Defect Detection Of Nuclear Fuel Assembly Based On Deep Learning

Nuclear fuel assembly may be damaged during the production,transportation,service and decommissioning,which will affect the safe and stable operation of the reactor.At present,manual visual inspection is one of the most commonly used methods for the inspection of appearance defects of fuel assemblies,which is a time-consuming,tedious and subjective process.In recent years,Deep Learning(DL)has been a research hotspot in the field of machine Learning,which is widely used in the fields of image recognition,speech recognition and natural language processing.This paper proposes a method of nuclear fuel assembly defect detection based on deep learning.With a false positive rate of 0.1 per frame,the true positive rate is as high as 0.98,which can significantly improve the detection efficiency and accuracy,and has certain theoretical significance and practical value for nuclear fuel assembly production and in-service inspection.The main content of the full text is as follows:1、This work focus on the basic theory of deep learning,and summarize the domestic and international research status and latest research result.Several classical deep neural networks and object detection frameworks are introduced.The existing technology of defect detection of nuclear fuel assembly is investigated and its limitations are analyzed,so as to make full preparations for the next step.2、Developing a nuclear fuel assembly defect detection system based on deep learning.For the hardware part,a 15×15 simulated nuclear fuel assembly is built according to the special geometry of the nuclear fuel assembly.To simulate the environment of the fuel assembly,a darkroom and a robotic arm are designed respectively to load the nuclear fuel assembly and a mobile camera.The software part realized the engineering transformation of the algorithm through programming,and successfully completed the detection task with the hardware part.3、In order to obtain the standard defect sample database,the destruction of the nuclear fuel assembly has been carried out randomly,and the required defect pictures are collected through the cooperation of the robotic arm and the camera.the collected defect images were firstly normalized clipped,then the number of defect images was expanded through data enhancement,and finally defect images were annotated.A variety of deep neural network and target recognition framework are combined to train the prepared defect sample database and check using that has not participated in the training.Moreover,the true positive rate and false positive rate of the detection task are counted.At a false positive rate of 0.1 per frame,the true positive rate reached 98%.The experimental results prove the feasibility of this method.