A deep learning-based concrete strength prediction method

Abstract

The application provides a concrete strength prediction method based on deep learning, comprising: acquiring a coarse aggregate surface image by using a three-dimensional scanning method, identifying coarse aggregate surface roughness, and extracting pore distribution characteristics of the coarse aggregate surface in combination with an image segmentation algorithm; performing feature extraction on an interface transition zone by using wavelet transform to obtain local features of the interface transition zone, and performing weighted calculation on the local features in combination with an adaptive weighted fusion algorithm to obtain enhanced local feature representation; fusing labeled microstructure parameters and macroscopic parameters of the whole concrete to construct a concrete strength prediction model and output a prediction value of the overall strength of the concrete; and obtaining local features contributing to the target strength by using the weight of the region corresponding to different strengths obtained by the adaptive weighted fusion algorithm.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to a method for predicting concrete strength based on deep learning. Background Technology

[0002] In concrete strength prediction research, the surface roughness and porosity distribution of coarse aggregates significantly affect the strength characteristics of the interfacial transition zone. The interfacial transition zone is a weak area between aggregates and cement paste in concrete, and its uneven strength distribution mainly stems from the complexity of aggregate surface morphology and the randomness of porosity distribution. When the aggregate surface roughness is high, the microstructure of the interfacial transition zone forms more local stress concentration points, leading to significant local differences in strength distribution. Simultaneously, the uneven porosity distribution further exacerbates this fluctuation in strength distribution. Deep learning models typically employ attention mechanisms to capture key features when predicting concrete strength. However, when the aggregate surface morphology leads to uneven strength distribution in the interfacial transition zone, the dynamic adjustment mechanism of the attention weight matrix faces challenges. Existing attention mechanisms often struggle to accurately capture subtle differences in local features when dealing with such complex scenarios, resulting in biased predictions. Furthermore, the coupling mechanism between the dynamic adjustment strategy of the attention weight matrix and local feature extraction remains unclear, further limiting the model's performance in practical applications. Therefore, designing an attention mechanism within a deep learning framework that can adaptively adjust the region of interest and accurately extract local features has become the core challenge in solving this technical problem. Summary of the Invention

[0003] This invention provides a deep learning-based method for predicting concrete strength, mainly comprising:

[0004] A three-dimensional scanning method was used to obtain images of the coarse aggregate surface, identify the surface roughness of the coarse aggregate, and extract the pore distribution characteristics of the coarse aggregate surface by combining image segmentation algorithm;

[0005] Based on the surface roughness and pore distribution characteristics, and using a deep learning network to calculate the mapping relationship between interface strength and aggregate morphology, the strength distribution in the interface transition zone is analyzed, the non-uniformity characteristics of the strength distribution are output, and the interface transition zone of the non-uniform strength distribution area is determined.

[0006] Scanning electron microscopy or X-ray computed tomography is used to identify local intensity differences in the interface transition zone. The microstructure and macrostructure parameters inside the interface transition zone are analyzed. If significant local intensity differences are found, the corresponding microstructure parameters are marked.

[0007] Wavelet transform is used to extract features from the interface transition region to obtain local features of the interface transition region. The local features are then weighted and calculated using an adaptive weighted fusion algorithm to obtain an enhanced local feature representation.

[0008] The marked microstructural parameters are fused with the macroscopic parameters of the concrete as a whole to construct a concrete strength prediction model and output the predicted value of the overall concrete strength.

[0009] The weights of regions corresponding to different intensities are predicted by an adaptive weighted fusion algorithm, and the local features contributing to the target intensity are obtained.

[0010] If the accuracy and recall of local features in the interface transition zone are improved to the target value after the adaptive weighted fusion algorithm is introduced, a strength distribution map of the concrete interior is generated. Combined with the surface roughness and porosity data of coarse aggregate, the degree of agreement between the strength prediction results and the actual strength test values ​​is analyzed.

[0011] Furthermore, a 3D scanning method is used to acquire images of the coarse aggregate surface, identify the surface roughness, and extract the porosity distribution characteristics of the coarse aggregate surface using an image segmentation algorithm. This includes: adjusting the light source intensity and scanning angle of the 3D scanner according to the reflective properties of the coarse aggregate surface; performing layer-by-layer scanning of the coarse aggregate using a continuous scanning method; and generating a first set of 3D point cloud data using point cloud data processing software. Noise filtering and data smoothing are performed on the first set of 3D point cloud data, and surface fitting is performed using the least squares method to obtain a second set of 3D point cloud data. The second set of 3D point cloud data is segmented using a watershed algorithm based on local curvature values. The standard deviation of the Euclidean distance between adjacent points within each segmented region is calculated to obtain the region roughness value. The segmented regions are classified according to a preset roughness threshold. If the region roughness value is greater than the high threshold, it is marked as a pitted region; if the region roughness value is less than the low threshold, it is marked as a flat region. Morphological processing is performed on the marked pitted regions, and connected regions are extracted using a region growing algorithm. The area, depth, and spacing parameters of each connected region are calculated. Density clustering algorithm is used to cluster connected regions, and porosity regions are identified based on area and depth thresholds. The shortest distance matrix between porosity regions is calculated. Network connectivity analysis is then performed on the porosity regions based on the shortest distance matrix to extract the spatial distribution characteristics of the porosity group, including pore density, connectivity, and directionality parameters.

[0012] Furthermore, based on the surface roughness and pore distribution characteristics, and using a deep learning network to calculate the mapping relationship between interface strength and aggregate morphology, the strength distribution of the interface transition zone is analyzed, the non-uniformity characteristics of the strength distribution are output, and the interface transition zone of the non-uniform strength distribution region is determined. This includes: constructing a first morphological feature vector containing surface height distribution, mean roughness, pore depth, pore density, and pore connectivity based on the surface roughness parameters and pore distribution characteristics of the coarse aggregate; normalizing the feature data to the interval [0, 1] using the min-max normalization method; and extracting the deep representation of the morphological features through a convolutional neural network. For the normalized deep features, a three-layer fully connected network is used for feature transformation, with the number of nodes in each layer set to 128, 64, and 32, respectively. The ReLU function is used as the activation function, and the interface bonding strength is predicted through the residual connection structure to obtain the interface micro-region bonding strength value. Cubic spline interpolation is performed on the predicted interface micro-region bonding strength value, and the region smoothing is performed using the opening operation morphological operator. The strength difference between adjacent micro-intervals is calculated to obtain the interface bonding strength gradient distribution map. The mean and standard deviation of intensity in local areas are calculated based on the intensity gradient distribution map. If the standard deviation is greater than a preset threshold, the area is determined to be a region with uneven intensity distribution, and its spatial coordinates are obtained. Connectivity analysis is performed on the marked uneven intensity distribution regions to calculate the area and perimeter of each connected region. A shape factor is used to characterize the morphology of the uneven regions, obtaining the intensity distribution characteristic parameters of the interface transition zone. A weight matrix is ​​established based on the intensity distribution characteristic parameters of the interface transition zone to weight the intensity values ​​of the uneven regions, obtaining the quantitative characterization results of the intensity distribution of the transition zone.

