Reinforcement Learning-based Bridge Inspection Management

Abstract

Biannual inspections are required to assess the physical and functional condition of our nation’s bridges. The Federal Highway Administration (FHWA) and various Departments of Transportation (DOT) in the United States periodically update specifications and techniques to normalize and advance the bridge inspection procedures. However, some ambiguity remains in these inspection requirements. One relevant example relates to inspection intervals and techniques. Currently, FHWA requires routine bridge inspection at least every two years, and if necessary, inspectors can adjust the inspection frequency. The details of how one would adjust the inspection frequency is not specified. And while many advanced techniques, e.g., ultrasonic surface wave and AI-based image inspection methods, can be applied to inspect bridges, the rationale to use these techniques relies on bridge inspectors’ experience. This study focuses on developing a reinforcement learning-based method to assist inspectors in managing bridge inspection planning. In this method, a reinforcement learning algorithm is utilized to optimize the frequency of inspection and the selection of the inspection method. A physics-based damage development model is utilized to simulate the deterioration process of the bridge. The reward function designed in the reinforcement learning process considers both economic cost and inspection plan risk. After training, the reinforcement learning agent can rapidly determine an optimal bridge inspection policy based on a bridge’s state, which can minimize both the cost and the risk of bridge inspection work. Thus, inspectors can refer to this agent to make a specific inspection plan for each bridge based on a bridge’s design, history, and features.