VGAN: Generalizing MSE GAN and WGAN-GP for Robot Fault Diagnosis

Abstract

Generative adversarial networks (GANs) have shown their potential for data generation. However, this type of generative model often suffers from oscillating training processes and mode collapse, among other issues. To mitigate these, this work proposes a generalization of both mean square error (mse) GAN and Wasserstein GAN (WGAN) with gradient penalty, referred to as VGAN. Within the framework of conditional WGAN with gradient penalty, VGAN resorts to the Vapnik V-matrix-based criterion that generalizes mse. Also, a novel early stopping-like strategy is proposed that keeps track during training of the most suitable model. A comprehensive set of experiments on a fault-diagnosis task for an industrial robot where the generative model is used as a data augmentation tool for dealing with imbalance datasets is presented. The statistical analysis of the results shows that the proposed model outperforms nine other models, including vanilla GAN, conditional WGAN with and without conventional regularization, and synthetic minority oversampling technique, a classic data augmentation technique.