Physics Informed Deep Learning: Applications in Transportation

Development in deep learning (DL) neural networks [1] [2] benefits a wide range of engineering applications. The learning capability of a DL neural network helps practitioners in numerous fields such as transportation engineering and has been widely adopted in projects on object detection, autonomous driving, and estimations of system conditions. Traffic state estimation (TSE) is a crucial task for transportation planners in understanding travel demand and road infrastructure's level of service (LOS). Due to the cost constraints associated with installing sensing devices along freeways and arterial roads, traffic observations can solely be obtained at predetermined locations, leaving areas where traffic conditions are unperceived. Traffic states such as vehicle speed v, density ρ, and flow f in the unobserved regions need to be approximated by using the collected measurements of traffic at sparse locations [3]. Take loop detectors as an example: the signal indicating the passage of a vehicle can only be obtained at predetermined locations where the electrically conducting loops are planted. The number of detectors deployed considerably affects the quantity of traffic data collected from a highway system [4]. The task of TSE is further impeded by issues such as the measurement noise in detectors and data loss due to sensor malfunctions [5] [6] [7]. The inaccuracy in recorded traffic data and the limited data resolution during signal processing contribute to the challenges in precise TSE [8] [9]. Figure 1 illustrates the process of traffic state data acquisition: sensing devices collect information such as speed v and headway T at designated locations and broadcast the information to central cloud infrastructure for data processing and storage.