A universal fault classification for gas turbine diagnosis under variable operating conditions

Normally, industrial gas turbines operate without shutdowns for an extended period of use. During each operational cycle control and ambient conditions can vary considerably. Because gas turbine monitoring system has to be uninterrupted at any operational conditions it needs a method appropriate for this purpose. This paper introduces a universal gas turbine fault classification suitable for diagnosing at variable operating conditions. The concept of such a classification is thoroughly examined for a stationary power plant operating at steady states and transients. The gas path fault classes are simulated by using non-linear static and dynamic power plant models. Each class is represented by a sample of measured values (patterns) that include measurement errors. These samples arc fed to a neural network used later on to make a diagnosis. The trained neural network is then subjected to a statistical test that permits us to calculate the probabilities of a correct diagnosis. Based on these probabilities, the suggested classification is compared to a conventional approach formed under a fixed operating condition. This comparison is drawn under a variety of diagnostic conditions. The results go to show that the decrease in diagnosis reliability when switching to the universal classification is relatively low. On the other hand, it offers continuous gas turbine monitoring and substantially streamlines the diagnostic algorithms employed.