Abstract: The nuclear power plant is a complex technical facility with potential radioactive hazards. Therefore, timely identification and treatment of accidents are necessary to avoid serious consequences. In order to quickly obtain information on nuclear power plant accidents and provide a new auxiliary means for the analysis of operator safety, it is necessary to study the accident diagnosis method and establish an accurate and reliable accident diagnosis model suitable for nuclear power plants. However, it is tough to establish a complete and accurate model of the nuclear power plant based on the model-based diagnosis method. The method based on expert knowledge needs a lot of expert experience knowledge which is difficult to obtain. Based on continuous data accumulated, the data-driven method can adaptively map the nonlinear relationship between symptoms and accidents, which is more suitable for fault diagnosis of the nuclear power plants. The data obtained by online monitoring for nuclear power plants are time series data with high data dimensions. However, there are few accident data, which makes it challenging to meet the training requirements of standard deep neural networks such as convolution neural network and time convolution network. Therefore, according to the data characteristics and high accuracy requirements of the nuclear power plant, a nuclear power plant accident diagnosis method based on time convolution capsule network was proposed in this paper. Firstly, the high-dimensional data obtained from the real-time monitoring of the nuclear power plant were input into the time convolution kernel to extract the depth of time-series features information of the data and reduce the difficulty of calculating. Then, the depth vector features of the data were mined through the capsule network to maximize the use of the data features of the operational information, which could also get a good convergence effect in the case of fewer accident data. Finally, the accident simulation data of the nuclear power plant were obtained by Fuqing simulator to train and test the model. The simulation data of regular operations and nine other accidents were collected to test the accident diagnosis model built in this paper. When the training set only accounts for 20% of the data set, it can still quickly and stably converge and accurately diagnose the accident information, and the diagnosis accuracy is more than 99%. Finally, the method in this paper was compared with other conventional depth learning methods through experiments. The results show that the time convolution capsule network model has better convergence and diagnostic accuracy than traditional machine learning and in-depth learning methods when the accident data of the nuclear power simulation system are not enough. The research content of this paper can provide some theoretical basis and technical reference for the application of the fault diagnosis method in nuclear power plants.