Planning maintenance and rehabilitation activities for airport pavements: A combined supervised machine learning and reinforcement learning approach

Maintenance and Rehabilitation (M & R) of airport pavement assets involves considerable financial resources. As such, even modest improvement in M & R activity planning could lead to non-trivial savings. The state-of-the-practice for planning M & R activities mostly relies on condition thresholds and engineering rules, while the state-of-the-art often requires untested assumptions and relies critically on complete knowledge about pavement deterioration. This study proposes a machine learning (ML)-based approach that integrates pavement condition prediction using supervised ML with M & R activity planning empowered by reinforcement learning (RL), for which the Q-learning method is adopted. Supervised ML involves minimum a priori assumptions while characterizing the relationship between pavement deterioration and its influencing factors, and can yield higheraccuracy prediction of future pavement conditions than traditional methods. Similarly, RL learns the optimal action-value functions in a model-free environment. The integrated ML approach is implemented and demonstrated using real-world data from Chicago O'Hare International Airport. The results show the effectiveness of the proposed approach which can lead to reduced M & R cost compared to practice.& COPY; 2021 Tongji University and Tongji University Press. Publishing Services by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/ licenses/by-nc-nd/4.0/).

Automated summary

This study proposes a machine learning and reinforcement learning-based approach for maintenance and rehabilitation (M & R) activity planning of airport pavement assets. The supervised machine learning method is used for pavement condition prediction, which can predict future pavement conditions more accurately without relying on a lot of prior assumptions. Similarly, reinforcement learning learns optimal action-value functions in a model-free environment. The application of this approach using real-world data demonstrates its effectiveness in reducing M & R costs.