| description abstract | Gas turbine trip is an operational event that arises when undesirable operating conditions are approached or exceeded, and predicting its onset is a largely unexplored area. The application of novel artificial intelligence methods to this problem is interesting both from the computer science and the engineering point of view, and the results may be relevant in both the academia and the industry. In this paper, we consider data gathered from a fleet of Siemens industrial gas turbines in operation that includes several thermodynamic variables observed during a long period of operation. To assess the possibility of predicting trip events, we first apply a new, systematic statistical analysis to identify the most important variables, then we use a novel machine learning technique known as temporal decision tree, which differs from canonical decision tree because it allows a native treatment of the temporal component, and has an elegant logical interpretation that eases the posthoc validation of the results. Finally, we use the learned models to extract statistical rules. As a result, we are able to select the five most informative variables, build a predictive model with an average accuracy of 73%, and extract several rules. To our knowledge, this is the first attempt to use such an approach not only in the gas turbine field but also in the whole industry domain. | |
