Intelligent Design Method for the Profile of Concrete Faced Rockfill Dam Based on Data Feature Fusion

Abstract

The present invention relates to the field of artificial intelligence technology, and particularly to an intelligent design method for the cross-section of a concrete face rockfill dam based on data feature fusion. This method collects initial data and historical data such as characteristic water levels, seismic properties, and stone material parameters, constructs a learning and prediction model to predict the dam height and slope ratio, and then calculates the dam contour parameters and predicts the area ratio of each zone. Next, the dam contour parameters, characteristic water levels, seismic properties, and zone area ratio are fused into a multi-dimensional feature vector and input into a generative adversarial network model to generate a zoning design image of the concrete face rockfill dam. By realizing the intelligent design of the cross-section scheme of the concrete face rockfill dam, the present invention greatly improves the design efficiency and further enhances the intelligent level of the design of concrete face dams in China.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and particularly relates to an intelligent design method for the profile of a concrete face rockfill dam based on data feature fusion. Background Art

[0002] Dams are the core of hydropower project construction. Concrete face rockfill dams and core wall dams are the two main types of earth-rock dams. Among them, the concrete face rockfill dam has the characteristics of good safety, economy, and adaptability, and has become the mainstream dam type for water retaining structures in hydropower facilities such as conventional hydropower and pumped storage power stations. The conventional design of concrete face rockfill dams has gradually matured. Just in China, the number of existing dam design and construction has exceeded 98,000. However, the design mode of concrete face rockfill dams that highly relies on engineers' manual work faces the development problems of difficult inheritance of experience and difficult improvement of efficiency. With the advancement and implementation of pumped storage power stations and hydropower development projects in recent years, the digital and intelligent design of concrete face rockfill dams is an inevitable choice to achieve high-quality and efficient engineering design and meet the development needs of intelligent engineering construction, and an intelligent transformation of the design is awaited to solve the current problems.

[0003] With the development of artificial intelligence technology, cutting-edge AI technologies represented by generative AI are reshaping the engineering design industry, bringing efficient design tools, effectively reusing existing design data, and improving design quality. Generative AI has good capabilities in deeply extracting and learning existing data features and generating new designs. Therefore, generative AI learns the existing design data of concrete face rockfill dams by imitating human engineers, learns the past design experience of concrete face rockfill dams, analyzes complex drawings, combines with requirement texts, integrates data and empirical knowledge, and then realizes a new design.

[0004] The patent document with the publication number CN117094059A discloses a design method for a concrete face rockfill dam with a broken-line dam crest. The rockfill dam includes a dam body. The elevation of the panel top at different profiles of the dam body is the sum of the static water level under normal operation at that place and the reserved settlement amount. The upstream panel slope of the dam body is unique and fixed, and the front edge of the upstream panel intersects with the bottom plate of the wave wall in a broken line. The design designs the center line of the dam top road according to the reserved settlement amount of the dam top and the panel slope, establishes a broken-line wave wall and a dam top road model, establishes an overall model of the upstream panel, constructs a panel model according to the bottom plate surface of the wave wall, and constructs a concrete face rockfill dam with a broken-line dam crest.

[0005] It can be seen that there are the following problems: Due to many subjective judgments and diverse design methods in the existing technology design process, it is difficult to form a unified standard process, resulting in a low consistency and repeatability of the design results. Summary of the Invention

[0006] To this end, the present invention provides an intelligent design method for the cross-section of a concrete face rockfill dam based on data feature fusion to overcome the problems in the prior art that there are many subjective judgments and diverse design methods in the design process, it is difficult to form a unified standard process, and the consistency and repeatability of the design results are relatively low.

[0007] To achieve the above object, the present invention provides an intelligent design method for the cross-section of a concrete face rockfill dam based on data feature fusion, including:

[0008] Step S1, collect the initial data required for the design of the concrete face rockfill dam and the foundation surface parameters, the initial data includes characteristic water levels, seismic attributes, and stone material parameters, and also collect the historical data corresponding to the initial data;

[0009] Step S2, construct a learning and prediction model according to the initial parameters and the historical data to predict the dam height, upstream dam slope ratio, and downstream dam slope ratio of the rockfill dam;

[0010] Step S3, calculate according to the dam height, the upstream dam slope ratio, the downstream dam slope ratio, and the foundation surface parameters to obtain the dam profile parameters, and predict the target partition area ratios of the main rockfill area, secondary rockfill area, drainage area, and additional model area according to the dam profile parameters;

[0011] Step S4, draw an input image according to the dam profile parameters, and fuse the input image, the characteristic water levels, the seismic attributes, and the target partition area ratios to obtain a multi-dimensional feature vector;

[0012] Step S5, construct a generative adversarial network model with the multi-dimensional feature vector as the input of the model, and generate a partition design image of the concrete face rockfill dam according to the generative adversarial network model.

