Digital twinborn discrete element model construction method based on CT scanning

Abstract

The invention belongs to the technical field of discrete element simulation, and particularly relates to a digital twinborn discrete element model construction method based on CT (Computed Tomography) scanning, which comprises the following three steps of: in a CT scanning stage, acquiring a high-resolution cross-sectional image by optimizing scanning parameters, and carrying out image noise reduction and feature enhancement processing; in the structural analysis stage, a segmentation model is constructed based on a deep convolutional neural network, and automatic annotation of a scanned image and intelligent prediction of a three-dimensional space topological relation are realized; in the discrete element modeling stage, the reconstructed three-dimensional grid model is converted into a discrete element model with real physical attributes through parameter calibration, and discrete element simulation is achieved. According to the method, the non-destructive detection advantage of industrial CT and the feature extraction capacity of deep learning are combined, microstructure characterization is converted into macroscopic performance prediction, and a reliable simulation analysis method is provided for performance optimization of the composite material.

Description

Technical Field

[0001] The present invention belongs to the field of discrete element simulation technology, and specifically relates to a method for constructing a digital twin discrete element model based on CT scanning. Background Art

[0002] In the field of micromechanics research in civil engineering materials, discrete element simulation technology has become a core means of analyzing the response of multiphase composite materials such as concrete and rock. Traditional methods usually use 3D scanners to obtain the contour data of individual components such as coarse aggregates, and then generate approximate models through random permutation algorithms for mechanical simulation. However, this technology has significant limitations: first, random generation algorithms have difficulty reproducing the spatial distribution patterns of particles in real materials (such as aggregate packing density gradients and interface transition zone characteristics), resulting in systematic deviations between the simulation model and the actual material microstructure; second, limited sample scanning results in a single model morphology library, which cannot represent the synergistic effects of multi-scale and multi-morphological components in actual materials.

[0003] In recent years, breakthroughs in X-ray computed tomography (CT) technology have provided a new path for nondestructive testing of three-dimensional material structures. Some CT systems have achieved microscopic imaging of concrete with a resolution of ≤50μm, theoretically enabling the construction of a digital twin of the real material. However, practice has shown that the existing technology system still has key bottlenecks: first, the size of the standard specimens used in experiments is still larger than that of CT, and large-scale scanning requires sacrificing resolution; second, mainstream commercial software takes a long time to process standard specimen CT data and has difficulty accurately distinguishing multiphase material interfaces; more importantly, existing reconstruction models lack an adjustable morphological parameter library, making it impossible to reverse-engineer material optimization design through simulation.

[0004] In summary, there is an urgent need for a digital twin discrete element model construction method based on CT scanning to solve the problems existing in the existing technology. Summary of the Invention

[0005] The purpose of this invention is to provide a method for constructing a digital twin discrete element model based on CT scanning. The specific technical solution is as follows: The method for constructing a digital twin discrete element model based on CT scanning includes the following steps: S1: Acquire two-dimensional images through CT scanning and perform preprocessing to obtain two-dimensional image slices; S2: Based on deep learning technology, the 2D image slices are segmented and reconstructed into 3D to obtain a 3D mesh model. S3: Construct a digital twin discrete element model based on the three-dimensional mesh model.

[0006] Preferably, in S1, obtaining a two-dimensional image by CT scanning comprises the following steps: Place the sample on the CT scanner platform; Collect data through rotation or multi-angle acquisition; During the acquisition process, the parameters of the scanning instrument are adjusted to acquire a two-dimensional image.

[0007] Preferably, in S1, during the acquisition process, adjusting the parameters of the scanning instrument includes: Adjusting the voltage value of the scan based on the material penetration of the sample; Determine the scanning current value according to the optimal power of the scanning instrument; Determine the exposure time of the scan according to the optimal air gray value range of the scanning instrument; Determine the filter thickness based on the scanning power of the scanning instrument; The number of scan combinations and the rotation step size are determined based on the imaging clarity.

[0008] Preferably, in S1, the preprocessing includes: Select the data file that needs to be reconstructed for pre-reconstruction, and adjust the horizontal offset and deflection angle of the scanning instrument; Set the reconstruction size, adjust the reconstruction center, and rotate the angle so that the target area coincides with the sample boundary, and remove the background area outside the sample boundary; The parts with the same material in the edge area and the center area of the material are selected as control samples. The edge hardening effect produced during the CT scanning process is eliminated through gray value correction to obtain a three-dimensional gray volume. Perform denoising and enhancement processing on the three-dimensional grayscale volume.

