Decompose Auto-Transformer Time Series Anomaly Detection for Network Management
Bo Wu, Chao Fang, Zhenjie Yao, Yanhui Tu, Yixin Chen
Abstract
Time series anomaly detection through unsupervised methods has been an active research area in recent years due to its enormous potential for networks management. The representation and reconstruction of time series have made extraordinary progress in existing works. However, time series is known to be complex in terms of their temporal dependency and stochasticity, which makes anomaly detection difficult. To this end, we propose a novel approach based on a decomposition auto-transformer networks(DATN) for time series anomaly detection. The time series is decomposed into seasonal and trend components, and renovated as a basic inner block deep model. With this design, transformers can decompose complex time series in a progressive manner. We also design an auto-transfomer block that determines dependencies and representation aggregation at the sub-series level based on series seasonal and trend components. Moreover, the complex transformer decoder is replaced by a simple linear decoder, which makes the model more efficient. Extensive experiments on various public benchmarks demonstrate that our method has achieved state-of-the-art performance.