Research On Key Technologies For Identifying And Locating The Number Of First And Last Turns Of Wire Based On Deep Learning

Wire is a widely used round bar product,which occupies a high proportion in China’s steel production.The finished wire is delivered in the form of coils,which is also called coil or disc.Due to the problems of irregular size and poor mechanical property at both ends of the wire,in order to ensure the overall quality of the product,it is necessary to cut 3～4 turns at the head and end of the wire coil.At present,steel enterprises generally use manual cutting operations,due to the high temperature on the surface of the rolled wire rod and the complicate number of coils,as a consequence,it is difficult to identify the number of wire coils accurately and cut the wire efficiently in manual operations.Therefore,an improved model based on deep learning,R-YOLOv5 s,is designed to complete the identification and positioning of the number of first and last turns of the wire coil,which provides theoretical support for recognition automation.The research content of this paper is as follows:(1)In order to obtain the model training image data,AT-S1000-01 A 3D camera is used to acquire images of wire in the high-line production workshop of a steel group in Hebei.The acquired images were preprocessed and classified,and a small wire dataset was produced.(2)The deep learning object detection model YOLOv5 s has poor detection effect when the target is dense and oblique distribution,and cannot identify the coil position accurately.In this paper,the R-YOLOv5 s detection model of the rotating frame is designed by defining the rotation frame representation method,the target loss function of the rotation frame,and the discretization of the angle problem,Besides the small wire dataset is trained and predicted in R-YOLOv5 s,and the F1 score and detection speed are then used to evaluate the model,which basically realizes the identification of the inclined wire.(3)According to the recognition effect of R-YOLOv5 s,the performance improvement experiment of the model is carried out,and the improvement of K-means clustering a priori box,multi-scale fusion,attention mechanism,Bi FPN and other technologies is expounded,and finally it is proved that the addition of small target detection head P2 + CBAM attention mechanism + Bi FPN enhances the information transmission between network layers,which is helpful to improve the recognition accuracy of small targets in the network.(4)The improved R-YOLOv5 s recognition results are processed,and the 3D information collected by the 3D camera is used to locate the cutting position by using the neighbor point search.Finally,the robot is used to verify the coordinate position.Through the performance index of R-YOLOv5 s on the test suite,combined with the positioning results of the cutting points in the experimental environment,it is used to test the performance of the improved R-YOLOv5 s on the wire detection task.Finally,the prediction speed of the improved R-YOLOv5 s network on the test set is 230 ms,which is more accurate than the rectangular box recognition of traditional YOLOv5 s.The improved R-YOLOv5 s has an accuracy rate of 92.1%,a recall rate of 89.4%,and an F1 score of 90.7%,which can meet the basic needs of identification and positioning.This test proves the good performance of R-YOLOv5 s in the wire inspection task.