Research On Detection Method Of Train Bolt Looseness Based On Deep Learning And Image Processing

With the continuous development of Chinese railway,the use of trains is increasing,and the workload of maintaining the bolts at the bottom of the trains is also increasing,which puts forward higher requirements for bolt looseness detection.It is necessary to study automatic and intelligent bolt looseness detection methods.In the traditional bolt looseness detection method,the staff check the looseness of bolts by torque test and experience judgment.The method has large workload and low detection efficiency,and it is easy to cause visual fatigue,resulting in false detection and missed detection,leaving hidden dangers in operation.In recent years,various kinds of contact sensors have been developed to detect bolt loosening,the method needs to install a large number of sensors on the structure of bolt connection,and sensors are usually susceptible to complex environment such as temperature,humidity,electromagnetic field.With the rapid development of deep learning object detection technology and digital image processing technology,bolt looseness detection method based on computer vision technology is gradually emerging.Compared with traditional detection methods,the method of image detection based on computer vision technology has the advantages of non-contact measurement,high precision,high efficiency,low cost and small environmental interference.In the thesis,deep learning object detection technology and digital image processing technology are applied to the detection of bolt loosening at the bottom of railway trains.The main contents are as follows :(1)The imaging model of the binocular camera is described,and the parameter calibration,distortion correction and stereo correction of the binocular camera are completed.The feature matching and image fusion of the left and right bolt images collected by the binocular camera are realized,and a higher quality bolt image is obtained.(2)Aiming at the problem of bolt object detection,the network structure and working principle of the latest object detection algorithm YOLOX are preliminarily studied.Based on the YOLOX network,the feature extraction network of the YOLOX-X object detection algorithm is improved to improve the speed,accuracy and stability of object detection.A large number of bolt images under different loose states are collected for training the improved YOLOX-X network,and the optimal bolt object detection model is obtained,the model is used to identify the bolt in a picture.(3)The background and shape of the bolt are preliminarily analyzed,and the feature extraction algorithm of the bolt image is designed.The main features of the bolt are extracted by using the digital image processing technologies such as image enhancement,image segmentation,edge detection and Hough transform.According to the principle of bolt loosening and the difference of bolt state before and after bolt loosening,a bolt loosening discrimination algorithm is proposed,the algorithm is combined with the bolt image features to judge whether the bolt is loose and the degree of looseness within the range of 0?360°.(4)The effectiveness of the bolt looseness detection method is verified by experiments,the experimental results show that the improved YOLOX-X network has better object detection effect.The average accuracy of object detection reached 99.23%,which was11.44% higher than that of the original YOLOX-X.Image processing algorithm and loosening discrimination algorithm can effectively detect whether the bolt is loose and the degree of looseness.(5)Combined with Open CV,Python,QT and Pytorch,the software development of bolt looseness detection system is completed,which integrates deep learning object detection technology and image processing technology.The visual interface is designed to view the detection process and results in real time.