An unsupervised framework for drift-aware anomaly detection in streaming time series

Abstract

This paper presents an unsupervised adaptive drift-aware anomaly detection framework (ADA-ADF) designed to address the challenges of concept drift in time series data streams. ADA-ADF integrates a hybrid drift detection mechanism, combining statistical tests with performance-based metrics to accurately identify and distinguish between sudden and incremental drifts. To ensure effective adaptation, it employs a replay-based model update strategy that adjusts replay ratios in a drift-specific manner and incorporates representative historical data based on reconstruction errors. This approach allows the model to seamlessly adapt to evolving data distributions while maintaining high stability and accuracy. Extensive experiments on four diverse datasets demonstrate ADA-ADF’s superior performance in managing various drift and application scenarios. It consistently outperforms state-of-the-art methods, particularly in environments characterized by incremental or sudden drifts. With robust adaptability to changing data patterns and accurate anomaly detection capabilities, ADA-ADF provides a reliable solution for real-world applications, such as IoT and environmental monitoring.