[0013] Further, the process of step S2 includes:

[0014] Normalize the characteristic water levels, the seismic attributes, the stone material parameters, and the historical data to obtain a first processing result;

[0015] Extract features from the first processing result to obtain a first feature result;

[0016] Characterize the first feature result to obtain a characterization result;

[0017] Use the characterization result as the input of the learning and prediction model to predict the dam height, the upstream dam slope ratio, and the downstream dam slope ratio.

[0018] Further, the process of using the first feature result as an input to the learning prediction model to predict the dam height, the upstream dam slope ratio, and the downstream dam slope ratio includes:

[0019] Performing a correlation analysis on the first feature result to obtain an analysis result;

[0020] Taking the dam height and the slope ratio corresponding to the historical data as target variables;

[0021] Taking a part of the analysis result and the target variables as a training set, and taking the other part of the analysis result and the target variables as a validation set;

[0022] Training the learning prediction model according to the training set to obtain a first initial training model;

[0023] Validating the initial training model according to the validation set to obtain a first target training model;

[0024] Predicting the dam height, the upstream dam slope ratio, and the downstream dam slope ratio according to the target training model.

[0025] Further, the process of calculating the dam profile parameters according to the dam height, the upstream dam slope ratio, the downstream dam slope ratio, and the foundation surface parameters includes:

[0026] Taking the ratio of the dam height to the upstream dam slope length as the upstream dam slope ratio, taking the ratio of the dam height to the downstream dam slope length as the downstream dam slope ratio, and taking the width of the dam bottom in contact with the foundation as the foundation surface width;

[0027] Taking the sum of the upstream dam slope length, the downstream dam slope length, and the foundation surface width as the dam crest width;

[0028] Using geometric methods to determine the dam profile parameters according to the dam height, the dam crest width, the upstream dam slope ratio, and the downstream dam slope ratio.

[0029] Further, the process of predicting the area ratios of the main fill, secondary fill, drainage, and additional modeling dam body partitions according to the dam profile parameters includes:

[0030] Dividing the dam body into a main fill area, a secondary fill area, a drainage area, and an additional modeling area according to the dam profile parameters;

[0031] Calculating the regional areas of the main fill area, the secondary fill area, the drainage area, and the additional modeling area according to geometric methods;

[0032] Comparing the area of each region with the total area of the dam body to obtain the area ratio of each partition;

[0033] Collect the partition area ratio data of the historical data, and analyze the partition area ratio data to obtain a second analysis result;

[0034] Use the second analysis result as the input of the prediction model, and train the prediction model with the partition area ratio data to obtain the target partition area ratios.

[0035] Further, the process of step S4 includes:

[0036] Convert the characteristic water level, the seismic resistance attribute, and the partition area ratio into numerical features to obtain a conversion result;

[0037] Fuse the conversion result and the input image to obtain a fusion result;

[0038] Extract multi-modal features from the fusion result to obtain a second feature result;

[0039] Perform normalization and feature alignment processing on the second feature result to obtain a second processing result;

[0040] Stack the second processing result along the data channels to obtain the multi-dimensional feature vector.

[0041] Further, the process of step S5 includes:

[0042] Statistically analyze the profile contour parameters, the initial data, the coordinate parameters of each partition, and the profile area ratio parameters of the input image to obtain the distribution laws of different parameters;

[0043] According to the distribution laws, perform parametric adjustment to augment and expand the upstream dam slope ratio, the downstream dam slope ratio, and the area of each region to construct a multi-source heterogeneous training dataset and a test dataset;

[0044] Perform feature data fusion on the training dataset and the test dataset to construct corresponding input feature datasets and output feature datasets;

[0045] Use the input feature dataset as input parameters, and train the generative adversarial network model with the training dataset to obtain a second initial training model;

[0046] Verify the second initial training model according to the validation dataset to obtain a second target training model;

[0047] Generate the panel rockfill dam partition design image according to the second target training model.

