Research on bolt loosening detection method based on PointNet

Aiming at the demand for bolt loosening detection in rail defects, this paper proposes a detection method based on PointNet. Firstly, the improved YOLOv5s model is used to accurately locate the bolts in the two-dimensional rail image, and then a bolt point cloud dataset is constructed. The point cloud voxel downsampling technology is used to preprocess the data, which effectively reduces data redundancy and improves processing efficiency. Subsequently, the image information is fused with the point cloud data in combination with the traditional machine vision algorithm to extract the region of interest (ROI), providing reliable feature support for the accurate judgment of the loose state of the bolts. Furthermore, this paper introduces the PointNet model, uses the deep learning method to directly process the bolt point cloud data, realizes the binary classification detection of the bolt loose state through global feature learning, and uses the binary classification accuracy as the main performance evaluation indicator. The accuracy of the training set is finally stabilized at 81.6%, and the accuracy of the validation set is 75.5%. The experimental results show that the PointNet model has excellent adaptability and robustness when dealing with bolt loosening detection tasks, especially in complex scenarios.