关键词: 航空发动机, 剩余使用寿命预测, 多维度预处理, 深度均衡网络, 双输入注意力机制

Abstract: To address the core challenges arising from data distribution heterogeneity and strong noise interference induced by multi-operating-condition switching in aero-engines, as well as the gradient vanishing, excessive memory consumption, and insufficient generalization under complex operating conditions inherent in conventional deep networks for long-sequence remaining useful life (RUL) prediction, this paper proposes an aero-engine RUL prediction method that integrates multi-dimensional data preprocessing with an attention-enhanced deep equilibrium network. First, a multi-dimensional preprocessing pipeline comprising operating condition clustering, Z-Score grouped normalization, random forest feature selection, and exponential moving average temporal smoothing is constructed to purify and denoise degradation features under complex operating conditions, thereby mitigating the distribution shift caused by multi-operating-condition switching at the data level. Subsequently, built upon the deep equilibrium network framework, a dual-input attention mechanism is designed to achieve collaborative enhancement of temporal features and hidden layer states; an Anderson acceleration solver is introduced to optimize the convergence of fixed-point iterations; and a Bayesian regression module is incorporated to quantify the uncertainty of prediction results. These designs simultaneously alleviate long-sequence gradient vanishing and memory overload at their source while substantially improving the model's adaptability to complex operating conditions and prediction reliability. Simulation experiments are conducted on the NASA C-MAPSS aero-engine dataset. Ablation studies verify the necessity and synergistic benefits of the three core components: the dual-input attention mechanism, the Anderson acceleration solver, and data smoothing. Compared with 12 state-of-the-art RUL prediction methods, the proposed method achieves the best performance on FD001, FD002, and FD004 and ranks among the top on FD003; particularly on the complex multi-operating-condition datasets FD002 and FD004, the RMSE is reduced to as low as 11.91 and 11.97, respectively, demonstrating significant advantages. To further validate the cross-domain generalization capability and engineering adaptability, the proposed method is transferred to the N-CMAPSS high-fidelity aero-engine simulation dataset, where it maintains stable and excellent prediction performance across all 8 sub-datasets of graded complexity, fully demonstrating its generalization capability and reliability in complex engineering scenarios. The proposed method achieves the collaborative optimization of RUL prediction accuracy and computational efficiency within a purely data-driven paradigm, exhibiting superior robustness and generalization under both single- and complex multi-operating conditions, and can provide reliable technical support for aero-engine predictive maintenance.

Key words: aero-engine, remaining useful life prediction, multi-dimensional preprocessing, deep equilibrium network, dual-input attention mechanism