[0048] Further, the process of augmenting and expanding the upstream dam slope ratio, the downstream dam slope ratio, and the areas of the respective regions according to the distribution law by parametric adjustment to construct a multi-source heterogeneous training dataset and a test dataset includes:

[0049] Determine the parameter thresholds for the upstream dam slope ratio, the downstream dam slope ratio, and the areas of the respective regions;

[0050] Define the adjustment step sizes for the upstream dam slope ratio, the downstream dam slope ratio, and the areas of the respective regions;

[0051] Establish a combination rule according to the parameter thresholds and the adjustment step sizes;

[0052] Generate a series of parameter value combinations according to the combination rule;

[0053] Perform geometric transformation and simulation on the input image according to the generated parameter value combinations to generate a target profile image;

[0054] Assign the corresponding parameter values as label information to each of the target images;

[0055] Construct the training dataset and the test dataset according to the target input image and the label information.

[0056] Further, the process of validating the second initial training model using the validation dataset to obtain a second target training model includes:

[0057] Create a binary mask for the profile contour parameters of the input image, where the profile contour area is marked as 1 and the remaining contour areas are marked as 0;

[0058] Filter the output features according to the mask to obtain a filtering result;

[0059] Determine that the generated image is only valid within the panel profile contour according to the filtering result.

[0060] Further, the process of filtering the output features according to the mask to obtain a filtering result includes:

[0061] Wherein, is the original output height H, width W, and channels C of the generation network, represents the element-wise multiplication Hadamard product, is the valid generated image after mask filtering, is the binary mask matrix, satisfying: .

[0062] Compared with the prior art, the beneficial effects of the present invention are as follows. By means of automated and intelligent methods, the present invention reduces a large amount of manual calculation and drawing work in traditional designs, significantly shortening the design cycle. Using historical data and machine learning models for prediction enables design based on a large amount of case experience, reducing human errors and improving the accuracy and reliability of the design. By learning the prediction model, key parameters such as the dam height and dam slope ratio can be predicted more precisely, thereby optimizing the dam structure and enhancing the engineering stability. The design scheme can be quickly adjusted according to different initial data and foundation surface parameters to adapt to different engineering environments and requirements. It reduces labor costs and repetitive work in the design process, and lowers the overall design cost through intelligent design methods.

[0063] In particular, through normalization, the influence of different dimensions and data ranges is eliminated, enabling the model to better learn data features, thereby improving the accuracy of prediction. The normalized data can accelerate the convergence speed of the model, reducing the time and resources required for calculation. Design prediction based on historical data and machine learning models can reduce human errors and design risks, enhancing engineering safety.

[0064] In particular, through correlation analysis, the features most relevant to the target variables (dam height and slope ratio) are screened out, reducing the interference of noise and irrelevant variables, thereby improving the prediction accuracy of the model. Using the training set and validation set for model training and validation respectively helps to avoid overfitting and ensure that the model has good generalization ability. The intelligent prediction model can quickly output design parameters, greatly shortening the time required by traditional design methods and accelerating the project progress.

[0065] In particular, using geometric methods to calculate the dam profile parameters can ensure the accuracy of the calculation results and avoid human errors. Using geometric methods to calculate the dam profile parameters can ensure the accuracy of the calculation results and avoid human errors. By improving design efficiency and accuracy, the design cost can be reduced.

[0066] In particular, by predicting the partition area ratio, the material configuration can be optimized. By optimizing the material configuration, the material cost can be reduced and the profitability of the project can be improved. It can also enhance the stability of the dam and reduce engineering risks. By predicting the partition area ratio, the material composition of the dam can be better understood, thereby optimizing the dam structure and enhancing the engineering stability.

[0067] In particular, by fusing features of different modalities, the performance of the generative adversarial network model can be improved. By fusing features of different modalities, the robustness of the generative adversarial network model can be enhanced. By analyzing the influence of different modality features on the generated images, the interpretability of the generative adversarial network model can be improved.

[0068] In particular, through parametric adjustment and augmentation expansion, diverse training datasets and test datasets can be created, which helps improve the generalization ability of the model, enabling it to handle more diverse design scenarios. By performing feature fusion and parametric adjustment on the datasets, the amount of data required for model training can be reduced while improving the training efficiency. Data augmentation can increase the diversity of training data, thereby enhancing the generalization ability of the model. By training a generative adversarial network model, high-quality panel rockfill dam zoning design images can be generated.

