A method for predicting cracking performance of a road water-stable base based on neural network learning

Abstract

This invention discloses a method for predicting the cracking performance of road water-stabilized base courses based on neural network learning. The steps are as follows: First, CT scans are performed on cement-based specimens, and the images are segmented and reconstructed in three dimensions using software. The orientation distribution index and porosity of each particle size in the specimen are calculated. Second, the model is imported into ABAQUS for uniaxial tensile and compressive numerical tests to obtain its compression modulus and tensile modulus. Next, a neural network model is constructed, using the orientation distribution index and porosity of each particle size in the specimen as input data, and the compression modulus and tensile modulus as output data for training. Finally, core samples are drilled and reconstructed from cracked sections of the semi-rigid base course to calculate its material morphology distribution. This data is then input into the neural network model to predict the tensile and compressive properties of the material, determine its material performance, and predict the tensile and compressive resilient modulus of the semi-rigid base course of asphalt pavement through neural network learning, thereby guiding material design and construction.

Description

Technical Field

[0001] This invention relates to the fields of CT imaging three-dimensional reconstruction and neural networks, specifically a method for predicting the cracking performance of road water-stabilized base courses based on neural network learning. Background Technology

[0002] Currently, road engineers in my country often assume isotropic properties when analyzing the tensile and compressive properties of cement substrates. While this simplifies the calculation process, the results obtained by simply assuming isotropic cement substrates are questionable due to the varying distribution of aggregates, mortar, and voids within the substrate. There is currently limited research on the anisotropy of cement substrates both domestically and internationally, and no accurate standard exists to measure the magnitude of anisotropy. However, the maturity of non-destructive testing systems has made it possible to use X-ray machines for CT scans of cement substrates to understand their internal structure, providing a tool for future microscopic analysis of anisotropy. The current industry practice of assuming material isotropy makes it difficult to accurately determine the tensile and compressive anisotropy of water-stabilized materials through mechanical calculations, thus limiting the guidance for material design, construction, and maintenance decisions. Summary of the Invention

[0003] The technical problem to be solved by this invention is: to address the complex tensile and compressive properties of existing water-stabilized base course materials due to the distribution of aggregates and voids, this invention provides a method for predicting the cracking performance of road water-stabilized base courses based on neural network learning, so as to better guide the maintenance decisions of cracked base courses.

[0004] To address the above technical problems, this invention provides the following technical solution: a method for predicting the cracking performance of road water-stabilized base courses based on neural network learning, comprising the following steps:

[0005] S1. Scan the image of the cement-based specimen to obtain a three-dimensional reconstruction model;

[0006] S2. Based on the scanned images of cement-based events, the morphological distribution index of each particle size is calculated using the structure tensor theory; at the same time, the porosity of the three-dimensional reconstruction model is calculated.

[0007] S3. Based on the three-dimensional reconstruction model obtained in step S1, assign values ​​to the material properties to obtain a cylindrical cement-based finite element model with voids. Then, perform uniaxial tension and compression on the cylindrical cement-based finite element model to obtain the tensile and compressive strength and tensile and compressive modulus of the specimen.

[0008] S4. Construct a neural network for predicting the cracking performance of road water-stabilized base course. Take the morphological distribution index and porosity of each particle size of the specimen obtained in step S2 as input, and the tensile and compressive strength and tensile and compressive modulus as output, and train the neural network for predicting the cracking performance of road water-stabilized base course to obtain a model for predicting the cracking performance of road water-stabilized base course.

[0009] S5. Using the road water-stabilized base cracking performance prediction model obtained in step S4, the performance of cracked road sections of semi-rigid base asphalt pavement is predicted.

[0010] Furthermore, the aforementioned step S1 specifically includes:

[0011] S1.1. The cement-based specimens were scanned axially at equal intervals using X-rays, and noise in the CT images was filtered out.

