Real-Time Structural Health Monitoring of Bridges Using Deep Reinforcement Learning and IoT Sensors

Abstract

Bridge mechanical failures are incredibly dangerous, and by 2023, there were more than 10,000 structurally deficient U.S. bridges, a class that led to the 2021 Pittsburgh bridge collapse, resulting in the death of five people. Conventional structural health monitoring (SHM) is usually based on constant thresholds and periodic checks, which are likely to miss progressive damage due to dynamic loads, variability of weather, and traffic, leading to the generation of late warnings and losses worth more than 20 billion a year across the globe. This paper will present a new real-time SHM system of bridges that will combine Deep Reinforcement Learning (DRL) with the usage of IoT sensor networks and will be based on the Vänersborg Bridge dataset (DOI: 10.5281/zenodo.8300495) to facilitate adaptive anomaly detection. The Vanersborg dataset, which was taken in 2023 on a Swedish railway bascule bridge, has 64 real-world events at 200 Hz, including accelerations, strains, inclinations, temperature, and wind data- capturing known truss fracture progression before and after the March 2023 damage. The IoT-DRL system has an architecture that consolidates multi-modal sensor streams to Markov Decision Process states, with a Proximal Policy Optimization (PPO) agent responding to actions by optimizing the rewards associated with anomaly precision and disturbing sensor priorities and latency. Experiments on the dataset demonstrate superior performance: the DRL model achieves 95% F1-score, 0.7s mean latency, and detects damage 12 events earlier than LSTM baselines (89% F1, 2.1s latency), adapting 20% better to fracture-induced strain shifts. This outperforms prior deep learning approaches by enabling autonomous policy evolution without retraining, reducing false positives by 25% amid environmental noise. Some of the contributions are the initial use of DRR to publicly available bridge fracture data, megascale edge-cloud pipeline, inference under a second, and 8 or more sensor channel benchmarks. This framework enhances proactive maintenance, potentially averting failures and saving billions in repairs—increasing bridge lifespan by 15-30% per infrastructure reports. The further extension aims at multi-bridge federated learning.