Transformer-GAN architecture for anomaly detection in multivariate time series
美玲 蔡, 家喜 汪, 金平 刘, 朝晖 唐, 永芳 谢 · Scientia Sinica Informationis · 2022
Anomaly detection based on multivariate time series correlation data collected in real time during the process is one of the key aspects of preventing industrial process accidents and ensuring system safety. However, industrial multivariate time series anomaly detection still faces two major challenges: (1) the complex nonlinear correlation characteristics among time series data variables lack an effective representation method, and (2) the complex correlation among time series with highly unbalanced normal/abnormal distribution needs to be deeply explored. In this paper, we propose a Transformer generative adversarial networks (GAN)-based multivariate time series anomaly detection method (TGAN-MTSAD). TGAN-MTSAD employs transformer neural networks as the base model of GAN and introduces a graph attention layer to automatically learn complex dependencies among time series multivariate variables. It applies a patch trick to enable the model to effectively capture anomaly details within a time window. An anomaly score calculation method is proposed based on a combination of reconstruction and discrimination errors. An extensive performance validation and comparative experimental analysis of the proposed method were carried out using three real-world datasets. The results show that TGAN-MTSAD can effectively detect in-process timing anomalies, outperforming the baseline method in most cases, and has good interpretability for complex industrial anomaly detection.