[0012] S1.2. Use the graythresh function of Matlab software to divide the grayscale histogram and generate a binary image of voids and aggregates;

[0013] S1.3. Use the imregionalmin and bwdist functions to calculate the center position of different aggregates, form watershed ridges, and set the area threshold of the particle size to divide and delete fine aggregates.

[0014] S1.4. Import the binary images of aggregates and voids into MIMICS software, stack them according to their spatial positions, solidify the closed surface, and obtain the three-dimensional reconstruction model through Boolean operations.

[0015] Furthermore, in the aforementioned step S2, calculating the morphological distribution indices of each particle size based on the scanned images of cement-based events using the structure tensor theory includes the following sub-steps:

[0016] S2.1. Using Matlab, extract the aggregate information from the i-th slice image of the specimen, and identify and calculate the number M and particle area P of the aggregates at that cross-section. (k) Then calculate the volume content V. 1i ;

[0017] S2.2 Calculate the ratio λ of the major and minor axes of the equivalent ellipse of the kth aggregate particle. (k) ;

[0018] Calculate the major axis orientation angle λ of the particle in two-dimensional coordinates. (k) The cross-sectional morphology distribution index of aggregate particle size for the i-th cross section is calculated using the structure tensor theory, as shown in the following formula:

[0019]

[0020] The morphological distribution indices of each particle size in the 3D reconstructed specimen model are as follows:

[0021]

[0022] Where L represents the number of cross sections of the cement-based specimen.

[0023] Furthermore, in step S2 above, the porosity of the three-dimensional reconstructed model is calculated as follows:

[0024]

[0025] Among them, V 2i The porosity of the i-th cross-section of the cement-based material is calculated by extracting aggregate information from the cement-based material using Matlab.

[0026] Furthermore, the aforementioned step S3 includes the following sub-steps:

[0027] S3.1 Mesh the 3D reconstructed model: First, perform adaptive surface meshing; then, detect the surface mesh elements and perform meshing according to the preset mesh element quality threshold.

[0028] S3.2. In Solidworks software, convert the STL format file to complete the solidification of coarse aggregate and voids, and then input it into ABAQUS software;

[0029] S3.3. Cement mortar is generated on the basis of the reconstructed model through Boolean operations. At the same time, the void part is removed, and the material properties of cement-based coarse aggregate and mortar are assigned to obtain a cylindrical cement-based finite element model with voids.

[0030] S3.4. Apply uniform compressive and tensile loads to a porous cement-based finite element model and calculate the tensile strength f of the cement-based specimen. t With compressive strength f c As shown in the following formula:

[0031]

[0032]

[0033] in, The maximum tensile and maximum compressive loads of the test are represented in N; A represents the cross-sectional area of ​​the cylinder in mm². 2 ;

[0034] S3.5. Apply compressive and tensile loads continuously and uniformly at a preset loading rate until the specimen fails. Extract the load-strain response curve of the model and calculate the axial compressive resilient modulus E of the cement-based specimen using the following formula. c With tensile resilient modulus E s :

[0035]

[0036]

[0037] in, These represent the maximum tensile load and maximum compressive load of the test, respectively, in N; D represents the specimen diameter; ε3 represents the axial tensile and compressive strain of the specimen when the load reaches the preset value, ε3=Δl / L.

[0038] Furthermore, in step S4 above, the input data of the neural network for predicting the cracking performance of road water-stabilized base courses—the morphological distribution indices and porosity of each particle size in the specimens—are normalized using the maximum-minimum method, as shown in the following formula:

[0039]

[0040] Among them, y j x represents the result after normalization of a single data point. j Indicates the value of a single data point; x jmax x represents the maximum value in the column containing the data. jmin This indicates the minimum value in the column containing the data.

[0041] Furthermore, in step S4 above, when constructing the neural network for predicting the cracking performance of road water-stabilized base courses, the hidden layer neurons are selected using the following formula:

[0042]

[0043] Where n is the number of neurons in the input layer, i.e., the morphological distribution index and porosity of each particle size in the specimen; m is the number of neurons in the output layer, i.e., the tensile and compressive strength and tensile and compressive modulus; and a is a constant between [1, 10].

