# SHM-Traffic: DRL and Transfer learning based UAV Control for Structural Health Monitoring of Bridges with Traffic

**Authors:** Divija Swetha Gadiraju, Saeed Eftekhar Azam, Deepak Khazanchi  
**arXiv:** [2402.14757](https://arxiv.org/abs/2402.14757)  
**Date:** 2024-02-22  
**Summary page:** https://www.summarizepaper.com/en/paper/2402.14757v1/

## Key points

- Research work focuses on enhancing structural health monitoring (SHM) in bridges with ongoing traffic
- Utilization of deep reinforcement learning (DRL)-based control for UAV surveys of concrete bridge decks to detect cracks
- Challenge of detecting cracks while traffic is present and initial crack locations are unknown
- Use of canny edge detection method and Convolutional Neural Network (CNN) for crack detection
- Implementation of transfer learning using CNN with pre-trained weights for improved crack identification and localization
- Application of Proximal Policy Optimization (PPO) for UAV control during bridge surveys
- Evaluation of proposed methodology through various scenarios considering task completion time and reward convergence metrics
- Findings show Canny edge detector offers 40% lower task completion time, while CNN provides 12% better damage detection and 1.8 times better rewards
- System model simulates a bridge environment where UAV operates in two-dimensional space for crack detection amidst traffic
- Illustration of workflows using Canny Edge Detector technique and Transfer Learning based CNN for crack detection tasks

## Summary

The research work titled "SHM-Traffic: DRL and Transfer learning based UAV Control for Structural Health Monitoring of Bridges with Traffic" focuses on enhancing structural health monitoring (SHM) practices in bridges with ongoing traffic through the utilization of advanced techniques. The proposed approach involves employing deep reinforcement learning (DRL)-based control for Unmanned Aerial Vehicle (UAV) to conduct surveys of concrete bridge decks and detect cracks. The challenge lies in detecting cracks while traffic is present, with the initial location of cracks being unknown. Two edge detection techniques are utilized in this study - the canny edge detection method and a Convolutional Neural Network (CNN). Transfer learning is implemented using CNN with pre-trained weights from a crack image dataset, allowing the model to adapt and enhance its performance in identifying and localizing cracks. Proximal Policy Optimization (PPO) is applied for UAV control during bridge surveys. Various scenarios are experimented upon to evaluate the effectiveness of the proposed methodology, considering key metrics such as task completion time and reward convergence. The study findings indicate that the Canny edge detector offers up to 40% lower task completion time, while the CNN excels in up to 12% better damage detection and provides 1.8 times better rewards. The system model considers a bridge as a simulation environment where a UAV operates within a two-dimensional space for crack detection amidst ongoing traffic. The UAV surveys the bridge deck, captures crack images along with their locations, and transmits them back to base for analysis. The simulation environment replicates the structural layout of a bridge, offering a two-dimensional representation accommodating varying traffic densities. Two workflows are illustrated - one depicting the Canny Edge Detector technique resulting in highlighted edges against a black background, and another showcasing Transfer Learning based CNN where pre-trained models are adapted for specialized tasks like crack detection. Overall, this research contributes valuable insights into enhancing SHM practices in bridges with ongoing traffic through innovative approaches combining DRL-based UAV control, transfer learning techniques, and advanced image processing methods like CNNs for efficient crack detection and localization.