Towards Robust ML-Algorithms for the Condition Monitoring of Switchgear

Abstract

In this paper, we describe our work-in-progress regarding the use of artificially generated data for the training of classifiers in an industrial context. In particular, our goal is to classify faulty/healthy switchgear by using infrared images. The paper describes the use of Generative Adversarial Networks (GANs) for the generation of new infrared images in order to bolster the meagre, repetitive and unbalanced data available to us. The paper describes the obstacles encountered, potential solutions, and the results of multiple experiments to test the impact of synthetic data on training.