A Study On The Method Of BIW Solder Joint Detection Based On Deep Learning

In automobile manufacturing,the common connection method for body-in-white is welding,in which spot welding is the most widely used welding process.The quality of the solder joints has a crucial impact on the safety performance of the body-in-white,so the quality of the solder joints needs to be detected in real time during the body-inwhite production process.At present,most of the welding spot quality inspection in the production workshop is manually checked,which is time-consuming and laborintensive and cannot reflect the quality of the body-in-white welding spot in real time.Therefore,an automated welding spot quality inspection method is required.Since the position of the welding spot may deviate from the original welding position in the production process of the body-in-white,we have studied the detection method of the body-in-white solder joints based on computer vision to obtain the position information of the solder joints.The main work of this article is as follows:(1)This paper designs the traditional image processing method to detect the body-in-white solder joints.The main processes include gray-scale image processing,median filtering,Canny edge detection and Hough circle detection.And the actual welding spot images of the production workshop were collected and tested,and the limitations of traditional image processing methods were analyzed.(2)In order to solve the limitations of traditional image processing methods,this paper studies the body-in-white welding point detection method based on deep learning.Taking into account the hardware facilities of the production workshop and the requirements for detection speed and solder joint characteristics,YOLOv2 was selected as the basic solder joint detection algorithm.Collect welding spot photos in the production workshop to establish a welding point detection data set and perform anchor frame clustering through the K-Means algorithm to obtain anchors suitable for bodyin-white welding point detection.After that,the data set was trained and tested and found that using the YOLOv2 algorithm has a good effect.On the test set,the detection accuracy AP@[.5,.95] reached 72.75%,AP@.5 reached 99.27%,and AP@.75 reached 90.51%.The detection time of the pictures of solder joints is 0.0452s.(3)In order to further accelerate the detection speed and improve the detection accuracy,this paper improves the YOLOv2 algorithm through deep separation convolution,feature fusion,GIo U loss,and CIo U loss,and proposes the FCM＿YOLO solder joint detection method.Compared with the or iginal YOLOv2,the detection accuracy of AP@[.5,.95] is improved by 2.69%,AP@.5 is increased by 0.49%,AP @.75 is increased by 4.58%,and the detection speed is the original 2 times.Compared with traditional image processing methods,the FCM＿YOLO solder joint detection method proposed in this paper is more suitable for production workshops with complex lighting environment.