A Cross Domain Feature Extraction Method for Bearing Fault diagnosis based on Balanced Distribution Adaptation

Traditional intelligent fault diagnosis techniques for rotating machines have two limitations: 1) Big data with fault information is not available in some cases; 2) The training and testing data are often drawn under discrepant distribution. Thus, transfer component analysis (TCA) has been designed to reduce the distance of marginal distribution between domains. The joint distribution adaptation (JDA) was proposed to simultaneously reduced the difference between the conditional distribution and marginal distribution in source or target domains. However, these two distributions are often treated equally in these existing methods, which will lead to poor performance in practical applications. Therefore, a cross-domain feature extraction method based on balanced distribution adaptation algorithm(BDA) has been proposed, which can adaptively utilize the importance of difference between marginal distribution and conditional distribution. It should be noted that several existing cross domain feature extraction methods can be treated as special cases of BDA. As a new method in the field of transfer learning, BDA is an effective cross-domain feature extraction method. The validity of the BDA algorithm has been successfully evaluated in the actual data set in this paper.