A method for digital reconstruction of microstructure of cementitious materials and prediction of mechanical properties

Abstract

The application discloses a kind of cementitious material mesostructure digital reconstruction and mechanical property prediction method, belongs to building material nondestructive testing and computer vision analysis.This method is first to material section image pre-processing and extracts one-dimensional gray sequence, training is carried out using one-dimensional convolutional neural network (1D-CNN), realize aggregate, matrix and cementation surface pixel level identification and two-dimensional image reconstruction.On this basis, the area proportion of each component and the distribution uniformity such as physical characteristics are automatically counted, and the deep pattern features extracted by the network are fused to form a multi-dimensional input feature vector.Finally, by establishing a multi-modal fusion regression model, continuous prediction from mesostructure image to macroscopic compressive strength is realized, which provides an intelligent solution for rapid and objective evaluation of concrete quality.The method integrates image features and physical statistical information, improving the accuracy and intelligence level of prediction.

Description

Technical Field

[0001] This invention relates to the field of non-destructive testing and computer vision analysis of building materials, and in particular to a method for digital reconstruction of the microstructure and prediction of the mechanical properties of cementitious materials. Background Technology

[0002] As a heterogeneous multiphase composite material, the macroscopic mechanical properties and durability of concrete largely depend on its microstructural characteristics, particularly the distribution and bonding state of aggregates, cement mortar, and the bonding interface (interfacial transition zone) between them. Currently, the analysis and evaluation of concrete microstructure mainly rely on manual interpretation and measurement of microscopic images. However, this traditional manual approach has significant drawbacks: firstly, manual interpretation is extremely time-consuming and labor-intensive, making it difficult to meet the timeliness requirements of large-scale engineering testing; secondly, the interpretation results are greatly influenced by subjective factors, resulting in poor consistency among the outputs of different testing personnel, leading to a lack of standardized quantitative data.

[0003] With the development of computer vision technology, image segmentation techniques based on convolutional neural networks (CNNs) have been increasingly applied to the analysis of material microstructures. Although existing automatic segmentation methods based on ordinary 2D-CNNs have achieved a degree of automation, they still suffer from the following problems: In images of concrete microsections, the grayscale distributions of aggregates, matrix, and cementitious surfaces are extremely similar, and their texture features are highly similar. Traditional image processing algorithms and conventional CNN networks often struggle to extract discriminative deep features under low contrast, resulting in severely insufficient accuracy in identifying key weak areas such as cementitious surfaces. Cementitious surfaces in images typically appear as extremely long, thin, and spatially continuous microstructures. Conventional 2D-CNN architectures (such as networks containing numerous pooling layers) are prone to losing such high-frequency detail information during feature reduction, making the network insensitive to capturing the structural features of long, thin, and continuous cementitious surfaces. The segmented boundaries are often blurry or broken, failing to accurately reconstruct the material's contact state. To improve accuracy, existing methods often increase network depth and width, resulting in a huge number of model parameters and complex structures. This not only incurs huge computational overhead, making it time-consuming to process a single cross-sectional image, but also makes it difficult to deploy on portable devices or edge computing terminals, limiting its practical application in engineering fields.

[0004] In summary, there is a lack of a dedicated method in the field that can maintain lightweight design while achieving high-precision segmentation. Existing technologies struggle to meet the need for rapidly and accurately quantifying the area proportions of each component (especially minute cementation surfaces) from single or a small number of cross-sectional images, thereby assessing material homogeneity. Therefore, there is an urgent need for a digital reconstruction method that can effectively enhance the extraction of subtle features, reduce computational complexity, and improve the accuracy of cementation surface identification, in order to improve the quality and efficiency of concrete quality assessment. Summary of the Invention

[0005] The purpose of this invention is to address the technical deficiencies in the existing technology by providing a method for digital reconstruction of the microstructure and prediction of the mechanical properties of cementitious materials.

