【摘要】 多维时间序列异常检测是维持复杂工业系统有效运行的必要环节，如何准确识别大量设备中的异常模式是一项重要挑战。现有方案大多对多维时间序列下实体存在的动态依赖关系提取不足并且会受异常数据影响难以重构出正常的模式。为此，提出一种结合对比学习的双分支多维时间序列异常检测方法。首先，通过图结构学习和图特征增强得到实体之间的关联图以捕获动态变化的实体相关性，以及使用长短期记忆网络对时间依赖关系进行提取得到时间编码；接着，插入分块重组并采用图卷积操作提取不同尺度间的时空融合关系；最后，将融合后的关系特征进行联合对比训练得到正异常差异表示以评估异常。在SWaT、WADI、SWAP和MSL四个公开工业数据集上进行实验，与近年来的方法相比，所提方法取得了较好的F₁分数，分别为91.63%、90.60%、90.06%和93.69%,比MTGFLOW方法平均高出1.52百分点。实验结果表明，所提方法在提取动态依赖关系和区分正常与异常模式方面具有显著优势，验证了其在多维时间序列异常检测中的有效性和先进性，并显示出广泛的应用潜力。更多还原

【Abstract】 Multivariate time series anomaly detection is essential for maintaining the effective operation of complex industrial systems. Accurately identifying anomalous patterns across numerous devices presents a significant challenge. To address this challenge, this paper proposed a dual-branch multivariate time series anomaly detection method that incorporated contrastive learning. Firstly, it used graph structure learning and feature enhancement to construct relational graphs that captured dynamic correlations among entities. Long short-term memory（LSTM） networks were then employed to extract temporal dependencies and generate temporal encodings. Next, it introduced block reassembly and applied graph convolution operations to extract spatiotemporal relationships across different scales. Finally, the fused relational features underwent joint contrastive training to produce differential representations that effectively distinguished between normal and anomalous patterns. It validated the proposed method through experiments on four public industrial datasets: SWaT, WADI, SWAP, and MSL. The results demonstrate that this method achieves superior F₁ scores of 91.63%, 90.60%, 90.06%, and 93.69%, respectively, averaging 1.52 percentage points higher than the MTGFLOW method. The experimental results confirm that this method significantly enhances the extraction of dynamic dependencies and the distinction between normal and anomalous patterns. This validates its effectiveness and advancement in multivariate time series anomaly detection, indicating its broad potential for practical applications.更多还原