Based on time–frequency analysis and residual neural network technology, this paper studies the composite fault diagnosis method of rotating equipment in nuclear powerplant. Firstly, three commonly used time–frequency analysis methods are used to process the experimental simulation signal, and the time–frequency concentration, reconstruction error and other indicators are analyzed. Then, the time–frequency analysis results are combined with the residual network to construct a diagnostic model, and the accuracy, loss value, training and testing time of the model are comprehensively evaluated to select the optimal combination method. Finally, these lection of network structure and training hyperparameters such as convolution kernel size, residual network layer number, training batch size and learning rate is explored to determine the model structure and parameters. Experiments show that the accuracy of the model can reach 99.954%, and the average test time per round is 0.15689 s. In general, the designed composite fault diagnosis model has high accuracy and good real-time performance, which has certain significance for improving the stability and safety of rotating equipment operation in nuclear power plants.