Prediction method for microstructure and water permeability of pervious concrete based on artificial intelligence

Abstract

The invention discloses a pervious concrete microstructure and water permeability prediction method based on artificial intelligence, and belongs to the technical field of pervious concrete performance prediction.The pervious concrete microstructure and water permeability prediction method comprises the steps that S1, a characteristic coarse aggregate particle three-dimensional geometric model data set library is established; s2, establishing an aggregate three-dimensional close packing model in the pervious concrete structure; s3, re-importing the aggregate three-dimensional close packing model into the image modeling software MIMICS, and reversely generating a two-dimensional cross-sectional image which is a coarse aggregate image; s4, constructing a pervious concrete microstructure prediction model by adopting an artificial intelligence method; s5, inputting the coarse aggregate image into the pervious concrete microstructure prediction model built in the step S4 to obtain a pervious concrete prediction image and porosity information; s6, establishing a connected pore model required by permeation to realize the prediction of the water permeability performance; according to the artificial intelligence-based pervious concrete microstructure and water permeability prediction method provided by the invention, refined prediction of the water permeability coefficient of the pervious concrete is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of permeable concrete performance prediction, and in particular to a method for predicting the microstructure and permeability of permeable concrete based on artificial intelligence. Background Art

[0002] Pervious concrete is composed of three parts: coarse aggregate, cement matrix wrapped on the surface of coarse aggregate, and reserved permeable pores. Therefore, its various performances are directly dependent on the characteristics of coarse aggregate and cement-based wrapping layer. Since pervious concrete only has a small amount of cement matrix and is often formed by pressure, the coarse aggregate in its structure can be considered to be in a tightly packed state. For the selected coarse aggregate, its structure and performance in a tightly packed state are known, so the distribution characteristics and pore structure of the cement matrix, as well as the permeability coefficient determined by the pore structure, can be designed by adjusting the amount of cement paste in the pervious concrete mixture.

[0003] At present, traditional methods all adopt the volume method recommended in the national industry standard "Technical Code for Permeable Concrete Pavement" to design permeable concrete. This design method can only design the total porosity of permeable concrete and make a preliminary estimate of the permeability coefficient based on the single total porosity parameter. However, permeability does not rely solely on the total porosity parameter, but also depends on more complex pore structures such as pore size, pore connectivity, and connected porosity. The traditional method is a rough summary of engineering experience, resulting in a large deviation between the designed permeability and the actual performance, which makes it difficult to meet the requirements for refined design and prediction of permeability performance of permeable concrete. Summary of the invention

[0004] The purpose of the present invention is to provide a prediction method for the mesostructure and permeability of permeable concrete based on artificial intelligence to solve the problems existing in the above-mentioned background technology.

[0005] To achieve the above object, the present invention provides a method for predicting the mesostructure and water permeability of permeable concrete based on artificial intelligence, comprising the following steps: S1. Establish a three-dimensional geometric model data set library of characteristic coarse aggregate particles; S2. Customize aggregate gradation. According to the customized gradation, select aggregate from the characteristic coarse aggregate particle 3D geometric model data set library, import it into discrete element modeling software, and use the close packing control method to establish a 3D close packing model of aggregate in the permeable concrete structure. S3, re-importing the three-dimensional close-packed model of aggregate into the image modeling software MIMICS, and reversely generating a two-dimensional tomographic image, which is a coarse aggregate image; S4. Use artificial intelligence methods to build a mesostructure prediction model for permeable concrete; S5, inputting the coarse aggregate image into the permeable concrete mesostructure prediction model constructed in S4, and obtaining the prediction image and porosity information of the permeable concrete; S6. Establish the connected pore model required for infiltration to predict water permeability.

[0006] Preferably, step S1 specifically includes: obtaining tomographic images of single coarse aggregate particles of different particle sizes by X-ray tomography; performing three-dimensional reconstruction of the above images using image modeling software MIMICS to obtain a 3D model (STL format) of coarse aggregate of a single particle size, thereby establishing a three-dimensional geometric model data set library of aggregate particles of different particle sizes.

[0007] Preferably, step S4 specifically includes: S41. Establish a permeable concrete microscopic image dataset; S42. Build a mesostructure prediction model for permeable concrete.

