Aiming at information propagation shared, inference time and other issues in classic junction tree inference algorithm, an effective junction tree algorithm is presented. Based on evidence information obtained and query nodes, the original structure of a Bayesian network is simplified, and then the junction tree algorithm is used for inference on the simplified structure. During the information propagation, the effective junction tree algorithm can achieve information shared between different evidence variables. The result of simulations shows that the improved algorithm can guarantee high precision at the same time, and do the diagnosis and reasoning in a short period of time. Based on data collected in the field, a fault diagnosis model is built for the cement rotary kiln, and the effective junction tree algorithm is applied to the fault diagnosis of cement rotary, achieving a faster and accurate fault diagnosis.