A method for asphalt mixture aggregate motion tracking
By using cross-threaded nails and an improved deep learning model in asphalt mixtures, the problems of insufficient digital image segmentation accuracy and aggregate cross-state tracking were solved, achieving high-precision aggregate segmentation and continuous trajectory tracking, thus improving pavement performance.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHANGSHA UNIVERSITY OF SCIENCE AND TECHNOLOGY
- Filing Date
- 2026-03-13
- Publication Date
- 2026-05-29
Smart Images
Figure CN121861086B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of road engineering technology, and specifically relates to a method for tracking the movement of aggregates in asphalt mixtures. Background Technology
[0002] The compaction quality of asphalt mixtures is crucial to the service performance of pavements. In-depth research on the microscopic motion trajectory of aggregates during compaction is of significant theoretical value for revealing the compaction mechanism and optimizing gradation design. Currently, the analysis of the evolution of the internal structure of asphalt mixtures during compaction mainly relies on discrete element numerical simulation (DEM). However, simulation results are often limited by simplified particle morphology and contact model assumptions, making it difficult to accurately reflect the complex motion behavior of aggregates during actual compaction. While CT-based real aggregate tracking methods can obtain realistic evolution data of the internal structure of mixtures through multi-state compaction-scan experiments, the following technical problems remain in practical applications: First, the placement of specimens under different compaction states is difficult to accurately reconstruct during CT scanning, making direct comparison and analysis of multi-state images impossible; second, traditional image segmentation methods have poor adaptability to mixtures with different gradations and insufficient accuracy in handling aggregate adhesion and fine aggregate identification; third, morphological parameter fluctuations caused by image segmentation errors and the complex rotational behavior of aggregates make cross-state identification and continuous trajectory tracking of the same aggregate under different compaction states difficult to achieve. Summary of the Invention
[0003] This invention aims to address the technical problems of insufficient digital image segmentation accuracy and difficulty in continuously tracking and identifying real aggregates in existing technologies, and provides a method for tracking the motion of aggregates in asphalt mixtures. This method can achieve high-precision segmentation of aggregate morphology and realize reliable cross-state identification and continuous trajectory tracking of aggregates through a multi-feature weighted matching algorithm.
[0004] Therefore, the present invention provides a method for tracking the movement of aggregates in asphalt mixtures, characterized by comprising the following steps:
[0005] Step S1: Pre-embed a cross-shaped threaded nail at the center of the bottom of the sample as a measurement target point, perform multiple rotational compaction on the asphalt mixture sample, and perform a CT scan after each compaction to obtain a CT image sequence of the same sample under multiple different compaction states.
[0006] Step S2: Using the first-state CT image obtained from the first scan as a reference, the CT images of the remaining states are rotated and corrected by comparing the spatial positions of the measurement target points in the CT data of each state, so as to eliminate the spatial position deviation caused by repeated assembly and disassembly of the specimen.
[0007] Step S3: Segment the corrected CT images of each state, identify and extract the aggregate phase in the images;
[0008] Step S4: Based on the segmentation results, perform three-dimensional reconstruction of the CT images under each compaction state and extract the feature parameters of each aggregate; the feature parameters include spatial location features, morphological features and principal axis direction features;
[0009] Step S5: Based on the feature parameters, perform multi-feature weighted matching on aggregates in adjacent compaction states to establish a unique correspondence between aggregates across compaction states, and thereby construct a complete motion trajectory of a single aggregate throughout all compaction states.
[0010] Specifically, step S1 includes:
[0011] A cross-shaped threaded nail is pre-embedded at the bottom center of the rotary compaction mold as a measurement target point. The cross-shaped threaded nail maintains a stable position during the compaction process through the mechanical interlocking effect between the thread and the asphalt mixture.
[0012] The asphalt mixture is loaded into a mold and rotated for compaction. After reaching the first preset number of times, the machine is stopped, the mold is removed, and the first CT scan is performed, marked as state #1. After scanning, the surface of the specimen is coated with lubricating oil and then reloaded into the mold. The specimen is heated to restore the rheological properties of the asphalt mortar. The heated specimen is rotated and compacted again to the second preset number of times. After demolding, a second CT scan is performed, marked as state #2. The above compaction, heating, and scanning steps are repeated until the target number of CT image sequences of different compaction states are obtained.
