Hard task-based dual-aligned meta-transfer learning for cross-domain few-shot fault diagnosis

Mainstream transfer learning techniques are highly effective in addressing the issue of limited target domain samples in fault diagnosis. However, when there are insufficient samples in the source domain, the transfer results are often poor. Meta-learning is a method that involves training models by constructing meta-tasks and generalizing them to new unseen tasks, offering a solution to the challenge of limited training samples. To address the few-shot problem of poor transfer effect caused by limited source domain samples under variable working conditions, this paper proposes a hard task-based dual-aligned meta-transfer learning (HT-DAMTL) method. Firstly, a dual-aligned meta-transfer framework is proposed, which embeds the designed cross-domain knowledge transfer structure (CDKTS) into the outer loop of meta-learning to achieve external transfer of meta-knowledge. The CDKTS method combines the use of multi-kernel maximum mean discrepancy (MK-MMD) with a domain discriminator to extract features that are invariant across different domains. Secondly, a meta-training method called information entropy-based reorganization hard task (RHT) is introduced to enhance the meta-model's feature learning on hard samples, leading to improved fault diagnosis accuracy. Finally, HT-DAMTL's performance is validated on public and private bearing datasets, showing its superiority over other methods.