Deep Transfer Learning Network for Fault Diagnosis Under Variable Working Conditions

Abstract

Rotating machinery fault diagnosis method base on intelligent approach has made great progress recent years. All of these success were under the assumptions: 1) the training set and testing set are drawn from the same working condition. 2) training samples with laebls are sufficient. However, in actual scenarios, variable working conditions will cause differences in data distribution. At the same time, obtaining abundant labeled data is difficult in practical scenarios. The ability of unsupervised learning determines the performance of intelligent fault diagnosis under variable working conditions. Therefore, a deep convolutional transfer learning network is proposed to address these issues in this paper. The proposed method is composed of feature extraction and domain adaptation two parts. In feature extraction part, a convolutional neural network (1D CNN) is leveraged to capture features from input signals. The domain adaptation followed the 1D CNN to get the domain-invariant features by maximizing the distance between health conditions in source domain, maximizing the domain classification error between domains and minimizing the distribution discrepancy distance between source and target domain. Experimental results conducted on different working conditions of a standard bearing dataset demonstrated the effectiveness of the proposed method.