An Interpretable Ensemble Framework for Nonrecurrent Traffic Labeling

Abstract:

Nonrecurrent traffic disturbances such as crashes, work zones, and irregular demand surges can rapidly disrupt freeway operations, yet they remain difficult to label reliably in real time from noisy and incomplete information. In practice, agencies often rely on incident reports, which may be delayed, missing, duplicated, or weakly aligned with the onset and spatial footprint of measurable traffic impacts. This paper proposes a real time, traffic first framework that labels nonrecurrent conditions directly from speed measurements using an ensemble of simple, interpretable detectors. The ensemble combines a robust distribution based deviation statistic, a short horizon speed gradient, and a spatial slowdown indicator comparing each link to upstream conditions. Candidate anomalies are confirmed through persistence logic, with incident reports used only as optional supporting evidence by reducing the required persistence, and episodes are delimited using link specific speed confirmation and recovery rules. Synthetic speed trajectories are used as controlled design tests to illustrate three intended behaviors: suppressing false alarms from noisy incident records, labeling disturbances that occur without any report, and flagging emerging disruptions during early degradation. The resulting labels provide a transparent and transferable intermediate layer between raw speed feeds and downstream tasks including nonrecurrent aware prediction, reliability analysis, performance monitoring, traffic management decision support, and safety oriented analytics such as speed management and disturbance screening.