Cost-Effective Pavement Friction Management Using Machine Learning

Abstract

To reduce roadway crashes and fatalities due to poor pavement friction, a strategy adopted by most transportation agencies for pavement friction management is to maintain adequate pavement friction based on established thresholds. Although most methods to determine these thresholds are data-driven, empirical judgement is commonly involved; and a method that can maximize the cost-effectiveness of the pavement friction management process is still missing. To address this gap, this manuscript proposes a framework employing deep reinforcement learning to support network-level pavement friction management. The objective of the framework is to choose cost-effective pavement friction management strategies that can maximize long-term benefits brought by future crash reductions while minimizing the costs associated with treatments implemented to address low pavement friction. A case study using actual field data demonstrated that the proposed framework could achieve an 8.37% improvement in network pavement friction, and a 3.91% reduction in crash rates compared with current practice, proving that with the proposed framework, better network friction performance can be achieved than with the current practice.