Siamese Distinguishing Features Attentional Enhancement Transfer Fault Diagnosis Method for Variable Rotational Speed

Transfer learning is widely used in artificial intelligence fault diagnosis field because it can solve the problem of label missing in rotating parts at varying speeds. However, the domain adaptive method in transfer learning is not suitable for real transfer fault diagnosis scenarios, and the adaptive enhancement of fault characteristics is not realized in the transfer process. To solve these thorny problems, a novel method called Siamese Distinguishing features Attentional Enhancement Transfer fault diagnosis (SDAET) has been proposed. The body of SDAET model adopts the dual-branch convolutional neural network architecture with shared weights. It mainly uses the contrast loss function in the siamese feature contrast network to extract the domain invariant features at two rotational speeds, and then applies it to the transfer diagnosis at other rotational speeds, so as to meet the more real transfer diagnosis scenarios. The distinguishing features attentional enhancement network is designed to adaptively enhance the differentiated domain invariant features at different rotational speeds. Furthermore, various feature visualization techniques are used to further explain the features learned from the black box neural network. The diagnostic results on two kinds of test datasets show that the proposed model has higher diagnostic accuracy. The technique of siamese distinguishing features attentional enhancement provides a new and better way to solve the transfer diagnosis problem.