[0069] In particular, the automated parametric adjustment and simulation process significantly reduces the time for manual design, enabling designers to quickly explore multiple design options and optimize the design process. The multi-source heterogeneous datasets contain various possible profile designs, which helps avoid overfitting during model training and ensures that the model can handle diverse design requirements. By simulating different parameter combinations, the optimal or most unfavorable design parameters can be identified, thus guiding the actual engineering design and improving the quality and safety of the final design.

[0070] In particular, by only focusing on the area within the profile contour, the model can more accurately learn the design-related features, reducing the interference of external noise or irrelevant information. By excluding the area outside the contour, unnecessary computational effort is reduced, improving the computational efficiency and reducing resource consumption. Ensuring that the generated panel rockfill dam zoning design images only contain valid information within the profile contour helps generate design solutions that better meet the actual engineering requirements.

[0071] In particular, through mask filtering, it can be ensured that the generated images are only valid within the panel profile contour, avoiding the interference of generating irrelevant areas, thereby improving the accuracy and reliability of the generated images. Mask filtering can reduce the amount of image information that the generation network needs to process, thus reducing the computational effort and computational time. By adjusting the type and parameters of the mask, the area of the generated image can be controlled, such as generating areas of specific shapes or sizes. Description of the Drawings

[0072] Figure 1 It is a schematic flowchart of the intelligent design method for the panel rockfill dam profile based on data feature fusion provided by the embodiment of the present invention;

[0073] Figure 2 It is a schematic flowchart of step S2 in the intelligent design method for the panel rockfill dam profile based on data feature fusion provided by the embodiment of the present invention;

[0074] Figure 3 It is a schematic flowchart of step S4 in the intelligent design method for the panel rockfill dam profile based on data feature fusion provided by the embodiment of the present invention;

[0075] Figure 4Schematic flow chart of step S5 in the intelligent design method for the panel rock-fill dam profile based on data feature fusion provided by the embodiments of the present invention;

[0076] Figure 5 Text-image fusion and intelligent design generation results provided by the embodiments of the present invention. Detailed implementation manners

[0077] In order to make the objectives and advantages of the present invention clearer and more understandable, the present invention will be further described below in conjunction with embodiments; it should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0078] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are only used to explain the technical principles of the present invention and do not limit the protection scope of the present invention.

[0079] It should be noted that in the description of the present invention, the terms indicating the direction or positional relationship such as "upper", "lower", "left", "right", "inner", "outer", etc. are based on the direction or positional relationship shown in the drawings. This is only for the convenience of description and does not indicate or imply that the device or element must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation to the present invention.

[0080] In addition, it should also be noted that in the description of the present invention, unless otherwise clearly defined and limited, the terms "installation", "connection", and "connection" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be directly connected or indirectly connected through an intermediate medium, and it can be the communication inside two elements. For those skilled in the art, the specific meanings of the above terms in the present invention can be understood according to specific situations.

[0081] Please refer to Figure 1 As shown, an intelligent design method for the panel rock-fill dam profile based on data feature fusion provided by the embodiments of the present invention includes:

[0082] Step S1, collecting the initial data and foundation surface parameters required for the panel rock-fill dam design, where the initial data includes characteristic water levels, seismic attributes, and stone material parameters, and also collecting the historical data corresponding to the initial data;

[0083] Step S2, constructing a learning and prediction model based on the initial parameters and the historical data to predict the dam height, upstream dam slope ratio, and downstream dam slope ratio of the rock-fill dam;

[0084] Step S3: Calculate based on the dam height, the upstream dam slope ratio, the downstream dam slope ratio, and the foundation surface parameters to obtain the dam profile parameters, and predict the target partition area ratios of the main rockfill area, the secondary rockfill area, the drainage area, and the additional model area according to the dam profile parameters;

[0085] Step S4: Draw the input image according to the dam profile parameters, and fuse the input image, the characteristic water levels, the seismic properties, and the target partition area ratios to obtain a multi-dimensional feature vector;

[0086] Step S5: Use the multi-dimensional feature vector as the input of the model to construct a generative adversarial network model, and generate a panel rockfill dam partition design image according to the generative adversarial network model.

