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

Surface defect detection is a concern in many fields,and effective surface defect detection can control product quality to a large extent.Nuclear fuel pellets are the basic components of nuclear reactor fuel assemblies,which must be intact and free of defects.However,unforeseen defects will inevitably occur during the pressing and sintering process,which will affect the performance of the reactor and even lead to nuclear safety accidents.Therefore,accurate and efficient defect detection is extremely important during the production of nuclear fuel pellets.This subject takes the laboratory simulation nuclear fuel pellet as the object to study the detection algorithm of the end face defect of the nuclear fuel pellet.Due to the limitations of traditional defect detection methods and the dependence of deep learning methods on defect data,they can not well adapt to the special circumstances such as the scarcity of defect samples.Therefore,this study selects the unsupervised algorithm based on image restoration as the surface defect detection algorithm for research.The unsupervised algorithm does not need labeled defect image dataset,and only needs a small number of intact samples to realize the reconstruction training of the network.Thus,the defect detection is realized according to the difference between the reconstructed repaired image and the original image.The main research contents are as follows:(1)The establishment of unsupervised model dataset.The defect detection algorithm research of this subject only focuses on the end face defects of the pellet.Therefore,all the images in the dataset are the end view images of the pellet,and the pellet is displayed as a circular shape with a central hole,and the image format is an 8-bit gray image.The training dataset contains only good pellet images,and the test dataset contains both defect and good pellet images.(2)Research on defect detection algorithm based on autoencoder.An unsupervised surface defect detection network based on autoencoder is proposed.The autoencoder network is used to repair the defective image to the intact image,and then the defect threshold segmentation algorithm is used to locate the defect in the residual image.The autoencoder is composed of two parts: symmetrical encoder and decoder,which realizes feature extraction and reconstruction respectively.By setting the appropriate number of downsampling layers and hidden layer feature codes,the autoencoder network trained with hybrid loss function can extract the efficient feature codes of intact images.After the defect image is reconstructed into a intact image through the autoencoder network,the structural similarity calculation is made with the original image to obtain the structural similarity map with residual property.The defect in the feature map has a large contrast with the background.The defect area can be located through OTSU threshold segmentation.The experimental results show that the unsupervised defect detection model can achieve a certain degree of defect reconstruction,and the threshold segmentation algorithm can achieve high defect detection accuracy.(3)Research on defect detection algorithm based on autoencoder adversarial network.On the basis of the autoencoder network,the adversarial network structure is added to form the autoencoder adversarial network.The autoencoder network is trained in the adversarial training mode of GAN to generate a reconstructed image that is closer to the background of the input image,thereby enhancing the contrast between the defect and the background in the residual map,and making the defect segmentation more accurate.In the autoencoder adversarial network,the discriminator is the same as the encoder,and a loss function in the form of WGAN-GP is used.The experimental results show that the autoencoder adversarial network after redesigning the network structure and loss function generates clearer images and the defect segmentation accuracy is higher than the autoencoder network,and the F1 reaches 0.842,indicating that the proposed defect segmentation network is effective for the defect detection of nuclear fuel pellets.