[0009] Preferably, in S2, segmenting the two-dimensional image slices comprises the following steps: Export 2D image slices to unify the format and size; Representative slices were selected as training sets, and preliminary masks were generated using threshold segmentation and edge detection. The masks were then divided into three types: large particles, matrix, and pores through annotation. Select a deep convolutional network with an encoder and decoder structure, use the training set to train the deep convolutional network, and obtain a segmentation model after training; Call the segmentation model to process the two-dimensional image slices and obtain the slice mask.

[0010] Preferably, in S2, training the deep convolutional network includes: The loss function combines Dice coefficient, cross entropy loss and boundary loss; Use Adam or SGD optimizer; Set the learning rate decay and regularization strategy.

[0011] Preferably, in S2, the three-dimensional reconstruction process is as follows: According to the scanning order and spatial spacing of the two-dimensional image slices, the slice masks are superimposed in layers to construct voxelized three-dimensional data, and the three-dimensional mesh model is generated using isosurface extraction or voxel direct display methods.

[0012] Preferably, in S3, the process of building a digital twin discrete element model is as follows: Import the 3D mesh model into discrete element software to generate an initial model; Perform loading test simulation analysis on the initial model, set the loading rate, stress path and monitoring points, record the mechanical response, particle movement and crack evolution data of the initial model during the loading process, and obtain the experimental mechanical results; The parallel bonding model was used for discrete element simulation. Based on the experimental mechanics results, the discrete element simulation was adjusted to obtain the digital twin discrete element model.

[0013] Preferably, in S3, the discrete element simulation is adjusted as follows: First, adjust the deformation of the discrete element simulation to be consistent with the loading test simulation analysis, and then adjust the strength of the discrete element simulation to be consistent with the loading test simulation analysis. If the loading test simulation analysis adopts a triaxial loading test, then after the strength adjustment, adjust the strength envelope of the discrete element simulation to be consistent with the loading test simulation analysis.

[0014] Preferably, in S3, adjusting the discrete element simulation includes: Adjusting the deformation of discrete element simulation means adjusting the elastic modulus. The elastic modulus adjustment value = current elastic modulus × test elastic modulus / simulation result elastic modulus; Adjusting the strength of discrete element simulation is to adjust the cohesion. The cohesion adjustment value = current cohesion value × test cohesion / simulation result cohesion; Adjusting the strength envelope of the discrete element simulation is to adjust the friction angle. The friction angle adjustment value = current friction angle × test friction angle / simulation result friction angle.

[0015] The application of the technical solution of the present invention has the following beneficial effects: The present invention provides a method for constructing a digital twin discrete element model based on CT scanning. The method of the present invention combines high-quality CT grayscale volume with deep learning segmentation, which not only improves the reconstruction accuracy of the complex structure inside the rock, but also greatly shortens the manual segmentation time. Secondly, the present invention adopts a standardized grayscale slicing and voxelized grid generation process, seamlessly connects the discrete element software geometry import link, and significantly simplifies the model construction steps. In addition, the present invention adopts a phased iterative parameter calibration strategy to take into account elasticity and destructive behavior, so that the simulation results can accurately reflect the experimental mechanical properties under a variety of stress states. The present method also has good versatility and scalability, and can be applied to digital twins and numerical simulations of various multi-component and porous materials, providing an efficient and accurate technical means for rock mechanics research, mining, geological engineering, materials science and other fields.

[0016] In addition to the above-described objects, features and advantages, the present invention has other objects, features and advantages. The present invention will be further described in detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the embodiments of the present invention or the technical solutions of the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. 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.

[0018] Figure 1 It is a flowchart of the steps of the method for constructing a digital twin discrete element model in a preferred embodiment of the present invention; Figure 2 It is a detailed step diagram of the digital twin discrete element model construction method in the preferred embodiment of the present invention. DETAILED DESCRIPTION

[0019] To address the technical bottlenecks of existing rock and concrete structure digitization and mechanical simulation processes, such as insufficient 3D reconstruction accuracy, cumbersome model geometry generation processes, inefficient numerical simulation parameter calibration, and large deviations between simulation results and experimental data, the present invention provides a method for constructing a digital twin discrete element model based on CT scanning. This method, through three steps—CT image acquisition and preprocessing, deep learning-based 2D slice segmentation and 3D reconstruction, and discrete element model construction and iterative parameter calibration—achieves the digital twin transformation from raw rock or concrete samples to high-precision numerical models, significantly improving reconstruction accuracy and simulation accuracy while significantly simplifying the model generation and parameter calibration processes.