[0087] Specifically, collect characteristic water level data, including the design flood level, the normal storage level, the dead storage level, etc. Collect seismic property data, such as seismic intensity, ground motion parameters, etc. Collect stone material parameters, including the planned excavation volume of the material source, the elastic modulus, porosity, dry density, and water content of the rockfill and core wall clay. Collect parameters such as the topography, geology, and geotechnical characteristics of the foundation surface. Collect historical design case data corresponding to the above initial data, including the dam height, the dam slope ratio, the partition area ratio, etc. Determine the key characteristics affecting the dam height and the dam slope ratio. Select appropriate machine learning algorithms, such as random forest, support vector machine, neural network, etc. Use historical data to train the model, with the input features being the initial data and the foundation surface parameters, and the output being the dam height, the upstream dam slope ratio, and the downstream dam slope ratio. Verify the prediction accuracy of the model through methods such as cross-validation. Calculate the dam profile parameters according to the predicted dam height, the upstream and downstream dam slope ratios, and the foundation surface parameters. Use the calculated dam profile parameters to predict the target partition area ratios of the main rockfill area, the secondary rockfill area, the drainage area, and the additional model area through regression analysis or other prediction methods. Draw the dam profile input image according to the dam profile parameters. Digitally process the input image and extract the image features. Convert non-image data such as characteristic water levels, seismic properties, and target partition area ratios into numerical features. Combine all features into a multi-dimensional feature vector. Use the multi-dimensional feature vector as the input to construct a generative adversarial network (GAN) model. Design and train the generator and the discriminator. Through iterative training, enable the generator to generate realistic panel rockfill dam partition design images. Adjust the model parameters to optimize the quality and accuracy of the generated images. Use the trained generator to generate a panel rockfill dam partition design image according to the multi-dimensional feature vector. Compare the generated design image with the actual design standards to verify the rationality and accuracy of the design.

[0088] Specifically, through automated and intelligent methods, a large amount of manual calculation and drawing work in traditional design has been reduced, significantly shortening the design cycle. By using historical data and machine learning models for prediction, design can be based on a large number of case experiences, reducing human errors and improving the accuracy and reliability of design. By learning the prediction model, key parameters such as dam height and dam slope ratio can be predicted more accurately, thereby optimizing the dam structure and improving engineering stability. According to different initial data and foundation surface parameters, the design scheme can be quickly adjusted to adapt to different engineering environments and requirements. It reduces labor costs and repetitive work in the design process, and reduces the overall design cost through intelligent design methods.

[0089] Specifically, as Figure 2 shown, the process of step S2 includes:

[0090] Step S21, normalizing the characteristic water level, the seismic attribute, the stone material parameters, and the historical data to obtain a first processing result;

[0091] Step S22, performing feature extraction on the first processing result to obtain a first feature result;

[0092] Step S23, characterizing the first feature result to obtain a characterization result;

[0093] Step S24, using the characterization result as the input of the learning prediction model to predict the dam height, the upstream dam slope ratio, and the downstream dam slope ratio.

[0094] Specifically, perform normalization processing on the collected characteristic water level, seismic requirements, and rockfill material source parameters. For continuous numerical data (such as uniaxial compressive strength, porosity, etc.), use min-max normalization or Z-score normalization. For categorical data (such as seismic grade), use one-hot encoding for conversion. The normalized data set is denoted as the first processing result. Analyze the normalized data and select the key features that have a greater impact on dam design. Statistical tests, feature importance scoring, and other methods can be used to determine the key features. The selected set of key features is denoted as the first feature result. Use floating-point numbers to characterize the key parameters (such as uniaxial compressive strength, porosity, etc.). Use one-hot vectors to characterize the key categories (such as seismic grade). The refined feature characterization data set is denoted as the characterization result. Use the characterization result as the input of the learning prediction model. Select a suitable machine learning algorithm, such as a regression model, a neural network, etc. Input the characterization result into the model for training. Divide the data set into a training set and a test set. Use the training set to train the model. Use the test set to evaluate the model performance and perform model tuning. Use the trained model to predict new data to obtain the predicted values of the dam height, the upstream dam slope ratio, and the downstream dam slope ratio.

[0095] Specifically, through normalization, the influence of different dimensions and data ranges is eliminated, enabling the model to better learn data features, thereby improving the accuracy of prediction. The normalized data can accelerate the convergence speed of the model and reduce the time and resources required for calculation. Based on the design prediction of historical data and machine learning models, human errors and design risks can be reduced, and engineering safety can be improved.