[0013] Specifically, step S2 includes:
[0014] The spatial position of the cross-shaped threaded nail in the CT image of state #1 is used as the reference;
[0015] A spatial coordinate system is established to unify the geometric reference of images of different states. This invention defines the following: the axial direction (height direction) of the cylindrical specimen is the z-axis, and the center of the bottom of the specimen is the origin O; the xy plane is a horizontal plane perpendicular to the z-axis, i.e., the cross section of the specimen, where the x-axis and y-axis are mutually perpendicular directions in the xy plane, used to characterize the movement characteristics of the aggregate in the horizontal direction. By comparing the spatial position of the cross-thread nail in the CT image of state #N, the deflection angle of state #N relative to state #1 in the xy plane is calculated.
[0016] The CT image of state #N is rotated and corrected according to the deflection angle so that the position of the cross-shaped screw in the CT image of each state is consistent with the reference state.
[0017] Specifically, in step S3, an improved deep learning model is used to segment the CT image; the improved deep learning model introduces a feature enhancement mechanism in the encoding stage to improve the segmentation accuracy of aggregate edges and fine aggregates.
[0018] Specifically, the improved deep learning model is the Nested U-Net architecture, which introduces residual connections and channel attention mechanisms in the downsampling convolution module of the encoder. Specifically, the input feature map X is convolved, batch normalized, and ReLU activated in sequence, and then convolved and normalized again to extract features F(X). Then, F(X) is residually connected with the original input X to obtain Y. Y is then fed into the channel attention module to weight different feature channels and output the feature Yse enhanced by the attention mechanism.
[0019] Specifically, step S4 includes:
[0020] The segmented aggregate phase image is reconstructed in three dimensions to establish a three-dimensional voxel model of the aggregate;
[0021] Connectivity analysis was performed on the three-dimensional voxel model to identify and label each individual aggregate particle;
[0022] The characteristic parameters of each aggregate are extracted, including morphological characteristic parameters, spatial position characteristic parameters, and principal axis direction characteristic parameters.
[0023] Specifically, the principal axis direction characteristic parameters are obtained through principal axis inertia analysis, which includes: constructing the inertia tensor of aggregate voxels relative to their centroids, performing eigenvalue decomposition on the inertia tensor to obtain its eigenvalues and corresponding eigenvectors, and taking the eigenvector corresponding to the largest eigenvalue as the principal axis direction of the aggregate.
[0024] Specifically, the morphological features include at least one of volume, surface area, equivalent diameter, and sphericity; the spatial position features are the three-dimensional centroid coordinates of the aggregate; the principal axis direction features include the angle between the principal axis and the z-axis, i.e., the polar angle φ, and the angle between the projection of the principal axis onto the xy plane and the x-axis, i.e., the azimuth angle θ.
[0025] Specifically, in step S5, the multi-feature weighted matching includes:
[0026] A KD-Tree spatial index is constructed based on spatial location features to perform fast neighbor search for aggregates in adjacent compaction states; a comprehensive similarity evaluation function is constructed to perform weighted summation of differences in spatial location, morphological features, and principal axis direction features; and a one-to-one correspondence between aggregates in adjacent compaction states is established through a bidirectional matching strategy.
[0027] Specifically, the comprehensive similarity evaluation function is:
[0028] (1);
[0029] (2);
[0030] (3);
[0031] (4);
[0032] (5);
[0033] In the formula: S total D is the total score for similarity evaluation. morphology D is the normalized Euclidean distance of spatial location. spatial D represents the morphological differences constructed based on volume, surface area, equivalent diameter, and / or sphericity. orientation To differentiate the directional features based on polar angle and azimuth angle, w s w m w o For the corresponding weight coefficients, w s +w m +w o =1, S is the sphericity, D eq Where V is the equivalent particle size, A is the volume, and ΔS and S are the surface areas. max ΔD represents the difference and maximum value of the sphericity of the bone fragments to be matched between the two CT sequences. eq D eqmax Let ΔV and V represent the difference and maximum value of the equivalent particle size of the aggregate to be matched between the two CT sequences, respectively. eqmax Let ΔA and A represent the difference and maximum volume of the bone fragments to be matched between the two CT sequences, respectively. eqmax y1, y1, z1 and x2, y2, z2 respectively represent the difference and maximum surface area of the aggregate to be matched between the two CT sequences, and x1, y1, z1 and x2, y2, z2 respectively refer to the three-dimensional coordinates of the centroid of the aggregate to be matched in the two CT sequences. spatial Let d be the distance between the centroids. max The aggregate movement threshold is set to 50 mm here, φ is the polar angle, Δφ is the polar angle difference between the two CT sequences, θ is the azimuth angle, and Δθ is the azimuth angle difference between the two CT sequences.
