Least Information Spectral GAN With Time-Series Data Augmentation for Industrial IoT
Joonho Seon, Seongwoo Lee, Young Ghyu Sun, Soo Hyun Kim, Dong In Kim, Jin Young Kim
Abstract
In industrial Internet of Things (IIoT) systems, imbalanced datasets are prevalent because of the relative ease of acquiring normal operational data compared to abnormal or faulty data. An unbalanced distribution of data may lead to a biased learning problem, resulting in performance degradation of deep learning models. Data augmentation approaches based on generative adversarial networks (GAN) have been proposed to mitigate biased learning problems. However, GAN-based approaches constructed solely with convolutional neural networks may be incapable of extracting temporal properties from data. To utilize the temporal properties of data, a novel GAN structure consisting of an embedding network and recurrent neural networks is proposed in this paper. Additionally, in the novel GAN model based on mean-squared error, modified loss and mutual information terms are employed to improve training stability. From simulation results, it is confirmed that classification accuracy can be significantly improved by up to 54% based on the proposed method when compared with conventional fault diagnosis methods.