Regularized Optimal Transport Enabled Few-Shot Cross-Domain Fault Diagnosis

In industrial scenarios, collecting adequate labeled fault samples for diverse conditions is challenging. The few-shot cross-domain domain diagnosis is a more challenging and underexplored scenario, where the model needs to generalize effectively to the target domain with only limited training samples from source domain. This paper proposed a regularized optimal transport-enabled few-shot cross-domain (ROTFC) fault diagnosis method via designing spectral regularized augmentation optimal transport for this problem. Specifically, the Mixup strategy is introduced into optimal transport to facilitate few-shot adaptation by extending source distribution, aligning source and target domain distribution and minimizing the vicinal risk, while learning domain-invariant classifier and cross-domain alignment. Besides, spectral regularization is used to balance the transferability and discriminability of the proposed ROTFC method by penalizing the eigenvectors. Experiments on three-phase motor fault diagnosis show the effectiveness of the proposed ROTFC method in dealing with few-shot cross-domain diagnosis, the diagnosis accuracy can be improved by at least 3 % in one-shot diagnosis.