[0006] The technical solution adopted to achieve the purpose of this invention is: A method for digital reconstruction of the microstructure and prediction of the mechanical properties of cementitious materials includes the following steps: Step 1: Obtain a color microscopic image of the cementitious material, and after preprocessing, obtain a preprocessed image; Step 2: Label the preprocessed image to form a label map, scan and extract pixel gray values, and convert them into a one-dimensional gray value sequence and its corresponding label sequence; Step 3: Input the one-dimensional grayscale sequence and its corresponding label sequence into the one-dimensional convolutional neural network 1D-CNN model for training to obtain the fine component recognition model. Input the one-dimensional grayscale sequence and its corresponding label sequence into the fine component recognition model to output high-dimensional deep graphic features. Step 4: Convert the microscopic image of the cementitious material to be tested into a one-dimensional sequence, input it into the microscopic component identification model to obtain the prediction sequence, and reconstruct the prediction sequence in reverse according to the size of the image to be tested to reconstruct a two-dimensional segmented image containing component semantics. Step 5: Obtain the true values ​​of the macroscopic mechanical properties of the cementitious material sample corresponding to the two-dimensional segmented image; based on the two-dimensional segmented image, count the number of pixels of each component in the cementitious material, calculate the area ratio and distribution characteristics of each component in the two-dimensional segmented image, and obtain the multidimensional feature vector X. phy ; the multidimensional feature vector X phy After being concatenated and fused with high-dimensional deep graphic features, a multi-dimensional input feature vector X is formed. fusion And establish paired datasets by mapping them to the true values ​​of macroscopic mechanical properties; Step 6: After training the intensity prediction regression model using the paired dataset, input the multi-dimensional feature vector X of the fine-grained image to be tested. fusion Input into the strength prediction regression model, and output the predicted values ​​of macroscopic mechanical properties.

[0007] In the above technical solution, in step 1, a color microscopic image of the cut or polished surface of the cementitious material is obtained.

[0008] In the above technical solution, in step 1, the preprocessing is: converting the color detail image into a grayscale image; performing image enhancement and noise reduction processing on the grayscale image. Preferably, the image enhancement is performed using the Limit Contrast Adaptive Histogram Equalization (CLAHE) algorithm.

[0009] In the above technical solution, in step 2, a sliding window method is used to capture and convert the data into a one-dimensional grayscale sequence.

[0010] In the above technical solution, in step 3, the 1D-CNN model includes an input layer, a feature extraction layer, a fully connected layer, and an output layer. The feature extraction layer contains several cascaded one-dimensional convolutional modules, and the one-dimensional convolutional modules include convolutional layers, batch normalization layers, activation layers, and pooling layers. The 1D-CNN model is trained using a combination of weighted cross-entropy loss, which gives a larger weight penalty to the prediction error of the glued surface category, forces the model to focus on the feature learning of weak interfaces, accelerates training convergence, and prevents gradient vanishing.

[0011] In the above technical solution, step 4 involves introducing morphological opening and closing operations to smooth the component boundaries of the two-dimensional segmented image.

[0012] In the above technical solution, in step 5, the high-dimensional deep graph features are reduced to a one-dimensional implicit graph feature vector X by global average pooling. img The multidimensional feature vector X phy With implicit graphical eigenvector X img The multi-dimensional input feature vector X is formed by splicing and fusing along the channel dimension. fusion The multidimensional feature vector X phy =[ R ITZ , D agg , C ITZ ],in, R ITZ This represents the percentage of the area of ​​the bonded surface. D agg It is an index of uniformity. C ITZ This represents the average length of the connected path.

[0013] In the above technical solution, in step 6, the intensity prediction regression model includes an input layer, two cascaded fully connected hidden layers, and a linear output layer; the two hidden layers use the ReLU activation function, and the output layer uses the linear activation function, with the forward propagation relationship as follows: H (1) = ReLU (W (1) X fusion +b (1) ); H (2) = ReLU (W (2) H (1) +b (2) ); Y *=W out H (2) +b out ; Among them, W (1) W (2) and W out These are the weight matrices for the first hidden layer, the second hidden layer, and the output layer, respectively. (1) b (2) and b out For the corresponding bias, H (1) and H (2) Y is the output of two hidden layers. * These are predicted values ​​for macroscopic mechanical properties.

[0014] In the above technical solution, in step 6, the mean squared error loss function is used to accurately calculate the deviation between the predicted value and the true value, and the weights of the intensity prediction regression model are optimized separately through backpropagation, or joint optimization training is performed end-to-end with the micro-component identification model.

