CAAD detects anomalies via multi-scale alignment and structural causal consistency

The authors propose CAAD, a framework for multivariate time series anomaly detection that reframes the problem as verifying Granger causality consistency through exogenous variables.

• Models exogenous time-series variables as residuals to identify anomalies caused by external interventions.
• Uses multi-scale alignment to internalize system dynamics and a gradient-based matrix to monitor causal relationship breakdowns.
• Quantifies causal deviations in both dynamic evolution and relational topology to capture subtle shifts.
• Outperforms most state-of-the-art baselines on real-world industrial datasets.

CAAD achieves high-precision anomaly detection by addressing the disruption of internal causal relationships that existing methods often overlook.