Nondestructive Analysis Of TRISO Fuel Particles Based On X-ray Phase Contrast Imaging Modalities

The molten salt reactor(MSR)is a kind of generation IV nuclear reactors,fueled by TRISO particles with ceramic coating layers.Traditionally,MSR exists with its fuel in the fluid form.In recent years,researchers in America and China have come to pro-pose a new kind of MSR,whose fuel is in solid state.The solid fuel molten salt reactor has received considerable attention and its fuel elements,based on TRISO(Tristructural isotropic)fuel particles,can be designed as spheres,rods and plates.TRISO particles consist of fuel kernel(oxide or carbide of Uranium and Thorium)and four coating lay-ers which are referred to as buffer layer,IPyC layer,SiC layer and OPyC layer from inside to outside.Structural integrity of coating layers is an important aspect regarding Reactor-safety since it can efficiently prevent radioactive waste from releasing into the outer world.Appropriate thickness and intact inter-structure enable comparatively low failure rate,thereby ensuring safety and good-performance of nuclear reactors.Hence,coating thickness is a key parameter with respect to fuel quality in the preparation of the particle.Currently Ceramography is the most commonly-used method to inspect coating thickness,despite the disadvantages that this method would destroy the inter-structure and operators can only figure out how coating thickness fluctuates according to radius information of one certain cross-section.By contrast,X-ray imaging modalities can overcome such shortcomings by non-destructively detect inter-structures and obtain more comprehensive radius information of the spherical-shape particle by changing ob-serving perspective.On the other hand,up till now few effective detection methods have been found to classify fabrication-caused particle-defects or reactor-operating-induced particle-ruptures.Benefitting from the nature to measure X-ray phase shifts which mainly occur at where edges locate,X-ray phase contrast imaging(PCI)tech-nique can well capture edge features inside the object and this makes it possible to visualize coating-interfaces and ruptures inside TRISO particles.In this study,PCI modality was exploited for two tasks:(1)Coating-thickness of TRISO fuel particles are computed with less human intervention;(2)Defects and ruptures inside the particle are recognized in an automated way using supervised dictionary learning methodsPrevious thickness measurements methods include three steps and they are noise reduction,edge detection and thickness calculation.Nevertheless,they concentrated less on the effectiveness of denoising algorithm and the automation of edge detection was not good enough to reduce human involvements and promote calculation.In this study,some improvements have been made to solve these problems1.Total variation(TV)algorithm was utilized to rule out image noise owing to its sensitiveness to edge features,which allows for preservation of more edges in the denoising process.The denoising results were quantified with CNR(contrast to noise ration),with the aim to optimize parameters of TV algorithm2.Adaptive Canny operator was realized by making two important parameters of Canny algorithm,width of filter and percentage of non-edge points,automatically obtained.Thus the Canny operator can detect thresholds for contour evolution in an automated way.Such automation can not only reduce manual efforts but also improve detecting accuracy.3.Contour center was oriented with least square method.Mean value and stan-dard deviation of coating thickness were finally derived.To investigate whether the proposed method works well,mean thickness of the method was compared with that of the ceremography using the one-way anova analysis.Standard deviation of coat-ing thickness was used to analyze the homogeneity of coating thickness.This method has proved to yield results comparable to Ceremography method and particle with poor curvature would have comparatively higher standard deviationRupture-inspection is machine-learning assisted,which is quite different from pre-vious methods employing finite element modeling for defect analysis.Novelties of this study can be concluded as follows1.Particles on the same image can be separated from each other using Otsu's seg-mentation method which first assumes that an unknown threshold exists for dividing im-age pixels into two classes and later this threshold is derived by computing class variance and class probabilities according to the gray-level histogram,followed by maximizing the between-class variance.Threshold value that maximizes the inter-class variance is the optimized one and this threshold can be used to convert the original image into a binary one.Connection between different pixels on the binary image can be analyzed to locate and split up each particle2.HOG(histogram of oriented gradient algorithm)and LBP-HF(local binary pat-tern histogram Fourier)descriptors were exploited to extract features for each particle HOG extractor is robust to changes in illumination and image contrast and capable of capturing edge details since it uses oriented gradient histogram as image feature vectors LBP-HF extractor is invariant to image rotation and as a modified version of LBP(local binary patterns)descriptor,it describes images with the Fourier transformation of uni-form local binary pattern histogram,which enables better discrimination of pixel and its surroundings,as well as better description of texture information.Canonical correlation analysis method was used to fuse HOG and LBP-HF features,thereby making good use of different descriptors to better distinguish cracked particles3.Supervised dictionary learning(Label consistent K-SVD,LC K-SVD)was fi-nally exploited to learn a dictionary for sparse representation of image feature vectors The LC K-SVD outperforms some other dictionary learning methods since it is capable of sparse coding and automatic classification as a result of modifying objective func-tion with label information and classifier parameters.The mean accuracy can reach up to 83.1%,indicating that such defect-classification have desirable performance.The defect inspection can support failure rate analysis4.This study is aimed to automatically classify different kinds of TRISO particles,including intact ones and those with different defects.Hence,five kinds of particles were employed,involving those with integral structure,those whose kernel layers mi-grate,those missing their OPyC layers,those with their buffer layers inhomogeneous and those cracked ones.Such automatic classification has the benefits of distinguish-ing failed particles from intact ones and recognizing particles with different defects.It can support analyzing failure rate and can also be used for some other similar rupture-recognition.