Bolt Loosening Identification Based On Computer Vision Technology

Bolts are the most widely used metal material in the connection of steel structures such as railway bridges.During the service period of the bolts,different degrees of loosening damage will occur due to the influence of physical and chemical factors,leading to a reduction in the overall load-bearing capacity of the structure and the existence of great safety hazards.At present,bolt loosening damage detection is mainly carried out by means of regular professional inspections,but this type of detection requires a high level of skill and the accuracy of the results varies from person to person and is time-consuming and costly.The development of a new bolt loosening detection technology is therefore of great significance for the safety of steel structures and the control of maintenance costs.In order to meet the real-time and accuracy of the inspection work as far as possible,and to control the cost of inspection,this paper proposes to use computer vision technology to detect the bolts in which loosening damage occurs,taking the most widely used external hexagonal bolts in actual steel structures as the research object,and the specific research content is arranged as follows:1.Create a bolt image dataset.Investigate the types of bolt materials in actual steel structures,take the smaller size of common bolts and the larger size of high-strength bolts as the main research objects,combine the network retrieval images and field shooting images to make the dataset samples,and expand the data by means of image enhancement to meet the basic needs of subsequent model training.2.Train bolt target detection models.The YOLOv3 target detection algorithm,which is currently the most widely used in industrial embedded devices,is improved by replacing its backbone feature extraction network with a Squeeze Net lightweight network to further reduce model complexity and number of parameters while maintaining model accuracy.The bolt image dataset was partitioned and loaded into a lightweight YOLOv3 network for training of the model,and the bolt target detection model performed well as verified by the test set and network images.3.Design of bolt loosening recognition algorithms.The algorithm is written to automatically capture small target areas in the captured image,eliminating irrelevant scenes and thus minimising the amount of computing.Based on the movement characteristics of the feature points in the image before and after the bolt loosening,the ORB algorithm is used to detect and match the feature points,and the MLESAC algorithm is used to estimate the transformation matrix to identify the angle of the bolt loosening.4.Verify the feasibility and accuracy of the algorithm.Three different types of steel bolted plate elements and various types of bolting materials were prepared for the verification of the algorithm.The bolts are loosened to varying degrees and the camera deflection angle is set to 0°,10°,20° and 30°,and samples are taken for each working condition.The results show that the closer the camera/handset is to the component and the smaller the shooting deflection angle,the smaller the error fluctuations in the algorithm’s recognition of the angle.