Local-global augmented contrastive learning for multivariate time series anomaly detection

Abstract Learning temporal representations is a challenging task, especially when the labelled flight data is limited. Recently, contrastive learning has shown significant improvement in extracting representations by contrasting different augmented views of the data. However, existing approaches do not comprehensively exploit both local and global data information, resulting in an inability to learn the local-global temporal representations. Moreover, these methods still rely on small amount of labelled data to enhance the robustness of anomaly detection. To address the challenges, in this paper, we propose a novel contrastive learning anomaly detection framework for multivariate time series (TimeCoLAD2), which aims to detect anomalous sequences of the aircraft pneumatic system without any labelled data, thereby eliminating the need for manual labelling. Our framework first extracts information using local-global data augmentation techniques, incorporating contrastive losses to fully capture temporal representations. Next, we dynamically determine the threshold values based on the flight phases to more accurately detect anomalies. Comparative experimental results show that the proposed anomaly detection framework outperforms baseline methods on a multivariate time series pneumatic system dataset, especially TimeCoLAD2, which increases F1 scores by up to 59%. This research contributes to the development of an engine performance anomaly detection system across various flight phases without relying on manual labelling.