Computer vision-based assessment method of repairing failures for stop-hole in steel box girders

Abstract

Fatigue cracks and their corresponding repair status are crucial indicators of the health of steel bridges. Moreover, detailed crack identification is essential for intelligent bridge management and maintenance. Hence, this study proposes a method based on visual geometric characteristics to assess the status of stop-holes and quantify crack features. A multi-task deep learning network, Crack-Weld-Network (CWNet), is designed to simultaneously identify fatigue cracks and weld joints. CWNet consists of two branches: a boundary-guided segmentation branch and a weld joint detection branch, both of which embed feature pyramid networks to fuse features. During the training process, task-specific evaluation metrics were designed for different tasks, and the training results were compared with the baseline models to assess performance improvements. The proposed assessment and quantification method was validated through laboratory tests on specimens representing local details of a steel box girder. Experimental results demonstrated that, compared to the baseline models, CWNet improved the mean intersection over union and structural similarity by more than 7% and 9%. Additionally, the line similarity metric for weld joint detection achieved 84.7%. On images of structural specimens, CWNet achieved a segmentation performance exceeding 75% and a weld detection score exceeding 90%. In addition, the proposed crack length quantification method yielded an average relative error of approximately 3.4%, while the accuracy of the stop-hole repair status assessment exceeded 95%.