A Hybrid MTS Anomaly Detection Method Based on Reconstruction and Adaptive Spatial-Temporal Graph Network

Effective anomaly detection of multivariate time series data is crucial for modern industrial applications. However, existing works usually ignore spatial and temporal dependencies of series. The data with complicated spatial-temporal feature put forward higher requirements for accurate detection of anomalies. In this paper, we propose an anomaly detection method for multivariate time series based on spatial-temporal network (STN). Firstly, the time series are reconstructed by Generative Adversarial Networks (GAN). The reconstruction block is used to amplify the influence of the abnormal state and prompt the anomalies more detectable, as poorly reconstructed data of abnormal state further worsen the quality of the final predicted value. Secondly, an adaptive learning graph structure layer is proposed. The graph structure is learned by adaptive method without any prior knowledge and is used for reconstructed data prediction. Thirdly, we use STN to extract spatial and temporal dependencies of series. Especially, attention mechanism is introduced into graph convolution network to enhance the prediction performance by focusing on more important neighbor node message. Finally, our model is tested on 3 public datasets to evaluate performance. The results demonstrate that our method performs better than baselines.