A method and system for generating three-dimensional finite element models and numerical simulation based on XCT digital images of composite materials in water-related engineering projects

Abstract

The present invention discloses a three-dimensional finite element model generation and numerical simulation method and system for XCT digital images of composite materials for water-related engineering projects, belonging to the field of finite element analysis technology. The method includes: collecting composite XCT images in the initial state and preprocessing; performing statistical analysis on two-dimensional microstructural parameters; generating virtual XCT images based on the Pix2PixHD style transfer network; accurately segmenting the real XCT images into component regions based on the Swin‑Tranformer network; extracting and smoothing the contour features of the segmented images based on the mapping relationship between pixels and nodes; constructing a spatial coordinate system and performing discrete derivatives of all yarn center points in each segmented image; performing meshing and inserting global cohesion units based on the multi-material moving cube algorithm in the CGAL library; completing the allocation of material properties and local material directions; and customizing appropriate boundary conditions and performing numerical simulation based on the finite element method.

Description

Technical Field

[0001] The present invention belongs to the technical field of finite element analysis, and in particular relates to a method and system for generating a three-dimensional finite element model and performing a numerical simulation of XCT digital images of composite materials for water-related engineering projects. Background Art

[0002] Basalt fiber reinforced composites (BFRP) are high-performance composite materials whose core reinforcing component is basalt fiber, and are carefully manufactured through advanced molding technologies such as resin transfer molding (RTM). For the construction of infrastructure such as water-related projects, the requirements for materials are extremely strict, and multiple factors such as strength, durability, corrosion resistance and construction convenience must be taken into account. Basalt fiber reinforced composites can not only effectively reduce the weight of the structure and improve the overall bearing capacity, but also maintain stable performance in harsh hydrological environments, reduce maintenance costs and extend service life. Therefore, BFRP has significant application prospects in the hydropower industry, which is conducive to promoting the innovation and development of hydropower engineering technology, providing strong technical support for the green and lightweight transformation of the hydropower industry, and is of great significance to promoting the optimization and upgrading of my country's energy structure.

[0003] X-ray Computed Tomography (XCT) scanning is an X-ray-based nondestructive testing technique that provides high-resolution information about a material's internal structure. XCT scanning can reveal the microstructure of basalt fiber composites. XCT scanning allows for nondestructive characterization of the composite's internal microstructure. This provides an important means of acquiring image data for further characterization of the composite's internal microstructure.

[0004] However, due to the significant anisotropy and heterogeneity of composite materials, their macroscopic mechanical properties testing often requires a large number of experimental samples, complex testing equipment, high testing costs, and long test cycles. Therefore, when studying the deformation and damage of composite materials under external loads, traditional testing methods focus solely on characterizing the average mechanical properties at the macroscopic level, ignoring the mechanical properties and spatial distribution characteristics of the components within the composite material, and failing to establish a cross-scale connection between the micro- and macroscopic mechanical properties.

[0005] Therefore, in recent years, finite element numerical simulation methods have become a common means of studying the mechanical properties of composite materials. By qualitatively and quantitatively characterizing microstructural parameters, constructing a more realistic finite element model, setting material properties and boundary conditions, and simulating the deformation and damage evolution of materials under real experimental conditions, the fracture mechanism on the surface and inside of the material can be observed at a low cost. However, most finite element models of composite materials are often based on idealized or parameterized models, ignoring the heterogeneity of the microscopic characteristic structure within the material. Three-dimensional finite element models that only consider idealized microscopic characteristics are difficult to accurately express the mechanical response characteristics. Summary of the Invention

[0006] In response to the shortcomings of the existing technology, the present invention proposes a three-dimensional finite element model generation and numerical simulation method and system for XCT digital images of water-related engineering composite materials, which can be used to accurately construct a three-dimensional finite element model of the composite material and study its microstructural damage evolution process and rupture state.

[0007] To achieve the above object, the present invention provides the following solutions:

[0008] A method for generating a three-dimensional finite element model and numerical simulation based on XCT digital images of composite materials for water-related engineering projects comprises the following steps:

[0009] S1: Acquire the real XCT image of the basalt fiber composite in the initial state and perform manual segmentation;

[0010] S2: Perform parametric statistical analysis on the manually segmented images to quantify the microstructural parameters of the yarn inside the material and perform two-dimensional parametric random modeling to obtain a parametric random modeling image;

[0011] S3: learning the parameterized random modeling image based on the adversarial learning style transfer network Pix2PixHD to generate a virtual XCT image;

[0012] S4: Based on the Swin-Tranformer network and virtual XCT images, accurate segmentation of each component region of the real XCT image is achieved;

[0013] S5: Based on the segmented real XCT image, the microscopic features of the contour are extracted and the local contours are smoothed to build a yarn and pore feature database;

[0014] S6: Based on the yarn and pore feature database, extract the contour features of each yarn in each segmented real XCT image, calculate the coordinate value of the yarn contour center, establish a spatial coordinate system, and based on the spatial coordinate system, perform discrete derivative of the contour center of each yarn in each segmented image to construct a feature database;

[0015] S7: Based on the feature database, a three-dimensional voxel model is reconstructed, and meshing and cohesive elements are inserted based on the multi-material marching cube algorithm in the CGAL library to simulate the yarn separation and delamination phenomenon under load;

[0016] S8: Based on the divided grid, define the material properties of each component region and assign local material directions;

[0017] S9: Set boundary conditions and perform numerical simulation calculations.

