Detector Data AI

Project summary:

MnDOT operates an extensive network of traffic detectors across the Twin Cities metro freeway system. The traffic detector data collected from metro freeways provides critical real-time traffic data for congestion management, incident detection, and traveler information systems, as well as historical data for understanding traffic patterns, calculating travel time reliability, and supporting research and policy decisions. The accuracy and reliability of traffic detector data are critical for these applications, as faulty or inconsistent data can lead to incorrect conclusions, ineffective traffic management, and misinformed infrastructure planning. Given the large size and complexity of the traffic-detector network, manual monitoring and maintenance of data quality are both time-consuming and inefficient. Individual sensor failures, calibration issues, or environmental disturbances (e.g., snow- covered detectors) can degrade data quality over time. Without a systematic and automated approach, faulty sensors may go unnoticed for long periods, leading to persistent errors in operational traffic monitoring and analytics. This project will deliver MnDOT a practical, scalable, and ready-to-use fault detection toolkit that automates the identification and prioritization of traffic detector issues. The toolkit will combine straightforward, rule-based verifications with advanced, AI-based data-driven analytics. Rule-based verifications will capture obvious issues such as flatlines, negative values, or inconsistencies in flow between merge/diverge points. Data-driven analytics will learn normal traffic patterns and identify harder-to-spot issues, such as gradual or context-dependent undercounting and overcounting, that evade simple thresholds. The final deliverable will be a validated prototype toolkit ready for use on the pilot corridor. The toolkit will feature an interactive corridor map that displays severity levels for each detector station along with sensor status and detection confidence, a time-series viewer that overlays raw detector data with results from rule-based verifications and AI-predicted baselines for easy comparison, and a node-level view that presents Virtual Balance Sensor (VBS) results at merges and diverges to highlight flow inconsistencies and pinpoint potential fault locations.