Leveraging subdomain alignment for enhanced anomaly detection in time series

Abstract

Time series anomaly detection focuses on identifying anomalies in continuously collected data at each time step. Developing an effective detection model requires not only accurate anomaly identification but also adaptability to the dynamic changes inherent in time series data. Current research primarily employs deep learning networks with task-specific reconstruction or prediction objectives to learn the underlying patterns of normal data. However, the distribution of complex time series data often shifts subtly over time, leading to evolving normal patterns. These distributional shifts make it difficult for models to establish clear decision boundaries, as they often fail to recognize such changes. To address these challenges, this paper proposes Subdomain Alignment for Enhanced Anomaly Detection in Time Series (SA-EADTS). This method aligns the latent distributions of unknown subdomains using a sensitive distance, enabling anomaly detection on unseen data distributions. Extensive experiments on four real-world datasets demonstrate that SA-EADTS significantly outperforms state-of-the-art baseline methods.