A cross-domain intelligent fault diagnosis method based on deep subdomain adaptation for few-shot fault diagnosis

摘要

Most existing cross-domain intelligent fault diagnosis algorithms rely on many samples and only consider the global alignment of all faults. It is not practical to obtain numerous fault samples under actual working conditions. Therefore, it is a great challenge to fully utilize the fine-grained information of each type of fault to improve the fault diagnosis performance of the model under few-shot conditions. To ease this challenge, we develop a novel deep subdomain adaptation intelligent fault diagnosis (DSAIFD) model. First, a convolutional neural network is pretrained using the source domain to extract cross-domain invariant features. Second, we propose an effective class center alignment method to facilitate deep subdomain adaptation, which improves the fault diagnosis performance of the model. By combining the class center alignment method with conditional adversarial networks, the proposed model can fully utilize the fine-grained information of subclasses and reduce the distribution discrepancy between two domains. Last, high-confidence samples are selected from the target domain by setting an adaptive threshold. These selected samples are combined with the training samples from the source domain to train a classifier, which alleviates the problem of insufficient sample size. The experiments show that DSAIFD achieves significant results on two validation datasets.