A Multi-scale Parallel Unsupervised Model for Multivariate Time Series Anomaly Detection

Anomaly detection for multivariate time series data is of great significance for practical applications. The existing anomaly detection methods mainly adopt a fixed length sliding window to extract data features and perform deep learning training. However, a single fixed length window of data makes it difficult to simultaneously detect anomalies in different scale, such as small-scale point anomalies and large-scale contextual anomalies. Additionally, the patterns in multivariate time series data may be more complex and diverse. This paper proposes an unsupervised multi-scale model for multivariate time series anomaly detection (MMTSAD) to address the above problem. We downsample the original data to obtain coarse-grained and fine-grained sequences in different scales, and design two autoencoder modules based on attentionmechanism to learn time series patterns at different scales. In addition, we introduce a GAT module in the coarse-grained autoencoder to capture the correlation between different variables. At the detection stage, we propose an anomaly score fusion method to comprehensively fuse the anomaly scores from different scale models. We conduct experiments on five real-world public datasets. The results show that MMTSAD outperforms most existing models.