The traditional deep learning bearing fault diagnosis method has a complex network, many training parameters, and weak model generalization. In response to the above problems, under the background of industrial big data, a bearing fault diagnosis method combining the improved Incetpion V2 module and the CBAM attention mechanism is proposed. The improved Inception V2 module further broadens the branch network structure by adding an average pooling layer, thereby Improve network expression ability. First, the bearing vibration signal is converted into a time-frequency image through wavelet transform, which is used as the input of the convolutional neural network. The input features are adaptively extracted through the improved Inception V2 module, and the extracted features are organized across channels; then through CBAM The attention mechanism generates dual attention weights of channel and space, enhances the features with high correlation and suppresses the features with low correlation; finally, the generated feature data is input to the global average pooling layer and the fault diagnosis result is output. Experimental results show that this method can establish a "shallow" convolutional neural network model, reduce model parameters, speed up model convergence, and achieve an accuracy of 99.75%; at the same time, under different loads and high noise conditions, the model has good generalization. It is more suitable for application in industrial big data.