[0008] Preferably, step S41 specifically includes: S411, collecting real permeable concrete tomographic CT scanning images as original image data of the data set; S412, performing preprocessing on the collected original image data to improve the quality and efficiency of annotation, including processing to improve image contrast (taking contrast-limited adaptive histogram equalization as an example); S413, selecting an image segmentation task model, using transfer learning technology, roughly annotating the preprocessed image, and obtaining rough annotation results of coarse aggregate, pores, and cement matrix; S414, adjusting the coarse annotation results based on manual annotation and cross-validation, assigning different label colors, and calculating the porosity corresponding to each original image data according to the fine annotation results; S415. Based on the fine annotation results, the coarse aggregate image (the image of only coarse aggregate) and the meso distribution image (the image containing coarse aggregate, cement matrix and pores) are extracted respectively, and an image information pair of coarse aggregate image-meso distribution image-porosity is established.

[0009] Preferably, the preprocessing of step S412 includes enhancing the contrast of the original image data by contrast-limited adaptive histogram equalization, specifically: Image segmentation: The original image Split into small pieces; Calculate the grayscale histogram of each small block as follows: ; in, Grayscale The normalized frequency of Grayscale The number of times it appears in this small block; Contrast Limit: Set the contrast threshold , for each gray level , Less than ,for: ; After renormalization, we get ,as follows ; Perform histogram equalization on each small block as follows: ; Map pixel values ​​to: ; in, and They are The minimum and maximum values ​​of is the total number of gray levels in the image; The original pixel in the small block Gray value Updated to , get the pixel points of the equalized image , and then the small blocks of images are spliced ​​and interpolated.

[0010] Preferably, step S42 is specifically as follows: S421, selecting a deep neural network according to the permeable concrete mesostructure image prediction task; S422. Define the total loss function during model training by combining the task requirements of permeable concrete mesoscopic structure image prediction and structural physical constraints; S423, model hyperparameter optimization, randomly dividing the water permeable microscopic image dataset established in step S41 into a training set, a validation set, and a test set in proportion, and optimizing the model by considering hyperparameters such as learning rate, regularization strategy, optimizer selection, and data enhancement method; S424, model training.

[0011] Preferably, the total loss function in step S422 includes multiple loss functions by weighted combination, and the total loss function is as follows: ; in, To combat losses, is the L1 loss, is the physical constraint loss, , , are the weight coefficients of adversarial loss, L1 loss and physical constraint loss respectively; they are expressed as: ; in, , They are the generator loss and the discriminator loss respectively, which are obtained by the following formula: ; ; in, and are the true value and the predicted value respectively; , The discriminator The predicted probability of the corresponding image true value and predicted value given; is the expectation of the true value; is the expectation of the predicted value; is the joint expectation of the true value and the predicted value; The L1 loss is: ; The physical constraint loss is: ; in, and represent the predicted and true values ​​of porosity, respectively.

[0012] Preferably, the model training of step S424 is specifically as follows: the coarse aggregate image in the image information pair of coarse aggregate image-micro distribution image-porosity established in S415 is used as the input of the model, the micro distribution image and porosity are used as the true values, and the predicted micro distribution image is used as the output, and the training is performed through the loss function defined in S422 with the average intersection-union ratio as the evaluation indicator.

[0013] Preferably, step S6 specifically comprises: The pore 3D structure was reconstructed using mimics software. The volume of a single voxel was known. The total pore volume was obtained by accumulating the number of voxels of the pore 3D model. The connectivity of the voxels of the pore 3D model was determined by the program, and the connected pore model was extracted. The connected pore model was generated into a finite element model of a hexahedral mesh, and imported into the fluid simulation software Fluent_3D to calculate the permeable flow rate of the pore model per unit time, and converted into the permeability coefficient of permeable concrete to predict the permeability performance. The conversion formula is as follows: ; in, is the water permeability coefficient, is the volume flow calculated by Fluent_3D, is the height of the permeable concrete model, is the permeable cross-sectional area of ​​the upper surface of the permeable concrete model, is the pressure head on the upper surface of the permeable concrete model.

[0014] Therefore, the present invention adopts the above-mentioned artificial intelligence-based prediction method for the mesostructure and permeability of permeable concrete, and uses artificial intelligence methods to embed a specifically distributed cement matrix into the aggregate distribution model, thereby realizing the prediction of the three-dimensional complex mesostructure of permeable concrete, and predicting the permeability coefficient of the structure through the method of numerical simulation of seepage, thereby realizing the design and regulation of the mesostructure and permeability of permeable concrete.

