Adversarial Semi-Supervised Learning for Diagnosing Faults and Attacks in Power Grids

Abstract

This paper proposes a novel adversarial scheme for learning from data under harsh learning conditions of partially labelled samples and skewed class distributions. This novel scheme integrates the generative ability of the state-of-the-art conditional generative adversarial network with the semi-supervised deep ladder network and semi-supervised deep auto-encoder. The proposed generative-adversarial based semi-supervised learning framework, named GBSS, is a triple network that aims to optimize a newly defined objective function to enhance the performance of the semi-supervised learner with the help of a generator and discriminator. The duel between the generator and discriminator results in the generation of more synthetic minority class samples that are very similar to the original minority samples (attacks and faults). Meanwhile, GBSS trains the semi-supervised model to learn the general distribution of the minority class samples including the newly generated samples in contrast to other classes and iteratively adjusts its weights. Moreover, a diagnostic framework is designed, in which GBSS and several state-of-the-art semi-supervised learners are used for learning and diagnosing attacks and faults in power grids. These methods are evaluated and compared for diagnosing attacks and faults in two different power grid cases. The attained results demonstrate the superiority of GBSS in diagnosing attacks and faults under the harsh conditions.