A Memory-Guided Anomaly Detection Model with Contrastive Learning for Multivariate Time Series

• DOI: 10.32604/cmc.2023.044253
• Publication Date: 2023-11-29T09:18:00Z
• Authors: Wei Zhang; Ping He; Ting Li; Fan Yang; Ying Liu
• Keywords: Anomaly (physics); Benchmark (surveying); Feature (linguistics)

Some reconstruction-based anomaly detection models in multivariate time series have brought impressive performance advancements but suffer from weak generalization ability and a lack of identification. These limitations can result the misjudgment models, leading to degradation overall performance. This paper proposes novel transformer-like model adopting contrastive learning module memory block (CLME) overcome above limitations. The tailored for data learn contextual relationships generate temporal fine-grained representations. record normal patterns these representations through utilization attention-based addressing reintegration mechanisms. two modules together effectively alleviate problem generalization. Furthermore, this introduces fusion strategy that comprehensively takes into account residual feature spaces. Such enlarge discrepancies between abnormal data, which is more conducive proposed CLME not only efficiently enhances also improves detection. To validate efficacy approach, extensive experiments are conducted on well-established benchmark datasets, including SWaT, PSM, WADI, MSL. results demonstrate outstanding performance, with F1 scores 90.58%, 94.83%, 91.58%, 91.75%, respectively. findings affirm superiority over existing state-of-the-art methodologies terms its detect anomalies within complex datasets accurately.