[0034] Compared with the prior art, the present invention has the following beneficial effects:
[0035] 1. Using cross-shaped threaded nails as measurement targets, the mechanical interlocking effect between the threads and the mixture ensures the positional stability of the targets during compaction. Based on this, batch rotational correction is performed using the cross-shaped geometric center of the targets as a reference, achieving precise spatial registration of CT images under different compaction states. This lays a spatial consistency foundation for subsequent cross-state tracking of aggregates.
[0036] 2. An improved Nested U-Net model was constructed by introducing residual networks and channel attention mechanisms. The residual connections effectively alleviated the gradient vanishing problem of deep networks, enabling the model to learn richer aggregate morphological features. The channel attention mechanism automatically enhances the weights of key features such as aggregate edges and textures, and suppresses noise interference. Compared with the single U-Net model in the prior art, the improved model of this invention has higher segmentation accuracy and stronger generalization ability.
[0037] 3. A three-dimensional feature system integrating spatial location, morphological features and principal axis direction was constructed, and a KD-Tree fast index and multi-feature weighted matching algorithm were designed to achieve unique identification of aggregates across states. This method does not depend on specific gradation and is applicable to various asphalt mixture types such as AC, SMA, and OGFC, and has wide applicability.
[0038] 4. Precisely quantifying the migration and rotation patterns of aggregates reveals the response characteristics of mixtures to compaction loads at a microscopic level. This provides theoretical guidance for the placement of sensors and setting of detection parameters in intelligent compaction, and obtains the motion characteristic parameters of aggregates throughout the compaction process. It can provide quantitative reference for the gradation optimization and skeleton structure design of asphalt mixtures, which helps to improve the durability and performance of pavements.
[0039] In summary, this invention systematically solves the technical problems of difficult spatial registration of multi-state CT images, insufficient segmentation accuracy, and difficulty in cross-state tracking in the prior art by combining the techniques of target point measurement correction, improved deep learning segmentation, and multi-feature weighted matching. It has significant technical progress and practical value. Attached Figure Description
[0040] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0041] Figure 1 It measures the effectiveness of target placement;
[0042] Figure 2 This is a schematic diagram of the CT scanning process for asphalt mixtures;
[0043] Figure 3 This refers to the effect of CT sequence position calibration;
[0044] Figure 4 It is an outline label for aggregates;
[0045] Figure 5 It is the effect of aggregate splitting;
[0046] Figure 6 It is a three-dimensional reconstruction diagram of the aggregate;
[0047] Figure 7 This is a flowchart of the method of the present invention. Detailed Implementation
[0048] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0049] like Figure 7 As shown, a method for tracking the movement of aggregates in asphalt mixtures includes the following steps:
[0050] Step S1: Prepare the specimen and perform multi-state CT scanning.
[0051] This step aims to obtain a sequence of CT images of the same specimen under multiple different compaction conditions. Specifically, it includes the following sub-steps:
[0052] Step S11: Place a 6mm long cross-shaped threaded nail at the center of the bottom of the rotary compaction mold as a measurement target point. Fix it with padding paper to ensure the target point position remains stable during compaction. See the attached image for the effect. Figure 1 Cross-threaded studs enhance the shear resistance of markers through the mechanical interlocking effect between the threads and the asphalt mixture.
[0053] Step S12: The asphalt mixture, which is mixed evenly according to the target gradation, is loaded into a mold and placed in a rotary compactor for molding. The compaction parameters are set as follows: vertical pressure 600 kPa, rotation speed 30 r / min, and inner rotation angle 1.16°.
