ScholarWorks@성균관대학교

초록

Gas turbine power plants operate under high-temperature and high-pressure conditions, and failures in components such as combustors can lead to unplanned shutdowns and power supply disruptions. Developing fault diagnosis models to prevent such failures is challenging because of the limited availability of actual fault data. This study proposes a generative approach that synthesizes fault data from normal exhaust temperature data to train a supervised learning model. The exhaust temperatures measured at the turbine outlet reflect the thermal balance across the combustors. The data from the sensors arranged in a circular layout are transformed into two-dimensional images to represent the spatial distribution of the temperature. Synthetic fault data are generated by applying three types of designed imbalance patterns to normal data, simulating changes in the temperature distribution that occur under fault conditions. A convolutional neural network model was trained using normal data collected from operating power plants and the generated fault data, and then evaluated on actual fault cases. The performance was compared with that of unsupervised models trained only on the same normal data. To evaluate the generalization, the proposed method was also tested on data from another power plant that was not used during training. The results show that the model trained with synthetic fault data achieved improved diagnostic performance and generalization capability without requiring actual fault data for training.

키워드