Multi-Attention Integrated Convolutional Network for Anomaly Detection of Time Series

Time series containing abundant monitoring information can tell how a system is running, and anomaly detection of time series is closely related to the identification of potent fault and implementation of proper measurements. Therefore, accurate anomaly detection is of great significance to system stability. Anomaly detection of time series has been studied for decades, and various approaches have been reported for effective detection. In this paper, we propose a novel deep learning-based model for anomaly detection of time series. The proposed model consists of three parallel pipelines and each pipeline containing a convolutional unit in series connection with an amplified attention mechanism is responsible for both temporal and spatial feature extraction. The parallel design can help the model capture input features in a different perception field and the pipelines can work complementarily for a comprehensive understanding. The proposed model is then evaluated in multiple datasets including univariate and multivariate time series, and the results prove the effectiveness of the proposed compact model. An ablation study is also carried out to demonstrate the promotion of the proposed amplified attention mechanism.