Prediction of Gas Turbine Trip by Combining Gas Path Measurements and Vibration Signals

As well-known, gas turbine (GT) trip causes a reduction of GT lifespan and makes costs increase, because of unscheduled stops. Thus, predicting GT trip in advance would allow saving costs and maintaining high operational efficiency and availability. For these reasons, this paper presents the development of a model based on deep learning, i.e., a temporal convolutional neural network, aimed at GT trip prediction. The model exploits both vibration data collected from multiple sensors installed at different locations of the GT asset and gas path measurements. Its prediction capability is compared to that of the models fed only with vibration signals or gas path measurements. The simultaneous exploitation of vibration signals and gas path measurements, which is the main novel contribution of this paper, is aimed at enhancing the reliability of GT trip prediction, since trip causes are heterogeneous. The paper also investigates different time frames for predicting GT trip, by considering training and testing time windows taken at different time points from trip occurrence. The case study is represented by a collection of field measurements referring to real-world GT trips taken from multiple GTs during several years of operation. The field data used to validate the results cover multiple ambient and operating conditions. The results demonstrate that the simultaneous exploitation of both gas path measurements and vibration signals actually enhances the reliability of GT trip prediction in the majority of the considered cases, with accuracy values up to 73%.