Simplified Graph Convolution based Anomaly Detection for Multivariate Time Series

In industrial settings, multivariate time series generated from various monitoring metrics are abundant. Anomaly detection in these time series is crucial for applications such as fault diagnosis and root cause analysis. Recent advancements in unsupervised methods, particularly autoencoder (AE)-based reconstruction architectures, have made significant progress in this area. These systems learn the normal data distributions and produce substantial reconstruction errors when encountering anomalies. While AEs are effective at reconstructing subtle abnormal patterns due to their strong generalization capabilities, this can also result in a high false negative rate. Furthermore, AE-based models often fail to account for inter-variable dependencies across different time scales. In this paper, we propose an enhanced anomaly detection framework that builds upon the Multiscale Wavelet Graph Autoencoder (MEGA) by replacing the Graph Convolutional Network (GCN) with Simplified Graph Convolution (SGC) to improve model performance. The key idea is to utilize the spectral methods of SGC to process the multivariate time series data, integrating Discrete Wavelet Transform (DWT) into the AE. We conducted experiments on three public multivariate time series anomaly detection datasets, and the results demonstrate that the improved model using SGC outperforms existing methods.