Deep discriminative transfer learning network for cross-machine fault diagnosis

Many domain adaptation methods have been presented to deal with the distribution alignment and knowledge transfer between the target domain and the source domain. However, most of them only pay attention to marginal distribution alignment while neglecting the discriminative feature learning in two domains. Thus, they still cannot satisfy the diagnosis requirement in some cases. To enhance the distribution alignment and match the marginal distributions as well as conditional distributions of two domains, an improved joint distribution adaptation (IJDA) mechanism is proposed. In IJDA, to enhance domain confusion, maximum mean discrepancy and CORrelation Alignment (CORAL) are combined as a new distribution discrepancy metric. Furthermore, an improved conditional distribution alignment mechanism is constructed. To contribute to feature learning and learn more separable features, a new I-Softmax loss that can be optimized like the original Softmax loss and possesses a stronger classification ability is proposed. With the IJDA mechanism and I-Softmax loss, the deep discriminative transfer learning network (DDTLN) is built to implement fault transfer diagnosis. Under the unlabeled target-domain samples, the experimental results on six cross-machine diagnostic tasks verify that the proposed DDTLN has a higher performance of transfer fault diagnosis than other typical domain adaptation methods.