[0018] Preferably, in S3, the method of learning the parameterized random modeling image based on the adversarial learning style transfer network Pix2PixHD to generate a virtual XCT image includes:

[0019] S31: Generate warp yarn random model based on warp yarn parameterized statistical analysis;

[0020] S32: Generate a weft yarn random model based on weft yarn parameterized statistical analysis;

[0021] S33: generating an artificial synthetic label image based on the warp yarn random model and the weft yarn random model;

[0022] S34: Based on the artificially synthesized labeled image, an improved Pix2PixHD network is applied to generate a virtual image of the basalt fiber composite material similar to the real mesostructure, and a hybrid data set, i.e., a real-virtual XCT image, is constructed.

[0023] Preferably, in S4, the method for accurately segmenting each component region of a real XCT image based on the Swin-Tranformer network includes:

[0024] S41: Input the mixed dataset to the Swin-Tranformer network to obtain training images;

[0025] S42: Divide the training image into different small blocks by pixels through a block division module, and then convert it into a one-dimensional feature vector image;

[0026] S43: Accurately segment each component region in the one-dimensional feature vector image to obtain a semantic segmentation result image;

[0027] S44: Based on the semantic segmentation result image, a watershed segmentation algorithm is used to complete regional segmentation for each yarn.

[0028] Preferably, in said S5, based on the segmented real XCT image, the method of extracting contour micro features and smoothing the local contour to construct the yarn and pore feature database includes:

[0029] S51: selecting a resolution that meets preset requirements to downsample the segmented real XCT image;

[0030] S52: Perform feature screening on the real XCT image after downsampling;

[0031] S53: Based on the real XCT image after feature screening, extract the image contour features of the pores and yarns, and convert the coordinate values ​​according to the positional relationship between the pixels and the nodes to obtain the contour coordinate values;

[0032] S54: setting a convolution kernel that meets preset requirements, and achieving contour smoothing by performing an operation between the convolution kernel and the contour coordinate values;

[0033] S55: Process all segmented images according to S51-S54, build a yarn and pore feature database, and store the contour coordinate values ​​of each component after coordinate system transformation.

[0034] Preferably, in S6, based on the yarn and pore feature database, the contour features of each yarn in each segmented image are extracted, the coordinate values ​​of the yarn contour center are calculated, a spatial coordinate system is established, and based on the spatial coordinate system, the method for performing discrete derivation of the contour center of each yarn in each segmented image includes:

[0035] S61: extracting the coordinate value of the center point of each yarn contour in each segmented image from the yarn and pore feature database, and establishing a spatial coordinate system in the order of the segmented images;

[0036] S62: Based on the spatial coordinate system, using the first-order intermediate differential law, discretely derivate the center point of each yarn contour in each segmented image to obtain the tangent direction of the center point, that is, the 1 direction of the yarn;

[0037] S63: Obtain direction 2 from direction 1 obtained by calculation.

[0038] Preferably, in S7, based on the feature database, the three-dimensional voxel model is reconstructed, and the method for meshing based on the multi-material marching cubes algorithm in the CGAL library includes:

[0039] S71: Based on the feature database, the segmented XCT image is imported into Avizo 2019.1 software to complete the three-dimensional voxel model reconstruction and export the inr file;

[0040] S72: Mesh the reconstructed 3D voxel model using the multi-material marching cubes algorithm based on the CGAL library to obtain a mesh file;

[0041] S73: Use a python script to convert the mesh file into an inp file that can be imported into Abaqus.

[0042] Preferably, in S72, the method of meshing the reconstructed three-dimensional voxel model based on the multi-material marching cubes algorithm of the CGAL library to obtain the mesh file includes:

[0043] The reconstructed 3D voxel model is downsampled and the intersection edges of different components are extracted. The boundary features of the intersection edges are used as one-dimensional line features, and the one-dimensional line features are smoothed.

[0044] Delaunay triangulation is performed using the outer surfaces of voxels of different components to generate surface triangular meshes and then re-mesh the surface triangular meshes. The classic Laplacian smoothing algorithm is used to process non-manifold areas to avoid surface self-intersections and sharp dihedral angles.

[0045] One-dimensional line features are added to the re-divided surface triangular mesh, and the interior of the closed surface triangular mesh is divided into a tetrahedral mesh C3D4. Based on the node sharing relationship of the surface triangular mesh on the intersecting edges of the mutually contacting components, a global zero-thickness cohesion unit COH3D6 is inserted. Finally, the reconstructed three-dimensional voxel model is converted into a mesh model to obtain a mesh file.

[0046] Preferably, in said S8, based on the divided grid, the material properties of each component region feature are defined, and the method of allocating local material directions includes:

[0047] S81: Define the material properties and mechanical response characteristics of each component region;

[0048] S82: The calculated local material direction is gradually assigned to each unit of each yarn.

[0049] The present invention also provides a three-dimensional finite element model generation and numerical simulation system for XCT digital images of composite materials involved in water projects, including: an acquisition module, a modeling module, a learning module, a precise segmentation module, a first database construction module, a second database construction module, a partitioning module, an allocation module, a simulation module and a calculation module;

[0050] The acquisition module is used to acquire the real XCT image of the basalt fiber composite material in the initial state and perform manual segmentation;

[0051] The modeling module is used to perform parametric statistical analysis on the manually segmented image, quantify the microscopic structural parameters of the yarn inside the material and perform two-dimensional parametric random modeling to obtain a parametric random modeling image;

[0052] The learning module is used to learn the parameterized random modeling image based on the adversarial learning style transfer network Pix2PixHD to generate a virtual XCT image;

[0053] The precise segmentation module is used to accurately segment the component regions of the real XCT image based on the Swin-Tranformer network and the virtual XCT image;

