Aiming at the problem that it is difficult to obtain bearing fault data with labels in actual industrial operation, which leads to the poor effect of supervised learning fault diagnosis, a bearing fault diagnosis method combining unsupervised transfer learning(TL) and kernel norm maximization is proposed. This method extracts fault features through structure optimized deep convolutional neural networks(SOCNN), and uses the maximum mean difference(MMD) to improve the distribution similarity between source domain and target domain, The fast batch nuclear norm maximization(FBNM) is combined to improve the distin-guishability and diversity of the target domain batch output matrix. The experimental results show that the proposed method has high diagnostic accuracy in different noise environments, can accurately identify the fault types and hazard levels of bearings, and provides effective technical support for bearing fault diagnosis.