[0013] Furthermore, scanning electron microscopy (SEM) or X-ray computed tomography (CT) is used to identify local intensity differences in the interface transition zone, and to analyze the microstructural and macroscopic parameters within the interface transition zone. If significant local intensity differences are found, the corresponding microstructural parameters are labeled. This includes: performing layered scanning imaging of the interface transition zone using SEM; obtaining the first grayscale image sequence of the interface transition zone through backscattered electron signals; enhancing the image contrast using a histogram equalization method to obtain the second grayscale image sequence; performing image segmentation based on the second grayscale image sequence using the maximum entropy thresholding method; obtaining the binarized results of each layer; and using an iterative backprojection algorithm to perform three-dimensional reconstruction of the segmented image to obtain the reconstructed image of the interface transition zone structure. The reconstructed image is then labeled with voxels, with the voxel size set to cubic micrometers. Porosity is calculated through three-dimensional connected component analysis, and density clustering is used to identify the distribution of hydration product regions. For the hydration product regions, the contrast, entropy value, and second moment of the grayscale co-occurrence matrix are calculated, with orientation angles set to 0 degrees, 45 degrees, 90 degrees, and 135 degrees, to extract the morphological feature vectors of the hydration products. Based on morphological feature vectors, a support vector machine is used to classify hydration products, obtaining a distribution map of hydration product types and calculating the volume fraction of each type of hydration product. A local adaptive thresholding method is used to mark regions of intensity difference. If the intensity variance within a region is greater than twice the global variance, the porosity, hydration product type distribution, and volume fraction of that region are extracted as feature parameters. A feature mapping table is established for the feature parameters within the marked regions, recording the location coordinates of the intensity difference regions and their corresponding microstructural feature parameter values.

[0014] Furthermore, wavelet transform is used to extract features from the interface transition area to obtain local features. An adaptive weighted fusion algorithm is then used to weight these local features, resulting in an enhanced local feature representation. This includes: using Haar wavelets to decompose the interface transition area structure image into four layers, obtaining high-frequency and low-frequency sub-band images from layers one to four; acquiring detail feature maps in the horizontal, vertical, and diagonal directions from the high-frequency sub-bands; and acquiring approximate feature maps from the low-frequency sub-bands. The local variance and energy distribution of each sub-band image are calculated based on the wavelet coefficient amplitudes. Laplacian operator edge features and gray-level co-occurrence matrix texture features are extracted from the high-frequency sub-bands, while gray-level histograms and shape moment features are extracted from the low-frequency sub-bands. The information entropy value of each sub-band feature map is calculated, including gray-level entropy, texture entropy, and edge entropy. Weighting coefficients are set for the sub-band features based on the entropy value; the larger the entropy value, the higher the weighting coefficient. A bilateral filter is used to perform joint spatial and value domain filtering on the weighted sub-band feature maps, setting the spatial and value domain standard deviations to obtain the filtered feature maps. An adaptive thresholding method based on local means is used to enhance the filtered feature map. The gray-level mean of local regions in the image is calculated, and the threshold parameter is determined based on the mean. The feature weights are adjusted according to the variance of the local regions, and the variance contribution rate of each scale feature is calculated. The larger the variance contribution rate, the larger the corresponding weight. A weighted average algorithm is used to fuse multi-scale features, and the features are combined through weight normalization to generate a local feature description vector that includes edge, texture, and shape features.

[0015] Furthermore, the marked microstructural parameters are fused with the overall macroscopic parameters of the concrete to construct a concrete strength prediction model and output the predicted value of the overall concrete strength. This includes: normalizing the maximum and minimum values ​​of the marked interface transition zone microstructural parameters, converting porosity, hydration product distribution, and interface bonding degree to the [0, 1] interval, extracting the top three feature components by principal component transformation, and constructing the first feature vector. Based on the concrete mix proportion parameters, including water-cement ratio, cementitious material dosage, admixture dosage, and aggregate physical property parameters, including apparent density, water absorption rate, and crushing index, a second feature vector is established. The Euclidean distance between the first and second feature vectors is calculated using a hierarchical clustering algorithm to generate a clustering tree structure. The parameter correlation degree is determined based on the distance threshold between clustering tree nodes. A five-layer neural network structure is constructed, with the number of input layer nodes being the same as the feature dimension, a three-layer hidden layer structure, and a single output layer node. The ReLU function is used as the activation function. Feature data is input into the corresponding layers of the network according to micro and macro scales, with initial weight coefficients set for each scale feature, and the sum of the weight coefficients is 1. Stochastic gradient descent is used to optimize the network weights, with mean squared error as the loss function and a dynamically decaying learning rate, updating the weight values ​​in each iteration. A concrete strength prediction model is constructed based on the optimized network weights, with multi-scale feature vectors as input and predicted concrete strength values ​​as output.

[0016] Furthermore, by using an adaptive weighted fusion algorithm to predict the weights of regions corresponding to different strengths, local features contributing to the target strength are obtained. This includes: dividing the interface transition zone into regions based on the predicted concrete strength values; calculating the weight of each region using an adaptive Gaussian kernel function, with the kernel function bandwidth parameter adaptively adjusted according to the region size; iteratively optimizing the weights using gradient descent to obtain the contribution value of the first region. Local response normalization is performed on the contribution value of the first region, calculated by subtracting the local mean from the region contribution and dividing by the local standard deviation, to obtain the contribution value of the second region. Two thresholds are set for the contribution value of the second region: a high threshold of 80% of the maximum contribution of the local region and a low threshold of 20% of the minimum contribution of the local region. The weight variation region located between these two thresholds is extracted. Morphological closing operations are used to smooth the boundaries of the weight variation region, with a disk shape of 5 pixels as the structuring element. Contour curves are extracted from the smoothed region boundaries, and the curvature value and rate of change of curvature of the curves are calculated to obtain the region morphological feature values. The moving average method was used to calculate the surface roughness of the region, with the window size set to 5% of the region area and the sliding step size to 25% of the window size. A mapping table between morphological feature values ​​and weight changes was established, recording the correspondence between curvature values, curvature change rates, surface roughness, and weight adjustment magnitudes. Correlation analysis was used to calculate the correlation coefficient between morphological parameters and weight changes, and the sign and magnitude of the correlation coefficient were used to determine the direction and degree of morphological influence.

