In order to solve the complexity of NPC inverter modeling process and the uncertainty of fault diagnosis,bond graph theory was used to model NPC inverter,by which the modeling process could be greatly simplified and the complexity of the modeling process could be reduced.Considering the similarity of causality between bond graph theory and Bayesian network,Bayesian network was introduced to sort the fault components dynamically.The results show that by using the method,not only the fault components of the system can be located accurately,but also the influence degree of the fault components on the system can be obtained.The method is feasible and effective.