Active Learning Boosts Unsupervised Time Series Anomaly Detection

Detecting subtle and noisy anomalies in complex time series data remains a significant challenge, particularly in large-scale industrial applications where data labeling is prohibitively expensive. While unsupervised methods are widely adopted, they often struggle to differentiate near-normal anomalies from regular patterns and are vulnerable to noise within normal samples. This new research proposes a novel framework that integrates active learning to iteratively improve the performance of existing unsupervised anomaly detection models. The framework introduces two key contributions. First, a masked time-series reconstruction feedback strategy compels the model to learn robust temporal dependencies, making it more resilient to variations. Second, a minimax learning strategy is employed to enhance robustness by treating normal and abnormal samples differently, thereby encouraging the model to better capture the dynamics of subtle and noisy patterns. Evaluated across 28 test cases involving four multivariate time-series datasets and seven unsupervised backbone models, the proposed method demonstrated a substantial 12.39% improvement in AUC compared to the original models. This confirms its potential to be readily integrated into existing reconstruction-based anomaly detection systems, significantly boosting their ability to identify difficult-to-detect anomalies.