[0017] Furthermore, if the accuracy and recall of local features in the interface transition zone are improved to the target value after introducing an adaptive weighted fusion algorithm, a concrete internal strength distribution map is generated. Combined with coarse aggregate surface roughness and porosity data, the degree of agreement between the strength prediction results and the actual strength test values ​​is analyzed. This includes: based on the accuracy calculation results of local features in the interface transition zone, the feature weights are iteratively updated using an adaptive weighted fusion algorithm; the accuracy and recall of the updated features are calculated using cross-validation, with an accuracy threshold of 0.85 and a recall threshold of 0.80. If both indicators exceed the thresholds, a first strength distribution map is generated using bilinear interpolation. The first strength distribution map is then gridded, with the grid size set to a square with a side length of 5 mm. The predicted strength value is extracted at each grid node to generate a second strength distribution map. A strength distribution curve is established using coarse aggregate surface roughness data, and the curve is smoothed using a cubic spline function to obtain a third strength distribution map. A stratified sampling method is used to select test point locations, with a sampling ratio set to 30% of the total population, to obtain actual strength test data and construct a validation dataset. The least squares method was used to calculate the sum of squared residuals between the third intensity distribution map and the validation dataset. 1000 repeated samplings were performed using the Bootstrap method to obtain residual statistical data. Kernel density estimation was performed on the residual statistical data, and the bandwidth parameter was determined using the Silverman criterion. Residual probability density curves were plotted. A piecewise linear compensation function was established based on the surface porosity data of the coarse aggregate, with different compensation coefficients set for different porosity intervals. The compensation function was used to correct the third intensity distribution map, obtaining the final intensity distribution result. The corrected intensity values ​​and prediction errors were recorded.

[0018] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects:

[0019] This invention discloses a deep learning-based method for predicting concrete strength. The method acquires surface images of coarse aggregates using 3D scanning technology, identifies surface roughness and porosity distribution characteristics, and constructs a strength distribution model for the interface transition zone using a deep learning network. Further, scanning electron microscopy or X-ray tomography is used to analyze the microstructure of the interface transition zone, and wavelet transform and adaptive weighted fusion algorithms are employed to extract enhanced local features. Finally, microstructural parameters are fused with macroscopic concrete parameters to construct a multi-scale concrete strength prediction model. This invention achieves accurate prediction of the internal strength distribution of concrete by analyzing the influence of aggregate morphology on the interface transition zone, providing new ideas and methods for optimizing concrete performance. Attached Figure Description

[0020] Figure 1 This is a flowchart of a deep learning-based concrete strength prediction method according to the present invention.

[0021] Figure 2 This is a schematic diagram of a deep learning-based concrete strength prediction method according to the present invention. Detailed Implementation

[0022] To further understand the content of this invention, a detailed description of the invention is provided in conjunction with the accompanying drawings and embodiments. The specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention. It should also be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0023] like Figure 1 This embodiment of a deep learning-based concrete strength prediction method specifically includes:

[0024] S101. Use 3D scanning technology to collect image data of the coarse aggregate surface to identify its roughness characteristics. At the same time, combine image segmentation algorithm to extract pore distribution characteristics and generate spatial distribution parameters of pore group.

[0025] S1011. During the data acquisition process on the surface of coarse aggregate, the light source intensity and incident angle of the 3D scanner are adjusted according to its reflective characteristics to optimize the scanning effect. The point cloud data of the coarse aggregate is acquired layer by layer by continuous scanning. The initial 3D point cloud data set is generated by the point cloud processing software. The light source power of the light source on the light-colored surface is set to 350 milliwatts and the incident angle is 45 degrees. The light source power of the light source on the dark-colored surface is adjusted to 500 milliwatts and 30 degrees. The scanning interval is kept at 0.5 mm to ensure data accuracy. Then, the initial point cloud data is subjected to Gaussian filtering noise reduction processing. The filtering window is 5×5 pixels and the standard deviation is set to 1.5 to remove noise points caused by equipment shaking, while preserving the integrity of edge details, so as to obtain uniform and low-noise point cloud data.

[0026] S1012. Perform surface fitting and feature segmentation on the generated initial point cloud data. The least squares method is used to perform third-order surface fitting on the point cloud data, and the fitting residual is controlled within 0.1 mm to generate a smooth second set of 3D point cloud data. Then, based on the local curvature value, the watershed algorithm is used to divide the second set of point cloud data into regions. The neighborhood radius is set to 2 mm. The principal curvature and Gaussian curvature of each point are calculated. The segmentation threshold is set to 0.5 to separate different feature regions. The standard deviation of the Euclidean distance between adjacent points in each region is calculated as a roughness index. If the standard deviation is greater than 0.8 mm, it is marked as a pit region, and if it is less than 0.3 mm, it is marked as a flat region. Then, morphological processing is performed on the pit regions. Circular structuring elements with a radius of 3 mm are used to smooth the boundaries through dilation and erosion operations. The region growing algorithm is used to expand the connected regions using the pit regions as seed points. The stopping condition is that the curvature of the boundary points is less than 0.3. The area, depth and spacing parameters of the connected regions are extracted.

[0027] S1013. Further analysis of the marked pit areas to extract pore distribution features. Density clustering algorithm is used to cluster connected regions with a cluster radius of 10 mm and a minimum number of points of 3. Regions with an area greater than 25 square millimeters and a depth greater than 2 mm are identified as pore regions. The shortest distance matrix between pore regions is calculated, and a distance threshold of 15 mm is set to analyze the connectivity of the pore group. Finally, the spatial distribution parameters of the pore group are extracted, including pore density per unit area, connectivity between adjacent pores, and main distribution direction. These parameters provide key inputs for the strength modeling of the interface transition zone.

[0028] The application of 3D scanning technology allows for the efficient acquisition of high-precision morphological data of coarse aggregate surfaces. Combining watershed and density clustering algorithms, roughness and porosity distribution characteristics can be accurately identified, laying the foundation for subsequent deep learning model training. The obtained spatial distribution parameters not only reflect the physical properties of the coarse aggregate surface but also provide reliable data support for strength analysis of the interface transition zone.

[0029] It is understood that the embodiments of the present invention do not impose too many restrictions on the model of the 3D scanning equipment or the specific algorithm parameters. Technicians can adjust the light source settings or clustering conditions according to the actual application scenario to adapt to the feature extraction needs of different coarse aggregate types.

[0030] S102. Based on the surface roughness and pore distribution characteristics of coarse aggregate, a mapping relationship between interface strength and aggregate morphology is established using a deep learning network. An interface transition zone strength distribution model is constructed to output its non-uniformity characteristics and to determine the areas with non-uniform strength distribution.

[0031] In practical applications, key parameters such as height distribution, mean roughness, and pore depth are first extracted from the surface of coarse aggregate. Height distribution reflects the variations in surface undulations, mean roughness characterizes the overall surface roughness, and pore depth reveals the depth characteristics of the surface microstructure. To ensure data consistency, the min-max normalization method is used to normalize these parameters to a range of 0 to 1. Subsequently, a convolutional neural network (CNN) is used to extract the deep representation of morphological features. Through multiple convolution and pooling operations, the CNN can capture the spatial distribution patterns of roughness and pores. For example, the convolution kernel size can be set to 3×3 with a stride of 1, and the pooling layer can use 2×2 max pooling to progressively abstract higher-order information about surface features.

