Root Cause Analysis in Causal Anomaly Detection

Abstract

The rapid growth of big data has amplified the importance of reliable anomaly detection, particularly in complex technical systems susceptible to malign anoamlies. While most existing approaches rely heavily either on statistical properties and manual inspection often unfeasable for big data or opaque deep learning, we propose a novel framework for causal anomaly detection that integrates data-driven anomaly identification with causal reasoning. The method consists of three stages: (i) anomaly detection directly on time series of system variables, followed by flagging variables exhibiting anomalous behavior; (ii) causal discovery, where the underlying causal structure of the system is derived in the form of a causal graph; and (iii) root cause analysis, which traces the propagation of anomalies backward through the causal structure to identify their origins. We here concentrate on the third stage of this process and present results from a case study on satellite telemetry data to demonstrate the effectiveness of this approach. We highlight its potential for advancing anomaly detection in complex, safety-critical domains.