Learning Bayesian networks for systems diagnosis

This paper proposes the construction of a Bayesian network for failure diagnosis in industrial systems. We built this network considering the plant mathematical model and it includes parameters and structure learning through the Beta Dirichlet distributions. We experience the previous methodology by means of a case study, where we simulate some failures that can occurs in the valves used to interconnect a deposits system. With those failures information, we train the network and this way we learn the structure and parameters of the Bayesian network. Once obtained the network, we design the diagnosis probabilistic inference through the poly-trees algorithm. It will give us the valves failure probabilities according to the evidences that show up in our entrance sensors. In this work, we try the existent uncertainty in the diagnosis variables through the probabilistic and fuzzy approach. Since the information provided by our sensors (diagnosis variables) is, represented in a fuzzy logic form, for then to be converted to probability intervals, generalizing the Dempster-Shafer theory to fuzzy sets. After that, we spread this information in interval form throughout the diagnosis Bayesian network to get our diagnosis results. The probability interval is more advisable in the taking decisions that a singular value.