BO-CNN-LSTM-RSA: A Hybrid Deep Learning Model for Aircraft Engine Remaining Useful Life Prediction

Abstract

Aircraft engine performance degradation affects flight safety and requires regular maintenance, creating significant cost pressures for airlines. Remaining Useful Life (RUL) prediction, as a core task in Prognostics and Health Management (PHM), has attracted extensive research attention. This paper proposes a hybrid deep learning model named BO-CNN-LSTM-RSA for aircraft engine RUL prediction. The model integrates Convolutional Neural Networks (CNN) for local feature extraction, Long Short-Term Memory (LSTM) networks for temporal dependency capture, and a Residual Self-Attention (RSA) mechanism to enhance focus on critical degradation stages while alleviating gradient vanishing. Bayesian Optimization (BO) is employed for automatic hyperparameter tuning. Experiments on the NASA C-MAPSS FD001 dataset show that the proposed BO-CNN-LSTM-RSA model achieves an RMSE of 11.657 and a Score of 194.765, outperforming existing methods. Ablation studies validate the effectiveness of each module, with RSA improving both prediction accuracy and training stability. The proposed model provides an effective solution for aircraft engine RUL prediction.