[0096] Specifically, the process of using the first feature result as the input of the learning and prediction model to predict the dam height, the upstream dam slope ratio, and the downstream dam slope ratio includes:

[0097] Conduct a correlation analysis on the first feature result to obtain an analysis result;

[0098] Take the dam height and the slope ratio corresponding to the historical data as target variables;

[0099] Use a part of the analysis result and the target variables as the training set, and another part of the analysis result and the target variables as the validation set;

[0100] Train the learning and prediction model according to the training set to obtain a first initial training model;

[0101] Validate the initial training model according to the validation set to obtain a first target training model;

[0102] Predict the dam height, the upstream dam slope ratio, and the downstream dam slope ratio according to the target training model.

[0103] Specifically, methods such as the Pearson correlation coefficient are used to calculate the correlation between features. A correlation matrix is constructed to show the degree of correlation between different features. Identify feature pairs with high correlation (for example, the correlation coefficient is greater than a certain threshold, such as 0.8). Delete or combine highly correlated features to reduce feature redundancy. Perform principal component analysis on the cleaned feature data. Calculate the eigenvalues and eigenvectors to determine the principal components. Select the principal components according to the eigenvalues, usually choosing the principal components whose cumulative explained variance reaches a certain proportion (such as 95%). Project the original feature vector onto the selected principal components to obtain the feature vector after dimensionality reduction. Use the key features after PCA dimensionality reduction as the input of the model. Use the dam height, slope ratios of the upstream and downstream dam slopes as the output of the model. Extract the corresponding dam height, upstream dam slope ratio, and downstream dam slope ratio from historical data as the target variables. Divide the processed data into a training set and a test set, usually using methods such as 70-30, 80-20, or cross-validation. Build an ensemble learning model, which consists of an artificial neural network model, Bagging algorithm, random forest, and support vector regression. Determine the structure and parameters of the model, such as the number of neural network layers, the number of neurons, the number of decision trees, etc. The parameter prediction formula is as follows:

[0104]

[0105] In the formula is the predicted value, is the feature vector of the i-th sample, is the th decision tree, and K is the total number of all decision trees.

[0106] Specifically, through correlation analysis, the features most relevant to the target variables (dam height and slope ratio) are screened out, reducing the interference of noise and irrelevant variables, thereby improving the prediction accuracy of the model. Using the training set and the validation set for model training and validation respectively helps to avoid overfitting and ensure that the model has good generalization ability. The intelligent prediction model can quickly output the design parameters, greatly shortening the time required by traditional design methods and accelerating the project progress.

[0107] Specifically, the process of calculating the dam profile parameters according to the dam height, the upstream dam slope ratio, the downstream dam slope ratio, and the foundation surface parameters includes:

[0108] Take the ratio of the dam height to the upstream dam slope length as the upstream dam slope ratio, take the ratio of the dam height to the downstream dam slope length as the downstream dam slope ratio, and the width of the dam bottom in contact with the foundation as the foundation surface width;

[0109] The sum of the upstream dam slope length, the downstream dam slope length, and the foundation width is taken as the crest width;

[0110] Use geometric methods to determine the dam body profile parameters based on the dam height, the crest width, the upstream dam slope ratio, and the downstream dam slope ratio.

[0111] Specifically, use the following formula to calculate the upstream dam slope length (L1): L1 = H / m1;

[0112] Use the following formula to calculate the downstream dam slope length (L2): L2 = H / m2;

[0113] Use the following formula to calculate the crest width (W): W = L1 + L2 + B;

[0114] Use geometric methods to calculate the dam body profile parameters based on the dam height (H), the crest width (W), the upstream dam slope ratio (m1), and the downstream dam slope ratio (m2), for example:

[0115] Upstream dam slope line: A straight line segment from the crest to the foundation, with a length of L1.

[0116] Downstream dam slope line: A straight line segment from the crest to the foundation, with a length of L2.

[0117] Crest line: A straight line segment parallel to the ground, with a length of B1.

[0118] Bottom line: A straight line segment parallel to the ground, with a length of B2.

[0119] Dam body cross-sectional area: The area of the dam body cross-section, which can be calculated by integration.

[0120] Examples of geometric methods:

[0121] Triangle method: Consider the dam body cross-section as a triangle, calculate the base and height of the triangle, and then calculate the area of the triangle.

[0122] Trapezoid method: Consider the dam body cross-section as a trapezoid, calculate the upper base, lower base, and height of the trapezoid, and then calculate the area of the trapezoid.

