ADAD: Multivariate Time Series Anomaly Detection via Adaptive Series Decomposition

Time series anomaly detection aims to identify observations deviating from normal statues. As time series deliver complicated temporal patterns, many studies decompose time series into trend and seasonal components and then mine pattens in each component to simplify pattern learning process. However, mainstream methods often adopt fixed decomposition strategy such as moving average on the whole training set, ignoring the fact that time series evolve dynamically along time and each variate owns its unique evolutionary patterns. Some researchers try to conduct dynamic decomposition but treat pattern learning in each component as independent process, overlooking the guiding effect of learned features in one component on mining component-specific patterns in the other. To address the above issues, we propose anadaptivedecomposition-based multivariate time seriesanomalydetection algorithm, dubbed as ADAD. Specifically, we conduct time series decomposition in the frequency domain and formulate it as the spectrum splitting problem, with variate-specific splitting points dynamically learned for each sequence. Thus, decomposed trend and seasonal components could well reflect both variate-specific and time-specific features, enabling precise pattern mining in the following stages. Moreover, a cross-component connection is introduced to the pattern mining process, regarding patterns learned by one component as guidance for the pattern learning process in the other component. By excluding patterns that should not be learned, our design enables the model to learn distinct patterns in each component. Extensive experiments demonstrate that ADAD achieves superior performance over current state-of-the-art methods on public benchmarks.