Bridge Defect Detection via CNNs, Multimodal Data Fusion, and Predictive Maintenance Models

Abstract

Bridge structures are affected by weathering, load, and natural disasters over long-term use, leading to defects such as cracks, corrosion, and deformation. These defects directly threaten the safety and stability of bridges. Traditional inspection methods rely on manual processes, which are costly, inefficient, and prone to subjective errors. This paper proposes a new bridge defect detection method combining machine vision with big data analytics to improve accuracy and efficiency. High-resolution cameras and drones collect image data, which are analyzed using a VGG-based Convolutional Neural Network (CNN) with attention mechanisms to automatically detect and classify defects like cracks, spalling, and corrosion. The multimodal fusion strategy, utilizing graph neural networks, integrates image data with historical detection records, environmental conditions, and load information to predict the service life and maintenance needs of bridges. In experiments, this method achieves an accuracy of 95.8%, an F1 score of 93.5%, and a mean absolute error of ±10%. The accuracy of this method is approximately 30% higher than traditional techniques, and the data analysis reveals that bridges in high-humidity areas experience 23% higher corrosion rates than those in dry conditions. These results confirm the potential of the method for predictive maintenance and enhanced decision support in bridge management.