SFAFormer: Sampling Frequency-Aware Transformer Specialized for Unsupervised Anomaly Detection in Irregular Multivariate Time Series

초록

Anomaly detection in irregular multivariate time series (IMTS) plays a crucial role in diverse applications such as fault prediction in industrial systems, energy management, and medical diagnosis. However, most existing methods are developed under the assumption of regularly sampled data, which makes them insufficient for real-world scenarios characterized by irregular sampling and missing values. We aim to develop an unsupervised framework specifically tailored for anomaly detection in IMTS. We propose SFAFormer, a framework that adopts a sampling frequency-aware (SFA) embedding to convert irregular time series into fixed-length vectors and a dual transformer encoder architecture to jointly capture temporal dependencies and intervariable interactions. Furthermore, the input sequence is divided into patches, and both interpatch and intra-patch relationships are modeled to effectively identify anomaly patterns. Experimental evaluations on four benchmark datasets (PSM, SMD, SWAT and GECCO) show that SFAFormer consistently outperforms existing approaches, achieving F1-score improvements of up to 30%p and AUROC gains of up to 18.9%p while maintaining robustness under diverse irregular sampling conditions. These findings demonstrate that SFAFormer provides an effective and practical solution for anomaly detection in IMTS.

키워드