[0015] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a flow chart of the method for predicting the mesostructure and water permeability of permeable concrete based on artificial intelligence of the present invention; Figure 2 A real permeable concrete tomographic CT scan image according to an embodiment of the present invention; Figure 3 is a contrast-enhanced image according to an embodiment of the present invention; Figure 4 This is a schematic diagram of a rough labeling of an image according to an embodiment of the present invention; Figure 5 This is a schematic diagram of image fine-marking labels according to an embodiment of the present invention; Figure 6 Schematic diagram of image information pairs of coarse aggregate image-microscopic distribution image-porosity in an embodiment of the present invention, wherein (a) is a contrast-enhanced image; (b) is an image after fine annotation; (c) is a coarse aggregate image; and (d) is a microscopic distribution image. Figure 7 A schematic diagram of establishing a three-dimensional close-packed model of aggregates according to an embodiment of the present invention; Figure 8 Schematic diagram of establishment of connected pore model and numerical simulation according to an embodiment of the present invention. DETAILED DESCRIPTION

[0017] The following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention claimed for protection, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0018] See also Figure 1A prediction method for the mesostructure and water permeability of permeable concrete based on artificial intelligence is proposed. The method uses artificial intelligence to embed a cement matrix with a specific distribution into an aggregate distribution model, thereby realizing the prediction of the three-dimensional complex mesostructure of permeable concrete. The permeability coefficient of the structure is predicted by the method of numerical simulation of seepage, thereby realizing the design and regulation of the mesostructure and water permeability of permeable concrete. The method includes the following steps: S1. Establish a dataset library of three-dimensional geometric models of characteristic coarse aggregate particles: obtain the tomographic images of single coarse aggregate particles of different particle sizes through X-ray tomography; use the image modeling software MIMICS to perform three-dimensional reconstruction on the above images to obtain a 3D model of coarse aggregate of a single particle size (STL format), thereby establishing a dataset library of three-dimensional geometric models of aggregate particles of different particle sizes.

[0019] S2, custom aggregate gradation, select aggregate from the characteristic coarse aggregate particle 3D geometric model data set library according to the custom gradation, import it into the discrete element modeling software (PFC3D), and use the close packing control method to establish the three-dimensional close packing model of aggregate in the permeable concrete structure, such as Figure 7 As shown; S3, re-importing the three-dimensional densely packed model of aggregate into the image modeling software MIMICS, and reversely generating a two-dimensional tomographic image, i.e., a coarse aggregate image; S4. Use artificial intelligence methods to build a mesostructure prediction model for permeable concrete: S41. Establishing a permeable concrete microscopic image dataset: S411, collect real permeable concrete tomographic CT scan images as raw image data of the data set, such as Figure 2 As shown; S412, preprocessing the collected raw image data to improve the quality and efficiency of annotation, including image contrast enhancement processing (taking contrast-limited adaptive histogram equalization as an example), and the image after contrast enhancement is as follows: Figure 3 As shown; Since most original images are grayscale images, the contrast of the original images can be enhanced by contrast-limited adaptive histogram equalization, which includes the following steps: Image segmentation: The original image Split into small pieces; Calculate the grayscale histogram of each small block as follows: ; in, Grayscale The normalized frequency of Grayscale The number of times it appears in this small block; Contrast Limit: Set the contrast threshold , for each gray level , Less than ,for: ; After renormalization, we get ,as follows ; Perform histogram equalization on each small block as follows: ; Map pixel values ​​to: ; in, and They are The minimum and maximum values ​​of is the total number of gray levels in the image, usually 256; The original pixel in the small block Gray value Updated to , get the pixel points of the equalized image , and then the small blocks of images are spliced ​​and interpolated.

[0020] S413, select an image segmentation task model with better performance (Segment Anything is used as an example of this embodiment), use transfer learning technology to roughly annotate the preprocessed image, and obtain the rough annotation results of coarse aggregate, pores and cement matrix, such as Figure 4 As shown; S414, adjusting the rough labeling results based on manual labeling and cross-validation, the adjusted labels are as follows Figure 5 As shown, different label colors are assigned (taking coarse aggregate-white, cement matrix-gray, pore-black as an example), and the porosity corresponding to each original image data is calculated based on the fine annotation results; The difference between coarse labeling and fine labeling is that coarse labeling uses transfer learning technology to preliminarily identify aggregate distribution, which may have some recognition errors (for example: Figure 4 Therefore, in the fine annotation, manual error correction is used to remove images that are not aggregate distribution from the coarsely labeled images to achieve accurate division of the aggregate distribution area.

[0021] Attached Figure 4 and attached Figure 5 Difference: Roughly mark the aggregate distribution Figure 4 The cement matrix is ​​removed to form a fine mark Figure 5(Accurate distribution of aggregate).

[0022] S415, based on the fine annotation results, extract the coarse aggregate image (the image of only coarse aggregate) and the microscopic distribution image (the image containing coarse aggregate, cement matrix and pores), and establish the image information pair of coarse aggregate image-microscopic distribution image-porosity, such as Figure 6 shown.

