Research On The Detection Method Of End Face Defects Of Nuclear Fuel Pellets Based On Deep Learning

The nuclear fuel pellets are an important component of the nuclear fuel rods that generates energy to power the reactor through nuclear fission.Affected by the raw materials and processing technology,various types of defects may exist on the end face of the pellets.The defects may cause the pellets to interact with the nuclear fuel rod cladding,causing the nuclear fuel rod to rupture and leading to radioactive fission products entering the reactor coolant.This will not only affect the power generation efficiency of the reactor,but also cause serious safety accidents.Therefore,accurate and efficient automatic defect detection of nuclear fuel pellet end faces is required prior to the use of nuclear fuel pellets to ensure safe operation and maintenance of the reactor.In order to realize automatic detection of radioactive nuclear fuel pellets,machine vision technology has replaced manual visual inspection as an important means of detecting defects on the end face of nuclear fuel pellets.However,the pellet end face defect images have the characteristics of low contrast,complex image background,varying defect types and tiny morphology,which easily lead to the defect information being hidden by the complex noise background,which makes it difficult to detect defects in the future.When using the traditional image processing technology to detect pellet end face defects,the artificially set parameters and features cannot well characterize different types of defects due to the interference of noise background,resulting in defect misdetection and low defect detection accuracy.In contrast,deep learning technology learns abstract high-level features from training samples,which is easier to adapt to different types of defects,and has strong versatility and robustness.Based on this,this topic research the method of detection of end face defects of nuclear fuel pellets based on deep learning to achieve the high-quality detection task of the pellet end face defect.（1）In terms of image acquisition,a pellet end face image acquisition system is designed and built to acquire images that can reflect the real defects on the pellet end face and lay the foundation for the subsequent production of pellet defect data set.（2）To address the problem of insufficient training samples in the pellet defect dataset,this topic uses data augmentation technology to expand and optimize the dataset,increase the diversity of the dataset,and enhance the generalization ability of the network to improve the detection accuracy of the network.（3）To address the problem of low defect detection accuracy when detecting pellet end face defect images by traditional image processing technology,a defect detection method with SE-Defect Net convolutional neural network guided WOV（Weighted Object Variance）threshold is proposed.The method first uses a sliding window scanning technology and the SE-Defect Net classification network to locate defect candidate regions in the image.Then,the WOV threshold method is used to extract defects by threshold segmentation of defect candidate regions to achieve the pixel-level segmentation of the defect on the end face of the nuclear fuel pellet.The defect segmentation results show that the proposed method can overcome the problem of uneven grayscale of the defect image on the end face of the pellet,and the pixel-level F₁-measure of the method is about 97%,which can accurately identify the end face defect of the nuclear fuel pellet.（4）To address the problems of defect leakage and false detection and inaccurate defect edge segmentation in the original Mask-RCNN network for detecting pellet end face defects,a defect detection method based on Mask-RCNN optimization is proposed to improve the feature pyramid network and mask branch in the original Mask-RCNN.Firstly,on the basis of the original feature pyramid network,a bottom-up feature propagation path is added by lateral connection to enhance the global feature information and thus improve the defect segmentation accuracy.Secondly,edge detection is added at the output of the mask branch,and the difference between the edge detection results of the predicted mask and the label image is used as the loss function,so that the edge of the defect segmentation result is closer to the real defect edge.The defect segmentation results show that the pixel-level F₁-measure of the method is 93.4%,which is 4%better than the original Mask-RCNN network,and can effectively improve the detection accuracy of pellet end face defects in complex image backgrounds.This topic applies deep learning technology to the detection of end face defects of nuclear fuel pellets,and studies accurate and efficient detection methods for end face defects of nuclear fuel pellets.The experimental results show that compared with traditional image processing technology,the proposed method can accurately detect pellet end face defects and greatly reduce the defect false detection rate,meeting the quality detection requirements of the nuclear fuel pellet production line.At the same time,this method can also be extended to the quality detection of other industrial products.