[0054] Step S13: Based on the number of compaction times determined in the pre-test, stop the machine after the first number of compaction times is reached, demold the specimen and take it out for the first CT scan, and mark it as state #1.
[0055] Step S14: Apply lubricating oil evenly to the surface of the scanned specimen, reinstall it into the mold, and heat it in a 165℃ oven for 20 minutes to restore the rheological properties of the asphalt mortar.
[0056] Step S15: After heating, the specimen is rotated and compacted again to the second preset number of times. After demolding, a second CT scan is performed and marked as state #2.
[0057] Step S16: Repeat steps S14 to S15 until the target number of CT image sequences of the same specimen are obtained. See the operation procedure below. Figure 2 .
[0058] The asphalt mixture specimens are standard asphalt mixture specimens prepared by rotary compaction. The gradation type can be AC, SMA, OGFC, etc., and the maximum nominal particle size can be common specifications such as 13.2mm and 16mm. In this embodiment, three gradations of asphalt mixtures—AC-13, SMA-13, and OGFC-13—are selected for rotary compaction. The specimens are cylindrical, with a diameter of 100mm and a height of 100mm.
[0059] Step S2: Spatial position correction of CT images.
[0060] This step aims to eliminate spatial positional deviations between multi-state CT images caused by repeated assembly and disassembly of the specimen. Specifically, it includes:
[0061] Step S21: Slice the 3D CT scan data volume of each state obtained in step S1 to obtain a series of 2D cross-sectional images, and convert them into grayscale images in the range of 0~255.
[0062] Step S22: Establish a spatial coordinate system to unify the geometric reference of each state image. This invention defines: the axial direction (height direction) of the cylindrical specimen as the z-axis, and the center of the bottom of the specimen as the origin O; the xy plane is a horizontal plane perpendicular to the z-axis, i.e., the cross section of the specimen, where the x-axis and y-axis are mutually perpendicular directions in the xy plane, used to characterize the movement characteristics of the aggregate in the horizontal direction. Using the CT image of state #1 as the spatial reference, by comparing the geometric center position of the cross-shaped measurement target point in the CT slices of each state, the deflection angle of states #2 to #5 relative to the reference state in the xy plane is calculated. The deflection angle refers to the change in the angle between the projection of the line connecting the centers of the measurement target points on the xy plane and the x-axis.
[0063] Step S23: Use the Python programming language to call the image processing library and perform batch rotation correction on the CT images of states #2 to #5 according to the calculated deflection angle.
[0064] Step S24: Verify whether the target point position in the calibrated image is consistent with the baseline state, ensuring that voxel points at the same spatial location in CT images of each state have a one-to-one geometric relationship, laying a spatial consistency foundation for subsequent cross-state comparative analysis. Taking AC-13 as an example, its calibration results can be found in [link to calibration details]. Figure 3 .
[0065] Step S3: Aggregate segmentation based on an improved deep learning model.
[0066] CT images of asphalt mixtures are rich in detail, with complex aggregate morphology and irregular edges. Furthermore, the image features of different gradations of the mixtures vary greatly, making it difficult for traditional segmentation methods to meet the accuracy requirements.
[0067] This embodiment uses an improved deep learning model to achieve high-precision aggregate segmentation. Specifically, it includes the following sub-steps:
[0068] Step S31: Select 30 slices from different locations for manual annotation, construct an aggregate label dataset, and generate a binarized aggregate mask image. (See [link]). Figure 4 .
[0069] Step S32: Randomly divide the dataset into training set, validation set and test set in a ratio of 7:2:1, and perform data augmentation on the training set by flipping, mirroring, cropping and other methods to expand the training set by 20 to 60 times. In this embodiment, the training set is expanded to 1008 images.
[0070] Step S33: Construct an improved Nested U-Net convolutional neural network architecture that incorporates residual connections and channel attention mechanisms. The improvement mainly involves introducing residual connections and channel attention mechanisms into each convolutional module during the downsampling stage to enhance the learning ability of aggregate edges and fine aggregate detail features.
[0071] Specifically, in each convolutional module of the downsampling stage, the input feature map X is sequentially convolved, batch normalized, and ReLU activated. Then, convolution and normalization are performed again in the second convolutional layer to extract features F(X). F(X) is then residually concatenated with the original input X to obtain Y. Y is then fed into the channel attention module to weight different feature channels. Finally, the feature Yse enhanced by the attention mechanism is output through the ReLU activation function.
