Fault diagnosis is a key research part in the field of network fault management. In order to make effective fault diagnosis to the increasingly complicated distributed application systems(DAS) which are based on the computer network, Building an accurate and practicable fault propagation model(FPM) is generally the necessary prerequisite of the subsequent tasks such as probabilistic reasoning, fault recovery and failure prediction. In this paper, a method of constructing the FPM which combined sample datas and the expert knowledge was put forward based on Bayesian network. Firstly, an initial tree(T) including all the service nodes on the specific DAS was generated by the maximum weight spanning tree(MWST) algorithm with sample datas. Secondly, the initial tree(T) was revised according to expert experiences. Finally, the FPM of the DAS was learned using greedy search structure-learning algorithm with the revised structure(T') as its initial input model. In the end, the learned FPM using the proposed method was evaluated by calculating its BIC-score and comparing to the actual one. And the results show that the proposed method can give an accurate FPM of the distributed application system.