MSMFormer：一种基于多变量信号映射器的可变重构Transformer用于设备剩余使用寿命预测

Accurate remaining useful life (RUL) prediction is essential for predictive maintenance in aerospace, manufacturing, and energy industries. However, the complex nature of sensor signals, especially in turbofan engines, poses significant challenges to prediction accuracy. Transformer-based models have shown promise in RUL tasks but are limited by their reliance on fixed activation functions and their focus on dimension-specific information, which reduces their effectiveness in modeling nonlinear and heterogeneous features. To address these limitations, this article proposes a variable reconstruction transformer called MSMFormer, which is based on multivariable signal mappers for RUL prediction. The framework introduces three key innovations. First, the multidomain recursive decomposition method is designed. It reconstructs fine-grained features and retains the weaker degradation information in the signal. Second, the integrated feature extraction network adaptively weights to handle the distribution differences of multivariable data. It emphasizes key predictive features and alleviates the multidelay problem. Third, the Kolmogorov–Arnold network (KAN) is incorporated as a channel mapping module within the transformer. The dynamically optimized activation functions in KAN better capture complex nonlinear feature relationships, enhancing the fine-grained representation of degradation patterns in time series. MSMFormer exhibits average performance gains of 11.11%, 2.40%, and 35.7% on three widely used public datasets, respectively, highlighting its superior robustness and predictive accuracy, especially under complex operational conditions.