[0020] In order to enable those skilled in the art to better understand the present invention, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0021] like Figure 1 and Figure 2 As shown, this embodiment provides a method for constructing a digital twin discrete element model based on CT scanning, including the following steps: S1: Acquire two-dimensional images through CT scanning and perform preprocessing to obtain two-dimensional image slices; S2: Based on deep learning technology, the 2D image slices are segmented and reconstructed into 3D to obtain a 3D mesh model. S3: Construct a digital twin discrete element model based on the three-dimensional mesh model.

[0022] Preferably, in S1, obtaining a two-dimensional image by CT scanning comprises the following steps: The sample is placed on the CT scanner platform; this embodiment uses rock samples to construct a digital twin discrete element model, and the constructed digital twin discrete element model realizes the study of the mechanical properties of the rock. Specifically, before scanning the rock sample, this embodiment dries or vacuum-extracts the rock sample to be tested to eliminate the interference of free water in the pores on the grayscale distribution of CT imaging and ensure the distinguishability of different components in terms of grayscale values. It should be noted that the shape of the rock sample can be simply trimmed according to research requirements, but the selected area should be representative.

[0023] Collect data through rotation or multi-angle acquisition; During the acquisition process, the parameters of the scanning instrument are adjusted to acquire a two-dimensional image.

[0024] Specifically, during the acquisition process, the parameters of the scanning instrument are adjusted, including: adjusting the scanning voltage value based on the material penetration rate of the sample (penetration rate is generally greater than 10%, penetration rate = sample grayscale value / air grayscale value); determining the scanning current value according to the optimal power of the scanning instrument; determining the scanning exposure time according to the optimal air grayscale value range of the scanning instrument; determining the filter thickness according to the scanning power of the scanning instrument; and determining the number of scan merging and rotation steps according to the imaging clarity.

[0025] In this example, the processed rock sample is fixed to the CT scanner platform, and a series of two-dimensional projection images are acquired through rotational or multi-angle acquisition. During the acquisition process, instrument geometric parameters, such as the distance between the source and detector, the sampling field of view, and the sampling frequency, can be adjusted to balance resolution and scanning efficiency.

[0026] Preferably, in S1, the preprocessing includes: Select the data file to be reconstructed for pre-reconstruction, and adjust the horizontal offset and deflection angle of the scanner. It should be noted that the scanner software typically includes adjustment parameters for the detector's horizontal, vertical, and deflection angles. However, adjustments are generally only necessary for the detector's horizontal offset and deflection angle. The horizontal offset is generally adjustable within a range of ±50 pixels, and the detector angle is generally adjustable within a range of ±1°. Both adjustments are made using approximation.

[0027] The reconstruction size was set, the reconstruction center was adjusted, and the rotation angle was used to make the target area coincide with the sample boundary, and the background area outside the sample boundary was eliminated. The parts with the same material in the edge area and the center area were selected as control samples. The edge hardening effect produced during the CT scanning process was eliminated through grayscale value correction to obtain a three-dimensional grayscale volume. The three-dimensional grayscale volume was denoised and enhanced.

[0028] In this embodiment, denoising of a 3D grayscale volume primarily involves Gaussian filtering and non-local means filtering, thereby reducing image noise. This also adjusts the contrast between different components and highlights the boundaries between different materials. Furthermore, based on denoising, the method of this embodiment also exports 2D grayscale slices at evenly spaced intervals along defined cross-sectional directions (e.g., XY, XZ, YZ, etc.). This ensures uniform slice size, resolution, grayscale range, and file format (e.g., TIFF, PNG, etc.), providing standardized input for segmentation and 3D stitching.

[0029] Furthermore, in S2, the segmentation of the two-dimensional image slice is completed, including the following steps: Export 2D image slices to unify the format and size to facilitate deep learning prediction of different components.

[0030] Representative slices were selected as training sets, and preliminary masks were generated using threshold segmentation and edge detection. These masks were then annotated and classified into three main types: large particles, matrix, and pores. It should be noted that open-source annotation tools can be used to improve efficiency during the annotation process. Images are annotated for large particles, cracks, pores, and other structures without annotating the matrix. During 3D reconstruction, the volume of the matrix is calculated by subtracting the volume of other components from the total space. In this embodiment, geometric transformations (rotation, flipping, scaling, affine transformations), grayscale perturbations (brightness, contrast, noise simulation), and random cropping can also be applied simultaneously to the original image and the corresponding mask to generate diverse training samples, expand the training set size, and improve model robustness.

