Abstract: Root cause analysis is of critical importance in aiding rapid system recovery, accurately assessing risks, and ensuring production safety. Addressing the limitations of current root cause analysis methods, which struggle to adequately characterize dependencies among different sensors and fail to capture stochastic fluctuations in time series, we propose a novel approach for multi-dimensional time series root cause analysis using a probabilistic generative graph attention network. This method embeds dimension features as Gaussian distribution vectors to represent sensor latent characteristics, effectively capturing random fluctuations in multi-dimensional time series and enhancing model robustness against noise. Simultaneously, a deep probability generative graph attention network is constructed to learn the nonlinear dependencies between dimensions, effectively modeling dependencies within sensor networks. Finally, the root cause analysis is conducted by integrating the network topology causal score and the individual node causal score. The proposed method is experimented on two public datasets and one private dataset, achieving optimal values on certain metrics. Specifically, on the SWaT dataset, improves by 2.2%, 6.3%, and 9.3% on PR@5, MAP@5, and MRR metrics respectively compared to the second-best method. On the WADI dataset, this approach improves by 8.1%, 7.0%, and 10.9% on PR@5, MAP@5, and MRR metrics respectively compared to the second-best method. On the Private dataset, this approach improves by 3.6% and 2.3% on MAP@3 and MRR metrics respectively compared to the second-best method.

Key words: multidimensional time series, root cause analysis, probabilistic generative network, graph attention network, deep learning

摘要： 根因分析对于帮助快速恢复系统、精确评估风险和保障生产安全具有重要意义。针对当前的根因分析方法不能很好地表征不同传感器之间的依赖关系，难以捕获时间序列中存在的随机波动的问题，提出一种面向多元时间序列根因分析的概率生成图注意力网络方法。所提方法将维度特征嵌入定义为高斯分布向量，用于表征不同传感器的潜在特征，以捕捉多维时间序列中存在的随机波动，提高模型抗噪性。同时构建深度概率生成图注意力网络，学习维度之间非线性的依赖关系，从而很好地建模传感器网络中的依赖性，最后综合网络拓扑因果得分和节点个体因果得分进行根因分析。在2个公开数据集和1个私有数据集上开展了实验，部分指标下取得了最优值，在SWaT数据集上，所提方法在PR@5、MAP@5和MRR三个指标上比次优方法分别提升2.2%、6.3%和9.3%；在WADI数据集上所提方法在PR@5、MAP@5和MRR三个指标上比次优方法分别提升8.1%、7.0%和10.9%；在私有的数据集上所提方法在MAP@3和MRR两个指标上比第二名分别提升3.6%和2.3%。

关键词: 多元时间序列, 根因分析, 概率生成网络, 图注意力网络, 深度学习