[0015] Another aspect of the present invention includes an electronic device comprising: one or more processors; and a memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the method for digital reconstruction of the microstructure and prediction of the mechanical properties of cementitious materials.

[0016] Another aspect of the present invention includes a computer-readable storage medium storing computer-executable instructions, which, when executed, are used to implement the method for digital reconstruction of the microstructure and prediction of the mechanical properties of cementitious materials.

[0017] Another aspect of the present invention includes a computer program product comprising computer-executable instructions, which, when executed, are used to implement the method for digital reconstruction of the microstructure and prediction of the mechanical properties of cementitious materials.

[0018] Compared with the prior art, the beneficial effects of the present invention are: This invention utilizes the strong feature extraction capability of one-dimensional convolution kernels along the sequential scanning direction to keenly capture subtle gradient changes in grayscale on both sides of the cemented surface. Compared to 2D-CNNs, which are prone to losing elongated edge features during pooling, this method can more completely preserve the continuity and geometric shape of the cemented surface.

[0019] Transforming two-dimensional matrix operations into one-dimensional vector operations significantly reduces the number of parameters and floating-point operations in the network model. This results in a substantial reduction in model training time and faster inference speed, making it suitable for embedding into portable detection devices with limited computing power or real-time analysis systems in engineering sites.

[0020] This method automates the process from raw image acquisition to final quality assessment, eliminating subjective errors caused by human interpretation to a certain extent and providing rapid and objective data support for optimizing the mix proportions of cementitious materials and controlling construction quality.

[0021] This method employs an automated processing flow from original image acquisition to final quality evaluation, eliminating the need for manual intervention and removing subjective errors from human interpretation. It provides rapid and objective data support for optimizing the mix proportions of cementitious materials and controlling construction quality, while avoiding complex three-dimensional reconstruction and focusing on two-dimensional cross-sectional information that is more readily available in engineering practice, making it more practical. Attached Figure Description

[0022] Figure 1 This is a diagram of the overall architecture of digital reconstruction and intensity prediction of the present invention.

[0023] Figure 2 This is a flowchart of the concrete image processing of the present invention.

[0024] Figure 3 This is a flowchart of the intensity prediction process of the present invention.

[0025] Figure 4 This is a training framework diagram for the intensity prediction regression model of the present invention. Detailed Implementation

[0026] The present invention will be further described in detail below with reference to specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0027] like Figures 1-4 As shown, a method for digital reconstruction of the microstructure and prediction of the mechanical properties of cementitious materials includes the following steps: Step 1: Obtain a color microscopic image of the cut or polished surface of the cementitious material (such as concrete), and convert the color microscopic image into a grayscale image; perform image equalization enhancement and noise reduction processing on the grayscale image. Preferably, the image enhancement is performed using the Limit Contrast Adaptive Histogram Equalization (CLAHE) algorithm to obtain a preprocessed image.

[0028] Step 1 aims to acquire clear color detail images and suppress noise. The specific process is as follows: After cutting, grinding, and polishing the concrete test block, a high-resolution color detail image of the cross-section of the test block with a resolution of no less than 1200 dpi is acquired using an optical scanner. This image is then converted to a single-channel grayscale image. To address uneven cross-sectional illumination, a contrast-limited adaptive histogram equalization (CLAHE) algorithm is used for enhancement (setting the contrast limit threshold to 3.0 and the grid size to 8x8). This enhances local texture details while avoiding excessive noise amplification in uniform matrix areas, resulting in new grayscale values.

[0029] in, I new ( x,y () represents the enhanced new grayscale value; T 1, T 2, T 3, T 4 represent the transformation functions of the block center; dx and dy These are the calculations of the distance of pixel P relative to the center of the top-left corner block; I old This is the original grayscale value.

[0030] Step 2: Label the preprocessed image, marking the aggregate, matrix, and cementing surface areas to form a label image. Scan the preprocessed image and label image in a specific direction (such as row by row or column by column) to extract pixel gray values. Use a sliding window method to extract the two-dimensional image and label image and convert them into a series of one-dimensional gray-level sequences (feature sequences) and their corresponding label sequences to construct a dataset.

