Aiming at the difficulty of rolling bearing fault diagnosis under variable working conditions and noise background， a rolling bearing fault diagnosis method is proposed， which combines the attention mechanism， DropConnect and Dropout into the convolutional neural network-bidirectional long short-term memory network（CNN-BiLSTM） model. Firstly， the original vibration signals of various fault states of rolling bearings under different working conditions are preprocessed， and the labeled training data set and test data set are constructed. Secondly， the attention mechanism is introduced into BiLSTM to extract more important deep fault features. At the same time， DropConnect and Dropout are mixed to suppress the over fitting problem of CNN network and BiLSTM network in the training process， respectively， so as to obtain an improved CNN-BILstM model. Thirdly， the processed labeled data set is input into the improved model for training. Finally， Softmax classifier is used for fault diagnosis. By selecting different data sets， the method is proved to have good generalization and anti-noise capability.