Alignment subdomain-based deep convolutional transfer learning for machinery fault diagnosis under different working conditions

Abstract

In recent years, transfer learning (TL) methods have been extensively used in machinery fault diagnosis under different working conditions. However, most of these TL methods perform poorly in the actual industrial applications, due to the fact that they mainly focus on the global distribution of different domains without considering the distribution of subdomains belonging to the same category in different domains. Therefore, we propose an alignment subdomain-based deep convolutional transfer learning (AS-DCTL) network for machinery fault diagnosis. First, continuous wavelet transform is used to transform the original vibration signal into a 2D time-frequency image. Then, AS-DCTL uses a convolutional neural network as the feature extractor to extract the features of the source and target domain samples and introduces maximum mean difference (MMD) to align the global distribution of the extracted features. Simultaneously, we use local MMD as a metric criterion to align the distribution of related subdomains, by adding weights to similar samples in the source domain and target domain. The experimental results of the two case studies show that the proposed AS-DCTL network can achieve higher recognition accuracy and classification effect, in comparison with the current mainstream TL methods.