Classification of Bolt Corrosion Levels Combining Deep Learning and Multi-Feature Segmentation

Abstract

Many bolts are installed in subway tunnels, making manual inspection prohibitively costly, and deep learning models face difficulties in segmenting extremely small corroded regions, which results in low detection efficiency. To address these challenges, this study proposes a corrosion grade classification algorithm for subway tunnel bolts based on deep learning and multi-feature segmentation, which directly outputs the corrosion grade of each bolt to enhance maintenance efficiency. First, the YOLOv8 framework is improved using multi-scale channel group shuffle convolution (MSCGSC) and focal loss (FL) to develop the YOLO-MF (MSCGSC + FL) model for preliminary detection of corroded bolts. Second, the VGG16 network is employed as the backbone of U-Net, and channel shuffle is applied after the encoder–decoder concatenation to eliminate background noise of bolts using the VGG + channel shuffle (VCS)-Net model. Finally, the fusion of segmentation features, spatial features, and clustering features enables the accurate segmentation and grading of tiny corroded areas. Experiment results demonstrate that YOLO-MF and VCS-Net achieve higher accuracy in corroded-bolt detection and background noise removal. Compared with other segmentation approaches, the multi-feature fusion segmentation method improves the intersection over union by 0.1623. The corrosion grade results are directly printed on the images, facilitating maintenance operations, reducing the workload of tunnel maintenance personnel, and improving tunnel maintenance efficiency.