Abstract: In order to improve the fault diagnosis accuracy of nuclear power plant, neural network and expert systems were combined to give full play to their advantages. In this paper, the concept lattice was applied to get the object properties, extracting the core attributes, dispensable attributes and relative necessary attributes from a large number raw data of fault symptoms. Based on these attributes, neural networks with different levels of importance were designed to improve the learning speed and diagnosis accuracy, and the diagnosis results of the neural networks were verified by using rule-based reasoning expert system. To verify the accuracy of this method, some simulation experiments about the typical faults of nuclear power plant were conducted. And the simulation results show that it is feasible to diagnose nuclear power plant faults with the confederation diagnosis methods combined the neural networks based on the concept lattice theory and expert system, with the distinctive features such as the efficiency of neural network learning, less calculation and reliability of diagnosis results and so on.