Engine remaining useful life prediction method based on deep residual network and attention mechanism

Abstract Traditional RUL prediction models are often constrained by insufficient prediction accuracy, difficulties in feature extraction, and reliance on manually defined health indicators. To address these challenges, this study proposes an engine RUL prediction method that integrates an attention mechanism with a deep residual network. This approach first employs sliding window techniques to construct multi-source time series samples from the engine’s diverse state parameters, comprehensively characterizing its operational status. Building upon this foundation, a prediction model based on a one-dimensional separable convolutional network is developed. To enhance model performance, an attention mechanism layer is introduced to adaptively weight and fuse the importance of state parameters at different time points, thereby strengthening the model’s ability to capture critical degradation features. Simultaneously, residual modules are embedded within the network to connect convolutional layers, mitigating the vanishing gradient problem in deep network training and ensuring training stability. The proposed method was evaluated on the C-MAPSS dataset, achieving an average RMSE of 13.145 across all four test subsets (FD001–FD004). This performance significantly surpasses that of multiple established RUL prediction benchmarks. Additionally, the approach demonstrates strong generalization capability under diverse and variable operational conditions, enabling accurate tracking of degradation progression and reliable RUL estimation. The study offers an efficient and intelligent solution for predictive health management of mechanical systems, particularly aero-engines, with considerable practical applicability.