Addressing imbalance of sample datasets in dissolved gas analysis by data augmentation: Generative adversarial networks

Abstract Dissolved gas analysis (DGA) is an efficient technique to diagnose transformer faults; however, the imbalance of the sample datasets is frequently observed in DGA, which has become one of the main challenges for accurate fault diagnosis. A generative adversarial network (GAN) model is proposed to realize the data augmentation for addressing these problems. The GAN model is nested by two backpropagation neural networks (BPNN) containing a generator and a discriminator. The generator produces real‐like data and continuously shapes the real‐like data approximate to real data. The discriminator works to distinguish between the real‐like data and the real data until the real‐like data cannot be discriminated. Under this adversarial learning, the generator output obtains the real‐like DGA samples to supplement the minority samples. To verify the diagnosis accuracy after the DGA data augmentation, three typically intelligent models, that is, BPNN, support vector machine (SVM) and deep belief network (DBN) are introduced to establish the diagnostic models. The results show that the recognition accuracy of the minority samples acquires average 30%, 25% and 50% increases for the three models respectively. It is shown that the GAN model well balances the sample datasets and strengthens the performance of the fault diagnostic models.

Abstract licence: CC BY-ND

A person counts once a day per paper. We keep no cookie and no address: a reader is a hash that changes every day. Any month under five reads shows no number.