[0044] Furthermore, in step S4 above, when training the neural network for predicting the cracking performance of road water-stabilized base courses, the activation function of the hidden layer of the BP neural network is the Sigmoid function. The output layer uses the Leaky ReLU function: ReLU(x) = max(0,x).

[0045] Compared with the prior art, the beneficial technical effects of the present invention using the above technical solution are as follows:

[0046] (1) The image scanning analysis and three-dimensional reconstruction method of cement-based specimens has realized the progress of numerical model in engineering restoration and is more meaningful for guiding actual engineering.

[0047] (2) Based on the material orientation distribution analysis method of the structural tensor theory, the complex arrangement and interlocking relationship of aggregates and voids inside the material is defined and extracted into numerical indicators for research.

[0048] (3) Neural network technology can process a large amount of data, has a fast computing speed, and has self-learning, self-organization and self-adaptation. It can fully approximate any complex nonlinear relationship and realize the correlation between the orientation distribution of the internal structure of water-stabilized base material and its tensile and compressive modulus. Attached Figure Description

[0049] Figure 1 This is a flowchart of the method of the present invention.

[0050] Figure 2 This is a binary image of aggregates obtained by processing a scanned image of cement-based materials.

[0051] Figure 3 This is a morphological feature diagram of the k-th aggregate in the material.

[0052] Figure 4 The diagram shows the neural network node model provided by this invention. Detailed Implementation

[0053] To better understand the technical content of the present invention, specific embodiments are described below in conjunction with the accompanying drawings.

[0054] In this invention, various aspects of the invention are described with reference to the accompanying drawings, in which numerous illustrative embodiments are shown. Embodiments of the invention are not limited to those depicted in the drawings. It should be understood that the invention is implemented through any of the various concepts and embodiments described above, as well as the concepts and embodiments described in detail below, because the concepts and embodiments disclosed herein are not limited to any particular implementation. Furthermore, some aspects of the invention disclosed may be used alone or in any suitable combination with other aspects of the invention disclosed.

[0055] refer to Figure 1 A method for predicting the cracking performance of road water-stabilized base courses based on neural network learning includes the following steps:

[0056] S1. Scan the image of the cement-based specimen to obtain a three-dimensional reconstructed model; including the following sub-steps (S1.1 to S1.4):

[0057] S1.1. X-rays are used to scan the cement-based specimens axially at equal intervals and CT image noise is filtered out. In practice, 100 sets of 150mm×150mm cylindrical cement-based specimens are prepared in the laboratory. First, the first cylindrical sample is fixed on the support plate and placed in the middle of the X-ray emission port and the detection device. The X-ray beam detection device is then calibrated. The scanning interval is preset to 0.4mm. The height of the support plate is adjusted so that the center line of the X-ray emission port intersects perpendicularly with the axis of the cylinder in the same plane, and the emitted X-ray beam passes through the top surface of the specimen. The detection device is started, and the collimator is used to process and form a cone beam X-ray to scan the sample. Cross-sectional image information is recorded every 1mm, imported into the computer, numbered, and saved.

[0058] After scanning the mixture specimen from top to bottom, place it horizontally and perform a lateral scan using the same method. Then repeat the above steps to scan the second sample until all samples have been scanned and tested. Save all cement-based specimen samples according to their numbers.

[0059] S1.2. The grayscale histogram is divided using the graythresh function in Matlab software to generate a binary image reference of the voids and aggregates. Figure 2 Specifically:

[0060] The CT scan images of the cement-based materials obtained from the scans were converted into grayscale images in Matlab. The grayscale values ​​of these images range from 0 to 255, with higher grayscale values ​​corresponding to higher material density. Therefore, the pixel grayscale values ​​of aggregates are close to 255, while the pixel grayscale values ​​of voids and background are close to 0. The pixel grayscale values ​​of cement mortar fall between those of aggregates and voids.

