The work aims to propose a new fault diagnosis method based on time-frequency image training and fault feature learning, in order to solve the problem that the time-frequency image of bearing fault feature is difficult to recognize. MDCNet network was proposed, which was composed of Multi-Size Convolution Kernel Module, Dual-Channel Pooling Layer and Cross Stage Partial Network. Firstly, the acquired vibration signal was compressed and transformed synchronously to obtain the instantaneous frequency image of the signal. Finally, the fault diagnosis result was obtained by inputting the neural network. The prediction accuracy of the proposed method was 99.9% after applied to the bearing data set of Case Western Reserve University. Compared with AlexNet, VGG -- 16, Resnet and other traditional methods, MDCNet method realized a classification accuracy of 99.9%, which was higher than the classification accuracy of 95.70%, 98.51% and 97.64% of traditional methods. The results show that the prediction accuracy of the proposed method is higher than that of other methods, which verifies the feasibility of the proposed method in fault diagnosis of packaging machinery.