On line diagnosis of gas turbines using probabilistic and qualitative reasoning

Artificial intelligence (AI) techniques are becoming an active area of research for real applications. Power industry is one of the best examples of this. Different problems have been solved with these techniques, for example monitoring, alarm management, diagnosis, and network planning. This paper presents an on-line diagnosis system for gas turbines in power plants. Since this application deals with unexpected behavior, probabilistic reasoning and specifically Bayesian networks, offer a natural mechanism for diagnosis. However, the use of Bayesian networks in real applications presents two challenges. First, the acquisition of representative models of the process with and without faults. Second, dealing with continuous variables makes very expensive the computation for inferences. This project utilizes automatic learning algorithms, together with expert advice to determine the models of the most common faults in gas turbines. Also, a quantification of the behavior is used to minimize the cost of the probability propagation in Bayesian networks. These produces an original probabilistic and qualitative diagnosis of gas turbines. Experiments were carried out utilizing real data in a simulator. The results are presented and discussed