On the Effect of Loss Function in GAN Based Data Augmentation for Fault Diagnosis of an Industrial Robot

Intelligent fault diagnosis often requires a balanced dataset which is hard to be obtained in industrial equipments, often resulting in an imbalance between data in normal and data in the presence of faults. Data augmentation techniques are among the most promising approaches to mitigate such issue. Generative adversarial networks (GAN) are a type of generative model consisting of a generator module and a discriminator. Through adversarial learning between these modules, the optimised generator can produce synthetic patterns that can be used for data augmentation. We investigate the role of loss function in improving the training efficiency of GAN. We proposed a generalization of both mean square error (MSE GAN) and Wasserstein GAN with gradient penalty (WGAN-GP), referred to as VGAN (the V-matrix based GAN) to mitigate training instability. Also, we investigate the sliced Wasserstein distance (SWD) as the loss function of a cycle consistency generative adversarial network (CycleGAN), referred to as SW-CycleGAN. Both two models are evaluated on an industrial robot data set. Experimental results show that the proposed loss functions outperform other competitive models especially in terms of computational costs.