CMPASS Dataset

This paper introduces an aircraft engine remaining life prediction model based on an improved Transformer architecture (SBi-Transformer), addressing the computational inefficiencies and inadequate local temporal dependency capturing capabilities of the standard transformer model in processing long sequence data. The SBi-Transformer employs a dual-layer attention mechanism to reduce computational complexity and enhance local temporal dependencies. It eliminates the decoder component, directly using BiLSTM to model the time series of the encoder outputs, supplementing local temporal context information, and strengthening the forward and backward dependency modeling of degradation trends. Ultimately, the integrated global-local features are mapped to remaining useful life (RUL) prediction values through a fully connected layer. Experiments on NASA’s C-MAPSS dataset demonstrate that SBi-Transformer surpasses existing methods in prediction accuracy and uncertainty estimation. Ablation experiments validate the synergistic effects of the modules, indicating that the sparse attention mechanism and BiLSTM complementarily enhance model performance. The model exhibits strong adaptability and generalizability under complex operating conditions and multiple fault modes, with prediction result confidence intervals dynamically converging to the true degradation trajectory. This study provides an efficient and accurate solution for predicting the remaining life of aircraft engines, holding significant engineering application value