[0123] Numerical integration method: Divide the dam body cross-section into multiple small elements, such as triangles or trapezoids, then calculate the area of each element, and finally sum up the areas of all elements to obtain the dam body cross-sectional area. Use existing engineering software, such as AutoCAD or Civil 3D, to calculate the dam body profile parameters. It is also possible to write code using programming languages, such as Python or MATLAB, to calculate the dam body profile parameters, which will not be elaborated here.

[0124] Specifically, using geometric methods to calculate the dam body contour parameters can ensure the accuracy of the calculation results and avoid human errors. Using geometric methods to calculate the dam body contour parameters can ensure the accuracy of the calculation results and avoid human errors. By improving the design efficiency and accuracy, the design cost can be reduced.

[0125] Specifically, the process of predicting the area ratios of the main fill, secondary fill, drainage, and additional modeling zones based on the dam body contour parameters includes:

[0126] Dividing the dam body into a main fill zone, a secondary fill zone, a drainage zone, and an additional modeling zone according to the dam body contour parameters;

[0127] Calculating the area of the main fill zone, the secondary fill zone, the drainage zone, and the additional modeling zone according to geometric methods;

[0128] Comparing the area of each zone with the total area of the dam body to obtain the area ratio of each zone;

[0129] Collecting the area ratio data of the historical data, and analyzing the area ratio data to obtain a second analysis result;

[0130] Taking the second analysis result as the input of the prediction model, and using the area ratio data of each zone to train the prediction model to obtain the area ratio of each target zone.

[0131] Specifically, the main fill zone, as the main part of the dam body, undertakes the main water retaining function. The secondary fill zone is located upstream of the main fill zone and plays a role in transition and anti-seepage. The drainage zone is located upstream of the dam body and is used for drainage and reducing the pore water pressure. The additional modeling zone is located downstream of the dam body and is used to improve the stability of the dam body. According to the dam body contour line and the preset partition boundary, the dam body is divided into different zones. The dam body can also be divided into grids, and according to the position where the center point of the grid is located, the grid is assigned to different zones. Each zone is divided into multiple small units, such as triangles or trapezoids, and then the area of each unit is calculated. Finally, the total area of each zone is obtained by adding up the areas of all units. The area of each zone is compared with the total area of the dam body to obtain the area ratio of each zone. Collect the area ratio data of historical design cases. Analyze the distribution law of historical data, such as mean, variance, maximum value, and minimum value, etc. The key factors affecting the area ratio of each zone, such as dam height, slope ratio, stone material parameters, etc. Select a suitable machine learning algorithm, such as regression model, decision tree, or neural network. Divide the historical data into a training set and a test set. Use the training set to train the prediction model, with the input features being dam height, slope ratio, stone material parameters, etc., and the output being the area ratio of each zone. Use the test set to evaluate the performance of the model and perform model tuning. Use the trained prediction model to predict the area ratio of the target zone according to the new dam height, slope ratio, stone material parameters, etc.

[0132] Specifically, by predicting the partition area ratio, the material configuration can be optimized. By optimizing the material configuration, the material cost can be reduced and the project profitability can be improved. The stability of the dam body can also be enhanced and the project risks can be reduced. By predicting the partition area ratio, the material composition of the dam body can be better understood, thereby optimizing the dam body structure and improving the project stability.

[0133] Specifically, as Figure 3 shown, the process of step S4 includes:

[0134] Step S41, converting the characteristic water level, the seismic attribute, and the partition area ratio into numerical features to obtain a conversion result;

[0135] Step S42, fusing the conversion result and the input image to obtain a fusion result;

[0136] Step S43, extracting multi-modal features from the fusion result to obtain a second feature result;

[0137] Step S44, performing normalization and feature alignment processing on the second feature result to obtain a second processing result;

[0138] Step S45, stacking the second processing result along the data channels to obtain the multi-dimensional feature vector.

[0139] Specifically, convert the characteristic water level (such as the design flood level, normal storage level, dead storage level), seismic attribute (such as seismic intensity, ground motion parameters), and partition area ratio into numerical features. For features that can be directly converted into numerical values, such as water level and area ratio, their numerical values can be directly used. For categorical features, such as seismic grade, one-hot encoding can be used to convert it into a numerical vector. For numerical features, normalization methods can be used to scale them to a specific range, such as between 0 and 1. Fuse the numerical features with the input image. Concatenate the numerical feature vector and the image feature vector to form a longer feature vector. Perform weighted summation on the numerical feature vector and the image feature vector to generate a new feature vector. Extract multi-modal features from the fused feature vector. The following methods can be adopted:

[0140] Convolutional neural network: Use a convolutional neural network to extract image features, such as edges, textures, and shapes, etc.

