A New Bearing Fault Diagnosis Method Based on Deep Transfer Network and Supervised Joint Matching

Abstract

In practical industrial environment, variable working condition can result in shifts in data distributions, and the labeled fault data in various working conditions is difficult to collect because rotating machines often works in normal status, and the insufficient labeled fault data brings data samples imbalance and performance degradation of intelligent fault diagnosis model. To overcome these problems, by integrating the superiority of deep learning method and feature-based transfer learning method, this work proposes an innovative cross-domain fault diagnosis framework based on deep transfer convolutional neural network and supervised joint matching. First, the continue wavelet transform is used to process original bearing vibration signals and extract time-frequency images. Second, a deep transfer convolutional neural network is built by the way of fine-tuning, and the trained network is used to extract deep features from different domains. Third, a new domain adaptation approach, supervised joint matching, is developed to conduct joint feature distribution matching and instance reweighting with the consideration of maximum marginal criterion. The intelligent bearing fault diagnosis model is then trained to predict the labels of the target domain's feature data. To verify the performance of the proposed approaches, this study uses two distinct datasets pertaining to bearing defects for conducting cross-domain fault diagnosis in the presence of balanced and imbalanced data. The experimental analysis indicates that the designed methods can achieve desirable diagnostic accuracy and possess robust generalization ability.

Keywords

• Deep learning
• domain adaptation
• fault diagnosis
• time-frequency images