[0072] Step S34: Train the improved Nested U-Net model in Python and PyTorch environments using the Adam optimizer, an initial learning rate of 0.0001, a cosine annealing scheduling strategy, and the Softmax cross-entropy loss function. Train for 200 epochs with a batch size of 2. Preliminary training verification shows that these hyperparameter settings can ensure effective and stable convergence of the model.
[0073] Step S35: Use the trained improved Nested U-Net model to segment the calibrated CT images, and output the aggregate segmentation results in batches. See [link to aggregate segmentation results] for details. Figure 5 .
[0074] Step S4: 3D reconstruction and aggregate feature extraction.
[0075] This step aims to reconstruct a three-dimensional aggregate model from the two-dimensional segmentation results and quantify the geometric and physical properties of each aggregate. Specifically, it includes:
[0076] Step S41: Import the aggregate phase mask image sequence obtained from step S3 into Avizo software, set the voxel size parameters so that the slice spacing is consistent with the physical size of the pixels within the slice, and reconstruct the three-dimensional voxel model of the aggregate through the Volume Rendering module.
[0077] Step S42: Apply the Label Analysis module to perform connected component analysis on the reconstructed 3D aggregate model. Use the 26-neighborhood connected component labeling algorithm to identify and label each individual aggregate particle. See the results below. Figure 6 .
[0078] Step S43: Extract the morphological characteristics of each aggregate using Avizo's measurement module, including volume, surface area, equivalent diameter, and sphericity. In this embodiment, aggregates with a particle size smaller than 4.75 mm were screened out.
[0079] Step S44: Use Avizo's centroid calculation function to extract the three-dimensional centroid coordinates of each aggregate as its spatial position feature. The three-dimensional centroid is used to characterize the average position of the entire aggregate in space. The centroid can be used to calculate the absolute displacement of the aggregate movement and the distance between it and other aggregates.
[0080] Step S45: Calculate the orientation characteristics of the aggregate based on principal axis of inertia analysis. Construct a 3×3 inertial tensor of the aggregate voxel relative to its centroid and perform eigenvalue decomposition. Take the eigenvector corresponding to the largest eigenvalue as the principal axis orientation of the aggregate. Calculate the angle between the principal axis and the z-axis, i.e., the polar angle φ, ranging from [0, 180°]; calculate the angle between the projection of the principal axis onto the xy-plane and the x-axis, i.e., the azimuth angle θ, ranging from [-180°, 180°]. φ and θ together form a spherical coordinate system to characterize the orientation characteristics of the aggregate itself.
[0081] Step S46: Export all aggregate characteristic parameters for subsequent steps.
[0082] Step S5: Aggregate tracking based on multi-feature weighted matching.
[0083] Under the influence of compaction loads and internal interactions, aggregates undergo mesoscopic movements primarily involving spatial displacement and orientation adjustments. Since the strength and modulus of aggregates are much higher than those of asphalt mortar and the local stresses generated by compaction loads, aggregates themselves do not undergo plastic deformation or breakage under normal compaction conditions, providing a fixed identifier for the aggregates. However, actual CT image segmentation results inevitably contain errors, leading to inconsistent morphological parameters measured for the same aggregate under different compaction conditions. Therefore, a multi-feature matching mechanism is needed to achieve aggregate identification.
[0084] This step aims to establish a unique correspondence between aggregates in adjacent compaction states through multi-feature matching, thereby constructing their complete motion trajectory. Specifically, it includes:
[0085] Step S51: Load the aggregate feature datasets of five consecutive compaction states, including the spatial coordinates (three-dimensional centroid), morphological feature parameters (volume, surface area, equivalent diameter, sphericity) and orientation feature parameters (φ and θ) of each aggregate.
[0086] Step S52: Based on spatial location features, construct KD-Tree spatial indexes for aggregates in adjacent compaction states to achieve fast proximity search, limit the search range of aggregates to a local space, and improve matching efficiency.
