Fault Diagnosis Using Unsupervised Transfer Learning Based on Adversarial Network

The fault diagnosis is very important for the modern industry. Due to machine working conditions changing frequently, most of current fault diagnosis models built on the training (source) domain can't perform well in test (target) domain. In addition, in test domain, there are few labeled data to adjust model to be adaptive to test working conditions. Domain adaptation, as one type of transfer learning, can be used to solve this problem. This paper proposes a novel fault diagnosis method using unsupervised transfer learning based on adversarial network. In this method, deep neural network is used to extract feature of fault signal while the adversarial network is used to accomplish the transfer learning process. Firstly, the fault signal is converted into RGB images as inputs of networks. Then, the adversarial training methods are used, which includes three training processes: the regular training process using source data, the maximum discrepancy training process and the minimum discrepancy training process. These three steps are adversarial to each other to adjust the model to be more adaptive. The method is tested on motor bearing dataset provided by Case Western Reserve University (CWRU). The prediction accuracies are better than other four comparison methods.