In order to systematically explore the causes of bus accidents under the joint action of vehicles, roads and environment, the data of 259 bus accidents in A city were used to establish a Bayesian network structure based on experts knowledge and data fusion method. The parameter learning is carried out by using the Bayesian method which obeys the Dirichlet distribution. After verifying the effectiveness of the model, combined with the Bayesian network model, the cluster tree propagation algorithm was used to infer the relationship between the variables. The research results show that weather, time, road alignment, and location may all cause bus accidents; according to the probability from high to low, the types of accidents caused are non-collision accidents, collisions with cars, non-motor vehicles, and buses; the casualties and the strongest causal factors of each type of accident are different. In addition, in the different circumstances of each cause, the probability of the type of accident and the probability of casualties are also different. The relevant results can provide certain basis for government and enterprises to establish an accident management system and reduce the occurrence of bus accidents.