Uncertainty-Gated Mixture Modeling for Anomaly Detection in Human-in-the-Loop Vehicle Systems

Anomaly detection in human-driven vehicle telemetry is complicated by mixed uncertainty: nominal deviations may arise either from stochastic driver behavior or from genuine departures from learned vehicle dynamics. Conventional forecastingbased detectors typically treat both as predictive error, which can produce heavytailed anomaly-score distributions and elevated false-positive rates under unseen driver behavior. To address this limitation, we propose the Uncertainty-Gated Mixture Model (U-GMM), a feature-wise anomaly-scoring framework that combines conditional probabilistic forecasting with marginal plausibility estimation through an uncertaintyaware gating mechanism. The conditional component captures temporal consistency with recent history, while the marginal component evaluates whether an observation remains plausible under the broader nominal feature distribution. The learned gate then uses predictive uncertainty to adaptively balance these two sources of anomaly evidence, reducing undue score inflation in nominally stochastic channels while preserving sensitivity to dynamically inconsistent or globally implausible deviations. Experiments on real-world vehicle telemetry datasets show that the proposed framework improves threshold transfer under unseen-driver evaluation, achieving up to a 2.5× reduction in extreme false-positive rate while maintaining competitive fault detection performance under injected anomalies. These results indicate that reliable anomaly detection in human-in-the-loop systems depends not only on predictive model capacity, but also on uncertainty-aware score construction that distinguishes difficult-to-predict nominal behavior from genuinely abnormal system dynamics.