Research On Surface Defect Detection Technology Of Aerospace Sealing Ring Based On Deep Learning

Sealing rings are widely used in the sealing systems of aerospace vehicles,and their surface defects are one of the important reasons for seal failure.Therefore,it is particularly important to ensure the surface quality of aerospace sealing rings.Aiming at the problems of low efficiency of manual detection and poor generality of traditional image processing detection algorithm,this paper studies the surface defect detection method of aerospace sealing ring based on deep learning.By analyzing the cause of the defect and its image characteristics,it is targeted to improve and propose a network based on RetinaNet.The detection algorithm realizes the efficient and accurate detection of surface defects of the sealing ring.The thesis first constructed a high-quality seal ring surface defect image data set,and analyzed the defect characteristics,and found that the aerospace seal ring surface defects have the characteristics of small targets,and the defect area has a concentrated aspect ratio.Based on the above analysis,the deep learning modularization technology suitable for this type of defect features is researched,and the anchor frame parameter setting,feature extraction module selection and loss function design are studied separately.In order to ensure high-precision detection of seal ring surface defects and real-time detection,lightweight models and methods are studied,and based on the lightweight network MoGaA,combined with the network structure characteristics and seal ring surface defect characteristics,the decomposition convolution module is used instead the deep convolution in MoGaA builds the newMoGaA network.The experimental results show that compared with the mainstream backbone networks Vgg16 and Res Net,newMoGaA has a higher detection accuracy,reaching 72.3%,with fewer parameters and calculations.Aiming at the characteristics of most defects with small targets,the RetinaNet network,which is more sensitive to small targets,is selected as the basic architecture of the defect detection algorithm.By introducing MoGaA and newMoGaA networks into RetinaNet,the MoGaA-RetinaNet and newMoGaA-RetinaNet algorithms are constructed.In view of the size characteristics of the defect,the frame regression loss function and anchor frame parameters are improved.The experimental results on the sealing ring surface defect test set show that the newMoGaA-RetinaNet algorithm with improved loss function and anchor frame parameters has the highest detection accuracy,which is higher than the basic network RetinaNet The accuracy of the algorithm is increased by 7% to 94.5%,the detection speed is increased by 55% to 31frame/s,and the amount of network parameters is reduced by 50%.Finally,combined with hardware characteristics and software architecture,the best detection algorithm is programmed and transplanted to the sealing ring surface defect detection equipment,and its performance is tested.The results show that the newMoGaA proposed in this paper improves the loss function and anchor frame parameters.-RetinaNet algorithm can efficiently and accurately detect the surface defects of the sealing ring.