Title: Cross-Channel Anomalies in Multivariate Time Series Benchmarks Are Rarely Present

Abstract: Recent advancements in multivariate time series anomaly detection (MT-SAD) have increasingly relied on cross-channel modeling, predicated on the assumption that anomalous patterns often manifest across multiple data channels simultaneously. To test this premise, we applied a per-segment diagnostic framework to eight prominent public benchmarks. This framework categorizes each labeled anomaly by determining whether the deviation stems from a single channel breaking its historical norm, a disruption in the relationships between channels, or a combination of both. Our analysis indicates that cross-channel disruptions never occur in isolation; they are invariably accompanied by univariate deviations within reasonable threshold ranges.

Furthermore, a secondary metric demonstrates that on six out of the eight benchmarks, at least half of the labeled anomaly segments exhibit univariate deviations across 79% to 100% of their timesteps, with three datasets showing this trait in every single timestep. To ensure our diagnostic tool could accurately identify cross-channel structures when they do exist, we generated synthetic datasets consisting of phase-shifted sinusoidal waves with shared noise. We introduced specific corruptions that preserved individual channel distributions while destroying cross-channel correlations. Our framework successfully identified these as pure cross-channel anomalies. On this synthetic data, channel-dependent (CD) models effectively utilized the cross-channel signals, whereas channel-independent (CI) models failed to perform. However, when comparing CI and CD versions of a leading state-of-the-art detector on real-world benchmarks, we found no significant performance advantage for CD modeling. These findings suggest that existing MTSAD benchmarks are ill-equipped to validate the efficacy of cross-channel modeling, highlighting an urgent need for the creation of evaluation sets with greater structural diversity. The code supporting this research is publicly accessible.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC