A meta-transfer learning-guided approach for remaining useful life prediction of rolling bearings with small-sample data

Abstract Rolling bearings, as a key transmission component of rotating machinery, require accurate life prediction and effective health management to enable intelligent operation and maintenance and ensure system reliability. A small-sample remaining useful life (RUL) prediction approach for rolling bearings based on meta-transfer learning is proposed in this paper. By fusing model-agnostic meta-learning (MAML) and domain adversarial neural networks (DANN), a MAML-DANN transfer learning (MDTL) framework is constructed to address the dual challenges of few-shot adaptation and cross-domain alignment. To enhance MAML’s small-sample adaptability, an outer-loop cosine annealing weight allocation strategy is designed to dynamically balance training priorities between task adaptation and domain alignment. For DANN, a feature spectrum penalty (FSP) regularization is introduced to constrain singular values of source/target domain features, preserving domain-invariant degradation information without compromising prediction performance. Combined with the maximum mean discrepancy (MMD) loss function, the model further reduces cross-domain distribution differences. Validated on IEEE PHM 2012 and ABLT-1A datasets, the proposed MDTL method reduces average RMSE by at least 31.45% compared to baselines (e.g., MAML, MAML-MMD). The results demonstrate its superiority in small-sample and varying-operating-condition bearing RUL prediction, providing a practical solution for industrial health management.