TGAN-AD: Transformer-Based GAN for Anomaly Detection of Time Series Data

• Published in: Applied sciences Vol. 12; no. 16; p. 8085
• Main Authors: Xu, Liyan, Xu, Kang, Qin, Yinchuan, Li, Yixuan, Huang, Xingting, Lin, Zhicheng, Ye, Ning, Ji, Xuechun
• Format: Journal Article
• Language: English
• Published: Basel MDPI AG 01.08.2022
• Subjects: Anomalies; anomaly detection; Artificial intelligence; Clustering; Datasets; Deep learning; Electric power grids; Fault diagnosis; generative adversarial network; Generative adversarial networks; Generators; Information processing; Machine learning; Normal distribution; Statistical methods; Stochastic models; Time series; time series data; transformer
• ISSN: 2076-3417; 2076-3417
• DOI: 10.3390/app12168085

Anomaly detection on time series data has been successfully used in power grid operation and maintenance, flow detection, fault diagnosis, and other applications. However, anomalies in time series often lack strict definitions and labels, and existing methods often suffer from the need for rigid hypotheses, the inability to handle high-dimensional data, and highly time-consuming calculation costs. Generative Adversarial Networks (GANs) can learn the distribution pattern of normal data, detecting anomalies by comparing the reconstructed normal data with the original data. However, it is difficult for GANs to extract contextual information from time series data. In this paper, we propose a new method, Transformer-based GAN for Anomaly Detection of Time Series Data (TGAN-AD), The transformer-based generators of TGAN-AD can extract contextual features of time series data to prompt the performance. TGAN-AD’s discriminator can also assist in determining abnormal data. Anomaly scores are calculated through both the generator and the discriminator. We have conducted comprehensive experiments on three public datasets. Experimental results show that our TGAN-AD has better performance in anomaly detection than the state-of-the-art anomaly detection techniques, with the highest Recall and F1 values on all datasets. Our experiments also demonstrate the high efficiency of the model and the optimal choice of hyperparameters.