[0031] Furthermore, step 2 aims to reduce the dimensionality of the preprocessed image (two-dimensional image) segmentation to a one-dimensional sequence classification problem. Relevant software is used to perform pixel-level semantic annotation on the enhanced grayscale image, classifying it into three categories: aggregate, cement matrix, and cementitious surface. A sliding window length of 32 pixels is set, and a one-dimensional grayscale vector is extracted along the horizontal or vertical scanning direction centered on any pixel to form a one-dimensional grayscale sequence. The category of the center pixel serves as the corresponding label sequence for the one-dimensional grayscale sequence. This method effectively transforms the elongated two-dimensional structure of the cemented surface into a grayscale abrupt change signal in a one-dimensional sequence.

[0032] Step 3: Input the dataset into a one-dimensional convolutional neural network (1D-CNN) model (including an input layer, a feature extraction layer, a fully connected layer, and an output layer). The feature extraction layer contains several cascaded one-dimensional convolutional modules, each including a convolutional layer, a batch normalization layer (BN layer), an activation layer (RELU), and a pooling layer. The convolutional layers include small-sized convolutional kernels. The 1D-CNN model is trained using a combination of weighted cross-entropy loss, which gives a larger weight penalty to the prediction error of the cemented surface category, forcing the model to focus on the feature learning of weak interfaces, accelerating training convergence, and preventing gradient vanishing, thus obtaining a fine component recognition model. Input the one-dimensional grayscale sequence into the 1D-CNN model to output high-dimensional deep graphic features. This step aims to achieve pixel-level high-precision classification and 2D image reconstruction, while extracting implicit graphic features to lay the foundation for subsequent fusion of explicit area proportions and macroscopic mechanical strength prediction. Feature extraction is closely linked to the final strength prediction, and its overall implementation process is as follows: Figure 2 As shown.

[0033] Preferably, to balance macroscopic aggregate morphology and microscopic cementation surface details, a multi-parallel branch structure is introduced, using convolutional kernels of different sizes in parallel to simultaneously extract local high-frequency grayscale abrupt changes and global macroscopic texture trends, which are then spliced ​​and fused along the channel dimension. The activation function employs a modified linear unit (MLU). ReLU Its mathematical expression is: ; in, x is any value in the feature vector output by the batch normalization layer.

[0034] To address the sample imbalance problem caused by the extremely low pixel proportion of the cemented surface, a weighted cross-entropy loss function is used during training. This function assigns a higher weight coefficient to the cemented surface category, forcing the network to focus on the classification error of the weak interface, thus significantly improving the sensitivity to the recognition of fine structures.

[0035] After the model training is completed and converged, the system enters the inference and reconstruction stage. The two-dimensional image to be tested is converted into a one-dimensional feature sequence according to the set scanning strategy, and batch input into the model for inference and output pixel-level predicted labels. The predicted sequence is reconstructed in reverse according to the original coordinates to restore the initial two-dimensional segmentation map containing component semantics. Finally, morphological opening and closing operations are performed to remove isolated noise, fill holes and smooth boundaries to obtain a high-precision two-dimensional digital reconstructed image.

[0036] Step 4: Convert the microscopic image of the cementitious material to be tested into a one-dimensional sequence, input it into the microscopic component identification model to obtain the prediction sequence, and reconstruct the prediction sequence in reverse according to the size of the image to be tested to reconstruct a two-dimensional segmented image containing component semantics. Morphological opening and closing operations are introduced to smooth the component boundaries of the two-dimensional segmented image in order to reconstruct the two-dimensional segmented image. The specific process is as follows: The microscopic image of the cementitious material to be tested is converted into a one-dimensional input sequence covering the entire image using the same scanning direction and serialization method as in step 2; this sequence is then batch-input into the microscopic component recognition model for inference, obtaining the classification prediction sequence for each pixel; based on the pixel's coordinate position in the original image, the one-dimensional prediction sequence is reverse-reconstructed and mapped into a two-dimensional matrix to generate a complete two-dimensional segmentation map. Furthermore, morphological opening and closing operations are used to remove isolated noise points and smooth component boundaries, improving the quality of the final reconstructed image.

[0037] Based on this, morphological opening and closing operations can be further used to remove isolated noise points, smooth component boundaries, and improve reconstruction quality.

