A Deep Reinforcement Learning based Approach for Bridge Health Maintenance

This work proposes an DRL based solution to the bridge health monitoring problem to determine cost-effective and hazard-reducing repair methods for deteriorated bridges. There is a need for efficient maintenance schedules that utilize the available offline data. To address this, a deep reinforcement learning (DRL)-based model is introduced to enhance Nebraska bridge maintenance throughout the bridge life-cycle. The DRL agent utilizes the provided offline data to predict optimal maintenance activities. By considering budget limitations, the DRL algorithm generates an optimized maintenance plan, aiming for maximum cost-effectiveness. This approach incorporates structural degradation and the impact of maintenance operations over time by employing probabilistic models to simulate the stochastic process. We propose an algorithm using reinforcement learning for bridge maintenance. We leverage Double Deep Q-Learning Network with Dueling Architecture called D3QN for our DRL approach. The results show that the best maintenance techniques learned are within the specified budget limits and maximize the life-cycle cost-effectiveness of maintenance operations. Furthermore, the proposed D3QN outperforms traditional techniques like Dueling Deep Q networks (DDQN) and heuristic algorithms. D3QN can achieve faster reward convergence and can achieve 75% better life cycle cost utilization compared to the heuristic algorithm.