[0023] S42. Build a mesostructure prediction model for permeable concrete: S421. Select a deep neural network based on the permeable concrete mesostructure image prediction task (taking pix2pix as an example); S422. Define the total loss function during model training in combination with the task requirements of permeable concrete mesoscopic structure image prediction and structural physical constraints; consider multiple loss functions by weighted combination, and the total loss function is as follows: ; in, To combat losses, is the L1 loss, is the physical constraint loss, , , are the weight coefficients of adversarial loss, L1 loss and physical constraint loss respectively; they are expressed as: ; in, , They are the generator loss and the discriminator loss respectively, which are obtained by the following formula: ; ; in, and are the true value and the predicted value respectively; , The discriminator The predicted probability of the corresponding image true value and predicted value given; is the expectation of the true value; is the expectation of the predicted value; is the joint expectation of the true value and the predicted value; The L1 loss is: ; The physical constraint loss is: ; in, and represent the predicted and true values ​​of porosity, respectively.

[0024] S423, model hyperparameter optimization, randomly dividing the water permeable microscopic image dataset established in step S41 into a training set, a validation set, and a test set in proportion, and optimizing the model by considering hyperparameters such as learning rate, regularization strategy, optimizer selection, and data enhancement method; S424, model training: the coarse aggregate image in the image information pair of coarse aggregate image-mesoscopic distribution image-porosity established in S415 is used as the input of the model, the mesoscopic distribution image and porosity are used as the true values, and the predicted mesoscopic distribution image is used as the output. The training is performed through the loss function defined in S422 and the average intersection-over-union ratio is used as the evaluation index.

[0025] S5, inputting the coarse aggregate image into the permeable concrete mesostructure prediction model constructed in S4, and obtaining the prediction image and porosity information of the permeable concrete; S6. Establish the connected pore model required for infiltration: Figure 8 As shown in the figure, the pore 3D structure is reconstructed by using mimics software. The volume of a single voxel is known. The total pore volume is obtained by accumulating the number of voxels of the pore 3D model. The connectivity of the voxels of the pore 3D model is determined by the program, and the connected pore model is extracted. The connected pore model is generated into a finite element model of a hexahedral grid, and imported into the fluid simulation software Fluent_3D to calculate the permeable flow rate of the pore model per unit time, and converted into the permeability coefficient of permeable concrete to predict the permeability performance. The conversion formula is as follows: ; in, is the water permeability coefficient, is the volume flow calculated by Fluent_3D, is the height of the permeable concrete model, is the permeable cross-sectional area of ​​the upper surface of the permeable concrete model, is the pressure head on the upper surface of the permeable concrete model.

[0026] Therefore, the present invention adopts the above-mentioned artificial intelligence-based prediction method for the mesostructure and permeability of permeable concrete. Compared with the existing technology, this method is based on the mesostructure characteristics of permeable concrete, combined with artificial intelligence image generation, three-dimensional reconstruction and numerical simulation technology, to realize the prediction of the mesostructure and permeability of permeable concrete, thereby meeting the requirements of refined prediction of permeable concrete.

[0027] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A method for predicting the mesostructure and water permeability of permeable concrete based on artificial intelligence, characterized in that: The following steps are involved: S1. Establish a three-dimensional geometric model data set library of characteristic coarse aggregate particles; S2. Customize aggregate gradation. According to the customized gradation, select aggregate from the characteristic coarse aggregate particle 3D geometric model data set library, import it into discrete element modeling software, and use the close packing control method to establish a 3D close packing model of aggregate in the permeable concrete structure. S3, re-importing the three-dimensional close-packed model of aggregate into the image modeling software MIMICS, and reversely generating a two-dimensional tomographic image, which is a coarse aggregate image; S4. Use artificial intelligence methods to build a mesostructure prediction model for permeable concrete; S5, inputting the coarse aggregate image into the permeable concrete mesostructure prediction model constructed in S4, and obtaining the prediction image and porosity information of the permeable concrete; S6. Establish the connected pore model required for infiltration to predict water permeability.

2. The method for predicting the mesostructure and water permeability of permeable concrete based on artificial intelligence according to claim 1 is characterized in that: Step S1 specifically includes: obtaining a tomographic image of a single coarse aggregate particle of different particle sizes by X-ray tomography; performing three-dimensional reconstruction on the above image using image modeling software MIMICS to obtain a 3D model of a coarse aggregate of a single particle size, thereby establishing a three-dimensional geometric model data set library of aggregate particles of different particle sizes.

