In Flexible Manufacturing System (FMS) fault diagnosis, there are some problems hard to tackle, such as fuzziness, polymorphism, etc. So this paper proposes an improved Bayesian network (BN) approach. By first introducing BN and describing the transformation process from FT to BN. In addition, this paper applies fuzzy theory to set up conditional probability table of BN, and proposes observing nodes used to describe symptom information. Finally, by analyzing the fault of the numerical control processing unit of FMS, results indicate that this approach can improve the efficiency and accuracy of reasoning for fault diagnosis.