Online Defect Detection Of Support Plastic Caps In Automotive Rear Wiper Assembly

In recent years,the domestic car production has been growing,in the car market,cars occupy more than 85%,SUV occupy more than 60% of the market share,and the rear wiper is the standard equipment of cars,SUVs,MPVs and other cars.Support plastic cap is an important component of the rear wiper transmission device,this thin-walled,small parts in the automotive rear wiper gear linkage mechanism automatic assembly there are cracks,breakage,leakage and other types of defects,for this type of small parts of multiple defects in the assembly line mostly rely on manual full inspection or simple gray-scale calculation based on the machine vision way online detection,this detection method is low efficiency,high leakage rate,it is easy to This inspection method is inefficient and has a high leakage rate,which can easily cause a large quality hazard by flowing out defective products.Therefore,this paper selects the plastic cap in the assembly process and designs and implements an online defect detection system for plastic caps using deep learning target detection technology and 3D reconstruction technology.The main work of this paper is as follows.（1）Firstly,by extracting the plastic cap defect images from the industrial production environment at the assembly line site,the plastic cap images are tested and analyzed to design a processing flow that is consistent with the detection of plastic cap defects.Through the study,it is found that there are some obvious surface defects as well as internal invisible cracks caused by press-fitting.In order to improve the number of training samples,this paper uses Wasserstein generative adversarial network under the premise of obtaining real defect samples,and expands the defect detection data set to 40,000 in order to meet the later YOLO＿v3 network training defect samples,in view of the objective problem that the defect samples in the production are small.The number of defect samples is expanded to 40,000 to meet the demand of the later YOLO＿v3 network training.（2）To address the problem that the a priori frames used in YOLO＿v3 are not ideal for detection,the actual bounding boxes in the training set are clustered again using the k-means algorithm,and the candidate frames with objectivity and representativeness are selected based on the clustering results.In the training of YOLO＿v3 network,the original activation function was improved by using Mish’s method instead to address the shortcomings of the original activation function,so that the performance of the network was better improved.Finally,experiments and analysis were conducted on the basis of the improved YOLO＿v3 network,and the detection effect reached the expectation,and the mean average accuracy m AP increased from 67.49% to 68.64% of the original network,and the detection speed also reached60ms/sheet,and the smaller defects of the improved YOLO＿v3 network were enough to achieve better detection results.（3）For the invisible crack defects that may appear in the inner thin wall of the plastic cap after press-fitting,the first use of finite element analysis technology to simulate the size and outer contour changes of the good product and the plastic cap containing invisible cracks after assembly,and found that the largest change in the outer contour size of both is at the bottom,where the deformation of the good plastic cap is about 0.109 mm,and the plastic cap containing invisible cracks is about 0.133 mm.The minimum deformation is about 0.133 mm,and the dimensional deformation between the two is 0.133mm;based on the change of outer contour,the roundness and outer contour curve segment circle fitting method is used to judge whether the plastic cap contains invisible cracks,in which the roundness of the good plastic cap after assembly is 0.01 mm,the number of outer contour curve segments is 4 to 5,and all can be fitted as a circle,while the plastic cap containing invisible cracks has the minimum The minimum roundness of the plastic cap with hidden cracks is 0.02 mm,and the minimum number of outer profile curve segments is 6,and at least 1 curve segment cannot be fitted as a circle.Then the plastic cap was reconstructed in three dimensions.Since the plastic cap is fixed in actual production,in order to reconstruct the whole structure of the plastic cap,four cameras are chosen to take pictures of the plastic cap in a circular array,and then the difference between the reconstructed plastic cap size and the good plastic cap size is calculated and compared with the deformation size,based on whether the difference between the two is within the allowable range and the change of roundness and the circle of the curve segment after decomposition of the outer contour.The presence of invisible cracks is determined based on whether the difference between the two is within the allowable range and the change of roundness and the circle of the curve segment after the decomposition of the outer profile.（4）Based on the object of the inspection device,the design of the online inspection system for plastic caps was completed.The system included the selection of hardware,the interface design of the system was completed using QT5 software,and the neural network detection algorithm and 3D reconstruction algorithm were encapsulated to make it easy to use in the production environment.After further practical production tests,the system was found to be very effective in detecting the defects that exist in plastic caps,achieving,to a certain extent,the identification of various types of defects and achieving the expected results,verifying the feasibility of the online defect detection system for plastic caps covered by this topic.