PAMA: Dual-Memory Augmentation Assisted Pseudo-Anomaly Contrastive Learning for Multivariate Time Series Anomaly Detection

Tang Qian, Hong-Jie Xing, Chun-Guo Li

Keywords: Multivariate time series anomaly detection, Contrastive learning, Dual-memory augmentation, Pseudo anomaly, Uncertainty learning

TL;DR: Dual-Memory Augmentation Assisted Pseudo-Anomaly Contrastive Learning for Multivariate Time Series Anomaly Detection

Abstract: For multivariate time series anomaly detection, most methods assume that training data are clean and ignore the characteristics of anomalous data. They often suffer from the overgeneralization problem during the reconstruction process. To address these problems, dual-memory augmentation assisted pseudo-anomaly contrastive learning for multivariate time series anomaly detection (shorted as PAMA) is proposed. First, the prior knowledge of anomalies is utilized to generate pseudo anomalies from the original time series. The normal and pseudo-anomalous feature representations that contain global and local information are respectively achieved by the global and local encoders. Two independent memory modules are constructed to further memorize normal and pseudo-anomalous prototypes. Second, a dual-memory augmentation mechanism is proposed to conduct data augmentation upon the normal and pseudo-anomalous feature representations and then obtain the memory-augmented feature representations. Third, the pseudo-anomaly contrastive learning is proposed to perform temporal contrastive learning and instance contrastive learning on the obtained memory-augmented representations. Compared with the fourteen baseline methods, the experimental results demonstrate that PAMA achieves the optimal detection performance.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 11843