[0038] Step 5: Obtain the true values ​​of the macroscopic mechanical properties of the cementitious material sample corresponding to the two-dimensional segmented image; based on the two-dimensional segmented image, count the number of pixels of the three components—aggregate, matrix, and cementing surface—and calculate the area ratio and distribution characteristics of each component in the two-dimensional segmented image to obtain the multidimensional feature vector X. phy ; the multidimensional feature vector X phy After being concatenated and fused with high-dimensional deep graphic features, a multi-dimensional input feature vector X is formed. fusion And establish paired datasets by mapping them to the true values ​​of macroscopic mechanical properties; Step 5 aims to establish a quantitative correlation between the detailed structure and macroscopic performance, and to extract the multidimensional feature vector X. phy It forms a multimodal fusion input with high-dimensional deep graphic features.

[0039] Preferably, the area ratio of each component in the two-dimensional segmented image is calculated to obtain the proportion of each component: ; in, R i Indicates the first i Components ( i The area ratio of ∈{aggregate, matrix, cementing surface} N i This indicates the total number of pixels in the image for that component. N total Represents the total number of pixels in an image; when i When it is an adhesive surface, R i That is, the proportion of the cemented surface area. R ITZ ,therefore R ITZ It is directly obtained from the ratio of the number of pixels on the adhesive surface to the total number of pixels in the image, and its calculation formula is: ; in, This represents the total number of pixels on the adhesive surface.

[0040] Furthermore, based on the obtained component proportions, a "microstructure-macroperformance" correlation database is established. First, the distribution characteristics of each component are evaluated: using a connected component analysis algorithm, the set of centroid coordinates of all aggregate particles in the image is extracted. ,in, m This represents the total number of aggregate particles detected in the image. Let be the coordinates of the centroid of the first aggregate particle in the two-dimensional image. For the first m The coordinates of the centroids of each aggregate particle in the two-dimensional image are then calculated. The Euclidean distance d from each centroid to its nearest neighbor centroid is then calculated.k Calculate all distances d k The coefficient of variation (CV) is calculated, and the distribution characteristics of each component are evaluated using the CV values. This process can be formulated as follows: ; in, σ d The standard deviation of the distance. μ d The mean of the distances. D agg The uniformity index is defined as such; the closer it is to 1, the more uniformly the aggregate is distributed in the matrix.

[0041] For the cemented surface, its connectivity and defects are evaluated: Based on the 2D segmented image reconstructed in step 4, all pixels classified as "cemented surfaces" are extracted to construct a binary image B(x,y) of the cemented surface. Cemented surface pixels are assigned a value of 1, while aggregate and matrix pixels are assigned a value of 0. An image thinning algorithm is used to skeletonize the binary image B(x,y). This process iteratively strips the edge pixels of the cemented surface region, shrinking the strip-shaped cemented surface into a single-pixel-width skeleton line S(x,y) while maintaining the original topological structure (connectivity). Based on this, the skeleton image S(x,y) is labeled with its neighboring connected components, and the number of pixels contained in each connected skeleton segment i is counted, defined as the connected path length L of that cemented surface segment. i By judging defects in the connectivity of the cemented surface, a definition is made. C ITZ The average connected path length is calculated using the following formula: ; in, N The total number of connected segments. L i For the first i The length of each segment.

[0042] The above three types of interpretable physical quantities are combined to form an explicit physical feature vector X. phy =[ R ITZ , D agg , C ITZ ].

[0043] Meanwhile, the high-dimensional deep graphic features output from the feature extraction layer of the 1D-CNN model in step 3 are reduced to a one-dimensional implicit graphic feature vector X by global average pooling. img X phy With X imgThe input feature vector X is formed by concatenating and fusing the data along the channel dimension. fusion The measured compressive strength of specimens from the same batch as the cross-section is used as the true label Y to establish a paired dataset.

[0044] Step 6: After establishing and training the intensity prediction regression model using the paired dataset, the corresponding features of the microscopic image to be tested are input into the intensity prediction regression model, which outputs continuous macroscopic mechanical performance prediction values. The intensity prediction regression model includes an input layer, two cascaded fully connected layers, and a linear output layer. The fully connected layers use the ReLU activation function, and the output layer uses a linear activation function. Their forward propagation relationship is as follows: H (1) = ReLU (W (1) X fusion +b (1) ); H (2) = ReLU (W (2) H (1) +b (2) ); Y * =W out H (2) +b out ; Among them, W (1) W (2) and W out These are the weight matrices for the first hidden layer, the second hidden layer, and the output layer, respectively. (1) b (2) and b out For the corresponding bias, H (1) and H (2) Y is the output of two hidden layers. * These are predicted values ​​for macroscopic mechanical properties.