[0054] The first database construction module is used to extract contour microscopic features and smooth local contours based on the segmented real XCT image to construct a yarn and pore feature database;

[0055] The second database construction module is used to extract the contour features of each yarn in each segmented real XCT image based on the yarn and pore feature database, calculate the coordinate values ​​of the yarn contour center, establish a spatial coordinate system, and perform discrete derivative of the contour center of each yarn in each segmented image based on the spatial coordinate system to construct a feature database;

[0056] The partitioning module is used to reconstruct a three-dimensional voxel model based on the feature database, perform meshing and insert cohesive units based on the multi-material marching cube algorithm in the CGAL library, and simulate the yarn separation and delamination phenomenon under load;

[0057] The allocation module is used to define material properties for each component region feature based on the divided grid and allocate local material directions;

[0058] The calculation module is used to set boundary conditions and perform numerical simulation calculations.

[0059] Compared with the prior art, the present invention has the following beneficial effects:

[0060] The present invention discloses a method and system for generating a three-dimensional finite element model and numerical simulation of XCT digital images of composite materials for water-related engineering projects. The method includes collecting initial XCT images of the composite materials and preprocessing them, performing statistical analysis on two-dimensional microstructural parameters, generating virtual XCT image-virtual label image pairs based on the Pix2PixHD style transfer network for data enhancement, accurately segmenting the internal components of the composite materials based on the Swin Transformer network, extracting and smoothing the contour features of the segmented images based on the mapping relationship between pixels and nodes, constructing a spatial coordinate system and performing discrete derivatives of all yarn center points in each segmented image, reconstructing the segmented three-dimensional model, meshing based on the multi-material marching cubes algorithm of the CGAL library, inserting global cohesion units, defining the material properties of each component, assigning local material directions, setting reasonable boundary conditions, and completing numerical simulation calculations. The present invention can be used to accurately establish a three-dimensional finite element model of a composite material based on XCT images, helping to reduce experimental costs in real-world situations, study the microstructural damage evolution process and rupture state, understand the mechanical properties of the material, and achieve rapid optimization design and application of the material structure. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0062] Figure 1 This is a flow chart of a method for generating a three-dimensional finite element model and performing numerical simulation of XCT digital images of composite materials for water-related engineering projects according to an embodiment of the present invention;

[0063] Figure 2 This is a schematic diagram of XCT image segmentation results based on deep learning according to an embodiment of the present invention;

[0064] Figure 3 A schematic diagram of the process of generating a 3D finite element model according to an embodiment of the present invention;

[0065] Figure 4 This is a schematic diagram of the size of the segmented sample and the mesh division results of its internal components according to an embodiment of the present invention;

[0066] Figure 5 A schematic diagram of the generation results and boundary condition settings of a three-dimensional finite element model according to an embodiment of the present invention;

[0067] Figure 6 Schematic diagram of the simulation results of a three-dimensional finite element model of an embodiment of the present invention. DETAILED DESCRIPTION

[0068] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0069] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0070] Example 1

[0071] like Figure 1 、 Figure 2 、 Figure 3 、 Figure 4 、 Figure 5 and Figure 6 As shown, this embodiment provides a flow chart of a method for generating a three-dimensional finite element model and a numerical simulation method for XCT digital images of composite materials involved in water engineering, including the following steps:

[0072] S1: Acquire XCT images of the initial state basalt fiber composite and perform manual segmentation;

[0073] S2: Perform parametric statistical analysis on the manually segmented images to quantify the microstructural parameters of the yarn inside the material and perform two-dimensional parametric random modeling to obtain a parametric random modeling image;

[0074] S3: learning the parameterized random modeling image based on the adversarial learning style transfer network Pix2PixHD to generate a virtual XCT image;

[0075] S4: Accurately segment each component region of real XCT images based on the Swin-Tranformer network;

[0076] S5: Based on the segmented real XCT image, the microscopic features of the contour are extracted and the local contours are smoothed to build a yarn and pore feature database;

[0077] S6: Based on the yarn and pore feature database, extract the contour features of each yarn in each segmented real XCT image, calculate the coordinate value of the yarn contour center, establish a spatial coordinate system, and based on the spatial coordinate system, perform discrete derivative of the contour center of each yarn in each segmented image to construct a feature database;

[0078] S7: Based on the feature database, a three-dimensional voxel model is reconstructed, and meshing and cohesive elements are inserted based on the multi-material marching cube algorithm in the CGAL library to simulate the yarn separation and delamination phenomenon under load;

[0079] S8: Based on the divided grid, define the material properties of each component region and assign local material directions;

[0080] S9: Set boundary conditions and perform numerical simulation calculations.

[0081] The specific implementation process of this embodiment is as follows:

[0082] S1, collecting an initial state XCT image of the composite material, and manually segmenting the image according to the internal component categories of the material based on the XCT image;

[0083] Specifically, the following steps are included:

[0084] Further preferably, S1 includes:

[0085] S11, obtaining complete XCT scan source image data to obtain 1450 source images, appropriately cropping and filtering the source image data, and using a cropping tool to remove segmented images with blurred edge details;

[0086] S12, extracting one segmented image every 10 segmented images, manually labeling each component in each extracted segmented image, and completely segmenting to obtain four components: resin base, warp yarn, weft yarn, and pores;

[0087] S13. Perform feature extraction on the annotated image set to obtain a true label set, which includes 139 true label images.

[0088] S2. Perform parametric statistical analysis on the manually segmented images to quantify the microstructural parameters of the yarns inside the material and perform two-dimensional parametric stochastic modeling.

