Deep convolution adaptive network-based bearing fault diagnosis under variable conditions

Abstract

Recently, significant research efforts have been made to develop bearing fault diagnosis methods based on deep learning. However, these diagnosis methods assume that training and testing data are drawn from the same input feature space with the same distributions, which may not be effectively applied for a complex process that has multiple operating modes due to the differences between the data distributions of the new and preexisting mode. To solve this issue, this paper proposes a novel deep convolution adaptive network for bearing fault diagnosis. In the method, the deep convolution adaptive network is constructed to extract transferable features from two different domains. And, Multi-kernel maximum mean discrepancy has been used to optimize the transfer performance and alleviate the parameter sensitivity problem. Three transfer experiments on four working conditions of bearing have verified the superior performance of DCAN. A significant improvement has been made with more than 30% classification accuracy improvement as compared with the bearing fault diagnosis method without transfer learning trick.