A Multi-Source Domain Generalization Method for Bearing RUL Prediction Based on Time-Variant and Time-Invariant Feature Extraction

Deep learning-based remaining useful life (RUL) prediction methods have demonstrated significant advantages due to their powerful feature representation capabilities. Although existing learning models have made progress in distinguishing between the time-varying and time-invariant characteristics of vibration signals, challenges remain in fully capturing these aspects, which can hinder their generalization performance. To address this issue, we propose a domain generalization method based on the extraction of time-varying and time-invariant features, tailored to the characteristics of bearing vibration time series signals. Firstly, Fourier transform and inverse Fourier transform are employed to isolate high-frequency features, obtaining both time-varying and time-invariant signals. The time-varying signals are then processed using an attention mechanism to extract time-varying features and learn the bearing degradation trend. Concurrently, the time-invariant signals are fed into an encoder-decoder structure to extract their invariant features. Finally, both sets of features are input into a Gate Recurrent Unit (GRU) module for RUL prediction. Experimental results across three tasks demonstrate that our model achieves excellent generalization capabilities.