predicting degradation trends of multivariate sensor data, in most cases, they are not able to simultaneously learn about immediate degradation motives and longer-term temporal dependencies and still be interpretable. In this paper, explainable hybrid CNN-Transformer-BiLSTM RUL prediction framework was proposed with the help of the NASA C-MAPSS benchmark datasets. The architecture employs convolutional front-end to model local features, Transformerbased self-attention layers to model long-range dependence and a Bi-directional LSTM layer to sequentially refine. In order to promote robustness, the training procedure includes the onflow balancing of data and the development of a client-specific asymmetric loss advantage as per the NASA scoring framework, which induces considerable abuse to dismal late-phase forecasting mistakes. The framework is confirmed on all four C-MAPSS subsets, that are more and more challenging fault scenarios. RMSE, MAE, R2and NASA asymmetric score have been used to assess the performance that has shown a steady convergence and predictive performance among various operating regimes. Moreover, temporal steps and sensor contributions in models are clarified and improved through Integrated Gradients and SHAP analysis, which explain the contribution of individual input factors. In turn the suggested framework provides a strong, precise, and clear methodology of prognostic health management.