[0032] S1021. Based on the measurement data of the coarse aggregate surface, a feature vector is constructed that includes height distribution, mean roughness, pore depth, pore density, and pore connectivity. The height value typically ranges from -2.5 mm to 2.5 mm, the mean roughness is approximately 0.85 mm, the pore depth distribution ranges from 0.5 mm to 3 mm, the pore density is 8 pores per square centimeter, and the connectivity is 0.65. Subsequently, the feature data is mapped to a unified interval through min-max normalization and input into a convolutional neural network. Deep features are extracted through three layers of convolution operations, with the number of kernels in each layer being 32, 64, and 128, respectively. The ReLU activation function is used to enhance the nonlinear expressive power, generating deep representation data of morphological features.

[0033] S1022. After obtaining deep features, a fully connected network is used to transform the features and predict the interface bonding strength. The fully connected network consists of three layers with 128, 64 and 32 nodes respectively. Nonlinearity is introduced between each layer through the ReLU activation function. At the same time, a residual connection structure is introduced to alleviate the gradient vanishing problem in the deep network. The residual connection ensures the effective transmission of information in the network by adding the input features to the transformed features. The output bonding strength value of the interface micro-region is typically between 3 MPa and 12 MPa, reflecting the strength characteristics of the local region of the interface.

[0034] Subsequently, the predicted intensity values ​​of the interface micro-regions are smoothed. A cubic spline interpolation algorithm is used to make the transition of intensity values ​​between adjacent sampling points more natural, eliminating the noise impact caused by abrupt changes. Next, morphological opening operations are used to smooth the intensity distribution region. The operator uses a circular structuring element with a radius of 5 pixels, and through erosion followed by dilation, small-scale noise is removed while preserving the main intensity distribution trend. Based on this, the intensity difference between adjacent micro-regions is calculated to generate an intensity gradient distribution map. Regions with high gradient values ​​often correspond to abrupt changes in roughness or dense porosity; for example, the maximum gradient value can reach 4 MPa per millimeter, indicating significant intensity changes in the interface transition zone.

[0035] For the strength gradient distribution map, the mean and standard deviation of the strength in local areas are further calculated to quantify its distribution characteristics. For example, the mean strength of a certain area might be 7.5 MPa, and the standard deviation might be 2.1 MPa. If the standard deviation exceeds a preset threshold, such as 1.8 MPa, the area is determined to be a region with uneven strength distribution, and its spatial coordinates are recorded. These uneven areas usually appear in depressions or areas with dense pores on the surface of coarse aggregate, and have a high risk of stress concentration.

[0036] S1023. For the marked areas with uneven intensity distribution, the morphological characteristics are calculated using connected component analysis. The area and perimeter of the connected components are obtained through pixel statistics and boundary tracing, respectively. For example, the maximum connected component area can reach 12 square millimeters, and the perimeter is 15 millimeters. Then, the shape factor is calculated as a morphological characterization index. The shape factor is defined as 4π area divided by the square of the perimeter, with a typical value of 0.72. A weight matrix is ​​constructed based on these characteristic parameters. The weight value is proportional to the area and shape factor of the connected components. The larger the area or the more irregular the shape, the higher the weight. The intensity distribution characteristic parameters of the interface transition zone are obtained through weighted calculation, showing that the intensity value gradually increases from 3.5 MPa to 9.8 MPa in the thickness direction.

[0037] The key to constructing an intensity distribution model for the interface transition zone lies in the deep learning network's ability to extract morphological features at a profound level. Convolutional neural networks can effectively capture the spatial correlation between roughness and porosity, while residual connection structures enhance the model's predictive ability for complex features. Furthermore, the combination of cubic spline interpolation and morphological processing ensures the continuity and accuracy of the intensity distribution map. The obtained non-uniform region feature parameters not only reveal the intensity variation patterns in the interface transition zone but also provide a reliable basis for multi-scale parameter fusion.

[0038] It is understandable that technicians can adjust the network structure or standardization method according to the specific type of coarse aggregate, such as increasing the depth of convolutional layers or modifying the weight matrix calculation rules, to adapt to the intensity prediction needs in different scenarios. These adjustments do not depart from the core technical solution of the embodiments of the present invention.

[0039] S103. Analyze the microstructure and macroscopic parameters of the interface transition zone using scanning electron microscopy or X-ray computed tomography, and mark the corresponding microscopic characteristic parameters when significant intensity differences are found.

[0040] First, high-precision imaging of the interface transition region was performed using a scanning electron microscope (SEM). Backscattered electron signals were acquired by adjusting the electron beam voltage to 20 kV and setting the working distance to 15 mm. The scanning process employed a layer-by-layer approach with an interval of 0.5 μm and a magnification range between 1000x and 10000x, ultimately generating a first grayscale image sequence with a resolution of 2048×2048 pixels. These images revealed the microstructural details within the interface transition region, such as the distribution of pores and hydration products. To improve image quality, histogram equalization was used to enhance contrast, increasing the number of grayscale levels from the initial 156 to 243, thereby highlighting the microscopic features.

[0041] like Figure 2As shown in step S1031, after preprocessing the first grayscale image sequence, the enhanced second grayscale image sequence is segmented using the maximum entropy thresholding method. The maximum entropy algorithm determines the segmentation threshold by iteratively calculating the maximum value of the image information entropy. For example, the threshold may be 127, dividing the image into foreground and background parts to generate a binarized image. Subsequently, the binarized image is reconstructed in three dimensions using an iterative back projection algorithm. During the reconstruction process, the projection angle interval is set to 1 degree, and the number of iterations is 50 to ensure reconstruction accuracy. The spatial resolution of the reconstructed image reaches 0.2 micrometers, clearly presenting the three-dimensional structural features of the interface transition area.

[0042] Next, the reconstructed image was voxelized. The voxel size was set to 0.5 × 0.5 × 0.5 cubic micrometers. Pore structure was identified and porosity was calculated using 26-neighborhood connectivity analysis. The results showed that the average porosity of the interface transition zone was 15.3%. Based on this, a density clustering method was used to identify hydration product regions. The cluster radius was set to 2 micrometers, and the minimum number of points was 5. This method effectively distinguished different hydration product types such as ettringite and hydrated calcium silicate gel. This clustering method is based on spatial density distribution and converges gradually through iterative cluster centering, ensuring the robustness of the identification results.

[0043] S1032. For the identified hydration product regions, a gray-level co-occurrence matrix is ​​calculated to extract morphological features, where the displacement distance is set to 1 pixel. Contrast, entropy, and angular second moment are calculated in four directions: 0 degrees, 45 degrees, 90 degrees, and 135 degrees, respectively, forming a 12-dimensional feature vector. Contrast reflects the texture coarseness, entropy measures the randomness of the distribution, and angular second moment represents uniformity. The feature vector is then input into a support vector machine for classification. The support vector machine uses a radial basis kernel function and learns the boundaries of hydration product types through training samples. The classification accuracy can reach 92.5%. A distribution map of hydration product types is output, and the volume fraction of each type of hydration product is calculated. For example, hydrated calcium silicate gel accounts for 32.5%, and unhydrated cement particles account for 18.6%.

