Abstract: Nuclear power plants (NPP) are very complex system, which need to collect and monitor vast parameters. It’s hard to diagnose the faults. A parameter reduction method based on neighborhood rough sets was proposed according to the problem. Granular computing was realized in a real space, so numerical parameters could be directly processed. On this basis, the decision tree was applied to learn from training samples which were the typical faults of nuclear power plant, i.e., loss of coolant accident, feed water pipe rupture, steam generator tube rupture, main steam pipe rupture, and diagnose by using the acquired knowledge. Then the diagnostic results were compared with the results of support vector machine. The simulation results show that this method can rapidly and accurately diagnose the above mentioned faults of the NPP.