[0087] Step S53: Construct a comprehensive similarity evaluation function, as shown in formula (1). Where D morphology The normalized Euclidean distance for spatial location is shown in formula (2); D spatial The differences in morphological characteristics are shown in formulas (3) and (4); D orientation The difference in directional characteristics is shown in formula (5);
[0088] (1);
[0089] (2);
[0090] (3);
[0091] (4);
[0092] (5);
[0093] In the formula: S total D is the total score for similarity evaluation. morphology D is the normalized Euclidean distance of spatial location. spatial D represents the morphological differences constructed based on volume, surface area, equivalent diameter, and / or sphericity. orientation To differentiate the directional features based on polar angle and azimuth angle, ws w m w o For the corresponding weight coefficients, w s +w m +w o =1, S is the sphericity, D eq Where V is the equivalent particle size, A is the volume, and ΔS and S are the surface areas. max ΔD represents the difference and maximum value of the sphericity of the bone fragments to be matched between the two CT sequences. eq D eqmax Let ΔV and V represent the difference and maximum value of the equivalent particle size of the aggregate to be matched between the two CT sequences, respectively. eqmax Let x1, x2, y1, y2, z1, and z2 represent the difference and maximum volume of the aggregate to be matched between the two CT sequences, respectively. Let x1, x2, y1, y2, z1, and z2 refer to the three-dimensional coordinates of the centroid of the aggregate to be matched between the two CT sequences, respectively. Let ΔA, A eqmax d represents the difference and maximum surface area of the aggregate to be matched between two CT sequences, respectively. spatial Let d be the distance between the centroids. max The aggregate movement threshold is set to 50 mm here, φ is the polar angle, Δφ is the polar angle difference between the two CT sequences, θ is the azimuth angle, and Δθ is the azimuth angle difference between the two CT sequences.
[0094] Step S54: Manually match 100 aggregates in the 3D reconstruction models of each gradation state #1 and state #2, extract the corresponding feature parameters and organize the variable statistics. Test 66 weight combinations through grid search to determine the optimal weight combination w. s =0.2、w m =0.7、w o =0.1, at which point the matching success rate of all three gradations reaches 100%.
[0095] Step S55: Set a feature difference threshold and use a two-way matching strategy to match aggregates in adjacent compaction states to ensure the uniqueness of the aggregate correspondence and avoid multiple aggregates being misidentified as the same aggregate.
[0096] Step S56: Starting from the aggregate in state #1, propagate backward based on the matching relationship between adjacent states to construct a complete motion trajectory that runs through all five compaction states.
[0097] Step S57: Introduce a reuse detection mechanism to interrupt invalid paths and retain only aggregates that can be successfully identified in all compaction states, forming a complete motion trajectory of continuous state.
[0098] Step S58: Output the full-chain tracking results containing motion parameters such as spatial displacement distance and direction angle change of each aggregate, and verify the absence of duplicate identification through set operation, and calculate the matching rate of each state to evaluate the tracking reliability.
[0099] Thus, this embodiment completes the precise tracking of aggregates in asphalt mixtures under different compaction states, obtaining motion parameters such as migration distance and rotation angle of each aggregate throughout the entire process from the initial compaction state to the final compaction state. These tracking results can be directly used for asphalt mixture compaction mechanism analysis, gradation optimization design, and to provide sensor deployment basis and detection parameter guidance for intelligent compaction detection technology.
