| Under the background of increasingly severe environmental problems,clean energy has become an essential direction for energy development.Nuclear power has the advantages of being clean,efficient,economical and dispatchable,and plays an important role in optimizing the energy structure,promoting energy conservation and emission reduction,and achieving sustainable development.However,the safe operation of nuclear reactors is fundamental.Nuclear fuel pellets are core parts of nuclear reactor fuel assemblies.Defects on the pellet surfaces will affect the reactor operation efficiency and even threaten the reactor safety,causing safety accidents.Therefore,surface defects need to be strictly detected before putting into use.Surface cracks are typical manifestations of pellet defects,which are usually subtle and hidden against a complex background,making them difficult to achieve accurate detection.Thus,it is of great value to study a crack detection algorithm with high accuracy for the characteristics of surface cracks on pellets.In recent years,with the continuous promotion of industrial automation and intelligence,machine vision technology has been developed rapidly,and crack detection methods based on machine vision technology have been optimized.However,due to the specificity of nuclear fuel pellets,the crack image data are complex and variable,so it is difficult to achieve ideal detection results using traditional image processing methods.In contrast,deep learning methods,which have the advantages of autonomous learning and the ability to generate abstract features,have attracted much attention in the field of product appearance quality detection,and many scholars have studied them in depth.Thus,the fully convolutional network architecture with end-toend training and pixel-level segmentation was chosen in this study,and optimized to achieve accurate detection of crack regions.The main research contents of this paper are as follows.(1)The pellet surface crack image dataset was established.Firstly,the image acquisition system was built for the cylindrical surface features of the pellet,using a high-precision,high-speed laser profiler and a double roller device with controlled speed and stable operation to realize image acquisition.Then,the acquired images were pre-processed and expanded to obtain 3200 crack images.Finally,the corresponding labels were made by manual labeling.This dataset was used for training and performance testing of crack detection networks in this study.(2)The effectiveness validation of the fully convolutional network for the detection of crack images and the determination of the basic network architecture.Three classical fully convolutional network architectures(FCN-VGG16,FCNGoog Le Net and FCN-Res Net50)were built,trained and tested respectively in this study.The experimental results showed that the F1-Score and the intersection over union of the above networks were higher than 80%,and the crack regions could be detected basically.Thus,the feasibility for the detection of pellet surface crack images using fully convolutional networks was verified,and the best-performing network among them,FCN-VGG16,was selected as the basic network architecture for this study,providing the basis for the optimization of subsequent crack detection networks.(3)The optimization of the fully convolutional network and the proposal of the FPCDNet architecture in this paper.Firstly,the shortcomings of FCN-VGG16 were analyzed,and corresponding improvement schemes were proposed from two aspects of the encoder structure and the feature fusion approach.In the encoder part,first,the image detail information was retained by modifying the numbers of convolution kernels in some convolution layers and reducing the number of down-sampling.Second,the pyramid pooling module was introduced to capture multi-scale feature information,increasing the receptive field of the network and enabling the network to make full use of the contextual information of different regions.Third,on the basis of the second point,the channel attention mechanism was introduced to improve the pyramid pooling module to realize the automatic evaluation of the importance of features at each scale.In the feature fusion part,first,the numbers of channels of feature maps were increased in the upsampling process to enrich the feature information used in the fusion process.Second,on the basis of the first point,the concat fusion approach was used instead of the add,allowing the network to autonomously decide the importance of each part of the fused features.The feasibility of the above improvement schemes was verified through experiments,and then an improved fully convolutional network architecture,FPCDNet,was proposed by combining the advantages of multiple improvement schemes.The experimental results showed that FPCDNet presented the best results in terms of evaluation metrics,with the precision rate,recall rate,F1-Score and intersection over union of 95.01%,94.86%,94.75% and 90.30%,respectively.Compared with FCN-VGG16,each metric increased by 2.39%,3.09%,2.95% and5.01%,with overall improvement in network performance. |
PDF Full Text Request