[0089] Further preferably, S2 includes:

[0090] S21. Extract the microstructural parameters such as yarn length and average width of each yarn in each segmented image based on the manually annotated microstructural features of the dataset.

[0091] S22. Qualitatively and quantitatively express the microstructural characteristics and apply statistical methods to analyze the statistical results;

[0092] S23. Based on the results of the parametric statistical analysis, the data set is expanded by parametric random modeling to generate a virtual data set;

[0093] The statistical analysis includes: statistical analysis of warp yarn geometric parameters and statistical analysis of weft yarn geometric parameters.

[0094] The method for statistical analysis of warp yarn geometric parameters is as follows: the average width, width standard deviation, length and other information of the marked warp yarn area are counted, the centroid of each warp yarn is calculated, and the distance between the centroids of each warp yarn is read to generate multiple warp yarns.

[0095] The statistical analysis method for weft yarn geometric parameters involves fitting an ellipse using the fitellipse function in the OpenCV library, calculating the weft yarn inclination angle, manually marking the maximum distance of each weft yarn in the image and using it as the major axis. The minor axis length is calculated by calculating the two intersection points of the lines perpendicular to the major axis, and the major-to-minor axis ratio is obtained. The collected data for each weft yarn is then substituted into the ellipse representation equation with the same center point. The ratio of the actual pixel radius to the ideal radius at each angle is calculated to achieve irregularity in the ellipse.

[0096] Statistical analysis of warp yarn geometric parameters and weft yarn geometric parameters: Use the fitter library to automatically search for distributions supported by the scipy library for fitting, and automatically fit the mean width, standard deviation, length, etc. of the warp yarn, the major diameter, aspect ratio, inclination angle and other data of the weft yarn.

[0097] S3, the adversarial learning style transfer network Pix2PixHD learns parameterized random modeling images to generate virtual XCT images;

[0098] Further preferably, S3 includes:

[0099] S31, generating a warp yarn random model according to warp yarn parameterized statistical analysis;

[0100] S32, generating a weft yarn random model according to weft yarn parameterized statistical analysis;

[0101] S33, generating an artificial synthetic label image based on the warp random model and the weft random model;

[0102] S34. Apply the improved Pix2PixHD network to generate basalt fiber composite images similar to the real mesostructure and construct a hybrid dataset;

[0103] The specific implementation method of S31 is:

[0104] First, a straight line is generated as the centerline, and then the centerline is padded with width. The centerline is constructed by selecting four to six nodes from the line, including the upper and lower endpoints. The coordinates of the nodes in the middle are randomly offset, with the maximum offset determined empirically. The generated points are then fitted to a curve using spline interpolation, which serves as the centerline. A corresponding number of widths are generated based on the length of the generated centerline. One to three nodes are randomly selected from the centerline, and the generated widths are sorted according to the nodes in an ascending and descending order to approximate the uneven width of the actual warp yarn. This method is also used to generate the tips required for other categories. The generated curved irregular rectangle is then subjected to random elastic deformation. This random elastic deformation involves randomly generating two matrices of the same size as the image, representing the horizontal and vertical offsets of each pixel's coordinates. These matrices are then smoothed using a Gaussian filter to make them continuous, forming a coordinate offset matrix for both directions. The original image coordinates are then offset in two different directions. The offset pixels are interpolated to obtain the grayscale values ​​of the original image coordinates, reconstructing the image to achieve the effect of random elastic deformation. In addition, the median filter method is used to remove the burrs generated during the width sorting process. Different types of warp yarns are arranged at a certain interval.

[0105] The specific implementation method of S32 is:

[0106] A basic elliptical shape is generated using the fitted minor axis length and major-minor axis ratio. Nodes are set at equal lengths on the circumference of the ellipse. Then, offsets are generated for all nodes based on the fitted weft offsets and applied to the lengths of all nodes. 720 nodes are collected during generation, and a special smoothing method is used to take the midpoints of adjacent nodes of the ellipse. After multiple smoothing steps, the entire ellipse is median filtered to eliminate burrs, and finally the ellipse is tilted according to the tilt angle generated by the fitting.

[0107] The specific implementation method of S33 is:

[0108] The generated image is scaled to the original image size, with the upper and lower boundaries of the resin region artificially defined. Warp yarns are prioritized and arranged using the generated spacing until the remaining width in the image no longer accommodates the next generated spacing and warp width. The ends of the warp yarns are then selected as the lateral boundaries of the resin region. Weft yarns are then generated and placed at randomly selected locations, with the placement area expanding upward and downward. This creates a truncated weft shape at the boundaries of the original image. The decision to place a weft is then made based on the overlap with already placed fibers. Weft placement is terminated after a certain number of failed placements or when the total fiber area of ​​the image reaches a certain size.

[0109] The specific implementation method of S34 is:

[0110] The Pix2PixHD network is a conditional generative adversarial network (cGAN). Through supervised learning, it generates images that more accurately resemble the grayscale distribution of the input target image. Its primary goal is to train a generator G, which passes an original image x through the generator network G to produce a fake image G(x). Simultaneously, a discriminator D is trained, which feeds the fake image G(x) and the real image into the discriminator D. Through training, the features of the original image and the generated image are similar, making it impossible for the discriminator D to distinguish whether the image is forged. By modifying the generative portion of the Pix2PixHD network, it simultaneously trains two generator models at different scales and fuses the features of these two generator models. Based on the fused features, the generator model generates images of fiber composite materials that have the same mesoscopic structure as the real one. A synthetic dataset was formed by manually selecting 120 images with good synthesis quality from the 200 images. These images were then combined with the original dataset of 139 images to form the final dataset. We randomly selected 15 images from the 139 original datasets as the validation set, and combined the remaining 124 original datasets and 120 synthetic datasets into the training set.

