Bayesian network has been widely used as a powerful reasoning and knowledge expression tool to analyze toot causes. With the increasing scale and integration of modern process industries, building an accurate Bayesian network is more and more difficult but important. To solve this problem, a novel multiblock transfer entropy based Bayesian network model is proposed to make root cause analyses. Developing the proposed methodology consists of three simple steps: modular decomposition of processes based on process knowledge, Bayesian network structure learning using the proposed multiblock transfer entropy, and root cause analysis. The accuracy of structure learning is influenced by the transfer entropy between the alarm variable and itself at the previous sampling time. To solve this problem, an improved transfer entropy method is presented, where the influence of the variable itself between sampling time is eliminated for correct Bayesian network structure scoring. Moreover, a regular penalty item is adopted to avoid overfitting during training the Bayesian network structure. To validate the performance of the proposed methodology, simulations on the Tennessee Eastman Process are carried out. Simulation results confirm the effectiveness of the proposed methodology.