CAAD: Causality-Aware Multivariate Time Series Anomaly Detection via Multi-Scale Alignment and Structural Causal Consistency

[Paper]

Abstract (Brief Summary)

CAAD reframes multivariate time series anomaly detection as the continuous verification of Granger causality consistency through exogenous variables. It models exogenous time-series variables as residuals, identifying anomalies as significant deviations caused by external interventions. The framework leverages multi-scale alignment to internalize system dynamics and a gradient-based causal matrix to monitor internal causal relationship breakdowns. By quantifying causal deviations of both dynamic evolution and relational topology, CAAD captures subtle causal shifts for precise anomaly detection in complex industrial systems.

Main Contributions

• Causality-Aware Reframing: Identifies anomalies as breakdowns in Granger causal consistency rather than magnitude deviations, targeting stealthy failures missed by conventional methods.
• Multi-Scale Alignment: Asymmetric self-supervised alignment across fine and coarse temporal scales filters transient noise and isolates stable causal dynamics.
• Dual Causal Scoring: Combines a Dynamic Causal Score (prediction residuals) and a Relational Causal Score (gradient-based Granger matrix) via an asymmetric gating mechanism to suppress false alarms.