3. The method for predicting the mesostructure and water permeability of permeable concrete based on artificial intelligence according to claim 1, characterized in that: Step S4 specifically includes: S41. Establish a permeable concrete microscopic image dataset; S42. Build a mesostructure prediction model for permeable concrete.

4. The method for predicting the mesostructure and water permeability of permeable concrete based on artificial intelligence according to claim 3 is characterized in that: Step S41 specifically includes: S411, collecting real permeable concrete tomographic CT scanning images as original image data of the data set; S412, preprocessing the collected original image data; S413, selecting an image segmentation task model, using transfer learning technology, roughly annotating the preprocessed image, and obtaining rough annotation results of coarse aggregate, pores, and cement matrix; S414, adjusting the coarse annotation results based on manual annotation and cross-validation, assigning different label colors, and calculating the porosity corresponding to each original image data according to the fine annotation results; S415. Based on the fine annotation results, the coarse aggregate image and the meso distribution image are extracted respectively, and an image information pair of coarse aggregate image-meso distribution image-porosity is established.

5. The method for predicting the mesostructure and water permeability of permeable concrete based on artificial intelligence according to claim 4 is characterized in that: The preprocessing of step S412 includes enhancing the contrast of the original image data by contrast-limited adaptive histogram equalization, specifically: Image segmentation: The original image Split into small pieces; Calculate the grayscale histogram of each small block as follows: ； in, Grayscale The normalized frequency of Grayscale The number of times it appears in this small block; Contrast Limit: Set the contrast threshold , for each gray level , Less than ,for: ； After renormalization, we get ,as follows ； Perform histogram equalization on each small block as follows: ； Map pixel values ​​to: ； in, and They are The minimum and maximum values ​​of is the total number of gray levels of the image; The original pixel in the small block Gray value Updated to , get the pixel points of the equalized image , and then the small blocks of images are spliced ​​and interpolated.

6. The method for predicting the mesostructure and water permeability of permeable concrete based on artificial intelligence according to claim 4, characterized in that: Step S42 is specifically as follows: S421, selecting a deep neural network according to the permeable concrete mesostructure image prediction task; S422. Define the total loss function during model training by combining the task requirements of permeable concrete mesoscopic structure image prediction and structural physical constraints; S423, model hyperparameter optimization, randomly dividing the permeable concrete mesoscopic image dataset established in step S41 into a training set, a validation set, and a test set in proportion, and optimizing the model by considering the hyperparameters; S424, model training.

7. The method for predicting the mesostructure and water permeability of permeable concrete based on artificial intelligence according to claim 6, characterized in that: The total loss function in step S422 includes multiple loss functions by weighted combination, and the total loss function is as follows: ； in, To combat losses, is the L1 loss, is the physical constraint loss, , , are the weight coefficients of adversarial loss, L1 loss and physical constraint loss respectively; they are expressed as: ； in, , They are the generator loss and the discriminator loss respectively, which are obtained by the following formula: ； ； in, and are the true value and the predicted value respectively; , The discriminator The predicted probability of the corresponding image true value and predicted value given; is the expectation of the true value; is the expectation of the predicted value; is the joint expectation of the true value and the predicted value; The L1 loss is: ； The physical constraint loss is: ； in, and represent the predicted and true values ​​of porosity, respectively.

8. The method for predicting the mesostructure and water permeability of permeable concrete based on artificial intelligence according to claim 6, characterized in that: The model training of step S424 is specifically as follows: the coarse aggregate image in the image information pair of coarse aggregate image-micro distribution image-porosity established in S415 is used as the input of the model, the micro distribution image and porosity are used as the true values, and the predicted micro distribution image is used as the output, and the training is performed through the loss function defined in S422 with the average intersection-union ratio as the evaluation indicator.

9. The method for predicting the mesostructure and water permeability of permeable concrete based on artificial intelligence according to claim 1, characterized in that: Step S6 is specifically as follows: The pore 3D structure was reconstructed using mimics software. The volume of a single voxel was known. The total pore volume was obtained by accumulating the number of voxels of the pore 3D model. The connectivity of the voxels of the pore 3D model was determined by the program, and the connected pore model was extracted. The connected pore model was generated into a finite element model of a hexahedral mesh, and imported into the fluid simulation software Fluent_3D to calculate the permeable flow rate of the pore model per unit time, and converted into the permeability coefficient of permeable concrete to predict the permeability performance. The conversion formula is as follows: ； in, is the water permeability coefficient, is the volume flow calculated by Fluent_3D, is the height of the permeable concrete model, is the permeable cross-sectional area of ​​the upper surface of the permeable concrete model, is the pressure head on the upper surface of the permeable concrete model.

Images