Research On Surface Crack Detection Algorithm Of Nuclear Fuel Pellets Based On Weakly Supervised Learning

Nuclear power is an important way to optimize the energy structure and realize the supply of green energy.As the core component of nuclear reactor fuel assembly,the appearance quality inspection of nuclear fuel pellets is an important step to ensure the safe and efficient operation of the reactor.Cracks are a typical surface quality problem,which are usually very small and have different directions.Traditional image processing algorithms and machine learning algorithms are difficult to achieve accurate crack detection.Fully supervised deep learning crack segmentation algorithms rely on a large number of pixel-level annotations.However,due to the special chip production process and strong professionalism of crack detection,it is difficult to obtain a large number of pixel-level labels.Aiming at this problem,a weakly supervised learning pellet crack detection algorithm was proposed,which can achieve high-precision detection of nuclear fuel surface cracks when only box-level or image-level annotations are provided.The specific research contents are as follows:(1)The construction of the surface crack dataset of the pellet and the selection of the basic segmentation model.According to the surface characteristics of the pellet,a suitable crack acquisition system on the peripheral surface of the pellet is selected,and a crack data set including three annotation levels of pixel-level,box-level and imagelevel is established.Then,three fully supervised segmentation models(FCN,U-NET and Mask R-CNN)with different architectures were compared and tested to verify the feasibility of convolutional neural network to detect the surface cracks in pellets,and the Mask R-CNN model was selected as the basic segmentation model for the subsequent weakly supervised learning.(2)A box-level weakly supervised crack segmentation model based on local fusion segmentation was proposed.To solve the problem of lack of fine target annotation in box-level annotation,the crack foreground was extracted by using Grab Cut and WOV fusion method,and the pixel-level pseudo label of the crack was obtained by using Canny constraint on the foreground edge.Aiming at the problem of false detection in areas such as pores and potholes by Mask R-CNN,The feature extraction network was optimized and the spatial attention module was added to strengthen the local response to the cracked area.At the same time,the network loss function of the head is improved to further strengthen the identification ability of the network to the difficult example region,and reduce the model error detection and missing detection.(3)An image-level weakly supervised crack segmentation model based on constrained multi-scale class activation maps was proposed.On the basis of box-level annotation,the supervision information is further weakened to improve the model generalization ability.Aiming at the problem that the traditional class activation map has low resolution and cannot provide a complete crack foreground,a constrained multi-scale class activation maps pseudo-label generation strategy based on Res Net50 was proposed.The relatively complete crack prospect was obtained by fusing multiscale activation maps,and the Highlight loss was introduced to constrain the target edge activation value of activation maps,so as to alleviate the problem of pseudo label over encirclement.On this basis,two pseudo label generation strategies of constrained multi-scale class activation image and local fusion segmentation are integrated with the ensemble method,and various features of images are fully utilized to obtain highquality pseudo-labels,and then high-precision crack segmentation model training and detection is completed.A series of ablation experiments show that Mask R-CNN in the fully supervised segmentation model obtains 86.04% Io U in the test set,and can segment the main area of the crack.On this basis,the optimization of feature extraction network and loss function to a certain extent compensated for the model uncertainty caused by the pseudo labels generated by the box-level annotation,and finally the Io U of the weakly supervised crack segmentation model based on local fusion segmentation reached81.91%.Further,in the image-level weakly supervision,the constraint multi-scale class activation maps module uses the image middle layer and deep feature map when generating pseudo labels,and the local fusion segmentation uses the shallow feature map.The integration of the two can make full use of the image information.Finally,the ensemble model obtains 82.52% IOU,which is 95.91% of the full supervision model.It verifies the feasibility of weakly supervision crack segmentation based on box-level and image-level annotation.