[0045] Furthermore, the mean squared error loss function is used to accurately calculate the deviation between the predicted and true values. The weights of the intensity prediction regression model are optimized separately through backpropagation, or they are jointly optimized and trained end-to-end with the micro-component identification model.

[0046] Step 6 aims to construct and train a deep multimodal regression model to fully learn the nonlinear mapping relationship between the microscopic multidimensional fusion characteristics of cementitious materials and their macroscopic mechanical properties, ultimately achieving the prediction of material strength from microscopic structural images.

[0047] The specific steps are as follows: Construct a regression model containing an input layer, two cascaded fully connected layers, and an output layer. The fused feature vector X constructed in step S5 is used as the starting point. fusionAs input, linear weighting and non-linear activation are performed layer by layer in the hidden layers. After deep feature mapping, the last layer (output layer) of the network uses a linear activation function, without limiting the output amplitude, and directly outputs a continuous macroscopic mechanical strength prediction value Y. * This multilayer perceptron is a general regression structure. This invention constructs its input by fusing specific explicit physical features with implicit image features of 1D-CNN.

[0048] During the model training phase, the measured compressive strength of multiple pre-collected concrete test blocks is used as the true label Y, combined with the synchronously extracted multi-dimensional fusion feature vector X. fusion The model is trained using a multimodal pairing dataset. Mean squared error is used as the loss function to quantify the prediction intensity Y. * Deviation from the true intensity label Y: ; in, M This represents the total number of samples in the training batch. For the first j The true compressive strength of each sample For the first j The predicted compressive strength of each sample is calculated. The gradient is calculated using the backpropagation algorithm, and the weight parameters of the regression network are optimized individually. The specific training and optimization process of this model is as follows: Figure 3 As shown.

[0049] After model training is completed and convergence is verified, the system enters the automatic inference and prediction stage for unknown cementitious material samples. First, a microscopic digital reconstruction is performed on the microscopic image of the sample, and explicit physical features such as the proportion of cemented surface and aggregate distribution are automatically quantified and calculated. Simultaneously, high-dimensional deep graphic features extracted by one-dimensional convolution are combined to form a multi-dimensional input feature vector X for the sample. fusion Subsequently, the multi-dimensional input feature vector is directly input into the pre-trained deep multimodal regression model. The deep multimodal regression model performs rapid forward propagation calculations based on the solidified fully connected layer weights and biases, ultimately outputting a specific continuous value from end to end from the output layer, which is the predicted compressive strength of the cementitious material under test. The overall implementation process is as follows: Figure 4 As shown in Table 1: Table 1 Output Results

[0050] The above description is only a preferred embodiment of the present invention. It should be noted that, for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for digital reconstruction of the microstructure and prediction of the mechanical properties of cementitious materials, characterized in that, Includes the following steps: Step 1: Obtain a color microscopic image of the cementitious material, and after preprocessing, obtain a preprocessed image; Step 2: Label the preprocessed image to form a label map, scan and extract pixel gray values, and convert them into a one-dimensional gray value sequence and its corresponding label sequence; Step 3: Input the one-dimensional grayscale sequence and its corresponding label sequence into the one-dimensional convolutional neural network 1D-CNN model for training to obtain the fine component recognition model. Input the one-dimensional grayscale sequence and its corresponding label sequence into the fine component recognition model to output high-dimensional deep graphic features. Step 4: Convert the microscopic image of the cementitious material to be tested into a one-dimensional sequence, input it into the microscopic component identification model to obtain the prediction sequence, and reconstruct the prediction sequence in reverse according to the size of the image to be tested to reconstruct a two-dimensional segmented image containing component semantics. Step 5: Obtain the true values ​​of the macroscopic mechanical properties of the cementitious material sample corresponding to the two-dimensional segmented image; based on the two-dimensional segmented image, count the number of pixels of each component in the cementitious material, calculate the area ratio and distribution characteristics of each component in the two-dimensional segmented image, and obtain the multidimensional feature vector X. phy ; the multidimensional feature vector X phy After being concatenated and fused with high-dimensional deep graphic features, a multi-dimensional input feature vector X is formed. fusion And establish paired datasets by mapping them to the true values ​​of macroscopic mechanical properties; Step 6: After training the intensity prediction regression model using the paired dataset, input the multi-dimensional feature vector X of the fine-grained image to be tested. fusion Input into the strength prediction regression model, and output the predicted values ​​of macroscopic mechanical properties.

