In IP network, tomography method can perform fault diagnosis by analyzing the end-to-end properties with low costs. However, most existing tomography based techniques have the following problems: 1) the end-to-end detected number is not sufficient to determine the state of each link; 2) as the scale of the network goes up, the diagnosis time may become unacceptable. To address these problems, a new congested link diagnosis algorithm based on Bayesian model was proposed in this paper. This method firstly modeled the problem as a Bayesian network, and then simplified the network by two steps and limited the number of multiple congested links. Therefore, the proposed method could greatly reduce the computational complexity and guarantee the diagnostic accuracy. Compared with the existing diagnosis algorithm which is called Clink, the proposed algorithm has higher diagnostic accuracy and shorter diagnosis time.