Deep learning-based detection and condition classification of bridge steel bearings

Abstract

Regular inspection of bridge bearings plays a critical role in ensuring bridge safety. Traditional manual visual inspection is labor-intensive, time-consuming, and subjective. In light of these limitations, this study aims to achieve efficient detection of bridge bearings by leveraging advanced deep learning techniques. Two deep learning models, BearDet and BearCla, were proposed to detect bearings from inspection images and classify their conditions, respectively. BearDet demonstrated efficient detection capabilities for bearings of various scales, achieving a recall of 91.4% and an average precision of 81.7%. BearCla reached performance levels of 89.6%, 90.8%, and 90.1% in terms of precision, recall, and F1_score, respectively, in the bearing condition classification test. These outcomes highlight the models' potential to enhance the accuracy and automation of bridge bearing inspection. Future research can improve the model inference efficiency and integrate these automated techniques into existing bridge maintenance practices.