A transfer learning method: Universal domain adaptation with noisy samples for bearing fault diagnosis

Abstract

Under the influence of frequent start-stop driving and rail launching during the service of urban rail vehicles, the source domain samples contain a large number of noise labels and noise samples. Moreover, the feature distribution and sample categories of the target domain and source domain are different because the urban rail vehicles are affected by the fluctuation of passenger flow and long-term service. This paper summarizes this real task in rail transportation as universal domain adaptation with noisy samples (UDANS). A novel multibranch convolutional neural network is proposed to solve the above problem. By optimizing the divergence of the two classifier outputs, the following objectives can be achieved: detecting noisy source samples, finding private classes in the target domain, and aligning the distribution of the source domain and the target domain. Finally, the results of the wheelset bearing dataset show that the method has advantages in rail transportation fault diagnosis.