[0111] S4, based on the Swin-Tranformer network, achieves accurate segmentation of XCT images, with clear contour features between each component;

[0112] Further preferably, S4 includes:

[0113] S41, input the mixed data set to the Swin-Tranformer network;

[0114] S42, dividing the image into different small blocks by pixels through a block division module, and then converting them into a one-dimensional feature vector;

[0115] S43, network model training, segmenting the image into various component regions;

[0116] S44, completing region segmentation for each yarn based on a watershed segmentation algorithm;

[0117] The specific implementation method of S41 is:

[0118] The 1119×1431 image was randomly cropped to 448×448, and data augmentation was performed on the image by horizontal flipping, vertical flipping, and rotation. The random cropping was done because the original image was large. This was done to reduce computational effort, adapt it to the Swing Transformer, and speed up training. Effective data augmentation can enhance model robustness and prevent overfitting. The resulting image was then fed into the network for training.

[0119] The specific implementation method of S42 is:

[0120] The training images in the mixed dataset are input into the network and processed by the block partitioning module to obtain images. The images are divided into different small blocks by pixels, and the pixels of the small blocks on all channels are stretched into one-dimensional feature vectors. All one-dimensional feature vectors are combined into a feature vector map.

[0121] The specific implementation method of S43 is:

[0122] The feature vector map is processed through four stages, resulting in four feature vector maps at different scales. Each stage consists of block merging and a Swin-Transformer block. The Swin-Transformer block includes a layer normalization layer, a windowed multi-head self-attention layer, and a multi-layer perception layer. Finally, a softmax function is applied to obtain normalized class prediction probabilities of length 2. These probabilities are assigned to the four channels corresponding to pores, resin matrix, weft, and warp, respectively. The channel with the highest prediction probability is assigned to the corresponding category, resulting in the semantic segmentation result map for the entire input image.

[0123] The specific implementation method of S44 is:

[0124] Based on the above semantic segmentation results, such as Figure 2 As shown in (a), the XCT image segmentation results are extracted along the warp and weft directions respectively, and the yarns are binarized, and the binarization results are used as the labeled images; secondly, the distance transformation operation is performed on the binarized image to obtain the distance transformation mapping map; finally, watershed segmentation is performed based on the mapping map, the maximum value of the distance transformation is used as the water injection point, and the binary image of the yarn is used as a mask to limit the segmentation area. Considering that the watershed segmentation effect of the long yarn is not ideal, the binary image of the weft needs to be extracted and processed along the weft laying direction. After the segmentation is completed, a small number of misclassified areas are manually corrected to obtain the instance segmentation results, as shown in Figure 2 As shown in (b) in .

[0125] S5. Based on the segmented XCT image, extract the contour micro features and smooth the local contour;

[0126] Further preferably, S5 includes:

[0127] S51. Select an appropriate resolution to downsample the image, retaining the detailed features of the microstructure as much as possible, and reducing the cost of computing resources and the amount of calculation;

[0128] S52, performing feature screening on the downsampled image, filtering out a small portion of non-interested areas, and retaining essential feature areas such as pores and yarns;

[0129] S53, extracting the image contour features of the pores and the yarn, and establishing a coordinate value conversion based on the positional relationship between pixels and nodes.

[0130] S54. Set a suitable convolution kernel and use the operation between the convolution kernel and the extracted contour coordinate values ​​to achieve contour smoothing, thereby avoiding stress concentration and error interruption during finite element numerical calculation.

[0131] S55. Process all segmented images according to steps S51-S54, construct a yarn and pore feature database, and store the contour coordinate values ​​of each component after coordinate system transformation.

[0132] The specific implementation method of S51 is:

[0133] Based on the image that has completed the segmentation task for each component, a secondary cropping operation was performed to remove peripheral non-interested regions, resulting in an image with a length and width of 1350 × 1725 × 1725 pixels and a resolution of 4.9819 μm. While preserving the microstructural details as much as possible, the image was downsampled at every 5 pixels along the length, width, and thickness directions, resulting in an integrated image of 270 × 345 × 345 pixels and a resolution of 24.9095 μm. This reduced the number of pixels that could be processed and reduced the cost of computing resources.

[0134] The specific implementation method of S52 is:

[0135] Due to the geometric characteristics of the yarn itself, the segmented image inevitably shows a high degree of dispersion of multiple fibers of the yarn itself. In a single image, the characteristics of a single yarn are distributed in multiple areas. Therefore, the contourArea function in the OpenCV library is used to filter small areas composed of multiple fibers. By setting the screening threshold to 10, the area smaller than the fixed threshold will not be considered, and only the characteristics of the yarn area with a larger area will be retained.

[0136] The specific implementation method of S53 is:

[0137] The findContours function in the OpenCV library is used to extract the pixel coordinate values ​​of the contours of each yarn and pore, and the coordinates are converted based on the positional relationship between the pixels and the nodes. In addition, considering the difference in the orientation of the coordinate system in the OpenCV library and the coordinate system in Abaqus2021, the node coordinate values ​​converted from the pixel contour coordinates need to be reconstructed to ensure high consistency in the microstructure of the segmented image and the constructed 2D finite element model.

[0138] The specific implementation method of S54 is:

[0139] The extracted contour feature information of each component is processed twice, and a suitable convolution kernel is set. When the local contour coordinate values ​​vary greatly, contour smoothing is achieved by creating additional nodes.

