Aiming at the problems that the selection of the time-frequency domain characteristics has strong subjectivity and the feature information about the time series is underused,a fault diagnosis method was proposed based on an improved convolutional neural network (CNN) and bi-directional long short-term memory neural network (BiLSTM) for an early rolling bearing.The convolutional neural network was used to extract the features of the original vibration signal,and a batch regularization layer was introduced after the convolutional layer to eliminate the influence of data irregularity on weight optimization.Meanwhile,the feature extraction efficiency was improved by expanding the first convolutional layer and adjusting the step size of the CNN model.The bi-directional long and short-term memory neural network was introduced to remedy the insufficiency of the convolutional neural network in term of extracting the feature information of the time series.A batch regularization layer and the dropout layer were used to enhance the robustness and reduce the interaction and the dependencies between neurons of the proposed model.Finally,the proposed model was verified by the test data of rolling bearing.The results show that compared with the traditional methods,the proposed method not only has faster training speed,but also greatly improves the fault diagnosis accuracy.