Aiming at the problem that the accuracy of rolling bearing fault diagnosis is low due to the large difference in the distribution of sample data between the source domain and the target domain under variable working conditions, a new transfer learning method called convolutional attention based Feature Transfer Learning (CAFTL) is proposed and used for fault diagnosisof rolling bearings under variable working conditions. In the proposed CAFTL, the source and target domain samples are input to the convolutional neural network after multi-head self-attentive computation and normalization to obtain the corresponding source and target domain features; then, the two domain features are projected into the same common feature space by the domain adaptive transfer learning network; then, the classifier constructed from the source domain labeled samples is used for classification; finally, the Stochastic Gradient Descent (SGD) method is used to train and update the parameters of CAFTL, after obtaining the optimal parameter set of CAFTL, the optimized CAFTL is used for fault diagnosis of rolling bearing samples to be tested. A fault diagnosis example for rolling bearings verifies the effectiveness of the proposed method.