A Comparison of GANs-Based Approaches for Combustor System Fault Detection
Rui Xu, Weizhong Yan · 2020
In manufacturing industry, anomaly detection (AD) has been widely applied to monitor asset operation status and provide decision support for proactive maintenance. However, the extreme complexities of many industrial assets have raised many challenges to the classical anomaly detection approaches. For the past several years, generative adversarial networks (GANs) have achieved breakthrough in a variety of applications, such as image generation and video prediction. Some initial applications of GANs in image-related AD problems also show promising results. In this paper, we investigated the performance of three GANs-based anomaly detection approaches for a specific industrial use case - fault detection of the combustion system of gas turbines in power plants. The results show that under the framework of semi-supervised learning, the three approaches do not perform well on the one-year field data collected from a gas turbine. However, if a small portion of fault data is provided for training, we observe that the performance of GANs is significantly improved.