PhyRes-BiLSTM: Physics-Informed Residual Bidirectional LSTM for Aircraft Engine Remaining Useful Life Prediction

The prediction of an aircraft engine's Remaining Useful Life (RUL) is crucial for flight safety and maintenance cost control. In this paper, a novel physics-informed residual bidirectional LSTM model (PhyRes-BiLSTM) is proposed. Physical laws were integrated with deep learning to address the limitations of traditional methods. The model was constructed by 1) integrating monotonicity constraints into the Physics-Informed Neural Network (PINN) to enforce decreasing RUL predictions over time, and 2) establishing residual connections to optimize gradient propagation and alleviate challenges in training deep networks. Experiments using the NASA CMAPSS benchmark dataset demonstrate that the proposed PhyRes-BiLSTM significantly outperforms mainstream models such as GRU and TCN. In the FD001 subset, it achieves improvements of 8.0%, 4.4%, 2.6%, and 2.1% in RMSE, MAE, R2, and EVS, respectively, compared to the best baseline model (Transformer). The physical constraint mechanism effectively suppresses unreasonable predictions, providing a highly reliable solution for predictive maintenance of aircraft engines.