A Hybrid Adversarial Domain Adaptation Network for Bearing Fault Diagnosis Under Varying Working Conditions

The complex working conditions of mechanical equipment cause deviations in the distribution of training and test data, and thus the model trained in one working condition cannot predict the data in other domains. Previous approaches to this problem rely on global domain adaptation (DA), which could confuse the categories of test data. Therefore, subdomain adaptation methods that can align the distribution of source and target domains on each class are gaining interest. However, when dealing with complex and numerous fault-type diagnosis tasks, existing subdomain adaptation algorithms can only align part of the subdomains, resulting in degradation of generalization performance. To solve this problem, a hybrid adversarial DA network (HADAN) for cross-domain fault diagnosis is proposed. In this network, local maximum mean discrepancy (LMMD) is implemented as a subdomain adaptation to obtain fine-grained information. In addition, an adversarial learning method called kernel sensitivity alignment (KSA) is designed to overcome the shortcomings of subdomain adaptation. Based on the kernel matrix, kernel sensitivity is defined to represent the relationship between one sample and other samples in the same domain. Compared to the traditional adversarial DA methods, KSA is spatially location-sensitive and can significantly reduce the domain shift by aligning the kernel sensitivity distribution of the two domains. Two experimental scenario cases are designed, and a total of 20 bearing fault diagnosis transfer experiments are conducted on Case Western Reserve University (CWRU), Paderborn, and XJTU datasets to demonstrate the effectiveness and advantages of the proposed method.