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

What is it about?

Industrial systems like power grids and water treatment plants rely on dozens of interconnected sensors. Traditional anomaly detection watches for numbers that spike or drop unexpectedly. But real failures often start differently: the normal cause-and-effect relationships between sensors break down first, while the individual readings still look perfectly normal. For example, a valve opening should predictably raise water levels, but during certain failures this relationship silently breaks, even though the water level itself stays within a normal range. Our method, CAAD, detects anomalies by continuously checking whether these cause-and-effect relationships remain intact, rather than just watching for unusual numbers. It combines two complementary signals, one tracking numerical deviations and one tracking structural changes in how variables influence each other, then fuses them in a way that lets the more reliable signal correct the noisier one. On real-world industrial data, this approach catches subtle anomalies that numbers-only methods miss entirely, while also reducing false alarms after a system recovers from an incident.

Featured Image

Photo by Valyri Bosserman on Unsplash

Photo by Valyri Bosserman on Unsplash