A Deep Adversarial Diagnosis Method for Fault Line Detection with Imbalanced Small Sample Scenarios

Abstract

AbstractSingle-line-to-ground (SLG) fault line detection is crucial for electric utilities to achieve fault isolation and service restoration. The transient zero-sequence current (TZSC) was usually adopted as key quantity for indicating the SLG fault line. However, the waveforms of TZSC have noise interference in the engineering environment, which requires a lot of manpower to collect. What’s more, in the SLG fault, the imbalance phenomenon exists in which the data collected from the fault line is less than that of the normal lines, which causes poor performance for the data-driven methods. A deep adversarial diagnosis method is proposed to detect fault lines using an improved generative adversarial network to have improved accuracy in detecting fault lines in the case of small unbalanced samples. First, we collected a small amount of data as the original sample. Then, the original sample is obtained and labeled to construct training and testing datasets. Because the data collected from the fault line and normal line is imbalanced, this work utilized the generative adversarial network (GAN) to produce synthetic samples of fault line to realize the balance between two kinds of data. In attempt to improve the training process of the GAN, the Wasserstein distance was used to redesign the loss function of the model. Compared with mainstream data-driven approaches, the proposed approach realizes the data self-generation and facilitates subsequent classification work. The proposed method is superior to other data-driven methods in terms of detection accuracy under unbalanced small sample conditions.KeywordsDistribution systemsFault line detectionGenerative adversarial networkSmall sample scenario