Improved and accurate fault diagnostic model for gas turbine based on 2D-wavelet transform and generative adversarial network

Abstract Severe working environments cause gas turbines to break down, which can directly affect their performance. Research on the diagnostic methods for gas turbine faults, such as, gas path faults and sensor failures, has always raised concerns. However, traditional fault diagnosis algorithms mostly use instantaneous data rather than time-series data, because they cannot efficiently use time-series analysis to extract fault features and improve algorithm accuracy. Problems with sparse fault samples and categories are also encountered with these algorithms. In this study, a gas turbine fault diagnostic method based on a 2D-wavelet transform and generative adversarial network (GAN) was proposed. The data preprocessing method, 2D-wavelet transform, of multiple time series images was used to obtain fault features. Based on the Fréchet inception distance, a performance evaluation index, an optimal generator built from a deep convolutional GAN model was selected to solve sparse or imbalanced datasets. The classification accuracy of the four algorithms, namely, random forest, support vector machine, convolutional neural network, and deep neural network, verified the performance of the data preprocessing and dataset building methods mentioned earlier. Compared with the original data, the 2D wavelet transform effectively improved the model accuracy. The generated samples also improved the misclassification issue caused by the imbalanced dataset; however, the ratio of real and generated samples in datasets still requires more attention.