It still remains a significant challenge for intelligent fault diagnosis of rolling bearings with unlabeled samples, and some transfer diagnosis methods have been researched. However, most of traditional transfer diagnosis methods suffer from significant differences between target domain data and source domain data, as well as missing labels in target domain data. This will seriously affect the classification accuracy of transfer diagnosis methods. To solve these problems, a domain adaptive-based attention separation transfer neural network (DA-ASTNN) is proposed for intelligent fault diagnosis of rolling bearing with unlabeled data under variable operating conditions. The pro-posed method mainly includes two parts, namely domain adaptive auto-encoder clustering module (DAACM) and attention separation transfer neural network (ASTNN). Firstly, in DAACM, a hybrid optimized stacked autoencoder is adopted to extract features from target domain data. After dimension reduced, these features will be clustered. Secondly, a label alignment strategy is designed by matching the arrangements of clustering labels with the predicted labels, which can generate enhanced pseudo labels for target domain data. Finally, ASTNN is established by combining group convolution and attention mechanism. The source domain model is used as a pre trained model to achieve effective generalization of target domain data. The effectiveness of the proposed method is verified by the rolling bearing fault simulation experiments. The results show that the proposed DA-ASTNN method has higher diagnostic accuracy compared with the comparison methods.