[0044] When analyzing local intensity differences, a local adaptive thresholding method was used to mark significant regions. Local variance was calculated using a sliding window with a window size of 21×21 pixels. If the intensity variance of a region exceeded twice the global variance, it was marked as a region with significant intensity differences. These regions typically had high porosity, for example, reaching 23.7%, and uneven distribution of hydration products, characterized by low content of hydrated calcium silicate gel and high content of unhydrated particles. After marking, a feature mapping table was established to record the location coordinates and microstructural parameters of these regions, such as porosity and the volume fraction of hydration products. It was found that regions with intensity differences were mostly distributed within a range of 15 to 45 micrometers from the aggregate surface.

[0045] The combination of scanning electron microscopy (SEM) and X-ray computed tomography (XCT) enables efficient acquisition of microscopic information in the interface transition zone. The application of the maximum entropy thresholding method and iterative back-projection algorithm ensures the accuracy of 3D reconstruction, while the combined use of gray-level co-occurrence matrix (GLCM) and support vector machine (SVM) achieves precise classification of hydration product types. These steps not only reveal the causes of local intensity differences but also provide crucial input data for intensity prediction.

[0046] It is understood that technicians can adjust scanning parameters or clustering conditions according to sample characteristics, such as increasing voxel resolution or modifying kernel function type, to further improve the accuracy of feature extraction. These adjustments are all within the technical scope of the embodiments of this invention.

[0047] S104. Wavelet transform is used to decompose the structural image of the interface transition area into multiple scales to obtain details and approximate features. These features are then weighted using an adaptive weighted fusion algorithm to generate an enhanced local feature representation.

[0048] First, Haar wavelets are used to decompose the interface transition zone structure image into four levels. This decomposition process progressively breaks the image down into high-frequency and low-frequency sub-bands. The high-frequency sub-band captures detailed information in the horizontal, vertical, and diagonal directions, while the low-frequency sub-band retains the main contours and approximate information of the image. This multi-scale analysis reveals the microstructural characteristics of the interface transition zone at different resolutions, such as edge abrupt changes, texture distribution, and overall shape. After decomposition, detailed feature maps are extracted from the high-frequency sub-bands. For example, the horizontal high-frequency sub-band highlights vertical edges, the vertical high-frequency sub-band reflects horizontal edges, the diagonal high-frequency sub-band presents oblique details, and the low-frequency sub-band provides global structural information.

[0049] For the sub-band images obtained from the decomposition, the local variance and energy distribution of the wavelet coefficients are further calculated to quantify the feature intensity. The local window size is set to 8×8 pixels, and the fluctuation of coefficients within the window reflects the richness of detail in the sub-band. In the high-frequency sub-band, edge features are extracted using the Laplacian operator with a 3×3 pixel template, which can effectively detect regions of abrupt changes in intensity. Texture features are calculated using the gray-level co-occurrence matrix, considering four directions: 0 degrees, 45 degrees, 90 degrees, and 135 degrees, with a displacement distance set to 2 pixels to capture the texture directionality and coarseness of the microstructure. In the low-frequency sub-band, the gray-level histogram is quantized using a 256-level method to extract brightness distribution features, and seven invariant moments are calculated for shape moments to describe the overall geometric characteristics of the interface transition zone. These features together constitute the basis for multi-scale feature representation.

[0050] S1041. After feature extraction, the information entropy of each sub-band feature map is calculated to measure its information content, including gray-level entropy, texture entropy, and edge entropy. Gray-level entropy reflects the uniformity of brightness distribution, texture entropy represents the complexity of texture, and edge entropy quantifies the salience of edges. In actual calculations, the gray-level entropy value ranges from 2.5 to 4.2, the texture entropy value ranges from 1.8 to 3.5, and the edge entropy value ranges from 1.2 to 2.8. Then, the weight coefficient is dynamically set according to the entropy value. The larger the entropy value, the more effective information the sub-band contains, and the higher the corresponding weight coefficient is. For example, the weight of the first high-frequency sub-band can reach 0.35, while the weight of the fourth low-frequency sub-band may be 0.15. This adaptive weight allocation method ensures the targeting of feature fusion.

[0051] Subsequently, a bilateral filter is used to process the weighted sub-band feature map to smooth noise while preserving edge information. The bilateral filter combines spatial and numerical dimensions for filtering. The spatial standard deviation is set to 3 pixels to control the influence range of the spatial neighborhood, while the numerical standard deviation is set to 0.15 to adjust the weight of gray-level similarity. After filtering, the feature map achieves a balance between detail clarity and noise suppression. Next, an adaptive thresholding method based on local means is used to further enhance the features. The local region size is set to 16×16 pixels, and the gray-level mean within the region is calculated as the threshold. The threshold parameter typically fluctuates between 0.4 and 0.7, thereby highlighting salient features and reducing background interference.

[0052] S1042. To optimize feature weight allocation, the local variance contribution rate of features at each scale is calculated to further adjust the weights. The variance contribution rate reflects the degree of contribution of features to the overall information change. For example, the weight of regions with a variance contribution rate greater than 0.25 is increased by 0.15, and the weight of regions with a variance contribution rate less than 0.1 is decreased by 0.12. The weight distribution is dynamically optimized in this way. Then, a weighted average algorithm is used to fuse multi-scale features, integrating edge, texture, and shape features into a unified feature description vector. Finally, the weight normalization process ensures that the values ​​of each component of the vector are distributed between -1 and 1, generating a local feature representation with a dimension of 128, of which edge features account for 32 dimensions, texture features account for 64 dimensions, and shape features account for 32 dimensions, comprehensively representing the structural characteristics of the interface transition area.

[0053] Wavelet transform's multi-scale decomposition capability significantly improves the resolution and diversity of feature extraction. In particular, the high-frequency sub-band accurately captures the local abrupt changes in the interface transition region, while the low-frequency sub-band ensures the stability of the overall structure. Combined with an adaptive weighted fusion algorithm, not only is the robustness of feature representation enhanced, but also key regions are focused on through dynamic adjustment of entropy and variance. The resulting feature vector provides high-quality input data for subsequent intensity prediction, effectively reflecting the correlation between the microscopic and macroscopic characteristics of the interface transition region.

[0054] It is foreseeable that technicians can adjust the number of decomposition layers or filtering parameters according to actual needs. For example, they can increase the decomposition layers to five to obtain more detailed features, or modify the standard deviation of the bilateral filter to adapt to different noise levels. All these optimizations are within the technical framework of the embodiments of the present invention.

[0055] S105. The microstructural parameters of the marked interface transition zone are fused with the macroscopic parameters of the concrete. By constructing a multi-scale parameter prediction model, the predicted strength value of the concrete is output.

