VEMLAN: Imputing missing data for failed freeway traffic sensors

Traffic volume is essential for traffic professionals to maintain the performance of freeways and highways. Freeway volume data are typically collected by traffic sensors, the most common of which is the inductive loop detector. However, loop detectors can fail due to issues such as faulty loop cards, deficient roadbeds, improper installation, or communication loss. Sometimes these failures are intermittent, but in some cases the detector will fail altogether. In the case of a complete traffic sensor failure, it would be beneficial to use a method to impute or estimate missing volume data until the sensor can be repaired. Most studies focus on imputing intermittent missing volume data, but few studies evaluate imputation for a failed sensor. To address this research gap, this study introduces a Volume Estimation Machine Learning Agent Network (VEMLAN) algorithm to estimate volume data from failed traffic sensors using nearby traffic sensor data and crowdsourced data. The VEMLAN algorithm is an online, spatiotemporal volume estimation algorithm that can be used with several different machine learning methods and traffic sensors. Furthermore, it is modular, allowing it to adapt to several different failure network topologies. The VEMLAN algorithm is evaluated in conjunction with five different machine learning models for five different failure network topologies, and a sensitivity analysis of volume data aggregation levels is performed. When using VEMLAN with a dense neural network at an aggregation level of 5 min, a MAPE of 10.4% is achieved. Furthermore, it is observed that crowdsourced data improved the estimation accuracy.