Multilevel Monitoring System for Road Networks: Anomaly Detection at the Network and Road-Segment Levels

Abstract

This study introduces a novel multilevel disruption detection method for road networks. The proposed monitoring and disruption detection method can detect disruptions at both the network and road-segment levels simultaneously. The monitoring process begins with a short-term prediction of hourly traffic speed on each road segment of the network using long short-term memory (LSTM) artificial neural networks. The prediction errors on each road segment at each timestep are used as a proxy to detect disruptions. Network-level disruptions are detected using a multivariate cumulative sum (MCUSUM) control chart. Local disruptions at a road-segment level of granularity are detected by decomposing the monitoring statistic of the MCUSUM control chart that follows a quadratic form using the correlation-maximization (corr-max) transformation. The proposed method was applied to the road network of Manhattan in New York City to examine its performance in detecting disruptions caused by Hurricane Sandy in 2012. The outcomes indicated that the proposed method could detect disruptions precisely at both network and road-segment levels. Whereas existing solutions can either monitor the entire network as a whole or focus on one or a limited number of road segments, the proposed method in this study can recognize if the entire network has been disrupted and also can recognize the road segments that are experiencing unusual traffic patterns. The outcomes of this study set the stage for transportation agencies and decision makers to design adaptive traffic management systems using real-time disruption detection at the network and road-segment levels.