Research On Detection Algorithm Of End-face Defect Of Nuclear Fuel Pellet Based On Transfer Learning

As a clean energy with low carbon and environmental protection,nuclear energy is of great significance in China’s future energy structure transformation.The main source of nuclear energy is the thermal energy generated by the fission of nuclear fuel pellets,and there may be various types of defects on the surface of nuclear fuel pellets due to factors such as materials and craftsmanship.Defects can be divided into two categories: defects caused by large area falling blocks and cracks caused by extrusion,and among the various types of surface defects,the end face defect is the most harmful.For the detection of nuclear fuel pellet defects,due to the particularity of the nuclear fuel field,the commonly used method is to use the traditional image processing operator based on fixed threshold,however,it is not satisfactory in dealing with the complex and variable defect characteristics.The defect detection method based on deep learning technology is a new method which is expected to solve the above difficulties.The method copes with changeable defect morphology in the training stage by using a large number of raw data with balanced class distribution.However,the yield rate of the nuclear fuel pellets preparation process far exceeds the defective rate,there are inherent difficulties in obtaining large numbers of defective samples.In view of the existing problems in the field of nuclear fuel pellet appearance detection,a domain adaptive end face defect detection scheme based on transfer learning scheme is proposed to quickly complete the detection task in the case of lack of special case samples or sample labels.The main research work of this topic is as follows:Part I: Analyze the detection task,build the image acquisition system,and establish the migration data set.First,analyze the shape and structure characteristics of the pellet,and clarify the detection task.At the laboratory environment,set up image acquisition system to acquire different materials images of nuclear fuel pellets end face defects,and build transfer learning data set.Part II: Design and verify the defect detection network Defect Net.Firstly,analyze the relevant features of the original UNet structure and its advantages in this field Use Coordinate Attention,group normalization and other feature modules to improve problems such as weak recovery ability of UNet structural features and enhance its generalization performance,in preparation for the next step of transfer the network.The analysis revealed show that the F-score of using Defect Net on the source domain data set reaches 91.52%,and the ablation study shows that the characteristic module of feature recovery does not bring significant improvement to the network parameters.Part III: Design different transfer schemes and verify their performance.Firstly,compare the advantages and disadvantages of different transfer method,and then select domain adaptive scheme to solve the target domain problem.Based on Defect Net,multi domain adaptive layer are added to learn the difference transformation characteristics between the source domain and the target domain,add Covariance matrix with different weights to act as adaptive layer learning targets,and compare the effect of the number of target domain labels involved in the training on the network performance in the experiment.The experimental results demonstrate that,the F-score index reaches 97.23% when the network stops with 3.3% real labels are used to participate in the network training.Build a simulation detection platform to verify the online detection performance,Ada Defect Net finally achieved an average detection accuracy of 97.54%,The results also meet the needs of practical applications.This project is dedicated to the application of transfer learning technology to the field of end defect detection of nuclear fuel pellets.Using domain adaptive transfer to solve the related problems.The results demonstrate that: compared with the traditional deep learning algorithm,our method can accurately detect the end face defects of pellets with a very small number of real labels.Our scheme can reduce the requirements of network training,accelerate the implementation of deep learning technology,and propose a new solution to achieve the detection of nuclear fuel pellet.