2. The method for digital reconstruction of the microstructure and prediction of mechanical properties of cementitious materials according to claim 1, characterized in that, In step 1, a color microscopic image of the cut or polished surface of the cementitious material is obtained; The preprocessing involves converting the color detail image into a grayscale image; and performing image enhancement and noise reduction on the grayscale image, wherein the image enhancement employs a contrast-limited adaptive histogram equalization algorithm.

3. The method for digital reconstruction of the microstructure and prediction of mechanical properties of cementitious materials according to claim 1, characterized in that, In step 2, a sliding window method is used to extract and convert the data into a one-dimensional grayscale sequence.

4. The method for digital reconstruction of the microstructure and prediction of the mechanical properties of cementitious materials according to claim 1, characterized in that, In step 3, the 1D-CNN model includes an input layer, a feature extraction layer, a fully connected layer, and an output layer. The feature extraction layer contains several cascaded one-dimensional convolutional modules. Each one-dimensional convolutional module includes a convolutional layer, a batch normalization layer, an activation layer, and a pooling layer. The convolutional layer includes a small-sized convolutional kernel. The 1D-CNN model is trained using a combination of weighted cross-entropy loss.

5. The method for digital reconstruction of the microstructure and prediction of mechanical properties of cementitious materials according to claim 1, characterized in that, In step 4, morphological opening and closing operations are introduced to smooth the component boundaries of the two-dimensional segmented image.

6. The method for digital reconstruction of the microstructure and prediction of mechanical properties of cementitious materials according to claim 1, characterized in that, In step 5, the high-dimensional deep graph features are reduced to a one-dimensional implicit graph feature vector X by global average pooling. img The multidimensional feature vector X phy With implicit graphical eigenvector X img The multi-dimensional input feature vector X is formed by splicing and fusing along the channel dimension. fusion The multidimensional feature vector X phy =[ R ITZ , D agg , C ITZ ],in, R ITZ This represents the percentage of the area of ​​the bonded surface. D agg It is an index of uniformity. C ITZ This represents the average length of the connected path.

7. The method for digital reconstruction of the microstructure and prediction of mechanical properties of cementitious materials according to claim 1, characterized in that, In step 6, the intensity prediction regression model includes an input layer, two cascaded fully connected hidden layers, and a linear output layer; the two hidden layers use the ReLU activation function, and the output layer uses the linear activation function, with the forward propagation relationship as follows: H (1) = ReLU (W (1) X fusion +b (1) ); H (2) = ReLU (W (2) H (1) +b (2) ); Y * =W out H (2) +b out ; Among them, W (1) W (2) and W out These are the weight matrices for the first hidden layer, the second hidden layer, and the output layer, respectively. (1) b (2) and b out For the corresponding bias, H (1) and H (2) Y is the output of two hidden layers. * These are predicted values ​​for macroscopic mechanical properties; In step 6, the mean squared error loss function is used to accurately calculate the deviation between the predicted value and the true value. The weights of the intensity prediction regression model are optimized separately through backpropagation, or they are jointly optimized and trained end-to-end with the micro-component identification model.

8. An electronic device, characterized in that, include: One or more processors; A memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the method for digital reconstruction of the microstructure and prediction of the mechanical properties of cementitious materials as described in any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that, The device stores computer-executable instructions, which, when executed, are used to implement the method for digital reconstruction of the microstructure and prediction of the mechanical properties of cementitious materials as described in any one of claims 1 to 7.

10. A computer program product, characterized in that, The aforementioned computer program product includes computer-executable instructions, which, when executed, are used to implement the method for digital reconstruction of the microstructure and prediction of the mechanical properties of cementitious materials as described in any one of claims 1 to 7.

Images