[0140] The specific implementation method of S55 is:

[0141] Obtain the contour coordinate value information of each yarn and pore in each segmented image along the X-axis and Y-axis directions respectively, repeat S51-S54 operations, create a corresponding json file for each segmented image, store the information in the dictionary form in the python language, and build a feature database.

[0142] S6. extracting contour features of each yarn in each segmented image, calculating the coordinate value of the center of the yarn contour, establishing a spatial coordinate system, and performing discrete derivative of the center of the yarn contour in each segmented image;

[0143] Further preferably, S6 includes:

[0144] S61. Extract the coordinate value of the center point of each yarn contour in each segmented image from the feature database, and establish a spatial coordinate system in the order of the segmented images.

[0145] S62. Use the first-order intermediate differential law to perform discrete differentiation on the center point of each yarn contour in each segmented image to obtain the tangent direction of the center point, that is, the 1 direction of the yarn.

[0146] S63. Direction 2 can be obtained from the calculated direction 1.

[0147] The specific implementation method of S61 is:

[0148] Considering that the segmented images are converted into finite element models and that individual yarns are transversely isotropic, a local material orientation assignment algorithm is proposed to account for the influence of fiber orientation on mechanical properties. The contour feature coordinates of each yarn are extracted from the feature database along both the X and Y axes. A spatial coordinate system is established based on the order of the segmented images. Assuming the distance between adjacent segmented images is i*4.9819 μm, the contour feature coordinates of all yarns are mapped to this constructed spatial coordinate system.

[0149] The specific implementation method of S62 is:

[0150] Statistical calculation of the center point of the i-th yarn section Statistical analysis is performed on the warp or weft center points of each cross-sectional image in different directions, and then the discrete derivative of each center point is solved to obtain the value of each cross-sectional image. or direction.

[0151]

[0152] Where i is the segmented image number, v represents the XCT image resolution, T is the pixel length on the Y or Z axis, and p it is the corresponding grayscale value of pixel t on the i-th yarn cross section, and are the corresponding Y-axis and Z-axis coordinates of pixel t on the i-th yarn cross section, It corresponds to The partial derivative solution on the yarn cross section, h is the corresponding solution step size.

[0153] The specific implementation method of S63 is:

[0154] Since the yarn is transversely anisotropic, the principal direction is direction 1. Directions 2 and 3 satisfy the requirement as long as they are perpendicular to direction 1. In other words, the sum of the products of the corresponding components in each direction equals 0. Therefore, the component of direction 2 on the X-axis is defined as 0, allowing the component in direction 2 to be calculated. The calculated discrete derivative values ​​are then stored in the feature database.

[0155]

[0156] S7, meshing and inserting cohesive elements based on the multi-material marching cube algorithm in the CGAL library;

[0157] Further preferably, S7 includes:

[0158] S71. Based on S52, the pre-processed XCT image is imported into Avizo software to complete the three-dimensional voxel model reconstruction and export the inr file;

[0159] S72. Perform mesh division and insert cohesive force units based on the multi-material marching cube algorithm of the CGAL library to obtain a mesh file.

[0160] S73. Use Python script to convert it into an inp file that can be imported into Abaqus.

[0161] The specific implementation method of S72 is:

[0162] First, the 3D model is downsampled to a resolution of 24.9095um to reduce the amount of computation. By extracting the intersection edges of different components in the 3D reconstruction results, the boundary features of the intersection edges are used as one-dimensional line features, and the line features are smoothed. Secondly, based on the voxel model, the outer surfaces of the voxels of different components are used for Delaunay triangulation to generate surface triangular meshes and re-divide the surface mesh. The classic Laplacian smoothing algorithm is used to process non-manifold areas to avoid self-intersection and sharp dihedral angles. Finally, the line features are added to the re-divided surface mesh, and the interior of the closed surface triangular mesh is divided into a tetrahedral mesh (C3D4). Based on the node sharing relationship of the surface triangular mesh generated on the intersection edges of the mutually contacting components, a global zero-thickness cohesion unit (COH3D6) is inserted. Finally, the 3D model of the voxel is converted into a mesh model to obtain a mesh file, as shown in the following example. Figure 4 shown.

[0163] The specific implementation method of S73 is:

[0164] Mesh files contain various information, including the nodes and coordinates of the divided elements. However, this file format cannot be directly imported into Abaqus for further processing. By analyzing the data format between mesh files and inp files, a Python script was written to read the relevant information for each element in the mesh file and convert it to the format of the inp file.

[0165] S8. Define the material properties of each component region feature and assign local material directions;

[0166] Further preferably, S8 includes:

[0167] S81. Define the material properties and mechanical response characteristics of each component region.

[0168] S82. The local material direction calculated by S6 is gradually assigned to each unit of each yarn.

[0169] Accurately define material properties, such as density, modulus, and Poisson's ratio, and select an appropriate material damage constitutive model. For each yarn in the model, obtain the corresponding discrete derivative results from the S63 feature database. Calculate the distance from each unit's centroid to all discrete points on the segmented image. Select the directional derivative component at the discrete point with the shortest distance as the local material directional property for that unit.

[0170]

[0171] Among them, x ijk 、y ijk 、z ijkThey correspond to the X-axis, Y-axis, and Z-axis coordinate values ​​of the k-th node of the j-th unit on the ith yarn, respectively. They correspond to the X-axis, Y-axis, and Z-axis coordinates of the j-unit centroid on the i-th yarn, respectively.

