DRPAD: A Dynamic-Aware and Robust Paradigm for Time Series Anomaly Detection

Yujue Zhou, Kaixiang Yang, Hong Peng, Jiarong Liu, Shuanghua Yang

Keywords: Anomaly Detection;Time Series;Forecasting-Based Methods;Anomaly Propagation

Abstract: Forecasting-based methods dominate unsupervised time series anomaly detection but primarily emphasize feature extraction and prediction accuracy. In real-world applications, however, the distinctiveness of anomalies depends on additional critical factors. We identify three major challenges: (1) anomaly propagation, (2) distribution shifts, and (3) univariate anomalies—common phenomena that are often overlooked. To address these issues, we propose DRPAD (Dynamic-Aware and Robust Paradigm for Time Series Anomaly Detection), introducing three novel components: Dynamic Prediction Replacement, Segmentation-Based Normalization, and a Mean & Dimension Dual-Check Strategy. Extensive experiments on nine benchmark datasets demonstrate that DRPAD can significantly enhance the performance of a wide range of forecasting-based methods, achieving state-of-the-art results. The source code is publicly available at \url{https://anonymous.4open.science/r/DRPAD-BEC8/}

Primary Area: learning on time series and dynamical systems

Submission Number: 4612