AnomalyBERT: Self-Supervised Transformer for Time Series Anomaly Detection using Data Degradation Scheme

• DOI: 10.48550/arxiv.2305.04468
• Publication Date: 2023-01-01
• Authors: Yungi Jeong; Eun-Seok Yang; Jung Hyun Ryu; Imseong Park; Myungjoo Kang
• Keywords: Code (set theory); Anomaly (physics)

Mechanical defects in real situations affect observation values and cause abnormalities multivariate time series, such as sensor or network data. To perceive data, it is crucial to understand the temporal context interrelation between variables simultaneously. The anomaly detection task for especially unlabeled has been a challenging problem, we address by applying suitable data degradation scheme self-supervised model training. We define four types of synthetic outliers propose which portion input replaced with one outliers. Inspired self-attention mechanism, design Transformer-based architecture recognize detect unnatural sequences high efficiency. Our converts points into representations relative position bias yields scores from these representations. method, AnomalyBERT, shows great capability detecting anomalies contained complex series surpasses previous state-of-the-art methods on five real-world benchmarks. code available at https://github.com/Jhryu30/AnomalyBERT.