A meta-curriculum dynamic weighting network equipped with frequency-aware attention for bearing cross-domain remaining useful life prediction

Bearing function as the core transmission components in rotating machinery systems. Its operational status directly influences the reliability of critical equipment in fields like wind power generation and rail transportation. Accurate remaining useful life (RUL) prediction constitutes a crucial element for preventing unexpected failures and optimizing condition-based maintenance decisions. However, in practical engineering domains, bearing RUL often faces cross-domain scenarios with large distributional differences. Traditional transfer learning methods adopt the same weights for all source domain samples, ignoring the transfer benefit from sample variability in cross-domain scenarios, resulting in model validity that is not always guaranteed and triggering negative transfer effects. Meanwhile, end-to-end feature learning mechanisms tend to obscure physical degradation information across different failure modes, resulting in suboptimal prediction accuracy. To address these issues, this paper introduces an innovative framework integrating frequency-aware spectral attention with meta-curriculum transfer learning. Firstly, by designing a dynamic fault prototype parameter mechanism, the physical a priori knowledge of bearing failures is introduced to improve the physical interpretability and prediction accuracy of the model. Simultaneously, a weighted network based on meta-curriculum learning dynamically assigns multiple types of weights to the samples, enabling adaptive learning toward more favorable update directions. Experimental validation using the XJTU-SY dataset demonstrates satisfactory RUL prediction performance across diverse cross-domain scenarios. Compared with state-of-the-art models, our method significantly reduces the prediction error, validating the advantages of the proposed method for engineering applications.