When Model Meets New Normals: Test-Time Adaptation for Unsupervised Time-Series Anomaly Detection

• DOI: 10.1609/aaai.v38i12.29210
• Publication Date: 2024-03-25T11:11:34Z
• Authors: Dongmin Kim; Sunghyun Park; Jaegul Choo
• Keywords: Anomaly (physics)

Time-series anomaly detection deals with the problem of detecting anomalous timesteps by learning normality from sequence observations. However, concept evolves over time, leading to a "new normal problem", where distribution can be changed due shifts between training and test data. This paper highlights prevalence new in unsupervised time-series studies. To tackle this issue, we propose simple yet effective test-time adaptation strategy based on trend estimation self-supervised approach normalities during inference. Extensive experiments real-world benchmarks demonstrate that incorporating proposed into detector consistently improves model's performances compared existing baselines, robustness shifts.