LatentFlow: Discovering Latent Continuous Dynamics across Channels for Multivariate Time Series Anomaly Detection

Effectively modeling the complex and evolving dependencies among multiple variables is a key challenge in multivariate time series anomaly detection (MTSAD). Existing methods typically model channel dependencies in discrete time steps, either window-wise or point-wise. However, they face a granularity dilemma: window-wise approaches are too coarse to capture transient local changes, whereas fine-grained methods lack constraints on dependency continuity, making them susceptible to high-frequency noise and leading to dependency oscillation. Furthermore, since self-channel correlations typically dominate cross-channel signals, existing methods are biased towards self-dependencies, often overlooking subtle cross-channel deviations that indicate anomalies. In this paper, we propose LatentFlow, a novel framework that treats channel dependency evolution as a latent continuous dynamic process. Specifically, we model the evolution of channel dependencies using an Ornstein-Uhlenbeck (O-U) process. This introduces a mean-reverting property and structural inertia, allowing the model to capture smooth dependency shifts while maintaining robustness against structural noise. Additionally, we introduce a Dependency Decoupling Strategy to explicitly separate and rebalance self- and cross-channel patterns. Extensive experiments on multiple real-world datasets demonstrate that LatentFlow achieves state-of-the-art performance, validating the effectiveness of modeling the continuous dynamics of dependency evolution.