Research on Intelligent Diagnosis of Corrosion in the Operation and Maintenance Stage of Steel Structure Engineering Based on U-Net Attention

Abstract

Intelligent corrosion diagnosis plays a crucial role in enhancing the efficiency of operation and maintenance for steel structures. Presently, corrosion detection primarily depends on manual visual inspections and non-destructive testing methods, which are inefficient, costly, and subject to human bias. While machine vision has demonstrated significant potential in controlled laboratory settings, most studies have focused on environments with limited background interference, restricting their practical applicability. To tackle the challenges posed by complex backgrounds and multiple interference factors in field-collected images of steel components, this study introduces an intelligent corrosion grading method designed specifically for images containing background elements. By integrating an attention mechanism into the traditional U-Net network, we achieve precise segmentation of component pixels from background pixels in engineering images, attaining an accuracy of up to 94.1%. The proposed framework is validated using images collected from actual engineering sites. A sliding window sampling technique divides on-site images into several rectangular windows, which are filtered based on U-Net Attention segmentation results. Leveraging a dataset of steel plate corrosion images with known grades, we train an Inception v3 corrosion classification model. Transfer learning techniques are then applied to determine the corrosion grade of each filtered window, culminating in a weighted average to estimate the overall corrosion grade of the target component. This study provides a quantitative index for assessing large-scale steel structure corrosion, significantly impacting the improvement of construction and maintenance quality while laying a solid foundation for further research and development in related fields.