[0056] First, the microstructural parameters of the interface transition zone are obtained, including porosity, hydration product distribution, and interfacial binding degree. These parameters reflect the microscopic characteristics of the interface transition zone. For example, porosity typically fluctuates between 5% and 35%, the characteristic value of hydration product distribution ranges from 0.2 to 0.8, and the interfacial binding degree ranges from 0.3 to 0.9. To eliminate dimensional differences and facilitate subsequent processing, the maximum-minimum normalization method is used to map these parameters to the interval between 0 and 1. Subsequently, principal component analysis is used to extract the dominant components of the features. The cumulative contribution rate of the first three principal components reaches 87.5%. The first principal component is mainly related to porosity, the second principal component captures the hydration product distribution characteristics, and the third principal component is related to the change in interfacial binding degree. Based on this, a first feature vector is constructed to characterize the microscale characteristics.

[0057] S1051. For the overall characteristics of concrete, macroscopic parameters are collected to construct a second feature vector. The macroscopic parameters include mix proportion parameters and aggregate physical property parameters. The mix proportion parameters cover water-cement ratio, cementitious material dosage, and admixture dosage. For example, the water-cement ratio ranges from 0.35 to 0.65, the cementitious material dosage ranges from 350 to 550 kg per cubic meter, and the admixture dosage accounts for 0.8% to 2.5% of the cementitious material weight. The aggregate physical property parameters include apparent density, water absorption rate, and crushing index. The apparent density is usually between 2.6 and 2.8, the water absorption rate is between 0.5% and 2.8%, and the crushing index is between 8% and 16%. These parameters are also normalized to form a unified macroscopic feature vector, reflecting the potential influence of material composition and aggregate quality on strength.

[0058] Next, a hierarchical clustering algorithm was used to analyze the correlation between microscopic and macroscopic feature vectors. Hierarchical clustering merges similar features step by step by calculating Euclidean distance to generate a cluster tree structure. In actual calculations, the maximum hierarchical distance may be 0.85. If a distance threshold of 0.6 is set, reasonable parameter correlation groups can be identified. The correlation coefficient between water-cement ratio and porosity is as high as 0.82, indicating that the two have a strong synergistic effect in strength formation; the correlation coefficient between aggregate crushing index and interfacial bonding degree is 0.75, showing the direct influence of aggregate quality on interfacial strength. This correlation analysis provides a theoretical basis for feature fusion, ensuring that the model can capture the intrinsic relationships between multi-scale parameters.

[0059] S1052. Based on feature fusion, a five-layer neural network model is constructed to achieve intensity prediction. The number of nodes in the input layer is consistent with the feature dimension, for example, set to 12, which includes the total dimension of micro and macro features. The hidden layer is designed with three layers, with the number of nodes being 24, 16, and 8 respectively. The ReLU activation function is used to introduce non-linear mapping capability. The number of nodes in the output layer is 1, directly outputting the intensity prediction value. Micro and macro features are assigned initial weights, for example, the micro weight is 0.6 and the macro weight is 0.4, with a total of 1. Subsequently, the network weights are optimized using stochastic gradient descent. The mean squared error loss function is selected, and the initial learning rate is set to 0.01, which is dynamically adjusted by decaying to 0.8 times every 50 iterations. After training for 500 epochs, the loss value decreases from 0.45 to 0.08. The optimized weights show that the proportion of micro features increases to 0.65 and macro features to 0.35, highlighting the dominant role of micro structure in intensity.

[0060] In constructing the predictive model, the integration of microscopic and macroscopic features significantly improves the accuracy of predictions. For example, for concrete samples with the same water-cement ratio but different porosities in the interfacial transition zone, the predicted strength difference can reach 5 to 8 MPa, indicating the crucial role of microstructural features in strength variation. The final output strength prediction value, through comprehensive analysis of multi-scale features, can effectively reflect the coupling relationship between the internal microscopic properties and macroscopic performance of concrete.

[0061] The combination of hierarchical clustering and neural networks provides an efficient approach for multi-scale parameter fusion. Standardization of microstructural parameters and principal component extraction ensure the effectiveness of features, while the inclusion of macroscopic parameters enhances the model's adaptability to the overall properties of materials. Technicians can adjust the number of network layers or learning rate strategies according to actual needs, such as increasing hidden layers to four or employing adaptive learning rate optimization algorithms to further improve prediction accuracy. These improvements all conform to the technical framework of this invention.

[0062] S106. The weights of each region in the interface transition zone are calculated using an adaptive weighted fusion algorithm to obtain the contribution of local features, and the influence of aggregate morphology on the characteristics of the interface transition zone is determined by the mapping relationship between morphological features and weight changes.

[0063] First, the interface transition zone is divided into regions based on the predicted concrete strength values, and an adaptive Gaussian kernel function is used to calculate the weight of each region. The core of the Gaussian kernel function lies in controlling the distribution range of the weights through a bandwidth parameter. The bandwidth is adaptively adjusted according to the region area; for example, for a region with an area of ​​100 square micrometers, the initial bandwidth is set to 10 micrometers. Subsequently, the weights are iteratively optimized using the gradient descent method, with an initial learning rate of 0.01. Each iteration adjusts the weight by 10% of the current value, and the iteration stops when the weight change rate is less than 0.1%. This process typically requires about 50 iterations, and the final bandwidth may converge to 7.5 micrometers, yielding the contribution value of the first region, reflecting the initial influence of each region on the strength.

[0064] S1061. Local response normalization is performed on the contribution value of the first region by calculating the mean and standard deviation within the window, where the window size is set to 15×15 pixels. Specifically, the contribution value of each region is subtracted from the mean of the window and then divided by the standard deviation. The resulting values ​​are distributed between -2.5 and 2.5. Then, high and low thresholds are set for the contribution value of the second region to extract the weight change area. The high threshold is 80% of the local maximum contribution, for example, about 2.0, and the low threshold is 20% of the minimum contribution, for example, about -0.5. The extracted change area is concentrated in the middle part of the interface transition area. This normalization and threshold setting method effectively highlights the relative difference in contribution between regions, providing a clear basis for analysis.

[0065] Next, the extracted weight variation regions are smoothed. Morphological closing operations are used, with a disk with a radius of 5 pixels as the structuring element. Boundary noise is eliminated through a dilation-erosion process, making the region contour more continuous and smooth. Based on this, the smoothed boundary contour curve is extracted, and the curvature value is calculated using the three-point method. The sampling point spacing is set to 2 micrometers, the curvature value ranges from 0.05 to 0.8, and the curvature change rate ranges from 0.02 to 0.15. Furthermore, the moving average method is used to calculate the surface roughness of the region. The window size is dynamically adjusted according to the region area. For example, for a 200 square micrometer region, the window is set to 10 square micrometers, the step size is 2.5 square micrometers, and the roughness value is distributed from 0.2 to 1.5 micrometers. These morphological features provide multi-dimensional data support for the mapping relationship.

[0066] S1062. After obtaining the morphological features, a mapping table is established between curvature value, rate of curvature change, surface roughness, and weight adjustment range. When the curvature value exceeds 0.5 and the rate of curvature change is greater than 0.1, the weight adjustment range can reach more than 50%, indicating that the region with drastic morphological changes contributes significantly to the strength. Subsequently, the correlation coefficient between the morphological parameters and the weight change is calculated through correlation analysis. The results show that the curvature value is positively correlated with the weight change, with a correlation coefficient of 0.75, while the surface roughness is negatively correlated with the weight change, with a correlation coefficient of -0.62. This reveals the dual influence mechanism of morphological features on the strength of the interface transition zone, namely, high curvature regions may induce stress concentration, while surfaces with higher roughness improve interface stability by increasing the bonding area.

