Fault identification method based on generative adversarial network in distributed acoustic sensing

Abstract

With the rapid development of machine learning and deep learning, neural-network-based pattern recognition techniques have become a trend for distributed acoustic sensing (DAS) systems. However, in some cases, certain types of data are difficult to obtain, which leads to imbalanced sample sets. To solve this problem, a data augmentation method based on a generative adversarial network is proposed in this study. First, normal operation samples, including the normal operation training and testing sets are collected using the DAS system. The cyclegan algorithm is then used to generate fault operation samples, a part of which can be selected as the fault operation training set, whereas the others can be viewed as the fault operation testing set. Furthermore, an effective method of data augmentation, called k-means clustering-synthetic minority oversampling technique deep convolutional generative adversarial network, is proposed to enhance the fault operation training set. Finally, the proposed method is applied to the on-load tap-changer, and the experimental results show that the average accuracy of the validation set for the classification task can reach 97%.