Abstract: In order to solve the problem that the data distribution of the source domain and the target domain of rolling bearing is different in the variable working condition environment and the samples of the target domain do not contain labels, a fault diagnosis model of the rolling bearing based on the deep adaptive transfer learning network (DATLN) is proposed. Firstly, a domain-shared characteristic extraction network is built, and multiscale convolutional neural network (MSCNN) is used to suppress noise interference, so as to effectively extract local fault information contained in vibration signals. Secondly, combined with a bi-directional long short-term memory network (BiLSTM), the temporal characteristics in the local fault information are further learned. Finally, transfer learning is introduced to build a domain adaptive module with domain adversarial (DA) training combined with adaptive joint distribution (AJD) metrics. By maximizing the domain classification loss and minimizing the AJD distance, the source and target domain characteristic samples are aligned. The anti-noise experiment and transfer experiment are carried out on the open source CWRU data set and the measured data set of the mechanical fault platform respectively. The anti-noise experiments show the following points. ① The identification accuracy of MSCNN-BiLSTM network is above 99% in the noise-free environment, which shows that MSCNN-BiLSTM network has a good characteristic extraction capability. ② The identification accuracy of MSCNN-BiLSTM, LeNet-5, MSCNN and BiLSTM decreases with the increase of noise intensity. ③ Under the noise environment of 3, 5 and 10 dB, the average identification accuracy of MSCNN-BiLSTM network is higher than that of LeNet-5, MSCNN and BiLSTM networks, indicating that MSCNN-BiLSTM network has better anti-noise interference performance. ④ The MSCNN-BiLSTM network converges first with less fluctuation in both the noise-free environment and the 3 dB noise environment. The transfer experiments show the following points. ① The average identification accuracy of DA+AJD method is 97.36% on unlabeled target domain dataset, which is higher than that of Baseline, transfer component analysis(TCA) and domain adversarial neural network (DANN). ② On the test set confusion matrix, only one sample of the DA+AJD method is incorrectly identified, indicating that the DA+AJD method based on domain adaptation has better fault transfer diagnosis performance. ③ The t-SNE algorithm is used to visualize the processed source and target domain characteristic samples. The DA+AJD method only has a small number of rolling element fault and outer ring fault characteristic samples in the target domain that are incorrectly aligned to the inner ring fault characteristic samples area in the source domain. This result indicates that the DA+AJD method can effectively reduce the edge distribution and conditional distribution differences between the source domain and the target domain, and thus achieves better characteristic sample alignment.