Abstract: Fault diagnosis is an important way to improve the reliability of complex systems. Machine learning algorithm is an effective means to improve the efficiency of fault diagnosis and Bayesian networks is widely used in the fault diagnosis due to its advantages in uncertainty reasoning. Being unable to select the fault paths effectively, the existing fault diagnosis algorithm based on Bayesian network cannot detect faulty nodes accurately and has high computational complexity. In this study Bayesian networks sensitivity analysis is introduced into fault diagnosis and a kind of Bayesian network fault diagnosis algorithm, SA_FD, is presented in complicated system. First, the formal model of Bayesian fault diagnosis networks is given. Second, the model of how parent nodes influence their child nodes is built based on sensitivity analysis. Last, sensitivity analysis of the nodes are used to detect the faulty nodes based on heuristic path search method, to overcome the blindness of existing algorithm in searching important parent nodes and selecting the fault paths so as to improve performance of fault diagnosis effectively. Experimental results show that SA_FD is more efficient is than DFS and DFC obviously, although its complexity increases with the scale of the network.