[0061] The histogram equalization function built into Matlab was used to enhance image contrast. Histogram equalization uses a cumulative function to adjust grayscale values, thereby enhancing image contrast. When the contrast of the region of interest (ROI) is similar, histogram equalization can effectively expand the commonly used brightness, making the brightness distribution within the cement-based image more uniform, thus ensuring the goal of enhancing local contrast without changing the overall contrast. A 3×3 two-dimensional median filter with a sliding window was used for image enhancement to reduce the impact of noise on image quality. The filtered CT grayscale image consists of three colors: gray-white, gray-black, and black, representing aggregate, mortar, and voids, respectively.

[0062] In OpenCV, the Otsu's method (OSTU) is selected to perform thresholding on CT images. Using the graythresh function in Matlab, the grayscale range [60, 255] pixels is selected to calculate the optimal grayscale threshold between the target and the background. Based on this threshold, a grayscale histogram is created, converting the grayscale image into a binary image. Next, morphological processing is performed on the binary image to eliminate image adhesion, remove internal voids, and smooth aggregate boundaries by removing burrs. Simultaneously, voids and material are binarized in the same way, with the grayscale range [0, 60] pixels designated as voids, generating a binary image of the material and voids.

[0063] S1.3. The `imregionalmin` and `bwdist` functions are used to calculate the center positions of different aggregates, forming watershed ridges. An area threshold is set for aggregates smaller than 2.36 mm to segment and delete fine aggregates. The binary images of coarse aggregates and mortar after morphological processing are segmented. The `imregionalmin` function in Matlab is used to calculate the center positions of different aggregate regions in the image, and the gray-level gradient calculation function is used to solve for the markers of the target region. The `bwdist` function is used to calculate the midpoint connecting the inner markers of the coarse aggregates, and this is used as the pixel point of the outer marker. Connecting the outer markers of different regions forms watershed ridges. 2.36 mm is used as the dividing particle size between coarse and fine aggregates. Fine aggregates smaller than 2.36 mm are mixed with cement mortar to form cement mortar. An area threshold is set to segment and delete fine aggregates from the image.

[0064] S1.4. Import the binary images of aggregates and voids into MIMICS software, stack them according to their spatial positions, solidify the closed surfaces, and obtain the 3D reconstruction model through Boolean operations. Specifically, the segmented CT scan images are processed in batches using a Matlab software editing program. The segmented coarse aggregate binary images and void binary images are imported into MIMICS software in JPG format in sequence. The image sequences are stacked according to their spatial positions to form a microscopic model of coarse aggregates and voids. The closed surfaces are solidified, and the surfaces of the aggregates and voids are smoothed again. The adhering parts between the aggregates are removed, etc. Finally, the final 3D reconstruction model is obtained through Boolean operations.

[0065] S2. Based on the scanned images of cement-based events, the morphological distribution indices of each particle size are calculated using the structure tensor theory, with reference to... Figure 3 Simultaneously, the porosity of the 3D reconstructed model is calculated; step S2 includes the following sub-steps:

[0066] S2.1. Using Matlab, the aggregate information of the i-th slice image of the specimen is extracted. Considering the content and mix ratio of aggregates of a certain particle size, the number M and particle area P of aggregates of a certain particle size under that cross section are identified and calculated by the software. (k) And based on this, its volume content V is calculated. 1i (%)

[0067] S2.2 Considering the flatness and morphological characteristics of the particles, calculate the ratio λ of the major and minor axes of the equivalent ellipse of the kth aggregate particle. (k) Considering the orientation distribution of the particles, calculate the orientation angle θ of the major axis of the particles in two-dimensional coordinates. (k) The cross-sectional morphology distribution index of aggregate particle size for the i-th cross section is calculated using the structure tensor theory, as shown in the following formula:

[0068]

[0069] The morphological distribution indices of each particle size in the 3D reconstructed specimen model are as follows:

[0070]

[0071] Where L represents the number of cross sections of the cement-based specimen. The morphological distribution index η of aggregates with particle sizes of 2.35mm, 4.75mm, 9.5mm, 13.2mm, 16.0mm, 19.0mm, and 26.5mm was calculated using aggregate information extracted from cement-based materials via Matlab. 2.35 η 4.75 η 9.5 η 13.2 η 16.0 η 19.0 η 26.5 .

