A responsive approach to multivariate time-series anomaly detection with K-distance based calibrated reconstruction

Anomaly detection is of great importance in the Industrial Internet of Things (IIoT) as it enables intelligent process control, analysis, and management, etc. There is a growing demand for responsive anomaly detection models that exhibit high sensitivity and precision. Various models have been proposed to address these challenges, with reconstruction-based models currently dominating the field. These models concentrate on learning complex representations of time-series; however, representation learning can be negatively affected by anomaly contamination and may experience varying degrees of detection delay. In this paper, we introduce a responsive approach for multivariate time-series anomaly detection. We design a new distance function that emphasizes the importance of the current timestamp in calculating anomaly scores to achieve more responsive detection. Additionally, we employ a calibration method that allows the model to focus exclusively on reconstructing normal patterns and an integrated prediction mechanism to enhance detection precision. Comprehensive experiments conducted on five real-world datasets indicate that our approach surpasses the state-of-the-art methods in multivariate time-series anomaly detection, shortens detection delays, and provides support for responsive anomaly reporting.