Research On Detection And Recognition Method Of Rocket Bonding Defects Based On Deep Learning

Due to various uncontrollable factors,bonding defects such as debonding,cracking,and delamination may occur during rocket production.It is of great practical value to accurately detect and recognize the types of defects to ensure product quality and improve production technology.At present,the traditional method of defect detection and recognition is to manually observe the enhanced X-ray images,while the algorithm based on machine learning is to manually extract the defect features and use classifiers to realize the detection and recognition of defects.The above methods all rely on human participation and have certain subjectivity and limitations,which is difficult to guarantee the accuracy of defect detection and recognition.Therefore,deep learning-based bonding defects detection and recognition methods are studied in this paper.This paper focuses on the one-stage object detection model with fast detection speed represented by YOLOv5 s and SSD algorithms.However,the above model has some problems such as incomplete feature extraction and missing small objects,which makes the detection accuracy of the model for defects low.At the same time,with the increase in the rate of good rocket products,defect samples are decreasing and the distribution of defect types is unbalanced,which leads to the tendency of small samples to large samples and even model overfitting phenomenon in the training of deep learning model.To solve the above problems,the main work of this paper is as follows:(1)An improved DCGAN algorithm is proposed.By fine-tuning the structure and loss function of DCGAN,a large number of high-quality defect images are generated based on the original X-ray images.A rocket X-ray defect image data set with a relatively balanced distribution is constructed.(2)The YOLOv5 s defect detection algorithm is improved.The EIOU regression loss function is used to accelerate the convergence speed of the model and improve the recall rate of the model.The CBAM is added after the C3 module to enhance the connection of defect features in channel and space,and improve the accuracy of the model.The experimental results show that compared with the GIOU loss function in the original model,the recall rate is increased by 10.5% after introducing the EIOU loss function,but the precision is decreased.Then,the precision is increased by 2.7% through the CBAM module,which verifies the effectiveness of the improved method.(3)The SSD defect detection algorithm is optimized.To solve the problem that the SSD model is limited by the VGG depth,Dense Net is used to replace the VGG.In view of the fact that shallow feature maps contain image detail information and deep feature maps contain semantic information,a new feature fusion network based on CBAM is designed to improve the expression ability of the model.Aiming at the limitation of the Cross-Entropy Loss function,the Focal Loss function is used to control the proportion of positive and negative training samples,suppress the samples that are easy to be divided,and make the training samples have a better distribution.Experimental results show that the above three optimization methods all improve the defect detection effect of the model to different degrees,but using Dense Net as the backbone network of SSD plays a major role in improving the defect detection performance.(4)The improved YOLOv5 s and SSD algorithms are compared and analyzed from multiple perspectives.The experimental results show that the m AP,accuracy rate,and recall rate of the improved YOLOv5 s algorithm increased to 78.6%,77.2%,and 76%,respectively,and the improved SSD network increased to 75.9%,77.3%,and 75.6%,respectively.The evaluation indexes of the two algorithms both reach more than 75%,achieving the expected goals.In terms of specific defect detection,the improved YOLOv5 s has better detection effects on debonding,cracking,and delamination defects than the improved SSD.In addition,the defect image recognition accuracy of the improved YOLOv5 s is higher than that of the improved SSD algorithm,reaching 97.8%,which basically meets the requirements of the factory.Overall,the improved YOLOv5 s is more suitable for detecting the bonding defects of rocket.