[0067] The combination of adaptive Gaussian kernel function and gradient descent method ensures the flexibility and accuracy of weight calculation, while local response normalization enhances the comparability of regional contributions. Through morphological processing and morphological feature extraction, the complexity of the local structure in the interface transition zone is clearly characterized. The mapping relationship table and correlation analysis provide quantitative evidence for the mechanism by which aggregate morphology affects intensity, indicating that morphological features significantly affect the overall intensity distribution by modulating interface properties.

[0068] It is foreseeable that technicians can adjust the Gaussian kernel bandwidth or window size according to actual needs, such as setting the initial bandwidth value to 15 micrometers or increasing the window to 20×20 pixels, to adapt to feature analysis at different scales. These optimizations are all within the technical scope of the embodiments of the present invention.

[0069] S107. If the adaptive weighted fusion algorithm improves the accuracy and recall of local features in the interface transition zone to the target level, then generate a strength distribution map of the concrete interior, and combine the surface roughness and porosity data of coarse aggregate to analyze the consistency between the predicted results and the measured values, so as to verify the effectiveness of the model.

[0070] First, an adaptive weighted fusion algorithm is used to iteratively update the weights of local features in the interface transition region. The initial weight is set to 0.5, and it is dynamically adjusted in each iteration based on the change in accuracy. For example, when the accuracy improves by more than 2%, the weight is increased by 0.05, and when it is less than 1%, the weight is decreased by 0.03. After about 25 iterations, the accuracy increases from 0.78 to 0.87, and the recall increases from 0.75 to 0.83. Subsequently, the performance of the updated features is evaluated using cross-validation. The accuracy threshold is set to 0.85, and the recall threshold is set to 0.80. If both thresholds are met, bilinear interpolation is used to generate the first intensity distribution map, with a resolution of up to 1 square millimeter, ensuring that the spatial details of the distribution map are preserved.

[0071] S1071. The first intensity distribution map is gridded, with the grid size set to a square with a side length of 5 mm. Each grid contains multiple intensity prediction points. The intensity difference between adjacent nodes is usually between 0.5 MPa and 2 MPa, especially in the interface transition zone where there is a significant gradient change, with the maximum gradient reaching 4 MPa per millimeter. Then, the intensity distribution curve is constructed by combining the surface roughness data of the coarse aggregate. When the roughness is between 0.8 mm and 1.2 mm, the interface bonding strength is the highest, reaching 12 MPa. When the roughness is below 0.5 mm or above 1.5 mm, the strength drops significantly, down to a minimum of 6 MPa. The curve is then smoothed by a cubic spline function to eliminate local abrupt changes and maintain continuity, generating the third intensity distribution map.

[0072] Next, stratified sampling was used to select test points to obtain actual intensity data. The sampling ratio was set to 30%, collecting approximately 90 test points within a 100 square millimeter area, covering different depths of the interface transition zone. Based on this data, the sum of squared residuals between the third intensity distribution map and the measured values ​​was calculated using the least squares method. The Bootstrap method was used for 1000 repeated samplings to obtain residual statistics, with an average deviation of approximately 1.2 MPa and a standard deviation of 0.8 MPa. Kernel density estimation was further used to analyze the residual distribution. The bandwidth parameter was set to 0.35 according to the Silverman criterion. The resulting probability density curve approximates a normal distribution, with peak residuals concentrated between -0.5 MPa and 0.5 MPa, indicating that the predicted values ​​are generally close to the measured values, but local deviations still exist.

[0073] S1072. To further improve accuracy, a piecewise linear compensation function was designed based on the surface porosity data of coarse aggregate. The compensation coefficient was 1.05 when the porosity was less than 10%, increased to 1.15 when it was between 10% and 20%, and 1.25 when it exceeded 20%. This function was used to correct the third strength distribution map and generate the final strength distribution result. After correction, the average deviation between the predicted value and the measured value was reduced to 0.6 MPa, and the maximum error was reduced from 3.5 MPa to 1.8 MPa. Especially in the area with high porosity, the prediction accuracy was significantly improved, reflecting the accurate correction effect of the compensation function on the influence of microstructure.

[0074] The adaptive weighted fusion algorithm improves feature representation capabilities through iterative optimization, while gridding and cubic spline smoothing ensure the continuity and readability of the intensity distribution map. Combined with roughness and porosity analysis, the model's predictive ability for the intensity of the interface transition zone is verified, and segmented compensation achieves a high degree of agreement with measured data.

[0075] It is foreseeable that technicians can adjust the grid size or compensation coefficient according to specific needs, such as reducing the grid to 3 mm or refining the porosity range, to further optimize the accuracy of the distribution map. These adjustments are all within the technical scope of the embodiments of the present invention.

[0076] The above are only some preferred embodiments of the present invention, but the present invention is not limited thereto, and many improvements and modifications can be made. Any improvements and modifications made based on the basic principles of the present invention should be considered to fall within the protection scope of the present invention.

Claims

1. A method for predicting concrete strength based on deep learning, characterized in that, The method includes: A three-dimensional scanning method was used to obtain images of the coarse aggregate surface, identify the surface roughness of the coarse aggregate, and extract the pore distribution characteristics of the coarse aggregate surface by combining image segmentation algorithm; Based on the surface roughness and pore distribution characteristics, and using a deep learning network to calculate the mapping relationship between interface strength and aggregate morphology, the strength distribution in the interface transition zone is analyzed, the non-uniformity characteristics of the strength distribution are output, and the interface transition zone of the non-uniform strength distribution area is determined. Scanning electron microscopy or X-ray computed tomography is used to identify local intensity differences in the interface transition zone. The microstructure and macrostructure parameters inside the interface transition zone are analyzed. If significant local intensity differences are found, the corresponding microstructure parameters are marked. Wavelet transform is used to extract features from the interface transition region to obtain local features of the interface transition region. The local features are then weighted and calculated using an adaptive weighted fusion algorithm to obtain an enhanced local feature representation. The marked microstructural parameters are fused with the macroscopic parameters of the concrete as a whole to construct a concrete strength prediction model and output the predicted value of the overall concrete strength. The weights of regions corresponding to different intensities are predicted by an adaptive weighted fusion algorithm, and the local features contributing to the target intensity are obtained. If the accuracy and recall of local features in the interface transition zone are improved to the target value after the adaptive weighted fusion algorithm is introduced, a strength distribution map of the concrete interior is generated. Combined with the surface roughness and porosity data of coarse aggregate, the degree of agreement between the strength prediction results and the actual strength test values ​​is analyzed.