[0072] On the other hand, the porosity of the three-dimensional reconstructed specimen model is:

[0073]

[0074] Among them, V 2i The porosity (%) in the i-th cross section of the cement-based material is calculated by extracting the aggregate information of the cement-based material using Matlab.

[0075] S3. Based on the three-dimensional reconstruction model obtained in step S1, assign values ​​to the material properties to obtain a cylindrical cement-based finite element model with voids. Then, perform uniaxial tension and compression on the cylindrical cement-based finite element model to obtain the tensile and compressive strength and tensile and compressive modulus of the specimen.

[0076] S3.1 Mesh generation of the 3D reconstructed model: First, adaptive surface mesh generation is performed. While ensuring that the calculation results are reliable enough, the number of mesh elements is minimized. Therefore, the mesh element length is 1. Then, the surface mesh elements are inspected. The ratio of the height to the base of the triangular mesh element is used as the standard for judging the quality of the mesh element. The threshold is set to 0.3. When the quality of the surface mesh element reaches the set threshold, the volume mesh generation is performed.

[0077] S3.2 In Solidworks software, convert the STL format file to x_b (or x_t) format to complete the solidification of coarse aggregate and voids, and then input it into ABAQUS software;

[0078] S3.3. Cement mortar is generated on the basis of the reconstructed model through Boolean operation, while the void part is removed. The material properties of cement-based coarse aggregate and mortar are assigned to obtain a 150mm×150mm cylindrical cement-based finite element model with voids.

[0079] S3.4. Uniaxial tensile-compression tests were conducted on a porous cement-based finite element model. Uniform compressive and tensile loads were first applied, and the tensile strength f of the cement-based specimen was calculated. t With compressive strength f c As shown in the following formula:

[0080]

[0081]

[0082] in, These represent the maximum tensile load and maximum compressive load of the test, respectively, in N; A represents the cross-sectional area of ​​the cylinder, in mm². 2 ;

[0083] S3.5. Apply compressive and tensile loads continuously and uniformly at a loading rate of 1 mm / min until the specimen fails. Extract the load-strain response curve of the model and calculate the axial compressive resilient modulus E of the cement-based specimen using the following formula. c With tensile resilient modulus E s :

[0084]

[0085]

[0086] in, These represent the maximum tensile load and maximum compressive load of the test, respectively, in N; D represents the specimen diameter, 150 mm; ε3 represents the load reaching the preset value of 0.3F. max The axial tensile and compressive strain of the specimen is ε3=Δl / L.

[0087] S4. Construct a neural network for predicting the cracking performance of road water-stabilized base courses, referring to... Figure 4 Using the morphological distribution index and porosity of each particle size of the specimen obtained in step S2 as input, and the tensile and compressive strength and tensile and compressive modulus as output, the neural network for predicting the cracking performance of road water-stabilized base course is trained to obtain the road water-stabilized base course cracking performance prediction model.

[0088] Using the maximum and minimum value method Orientation distribution index η of 2.35mm particle size aggregate 2.35 Orientation distribution index η of 4.75mm particle size aggregate 4.75 Orientation distribution index η of 9.5mm particle size aggregate 9.5 Orientation distribution index η of 13.2mm particle size aggregate 13.2 Orientation distribution index η of 16.0mm particle size aggregate 16.0 Orientation distribution index η of 19.0mm particle size aggregate 19.0 Orientation distribution index η of 26.5mm particle size aggregate 26.5 Specimen porosity V2, compression resilient modulus E c With tensile resilient modulus E s Tensile strength f t With compressive strength f c The data is normalized.

