GUARD: Multigranularity-based Unsupervised Anomaly Detection Algorithm for Multivariate Time Series

Unsupervised anomaly detection on multivariate time series (MTS) is a valuable problem in practice because it significantly reduces labeling costs. Even though most classical methods begin to consider inter-series dependencies, the hierarchical relationships between and within-series data are often ignored. To overcome these limitations, in this paper, we propose a multiGranUlArity based unsupeRvised anomaly Detection (GUARD) algorithm that captures the hierarchical relationships in three granularities and merges these dependencies in the following anomaly detection task. Specifically, the proposed algorithm first captures the internal dependencies within the subsequence and learns a hidden representation based on a trained autoencoder. Then, a hierarchical mechanism is designed to capture the local and global dependencies between the given subsequence and the whole set and learn their hidden representation. The experimental results show that the proposed algorithm can effectively and stably exploit the potential relationship of MTS and detect anomalies more accurately than the basic approaches.