[0172] S9. Set reasonable boundary conditions and perform numerical simulation calculations;

[0173] The specific implementation method of S9 is:

[0174] Abaqus2021 software was used to impose displacement conditions (U1=0, U2=-0.24, U3=0, UR1=0, UR2=0, UR3=0) on the upper movable platform to simulate the continuous loading process of the sample in the vertical direction. At the same time, full constraint conditions (U1=0, U2=0, U3=0, UR1=0, UR2=0, UR3=0) were imposed on the lower fixed platform to ensure that the sample was subjected to the compression effect, such as Figure 5 The numerical simulation results of the three-dimensional finite element model of composite materials generated by this algorithm are shown in Figure 6 shown.

[0175] Example 2

[0176] The present invention also discloses a three-dimensional finite element model generation and numerical simulation method system for XCT digital images of composite materials involved in water projects, which is characterized by comprising: an acquisition module, a modeling module, a learning module, a precise segmentation module, a first database construction module, a second database construction module, a partitioning module, an allocation module, a simulation module and a calculation module;

[0177] The acquisition module is used to acquire XCT images of the basalt fiber composite material in its initial state and perform manual segmentation;

[0178] The modeling module is used to perform parametric statistical analysis on the manually segmented images, quantify the microstructural parameters of the yarns inside the material, and perform two-dimensional parametric random modeling to obtain parametric random modeling images;

[0179] The learning module is used to learn the parameterized random modeling image based on the adversarial learning style transfer network Pix2PixHD to generate a virtual XCT image;

[0180] The precise segmentation module is used to accurately segment each component area of ​​the real XCT image based on the Swin-Tranformer network;

[0181] The first database construction module is used to extract contour microscopic features and smooth local contours based on the segmented real XCT image to build a yarn and pore feature database;

[0182] The second database construction module is used to extract the contour features of each yarn in each segmented image based on the yarn and pore feature database, calculate the coordinate values ​​of the yarn contour center, establish a spatial coordinate system, and perform discrete derivative of the contour center of each yarn in each segmented image based on the spatial coordinate system to construct a feature database;

[0183] The partitioning module is used to reconstruct a three-dimensional voxel model based on the feature database and perform mesh partitioning based on a multi-material marching cube algorithm in the CGAL library;

[0184] The allocation module is used to define the material properties of each component region feature and assign local material directions based on the divided grid;

[0185] The simulation module is used to insert cohesive elements based on the assigned local material direction to simulate the yarn separation and delamination phenomenon under load;

[0186] The calculation module is used to set boundary conditions and perform numerical simulation calculations based on the simulated yarn load separation and delamination phenomena.

[0187] The embodiments described above are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by persons skilled in the art should fall within the scope of protection defined by the claims of the present invention.

Claims

1. A three-dimensional finite element model generation and numerical simulation method for XCT digital images of composite materials in water-related engineering, characterized in that: The following steps are involved: S1: Acquire the real XCT image of the basalt fiber composite in the initial state and perform manual segmentation; S2: Perform parametric statistical analysis on the manually segmented images to quantify the microstructural parameters of the yarn inside the material and perform two-dimensional parametric random modeling to obtain a parametric random modeling image; S3: learning the parameterized random modeling image based on the adversarial learning style transfer network Pix2PixHD to generate a virtual XCT image; S4: Based on the Swin-Tranformer network and virtual XCT images, accurate segmentation of each component region of the real XCT image is achieved; S5: Based on the segmented real XCT image, the microscopic features of the contour are extracted and the local contours are smoothed to build a yarn and pore feature database; S6: Based on the yarn and pore feature database, extract the contour features of each yarn in each segmented real XCT image, calculate the coordinate value of the yarn contour center, establish a spatial coordinate system, and based on the spatial coordinate system, perform discrete derivative of the contour center of each yarn in each segmented image to construct a feature database; S7: Based on the feature database, a three-dimensional voxel model is reconstructed, and meshing and cohesive elements are inserted based on the multi-material marching cube algorithm in the CGAL library to simulate the yarn separation and delamination phenomenon under load; S8: Based on the divided grid, define the material properties of each component region and assign local material directions; S9: setting boundary conditions and numerical simulation calculation; In S7, based on the feature database, the three-dimensional voxel model is reconstructed, and the method for meshing based on the multi-material marching cube algorithm in the CGAL library includes: S71: Based on the feature database, the segmented XCT image is imported into Avizo 2019.1 software to complete the three-dimensional voxel model reconstruction and export the inr file; S72: Mesh the reconstructed 3D voxel model using the multi-material marching cubes algorithm based on the CGAL library to obtain a mesh file; S73: using a python script to convert the mesh file into an inp file that can be imported into Abaqus; In S72, the method of meshing the reconstructed three-dimensional voxel model based on the multi-material marching cubes algorithm of the CGAL library to obtain the mesh file includes: The reconstructed 3D voxel model is downsampled and the intersection edges of different components are extracted. The boundary features of the intersection edges are used as one-dimensional line features, and the one-dimensional line features are smoothed. Delaunay triangulation is performed using the outer surfaces of voxels of different components to generate surface triangular meshes and then re-mesh the surface triangular meshes. The classic Laplacian smoothing algorithm is used to process non-manifold areas to avoid surface self-intersections and sharp dihedral angles. Add one-dimensional line features to the re-divided surface triangular mesh, divide the closed surface triangular mesh into tetrahedral mesh C3D4, generate node sharing relationships of the surface triangular mesh on the intersection edges based on the contacting components, insert global zero-thickness cohesion elements COH3D6, and finally convert the reconstructed 3D voxel model into a mesh model to obtain a mesh file; The specific implementation method of S73 is: The mesh file contains a variety of information about the nodes and node coordinates of the divided elements. However, this file format cannot be directly imported into the Abaqus software for further processing. By analyzing the data format files between the mesh file and the inp file, a Python script is written to read the relevant information of each element in the mesh file and convert it according to the format of the inp file.