[0089] With formula Based on the formula (n is the number of neurons in the input layer, m is the number of neurons in the output layer, and a is a constant between [1, 10]), the number of hidden layer neurons is selected. In this embodiment, 4 neurons are used.

[0090] Using Matlab, the activation function of the hidden layer was determined to be the Sigmoid function. The output layer is a PB network model with the Leaky ReLU function: ReLU(x) = max(0,x).

[0091] The training function was set to `trainingdx`, and the performance function to `mse`. The network parameters were set as follows: 5000 epochs, a target error (goal) of 0.00000001, and a learning rate (lr) of 0.01. After setting the parameters, 100 sets of 150mm × 150mm cylindrical cement-based specimens were used as training data for this model. These were input into the neural network to begin training, resulting in a predictive model for the cracking performance of road water-stabilized base courses.

[0092] S5. Using the road water-stabilized base course cracking performance prediction model obtained in step S4, the performance of cracked sections of asphalt pavement semi-rigid base course is predicted. In this embodiment, in the cracked sections of asphalt pavement semi-rigid base course, core samples are taken from water-stabilized materials at different stages of crack initiation, convergence, cracking, crack propagation, and penetration. The samples are scanned and the particle size distribution index η and porosity V2 are calculated. The data are input into a neural network model for calculation to predict the tensile strength f of the water-stabilized base course. t Compressive strength f c Axial compression resilient modulus E c With tensile resilient modulus E s .

[0093] By predicting data to determine the strength and stiffness properties of water-stabilized base course materials at each cracking stage, we can assess the subsequent actual performance of different road sections and provide a basis for maintenance decisions for semi-rigid base courses.

[0094] While the present invention has been described above with reference to preferred embodiments, it is not intended to limit the invention. Those skilled in the art can make various modifications and refinements without departing from the spirit and scope of the invention. Therefore, the scope of protection of the present invention shall be determined by the claims.

Claims

1. A method for predicting the cracking performance of road water-stabilized base courses based on neural network learning, characterized in that, Includes the following steps: S1. Scan the image of the cement-based specimen to obtain a three-dimensional reconstruction model; S2. Based on the scanned images of cement-based events, calculate the morphological distribution indices of each particle size using the structure tensor theory; simultaneously calculate the porosity of the three-dimensional reconstruction model; wherein, the calculation of the morphological distribution indices of each particle size based on the scanned images of cement-based events using the structure tensor theory includes the following sub-steps: S2.

1. Using Matlab, the aggregate information of the i-th slice image of the specimen is extracted, and the number M of aggregates with different particle sizes and the particle area under the cross section are identified and calculated. Then calculate the volume content. ; S2.2 Calculate the ratio of the major and minor axes of the equivalent ellipse of the kth aggregate particle. ; Calculate the orientation angle of the major axis of the particle in two-dimensional coordinates. The cross-sectional morphology distribution index of aggregate particle size for the i-th cross section is calculated using the structure tensor theory, as shown in the following formula: The morphological distribution indices of each particle size in the 3D reconstructed specimen model are as follows: Where L is the number of cross sections of the cement-based specimen; The porosity of the 3D reconstructed model is calculated as follows: in, The porosity of the i-th cross section of the cement-based material is calculated by extracting aggregate information of the cement-based material using Matlab. S3. Based on the three-dimensional reconstruction model obtained in step S1, assign values ​​to the material properties to obtain a cylindrical cement-based finite element model with voids. Then, perform uniaxial tension and compression on the cylindrical cement-based finite element model to obtain the tensile and compressive strength and tensile and compressive modulus of the specimen. S4. Construct a neural network for predicting the cracking performance of road water-stabilized base course. Take the morphological distribution index and porosity of each particle size of the specimen obtained in step S2 as input, and the tensile and compressive strength and tensile and compressive modulus as output, and train the neural network for predicting the cracking performance of road water-stabilized base course to obtain a model for predicting the cracking performance of road water-stabilized base course. S5. Using the road water-stabilized base cracking performance prediction model obtained in step S4, the performance of cracked road sections of semi-rigid base asphalt pavement is predicted.

