Confronted with the reality of engineering work in complex system involving multi-disciplines, rigid concealment and uncertain factors, a risk evaluation method based on Bayesian network is put forward. By analyzing the process of engineering work, all possible factors are determined for the establishment of Bayesian network model. Aiming at the problem of lacking data for multi-state root nodes, the probabilities are obtained by applying triangular fuzzy function. Aiming at the difficulty in obtaining conditional probability values for multi-state nodes, DS evidence theory is applied to fuse experts' information. Arrangement for missiles in combat duty is taken as an example to perform risk evaluation. The result shows that the risk probability of engineering work can be determined easily through reasoning forward. Through reasoning backward key factors can be pointed out when the risk happens and preventive maintenance policies can be made accordingly.