Abstract: Aiming at the problems of modal aliasing, endpoint effect and confusion of inner and outer ring fault features in traditional empirical mode decomposition in rolling bearing fault diagnosis, an intelligent diagnosis framework based on time-varying filtering empirical mode decomposition-convolutional neural network-convolutional attention module-bidirectional long short-term memory network （TVFEMD-CNN-CBAM-BiLSTM） is proposed. Firstly, the optimal intrinsic mode function generated by TVFEMD decomposition is selected and nine kinds of indexes with clear physical meaning of time domain statistics are calculated to construct a high discriminant feature set as input. Then, the CNN layer is used to extract the spatial local pattern. The channel-space dual attention mechanism of the CBAM module enhances the fault sensitive features and suppresses the noise. The BiLSTM layer establishes the long-term temporal dependence of the fault features. Finally, the Softmax function is used to generate the probability distribution of six types of fault states （health, single fault, composite fault） to complete the fault diagnosis. Based on the bearing data of Paderborn University in Germany, the model is verified. The results show that the fault diagnosis accuracy of TVFEMD-CNN-CBAM-BiLSTM model can reach more than 94 % under single working condition and variable working condition. Compared with the traditional CNN, BiLSTM and CNN-BiLSTM models, the accuracy is higher, the convergence speed is faster, and the fitting situation is better.