An Attention Mechanism Based Approach for Multivariate Time Series Anomaly Detection
Yihao Xiong, Jinbo Wang, Yuanlin Xin, Chi Zhang, Panpan Xue · 2023
Multivariate time series (MTS) anomaly detection is of utmost importance in contemporary industrial systems, such as power grids, distributed systems, and spacecraft, which generate extensive multivariate time series data through diverse sensor arrays and monitoring techniques. Despite the promising performance of deep learning-based approaches in MTS anomaly detection, effectively modeling monitored metrics and temporal relationships remains a significant challenge. Existing methods often fall short in capturing these dependencies explicitly, leading to limited performance. To address this issue, we propose a Graph Attention Network (GAT) based framework for MTS anomaly detection in this paper. Our approach incorporates a transformer encoder layer to obtain the MTS data representation, followed by two parallel GAT layers which capture relationships between different metrics and different timestamps simultaneously. Moreover, our method optimizes both a reconstruction model and a prediction model simultaneously, then calculates the anomaly score for each timestamp for anomaly detection. Experimental results on three real-world datasets demonstrate that our proposed approach outperforms state-of-the-art baselines in terms of F1-score.