Root-cause analysis for time-series anomalies via spatiotemporal causal graphical modeling

摘要

Modern distributed cyber-physical systems encounter a large variety ofanomalies and in many cases, they are vulnerable to catastrophic faultpropagation scenarios due to strong connectivity among the sub-systems. In thisregard, root-cause analysis becomes highly intractable due to complex faultpropagation mechanisms in combination with diverse operating modes. This paperpresents a new data-driven framework for root-cause analysis for addressingsuch issues. The framework is based on a spatiotemporal feature extractionscheme for multivariate time series built on the concept of symbolic dynamicsfor discovering and representing causal interactions among subsystems of acomplex system. We propose sequential state switching ($S^3$) and artificialanomaly association ($A^3$) methods to implement root-cause analysis in anunsupervised and semi-supervised manner respectively. Synthetic data from caseswith failed pattern(s) and anomalous node are simulated to validate theproposed approaches, then compared with the performance of vectorautoregressive (VAR) model-based root-cause analysis. The results show that:(1) $S^3$ and $A^3$ approaches can obtain high accuracy in root-cause analysisand successfully handle multiple nominal operation modes, and (2) the proposedtool-chain is shown to be scalable while maintaining high accuracy.