Condition-Based Maintenance of Highway Bridges Using Q-learning and Considering Component Dependency

A bridge comprises many structural components, most of which are critical to its safety and must be maintained periodically. Finding the optimal maintenance policy for a bridge is challenging as each component has a unique deterioration process, and component failures interdepend. Moreover, the components have specific repair schemes with different effects. Maintenance work may involve various combinations of these repair schemes. In this study, we develop a bridge management approach that recommends repair schemes based on routine inspection results. The proposed method uses Q-learning to determine optimal maintenance decisions. The objective is to minimize the total maintenance and user costs in a given planning horizon. Deterioration models for the structural components are obtained using routine inspection data. Q-learning intelligently discovers available repair schemes while repeatedly simulating possible trajectories of bridge states throughout the planning horizon, and eventually summarizes the optimal repair schemes for a given bridge state. This approach eliminates the need for elaborate modeling of structural degradations and repair effects. The proposed decision-making framework is illustrated using an example steel girder bridge.