Abstract
With a large number of cable-stayed bridges being built and applied, these cables need regular inspection and maintenance to ensure traffic safety. Cable inspection robots can effectively replace traditional manual detection, and the rapid identification of defects is very important to improve inspection efficiency. This paper proposes an intelligent robotic system for bridge cable defect identification and positioning. The 360-degree cable surface images are captured simultaneously by four cameras installed around the developed inspection robot. Based on the Mask R-CNN and image processing, the robotic system can quickly identify the surface defects of cables. The deep learning network is mainly composed of multiple layers of convolutional neural networks, which can achieve classification, regression and pixel-level mask generation of cable surface defect objects. The location information of these identified cable defects will be measured and recorded by ultra-wide band (UWB) positioning components. These positioning data are sent to Kalman filter for optimal estimation. Experimental results indicate that the inspection robotic system can successfully and quickly identify cable defects, and the average processing speed of each cable image is about 0.16s. The robot can realize high-precision defect positioning, and the positioning error is within 2cm. The application of the robotic system is conducive to confirming the safety of bridges, improving the efficiency of cable inspection, and providing support for the subsequent maintenance.
Introduction
Cable-stayed bridges are vital infrastructure components, relying on high-strength steel cables to support the deck. Over time, these cables are susceptible to corrosion, fatigue, and environmental degradation. Traditional manual inspection methods are not only labor-intensive and costly but also pose significant safety risks to personnel working at great heights. While various robotic solutions have been proposed, they often struggle with two primary challenges: reliable defect detection under varying environmental conditions and precise positioning on steep, slippery cable surfaces.
This research introduces an integrated robotic system that addresses these limitations through a combination of advanced mechanical design, deep learning-based instance segmentation, and high-precision Ultra-Wide Band (UWB) positioning.
Robotic System Design
The developed inspection robot is designed for stability and adaptability. It features an adjustable frame capable of accommodating cable diameters ranging from 50mm to 230mm. To ensure reliable movement on vertical or inclined cables, the robot utilizes V-shaped driving wheels and an elastic suspension system that maintains constant pressure against the cable surface, minimizing slippage.
The "eyes" of the system consist of four high-resolution cameras positioned symmetrically around the cable. This configuration allows for simultaneous 360-degree coverage of the cable surface, ensuring no defects are missed during the climbing process.
Intelligent Defect Identification
The core of the identification system is a deep learning framework based on Mask R-CNN. Unlike standard object detection models that only provide bounding boxes, Mask R-CNN performs instance segmentation, generating a pixel-level mask for each detected defect.
The network utilizes a ResNet-50 or ResNet-101 backbone combined with a Feature Pyramid Network (FPN) to extract multi-scale features. A Region Proposal Network (RPN) identifies potential defect areas, which are then processed through an RoIAlign layer. The final output includes a classification label, a bounding box, and a high-fidelity mask. This granular level of detail is crucial for assessing the severity of damage, such as the exact area of a rust patch or the length of a crack.
High-Precision Positioning and Data Fusion
To overcome the inaccuracies of traditional wheel encoders—which often fail due to slippage on steep cables—the system employs UWB positioning. Using a Double-sided Two-way Ranging (DS-TWR) algorithm, the system calculates the distance between the robot and base stations by measuring signal travel time, effectively eliminating clock synchronization errors.
To further refine the location data, a Kalman filter is applied. This recursive filter smooths out the inherent noise and jitter in raw UWB signals, providing a stable and accurate estimate of the robot's position. This ensures that any identified defect can be precisely located for subsequent maintenance.
Experimental Results and Discussion
The system was evaluated using a custom dataset of 300 high-resolution (640x480) cable images. Key performance metrics include:
- Identification Accuracy: The model achieved a peak accuracy of 94.9% (0.949) after 700 training steps, with a high probability of correct categorization (>97%).
- Processing Speed: The average inference time per image is approximately 0.16 seconds, enabling real-time inspection during the robot's ascent.
- Positioning Precision: The integration of UWB and Kalman filtering reduced positioning errors to within ±2cm, a significant improvement over raw UWB data which fluctuated by up to ±5cm.
Conclusion
The proposed intelligent robotic system offers a robust and efficient solution for bridge cable inspection. By integrating Mask R-CNN for precise defect segmentation and UWB-Kalman filtering for accurate positioning, the system significantly improves upon existing manual and robotic methods. Future work will focus on expanding the dataset to include more specific defect classifications and enhancing the robot's performance in extreme weather conditions.
Summary updated