[0100] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for tracking the movement of aggregates in asphalt mixtures, characterized in that, Includes the following steps: Step S1: Pre-embed a cross-shaped threaded nail at the center of the bottom of the sample as a measurement target point, perform multiple rotational compaction on the asphalt mixture sample, and perform a CT scan after each compaction to obtain a CT image sequence of the same sample under multiple different compaction states. Step S2: Using the first-state CT image obtained from the first scan as a reference, the CT images of the remaining states are rotated and corrected by comparing the spatial positions of the measurement target points in the CT data of each state, so as to eliminate the spatial position deviation caused by repeated assembly and disassembly of the specimen. Step S3: Segment the corrected CT images of each state, identify and extract the aggregate phase in the images; Step S4: Based on the segmentation results, perform three-dimensional reconstruction of the CT images under each compaction state and extract the feature parameters of each aggregate; the feature parameters include spatial location features, morphological features and principal axis direction features; Step S5: Based on the feature parameters, perform multi-feature weighted matching on aggregates in adjacent compaction states to establish a unique correspondence between aggregates across compaction states, and thereby construct the complete movement trajectory of a single aggregate through all compaction states. In step S3, an improved deep learning model is used to segment the CT image; the improved deep learning model introduces a feature enhancement mechanism in the encoding stage to improve the segmentation accuracy of aggregate edges and fine aggregates. The improved deep learning model is the Nested U-Net architecture, which introduces residual connections and channel attention mechanisms in the downsampling convolution module of the encoder. Specifically, the input feature map X is convolved, batch normalized, and ReLU activated in sequence, and then convolved and normalized again to extract features F(X). Then, F(X) is residually connected with the original input X to obtain Y. Y is then fed into the channel attention module to weight different feature channels and output the feature Yse enhanced by the attention mechanism. Step S4 specifically includes: The segmented aggregate phase image is reconstructed in three dimensions to establish a three-dimensional voxel model of the aggregate; Connectivity analysis was performed on the three-dimensional voxel model to identify and label each individual aggregate particle; Extract the characteristic parameters of each aggregate, including morphological characteristic parameters, spatial position characteristic parameters, and principal axis direction characteristic parameters; The principal axis direction characteristic parameters are obtained through principal axis inertial analysis, specifically including: Construct the inertial tensor of aggregate voxels with respect to their center of mass; The inertial tensor is subjected to eigenvalue decomposition to obtain its eigenvalues and corresponding eigenvectors; Take the eigenvector corresponding to the largest eigenvalue as the principal axis direction of the aggregate; The morphological features include at least one of volume, surface area, equivalent diameter, and sphericity; the spatial position feature is the three-dimensional centroid coordinates of the aggregate; the principal axis direction feature includes the relationship between the principal axis and... z The included angle of the shaft and the spindle at xy Projection of a plane and x The included angle of the axis; In step S5, the multi-feature weighted matching specifically includes: A KD-Tree spatial index is constructed based on spatial location features to perform fast nearest neighbor search for aggregates in adjacent compaction states; A comprehensive similarity evaluation function is constructed, which performs a weighted summation of spatial location differences, morphological feature differences, and principal axis direction feature differences; A one-to-one correspondence between adjacent compacted aggregates is established through a two-way matching strategy.
2. The method for tracking the movement of aggregates in asphalt mixtures according to claim 1, characterized in that, Step S1 specifically includes: A cross-shaped threaded nail is pre-embedded at the center of the bottom of the rotary compaction mold as a measurement target point; the asphalt mixture is loaded into the mold and rotary compacted. After reaching the first preset number of times, the machine is stopped, the mold is removed, and the first CT scan is performed, marked as state #1; after scanning, the surface of the specimen is coated with lubricating oil and then reloaded into the mold, and heated to restore the rheological properties of the asphalt mortar; the heated specimen is rotary compacted again to the second preset number of times, and after demolding, a second CT scan is performed, marked as state #2; the above compaction, heating, and scanning steps are repeated until a target number of CT image sequences of different compaction states are obtained.
3. The method for tracking the movement of aggregates in asphalt mixtures according to claim 1, characterized in that, Step S2 specifically includes: The spatial position of the cross-shaped threaded nail in the CT image of state #1 is used as the reference; Taking the axial direction of the cylindrical specimen as z A spatial coordinate system is established with the center of the bottom of the specimen as the origin O. The spatial position of the cross-shaped threaded nail in the CT image of state #N is compared with that of state #1. xy Deflection angle in the plane; The CT image of state #N is rotated and corrected according to the deflection angle so that the position of the cross-shaped screw in the CT image of each state is consistent with the reference state.
4. The method for tracking the movement of aggregates in asphalt mixtures according to claim 1, characterized in that, The comprehensive similarity evaluation function is: （1）； In the formula: S total D is the total score for similarity evaluation. morphology D is the normalized Euclidean distance of spatial location. spatial D represents the morphological differences constructed based on volume, surface area, equivalent diameter, and / or sphericity. orientation To differentiate the directional features based on polar angle and azimuth angle, w s , w m , w o For the corresponding weight coefficients, and w s + w m + w o =1.