A small-sample cross-domain bearing fault diagnosis method based on knowledge-enhanced domain adversarial learning

Abstract

Traditional domain adaptation methods often perform poorly in cross-device bearing fault diagnosis when the target domain contains incomplete labels or exhibits imbalanced data. To address this issue, we propose an Adaptive meta-domain transfer learning network (AMTLN), which integrates a self-weighted fusion (SWF) module and a knowledge-enhanced domain adversarial learning (KEDA) framework to improve accuracy and robustness. An AMK-Fast DTW algorithm aligns vibration signals across domains, and kernel density estimation minimizes distributional differences. KEDA introduces auxiliary knowledge and meta-learning to enhance transfer performance in small-sample scenarios and reduce catastrophic forgetting. SWF further strengthens the forward knowledge transfer. Experiments show that AMTLN achieves high accuracy and strong generalization across varying operational conditions, even with incompletely labeled target data.