Decomposed Transformer with Frequency Attention for Multivariate Time Series Anomaly Detection
Shuxin Qin, Jing Zhu, Dan Wang, Liang Ou, Hongxin Gui, Gaofeng Tao
Abstract
Unsupervised anomaly detection for multivariate time series has been an active area of research due to its enormous potential for industrial applications. Existing works have made extraordinary progress in time series representation, reconstruction and forecasting. However, the intricate temporal patterns of individual signals and the relationships between different signals are rarely explicitly considered, which limits the performance. To this end, we propose a novel approach based on Transformer and signal decomposition for time series anomaly detection. Specifically, we provide a multi-view embedding method to capture temporal and correlational features of signals. To take full advantage of temporal patterns, we design a frequency attention module to extract periodic oscillation features. Then, the signals are decomposed into seasonal, trend and remainder components for further representation separately. Additionally, we employ a optimization strategy to improve feature representation. Extensive experiments on various public benchmarks demonstrate that our method has achieved the state-of-the-art performance.