Aiming at the complex distribution difference caused by two domains under different working conditions, an adaptation regularization based transfer learning method for rolling bearing fault diagnosis under different working conditions is proposed. Firstly, the training base classifier predicts the pseudo label for the target domain, and the joint distribution is used to align two domain distributions to reduce the distribution difference. Secondly, the target domain data are further utilized through the manifold regularization to mine the potential distribution geometry of the data and learn the target domain data distribution information. Finally, the classifier is established under the framework of structural risk minimization combined with the above two-step learning principle. The optimal coefficient matrix is obtained by iteratively updating pseudo labels to complete the fault diagnosis of rolling bearing under different working conditions. The experimental validation is implemented on two rolling bearing datasets. Results show that the identification accuracy values of the proposed method are 96. 38% and 94. 18% , respectively. It shows that the method can effectively deal with the complex distribution differences caused by multiple working conditions, and has good effectiveness and feasibility.