In view of the existing fault diagnosis methods mostly focus on single fault and lack of corresponding diagnosis methods for composite fault under actual working conditions, a multi working condition composite fault diagnosis model of ConvNeXt rolling bearing based on supervised learning (TConvNeXt) was proposed. The rolling bearing data set was reconstructed into a balanced data set by synthetic minority oversampling technology, so as to improve the utilization of composite fault samples; the TConvNeXt network model was made to master the required partial weight to distinguish the composite fault information of rolling bearing by using transfer learning. The one-dimensional signal was converted into RGB image input model through Gramian angle field, and the residual weight of the model was trained; finally, the trained ConvNeXt network model was used for rolling bearing fault diagnosis, and the Grad-CAM method was used for visualization to analyze the causes of network diagnosis errors and adjust the network. The model with the highest training accuracy was applied to the fault diagnosis of rolling bearing to test its diagnosis ability under actual working conditions. The experimental results show that TConvNeXt network model has high diagnosis accuracy. It is not only outstanding in aliasing fault diagnosis, but also has advantages in single fault diagnosis. It can well adapt to the requirements of rolling bearing fault diagnosis of different fault types under multiple working conditions.