2. The method for generating a three-dimensional finite element model and numerical simulation of XCT digital images of composite materials for water-related engineering according to claim 1, characterized in that: In S3, the method of generating a virtual XCT image by learning the parameterized random modeling image based on the adversarial learning style transfer network Pix2PixHD includes: S31: Generate warp yarn random model based on warp yarn parameterized statistical analysis; S32: Generate a weft yarn random model based on weft yarn parameterized statistical analysis; S33: generating an artificial synthetic label image based on the warp yarn random model and the weft yarn random model; S34: Based on the artificially synthesized labeled image, an improved Pix2PixHD network is applied to generate a virtual image of the basalt fiber composite material similar to the real mesostructure, and a hybrid data set, i.e., a real-virtual XCT image, is constructed.

3. The method for generating a three-dimensional finite element model and numerical simulation of XCT digital images of composite materials for water-related engineering according to claim 1, characterized in that: In S4, the method for accurately segmenting each component region of a real XCT image based on the Swin-Tranformer network includes: S41: Input the mixed dataset to the Swin-Tranformer network to obtain training images; S42: Divide the training image into different small blocks by pixels through a block division module, and then convert it into a one-dimensional feature vector image; S43: Accurately segment each component region in the one-dimensional feature vector image to obtain a semantic segmentation result image; S44: Based on the semantic segmentation result image, a watershed segmentation algorithm is used to complete regional segmentation for each yarn.

4. The method for generating a three-dimensional finite element model and numerical simulation of XCT digital images of composite materials for water-related engineering according to claim 1, characterized in that: In S5, based on the segmented real XCT image, the method of extracting contour micro features and smoothing the local contour to construct a yarn and pore feature database includes: S51: selecting a resolution that meets preset requirements to downsample the segmented real XCT image; S52: Perform feature screening on the real XCT image after downsampling; S53: Based on the real XCT image after feature screening, extract the image contour features of the pores and yarns, and convert the coordinate values ​​according to the positional relationship between the pixels and the nodes to obtain the contour coordinate values; S54: setting a convolution kernel that meets preset requirements, and achieving contour smoothing by performing an operation between the convolution kernel and the contour coordinate values; S55: Process all segmented images according to S51-S54, build a yarn and pore feature database, and store the contour coordinate values ​​of each component after coordinate system transformation.

5. The method for generating a three-dimensional finite element model and numerical simulation of XCT digital images of composite materials for water-related engineering according to claim 1, characterized in that: In S6, based on the yarn and pore feature database, the contour features of each yarn in each segmented image are extracted, the coordinate values ​​of the yarn contour center are calculated, a spatial coordinate system is established, and based on the spatial coordinate system, the method for performing discrete derivation of the contour center of each yarn in each segmented image includes: S61: extracting the coordinate value of the center point of each yarn contour in each segmented image from the yarn and pore feature database, and establishing a spatial coordinate system in the order of the segmented images; S62: Based on the spatial coordinate system, using the first-order intermediate differential law, discretely derivate the center point of each yarn contour in each segmented image to obtain the tangent direction of the center point, that is, the 1 direction of the yarn; S63: Obtain direction 2 from direction 1 obtained by calculation.

6. The method for generating a three-dimensional finite element model and numerical simulation of XCT digital images of composite materials for water-related engineering according to claim 1, characterized in that: In S8, based on the divided grid, the material properties of each component region feature are defined, and the method for allocating the local material direction includes: S81: Define the material properties and mechanical response characteristics of each component region; S82: The calculated local material direction is gradually assigned to each unit of each yarn.

7. A three-dimensional finite element model generation and numerical simulation system for XCT digital images of composite materials involved in water engineering, the system is used to implement the method according to any one of claims 1 to 6, characterized in that: include: Acquisition module, modeling module, learning module, precise segmentation module, first database construction module, second database construction module, partitioning module, allocation module, simulation module and calculation module; The acquisition module is used to acquire the real XCT image of the basalt fiber composite material in the initial state and perform manual segmentation; The modeling module is used to perform parametric statistical analysis on the manually segmented image, quantify the microscopic structural parameters of the yarn inside the material and perform two-dimensional parametric random modeling to obtain a parametric random modeling image; The learning module is used to learn the parameterized random modeling image based on the adversarial learning style transfer network Pix2PixHD to generate a virtual XCT image; The precise segmentation module is used to accurately segment the component regions of the real XCT image based on the Swin-Tranformer network and the virtual XCT image; The first database construction module is used to extract contour microscopic features and smooth local contours based on the segmented real XCT image to construct a yarn and pore feature database; The second database construction module is used to extract the contour features of each yarn in each segmented real XCT image based on the yarn and pore feature database, calculate the coordinate values ​​of the yarn contour center, establish a spatial coordinate system, and perform discrete derivative of the contour center of each yarn in each segmented image based on the spatial coordinate system to construct a feature database; The partitioning module is used to reconstruct a three-dimensional voxel model based on the feature database, perform meshing and insert cohesive units based on the multi-material marching cube algorithm in the CGAL library, and simulate the yarn separation and delamination phenomenon under load; The allocation module is used to define material properties for each component region feature based on the divided grid and allocate local material directions; The calculation module is used to set boundary conditions and perform numerical simulation calculations.

Images