SGAD-GAN: Simultaneous Generation and Anomaly Detection for time-series sensor data with Generative Adversarial Networks

In recent years, mechanical sensor data anomaly detection has gained much attention in the machine learning and mechanical fault warning fields. Limited by the fact that there is far less anomalous data available for analysis than normal data, many machine learning methods fail to perform excellent detection results. In this paper, we propose a novel Simultaneous Generation and Anomaly Detection with Generative Adversarial Networks (SGAD-GAN) framework to tackle the challenge of anomaly detection under imbalanced datasets. In our framework, the generators accomplish transfer between sensor signals while synthesizing realistic data. In addition to the regular discriminators for identifying the authenticity of samples, we design a classification discriminator to facilitate data synthesis in the direction benefitting anomaly detection, which is trained to act as an anomaly detector in an identical way as other discriminators. We conduct extensive data synthesis and anomaly detection experiments on Hydraulic System Sensor (HSS) data from the Jiaolong deep-sea manned submersible, and show the generalization ability of our approach to different application domains on the public dataset KDDCUP99. The experimental results demonstrate that our proposed algorithm outperforms several state-of-the-art data augmentation and anomaly detection methods.