2. The method according to claim 1, characterized in that, The process involves acquiring a surface image of the coarse aggregate using a three-dimensional scanning method, identifying the surface roughness of the coarse aggregate, and extracting the pore distribution features of the coarse aggregate surface using an image segmentation algorithm, including: Based on the reflective properties of the coarse aggregate surface, a continuous scanning method is used to obtain layer-by-layer scanning data of the coarse aggregate, and the first set of three-dimensional point cloud data is obtained through point cloud data processing software. The first set of 3D point cloud data is subjected to surface fitting using the least squares method to obtain the second set of 3D point cloud data. For the second set of 3D point cloud data, the watershed algorithm is used to calculate the regional roughness value. If the regional roughness value is greater than a preset high threshold, it is marked as a pitted region. Density clustering algorithm is used to identify pore regions in the pitted area, and the spatial distribution characteristics of the pore group are obtained through the shortest distance matrix between the pore regions.

3. The method according to claim 1, characterized in that, The process involves analyzing the intensity distribution in the interface transition zone based on surface roughness and porosity distribution characteristics, and using a deep learning network to calculate the mapping relationship between interface strength and aggregate morphology. This analysis outputs the non-uniformity characteristics of the intensity distribution and identifies the interface transition zone where the intensity distribution is non-uniform, including: Based on the surface height distribution parameters, mean roughness parameters, and pore depth parameters of the coarse aggregate surface, the maximum-minimum normalization method is used for normalization, and the deep expression data of morphological features is obtained through a convolutional neural network. For the deep representation data of the morphological features, a fully connected network is used for feature transformation, and the interface bonding strength is predicted through the residual connection structure to obtain the interface micro-region bonding strength value. A cubic spline interpolation operation is performed on the bonding strength value of the interface micro-region, and a morphological operator is used for region smoothing. The bonding strength gradient distribution map of the interface is obtained by calculating the strength difference between adjacent micro-regions. The mean and standard deviation of the intensity in a local area are calculated based on the interface and the intensity gradient distribution map. If the standard deviation is greater than a preset threshold, the area is determined to be an area with uneven intensity distribution.

4. The method according to claim 1, characterized in that, The method employs scanning electron microscopy or X-ray computed tomography to identify local intensity differences in the interface transition zone, analyzes the microstructural and macroscopic parameters within the interface transition zone, and if significant local intensity differences are found, marks the corresponding microstructural parameters, including: The interface transition region is scanned in layers using a scanning electron microscope. The first grayscale image sequence of the interface transition region is obtained by backscattering electron signals. The second grayscale image sequence is obtained by histogram equalization based on the first grayscale image sequence. A binarized image is obtained by using the maximum entropy thresholding method based on the second grayscale image sequence, and a reconstructed image is obtained by performing three-dimensional reconstruction on the binarized image through an iterative back projection algorithm. The reconstructed image is labeled with voxels to obtain a labeled image, and the hydration product regions are identified in the labeled image by density clustering. For the hydration product region, the contrast, entropy, and second moment of the gray-level co-occurrence matrix are calculated to obtain the feature vector. Based on the feature vector, it is determined whether there are significant local intensity differences. If so, the corresponding microstructure parameters are marked.

5. The method according to claim 1, characterized in that, The wavelet transform is used to extract features from the interface transition region to obtain local features of the interface transition region. An adaptive weighted fusion algorithm is then used to weight these local features to obtain an enhanced local feature representation, including: The interface transition zone structure image is decomposed into four layers using Haar wavelet to obtain high-frequency sub-band image and low-frequency sub-band image. Detail feature maps in the horizontal, vertical and diagonal directions are obtained from the high-frequency sub-band image. The local variance and energy distribution of the sub-band image are calculated based on the wavelet coefficient amplitude. The Laplacian operator edge features and gray-level co-occurrence matrix texture features are obtained from the high-frequency sub-band image. The gray-level histogram and shape moment features are obtained from the low-frequency sub-band image. Calculate grayscale entropy, texture entropy, and edge entropy values ​​for each sub-band feature map, and set weight coefficients for the sub-band features; A weighted average algorithm is used to fuse multi-scale features, and the features are combined by weight normalization to obtain enhanced local feature representations.

6. The method according to claim 1, characterized in that, The process of fusing the marked microstructural parameters with the macroscopic parameters of the concrete as a whole to construct a concrete strength prediction model and outputting the predicted value of the overall concrete strength includes: The microstructure parameters of the interface transition zone are obtained, and the microstructure parameters are normalized by the maximum and minimum values ​​to obtain the first feature vector. The microstructure parameters include porosity, hydration product distribution and interface bonding degree. A second feature vector is constructed based on concrete mix proportion parameters and aggregate physical property parameters. The concrete mix proportion parameters include water-cement ratio, cementitious material dosage and admixture dosage. The aggregate physical property parameters include apparent density, water absorption rate and crushing index. A hierarchical clustering algorithm is used to calculate the Euclidean distance between the first feature vector and the second feature vector, and the parameter correlation degree is determined based on the distance threshold between clustering tree nodes. Using the first feature vector and the second feature vector as input data for the concrete strength prediction model, the neural network weights are optimized by the stochastic gradient descent method to obtain the predicted value of the overall concrete strength.

7. The method according to claim 1, characterized in that, The weights of regions corresponding to different intensities predicted by the adaptive weighted fusion algorithm are used to obtain local features contributing to the target intensity, including: The weights of each region in the interface transition zone are calculated using an adaptive Gaussian kernel function, and the contribution value of the first region is obtained by iterative optimization of the weights using the gradient descent method. The contribution value of the second region is obtained by subtracting the local mean from the contribution value of the first region and then dividing by the local standard deviation. For the contribution value of the second region, high and low thresholds are set to extract the weight change region, and morphological closing operation is used on the weight change region to obtain a smooth boundary; The curvature value and the rate of change of curvature are calculated based on the contour curve extracted from the smooth boundary. The local features contributing to the target intensity are obtained.

8. The method according to claim 1, characterized in that, If the accuracy and recall of local features in the interface transition zone are improved to the target value after introducing the adaptive weighted fusion algorithm, then an internal strength distribution map of concrete is generated. Combined with data on surface roughness and porosity of coarse aggregate, the degree of agreement between the predicted strength and the actual strength test values ​​is analyzed, including: The adaptive weighted fusion algorithm is used to iteratively update the local feature weights of the interface transition area. Based on the feature weights, the accuracy and recall of the updated features are calculated using cross-validation. If the accuracy exceeds the accuracy threshold and the recall exceeds the recall threshold, a first intensity distribution map is generated. The first intensity distribution map is gridded, and the intensity prediction value is extracted based on the grid nodes to obtain the second intensity distribution map; An intensity distribution curve is established based on the second intensity distribution map and the surface roughness data of the coarse aggregate. The intensity distribution curve is then smoothed using a cubic spline function to obtain a third intensity distribution map. A piecewise linear compensation function is established based on the surface porosity data of coarse aggregate. The third intensity distribution map is then corrected using the piecewise linear compensation function to obtain the final intensity distribution result.

Images