Efficient time series anomaly detection by multiresolution self-supervised discriminative network

Time series anomaly detection aims to identify abnormal subsequences in time series that are markedly different from the temporal behaviors of the entire sequence. Although previous density-based or proximity-based anomaly detection methods are usually used for anomaly detection, they are still suffering from high computational costs due to the need of traversing the whole training dataset during testing. Recently, reconstruction-based deep learning methods are popular for time series anomaly detection. However, they may not work well because their objective is to recover all information appeared in time series, including high-frequency noises. In this paper, we propose a simple yet efficient method called Multiresolution Self-Supervised Discriminative Network (MS2D-Net) for efficient time series anomaly detection. Specifically, the MS2D-Net includes a multiresolution downsampling module, a feature extraction module, and a self-supervised discrimination module. The multiresolution downsampling module generates some multiresolution samples by downsampling the original time series with different sampling rates and creates different pseudo-labels representing multi-scale behaviors in time series. Then, in the feature extraction module, a shallow convolution network is used to extract temporal dynamics in time series at multiple resolutions. Finally, the self-supervised discrimination module uses the pseudo-labels obtained from the multiresolution downsampling module as the self-supervised information to help separate anomalies from the normal time series samples. Experimental results show that the proposed MS2D-Net can outperform recent strong deep learning baselines on 18 benchmarks for time series anomaly detection with a much lower computational cost.