2. The method for predicting the cracking performance of road water-stabilized base courses based on neural network learning according to claim 1, characterized in that, Step S1 is as follows: S1.

1. The cement-based specimens were scanned axially at equal intervals using X-rays, and noise in the CT images was filtered out. S1.

2. Use the graythresh function of Matlab software to divide the grayscale histogram and generate a binary image of voids and aggregates; S1.

3. Use the imregionalmin and bwdist functions to calculate the center position of different aggregates, form watershed ridges, and set the area threshold of the particle size to divide and delete fine aggregates. S1.

4. Import the binary images of aggregates and voids into MIMICS software, stack them according to their spatial positions, solidify the closed surface, and obtain the three-dimensional reconstruction model through Boolean operations.

3. The method for predicting the cracking performance of road water-stabilized base courses based on neural network learning according to claim 2, characterized in that, Step S3 includes the following sub-steps: S3.1 Mesh the 3D reconstructed model: First, perform adaptive surface meshing; then, detect the surface mesh elements and perform meshing according to the preset mesh element quality threshold. S3.

2. In Solidworks software, convert the STL format file to complete the solidification of coarse aggregate and voids, and then input it into ABAQUS software; S3.

3. Cement mortar is generated on the basis of the reconstructed model through Boolean operations. At the same time, the void part is removed, and the material properties of cement-based coarse aggregate and mortar are assigned to obtain a cylindrical cement-based finite element model with voids. S3.

4. Apply uniform compressive and tensile loads to a porous cement-based finite element model and calculate the tensile strength of the cement-based specimen. With compressive strength As shown in the following formula: ， ， in, , These represent the maximum tensile load and maximum compressive load of the test, respectively, in N; A represents the cross-sectional area of ​​the cylinder, in N. ; S3.

5. Apply compressive and tensile loads continuously and uniformly at a preset loading rate until the specimen fails. Extract the load-strain response curve of the model and calculate the axial compressive resilient modulus of the cement-based specimen using the following formula. With tensile resilient modulus : ， ， in, , These represent the maximum tensile load and maximum compressive load of the test, respectively, in N; D represents the specimen diameter. This indicates the axial tensile and compressive strain of the specimen when the load reaches the preset value. .

4. The method for predicting the cracking performance of road water-stabilized base courses based on neural network learning according to claim 3, characterized in that, In step S4, the input data of the neural network for predicting the cracking performance of road water-stabilized base courses—the morphological distribution indices and porosity of each particle size in the specimens—are normalized using the maximum-minimum method, as shown in the following formula: ； in, This represents the result after normalization of a single data point. Indicates the value of a single data item; This represents the maximum value in the column containing the data. This indicates the minimum value in the column containing the data.

5. The method for predicting the cracking performance of road water-stabilized base courses based on neural network learning according to claim 4, characterized in that, In step S4, when constructing the neural network for predicting the cracking performance of road water-stabilized base courses, the following formula is used to select hidden layer neurons: ， in, The number of neurons in the input layer represents the morphological distribution indices and porosity of particles of various sizes in the specimen. The number of neurons in the output layer, i.e., tensile and compressive strength and tensile and compressive modulus. It is a constant between [1, 10].

6. The method for predicting the cracking performance of road water-stabilized base courses based on neural network learning according to claim 5, characterized in that, In step S4, when training the neural network for predicting the cracking performance of road water-stabilized base courses, the activation function of the hidden layer of the BP neural network is the Sigmoid function. The output layer uses the Leaky ReLU function: .

Images