Abstract: Nuclear power plant has complex structure, multiple operating parameters and high coupling degree. In abnormal operating conditions, there are extremely complex nonlinear relations among operating parameters. It is difficult to diagnose faults manually and take a lot of time, so it is easy to miss the best disposal opportunity. Therefore, in order to assist the operator’s judgment and take correct disposal measures in time, an intelligent technology that can efficiently identify abnormal operating conditions is urgently needed. Artificial neural network is one of the most popular intelligent technologies. It can simulate the thinking of human brain according to the structure and principle of human neural network. It can realize the information processing ability of human neural network in the form of mathematical model, and achieve specific goals and solve corresponding practical problems by designing appropriate algorithms. Probabilistic neural network (PNN) has good nonlinear mapping function, does not need the learning process, also does not need to set the initial weight, has strong adaptability, fault tolerance and robustness, and popularity in the field of pattern classification, so the PNN can be used in nuclear power plant parameters, strong coupling, more cases of abnormal operation condition identification. In this paper, six abnormal operating conditions of nuclear power plant were selected, and the simulation calculation was carried out based on the accident analysis platform of nuclear power plant, and the characteristic parameters that can characterize the operation state of nuclear power plant were extracted according to the response characteristics of abnormal operating conditions. Two methods, PNN and BP neural network, were used for comparative study. Abnormal operation status identification models were established in MATLAB environment and verified by experiments. Theoretical analysis was conducted from the level of network structure and operation mechanism. As the abnormal operating conditions of nuclear power system are not only related to the type and degree of equipment failure, but also to the initial operating state before the failure, the same abnormal operating conditions under different initial states were selected in this paper for the extended study of PNN. The results show that compared with the traditional BP neural network model, PNN model can accurately identify the randomly ordered test samples, and the use time is very short. Therefore, the identification method of abnormal operating conditions based on PNN is effective in the field of nuclear power plant anomaly identification. PNN has excellent performance both in the comparison of BP neural network and in the extended study of different initial working conditions. PNN has higher recognition accuracy and faster operation speed. PNN has the advantages of simple training, rapid convergence and good expansion performance. With the gradual accumulation of known anomaly sample data, the identification network can be continuously expanded and upgraded, thus continuously improving the accuracy of anomaly identification, and it has a good application prospect.