[0031] A deep convolutional network with an encoder-decoder structure is selected and trained using the training set. After training, a segmentation model is obtained. Optional deep convolutional networks in this embodiment include U-Net, SegNet, or Transformer-based segmentation models.

[0032] Call the segmentation model to process the two-dimensional image slices and obtain the slice mask.

[0033] Preferably, in S2, training the deep convolutional network includes: The loss function combines the Dice coefficient, cross entropy loss, and boundary loss. The Dice coefficient, cross entropy loss, and boundary loss are weighted accordingly. The ratio of the three losses is adjusted based on the training results to calculate the total loss. The boundary loss in this application adopts IoU (Intersection over Union). IoU is an evaluation indicator of the structural recognition accuracy of deep learning models. Its calculation formula is: IoU = the intersection area of the predicted structure and the true structure / the union area of the predicted structure and the true structure. The closer this value is to 1, the higher the spatial match between the 3D reconstruction result and the true structure. The IoU indicator is used to quantitatively verify the 3D reconstruction accuracy and ensure the authenticity of the geometric topology of the discrete element model. Use Adam or SGD optimizer; Set learning rate decay: OneCycleLR learning rate schedule, initial learning rate 10 -4 , peak learning rate 5×10 -4 , smooth adjustment within the cycle.

[0034] Furthermore, after obtaining the slice mask, this embodiment also performs morphological processing (such as opening and closing operations), connected domain analysis, and area screening on the slice mask to remove isolated noise areas and repair small holes to ensure that each category of mask is coherent and complete.

[0035] Preferably, in S2, the three-dimensional reconstruction process is as follows: Based on the scanning order and spatial spacing of the 2D image slices, the slice masks are layered to construct voxelized 3D data. A 3D mesh model is generated using isosurface extraction or voxel direct display. This embodiment exports the 3D mesh model to STL, OBJ, or other standard mesh formats.

[0036] Preferably, in S3, the process of building a digital twin discrete element model is as follows: Import the 3D mesh model into the discrete element software to generate the initial model. Specifically, the process of generating the initial model is as follows: In the PFC software, a geometry is generated. Based on this geometry, a rigid cluster template is created. Then, the "original-position" command within the "clump replicate" command is used to create a rigid cluster of large particles at the corresponding position. Matrix balls are then generated within the model geometry. At this point, matrix balls also exist within the rigid cluster of large particles. Balls that overlap with the clump are deleted using the fish function or command. This example only describes the generation of the initial model; other general steps are not described.

[0037] In this embodiment, after the initial model is generated, servo balancing is performed to make the simulation results close to the real physical process. That is, by adjusting the boundary conditions of the model, the contact between the particle system reaches the ideal state as quickly as possible, and then the loading analysis is carried out on this basis.

[0038] A loading test simulation analysis is performed on the initial model. The loading rate, stress path, and monitoring points are set. The mechanical response, particle movement, and crack evolution data of the initial model during the loading process are recorded to obtain experimental mechanical results. The loading test simulation analysis in this embodiment includes at least one of a uniaxial compression test, a triaxial loading test, or a cyclic loading test.

[0039] The parallel bonding model was used for discrete element simulation. Based on the experimental mechanics results, the discrete element simulation was adjusted to obtain the digital twin discrete element model.

[0040] Preferably, in S3, the discrete element simulation is adjusted as follows: First, adjust the deformation of the discrete element simulation to be consistent with the loading test simulation analysis, and then adjust the strength of the discrete element simulation to be consistent with the loading test simulation analysis. If the loading test simulation analysis adopts a triaxial loading test, then after the strength adjustment, adjust the strength envelope of the discrete element simulation to be consistent with the loading test simulation analysis.

[0041] Preferably, in S3, adjusting the discrete element simulation includes: Adjusting the deformation of discrete element simulation means adjusting the elastic modulus. The elastic modulus adjustment value = current elastic modulus × test elastic modulus / simulation result elastic modulus; Adjusting the strength of discrete element simulation is to adjust the cohesion. The cohesion adjustment value = current cohesion value × test cohesion / simulation result cohesion; Adjusting the strength envelope of the discrete element simulation is to adjust the friction angle. The friction angle adjustment value = current friction angle × test friction angle / simulation result friction angle.

[0042] In this embodiment, the adjustment of the above parameters can be performed according to the following formula: ; in, 、 、 denote elastic modulus, cohesion and internal friction angle respectively, and the subscripts current is the current analog value, sim is the simulation result value, exp is the experimental value, new The above iterative correction can ensure that the discrete element model accurately reflects the experimental mechanical behavior under various stress states.

[0043] It should be noted that the device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of this embodiment.

