Learning Unified Representations of Normalcy for Time Series Anomaly Detection

Researchers have introduced a novel framework called Unified Unsupervised Anomaly Detection (U²AD) to address challenges in identifying abnormal patterns within multivariate time series data without prior knowledge. Published on arXiv, this study tackles the limitation of existing methods that struggle to distinguish normal data distributions from anomalies. The U²AD approach utilizes score-based generative modeling to learn the underlying distribution of normal samples. It features a new time-dependent score network and a unified training objective that captures both local and global temporal contexts. Reconstruction is achieved through a deterministic sampling process using an ordinary differential equation solver. Extensive experimental evaluations indicate that U²AD surpasses current state-of-the-art methods in detection accuracy. Furthermore, the framework demonstrates the ability to identify anomalies at significantly earlier stages of their occurrence, offering improved performance for unsupervised anomaly detection tasks in complex temporal datasets.