Claims

1. A method for constructing a digital twin discrete element model based on CT scanning, characterized in that: The steps include: S1: Acquire two-dimensional images through CT scanning and perform preprocessing to obtain two-dimensional image slices; S2: Based on deep learning technology, the 2D image slices are segmented and reconstructed into 3D to obtain a 3D mesh model. S3: Construct a digital twin discrete element model based on the three-dimensional mesh model.

2. The method for constructing a digital twin discrete element model according to claim 1, characterized in that: In S1, a two-dimensional image is acquired by CT scanning, including the following steps: Place the sample on the CT scanner platform; Collect data through rotation or multi-angle acquisition; During the acquisition process, the parameters of the scanning instrument are adjusted to acquire a two-dimensional image.

3. The method for constructing a digital twin discrete element model according to claim 2, characterized in that: In S1, during the acquisition process, adjust the parameters of the scanning instrument, including: Adjusting the voltage value of the scan based on the material penetration of the sample; Determine the scanning current value according to the optimal power of the scanning instrument; Determine the exposure time of the scan according to the optimal air gray value range of the scanning instrument; Determine the filter thickness based on the scanning power of the scanning instrument; The number of scan combinations and the rotation step size are determined based on the imaging clarity.

4. The method for constructing a digital twin discrete element model according to claim 3, characterized in that: In S1, preprocessing includes: Select the data file that needs to be reconstructed for pre-reconstruction, and adjust the horizontal offset and deflection angle of the scanning instrument; The parts with the same material in the edge area and the center area of the material are selected as control samples. The edge hardening effect produced during the CT scanning process is eliminated through gray value correction to obtain a three-dimensional gray volume. Perform denoising and enhancement processing on the three-dimensional grayscale volume.

5. The method for constructing a digital twin discrete element model according to claim 4, characterized in that: In S2, the segmentation of the two-dimensional image slice is completed, including the following steps: Export 2D image slices to unify the format and size; Representative slices were selected as training sets, and preliminary masks were generated using threshold segmentation and edge detection. The masks were then divided into three types: large particles, matrix, and pores through annotation. Select a deep convolutional network with an encoder and decoder structure, use the training set to train the deep convolutional network, and obtain a segmentation model after training; Call the segmentation model to process the two-dimensional image slices and obtain the slice mask.

6. The method for constructing a digital twin discrete element model according to claim 5, characterized in that: In S2, the deep convolutional network is trained, including: The loss function combines Dice coefficient, cross entropy loss and boundary loss; Use Adam or SGD optimizer; Set the learning rate decay and regularization strategy.

7. The method for constructing a digital twin discrete element model according to claim 6, characterized in that: In S2, the 3D reconstruction process is as follows: According to the scanning order and spatial spacing of the two-dimensional image slices, the slice masks are superimposed in layers to construct voxelized three-dimensional data, and the three-dimensional mesh model is generated using isosurface extraction or voxel direct display methods.

8. The method for constructing a digital twin discrete element model according to claim 7, characterized in that: In S3, the process of building a digital twin discrete element model is as follows: Import the 3D mesh model into discrete element software to generate an initial model; Perform loading test simulation analysis on the initial model, set the loading rate, stress path and monitoring points, record the mechanical response, particle movement and crack evolution data of the initial model during the loading process, and obtain the experimental mechanical results; The parallel bonding model was used for discrete element simulation. Based on the experimental mechanics results, the discrete element simulation was adjusted to obtain the digital twin discrete element model.

9. The method for constructing a digital twin discrete element model according to claim 8, characterized in that: In S3, adjust the discrete element simulation as follows: First, adjust the deformation of the discrete element simulation to be consistent with the loading test simulation analysis, and then adjust the strength of the discrete element simulation to be consistent with the loading test simulation analysis. If the loading test simulation analysis adopts a triaxial loading test, then after the strength adjustment, adjust the strength envelope of the discrete element simulation to be consistent with the loading test simulation analysis.

10. The method for constructing a digital twin discrete element model according to claim 9, wherein: In S3, adjustments to discrete element simulations include: Adjusting the deformation of discrete element simulation means adjusting the elastic modulus. The elastic modulus adjustment value = current elastic modulus × test elastic modulus / simulation result elastic modulus; Adjusting the strength of discrete element simulation is to adjust the cohesion. The cohesion adjustment value = current cohesion value × test cohesion / simulation result cohesion; Adjusting the strength envelope of the discrete element simulation is to adjust the friction angle. The friction angle adjustment value = current friction angle × test friction angle / simulation result friction angle.

Images