[0141] Recurrent neural network: Use a recurrent neural network to extract sequence features, such as time series data or text data.

[0142] Use a normalization method, such as L2 normalization, to scale the feature vector to unit length. Use a feature alignment method, such as feature transformation or feature selection, to align the features of different modalities. Stack the normalized and feature-aligned multi-modal features along the data channels to form a multi-dimensional feature vector.

[0143] Specifically, by fusing the features of different modalities, the performance of the generative adversarial network model can be improved. By fusing the features of different modalities, the robustness of the generative adversarial network model can be enhanced. By analyzing the influence of different modality features on the generated images, the interpretability of the generative adversarial network model can be improved.

[0144] Specifically, as Figure 4 shown, the process of step S5 includes:

[0145] Step S51, statistically analyze the profile contour parameters of the input image, the initial data, the coordinate parameters of each partition, and the profile area ratio parameters to obtain the distribution laws of different parameters;

[0146] Step S52, according to the distribution laws, perform parametric adjustment on the upstream dam slope ratio, the downstream dam slope ratio, and the area of each region to augment and expand to construct a multi-source heterogeneous training dataset and a test dataset;

[0147] Step S53, perform feature data fusion on the training dataset and the test dataset to construct corresponding input feature datasets and output feature datasets;

[0148] Step S54, use the input feature dataset as input parameters and train the generative adversarial network model with the training dataset to obtain a second initial training model;

[0149] Step S55, verify the second initial training model according to the validation dataset to obtain a second target training model;

[0150] Step S56, generate the panel rockfill dam partition design image according to the second target training model.

[0151] Specifically, calculate the length, curvature, and direction of the profile contour, and record the coordinates of each contour point. Analyze the distribution of characteristic water levels, seismic attributes, and stone material parameters, such as mean, variance, maximum, and minimum values, etc. Record the coordinate positions of each partition area, such as vertex coordinates and center coordinates, etc. Calculate the area ratio of each partition and compare it with the total area of the dam body. According to the statistical distribution laws, perform parametric adjustment on the upstream dam slope ratio, the downstream dam slope ratio, and the area of each region, such as:

[0152] Random perturbation: Randomly change the slope ratio and area ratio within a certain range to simulate different design schemes.

[0153] Interpolation and Scaling: Interpolate and scale the profile contours to generate images of different sizes and shapes.

[0154] Rotation and Translation: Rotate and translate the images to increase the diversity of the data.

[0155] Existing data augmentation tools, such as ImageDataGenerator or albumentations, can be used to facilitate data augmentation. The augmented image data is fused with the corresponding initial data (characteristic water level, seismic resistance properties, and stone material parameters) to form an input feature dataset. Convolutional neural networks or other image feature extraction methods are used to extract feature vectors from the images. The image feature vectors are concatenated with the initial data vectors to form an input feature dataset. The augmented image data is fused with the corresponding partition area ratio to form an output feature dataset. Each partition area is labeled, for example, using pixel-level labels or instance segmentation labels. Feature vectors are extracted from the annotation information, such as the area ratio of each partition. The image feature vectors are concatenated with the partition area ratio vectors to form an output feature dataset. A suitable generative adversarial network model, such as BigGAN, is selected. A suitable loss function, such as L1 loss or L2 loss, and adversarial loss are selected. The generative adversarial network model is trained using the training dataset and validated using the validation dataset. The new input data (characteristic water level, seismic resistance properties, stone material parameters, and partition area ratio) is converted into a multi-dimensional feature vector. The trained generative adversarial network model is used to generate panel rockfill dam partition design images based on the input feature vectors. The quality and accuracy of the generated images are evaluated, for example, using manual visual inspection or quantitative metrics.

[0156] Specifically, through parametric adjustment and augmentation expansion, diverse training datasets and test datasets can be created, which helps improve the generalization ability of the model and enables it to handle more diverse design scenarios. By performing feature fusion and parametric adjustment on the dataset, the amount of data required for model training can be reduced while improving the training efficiency. Data augmentation can increase the diversity of the training data, thereby improving the generalization ability of the model. By training a generative adversarial network model, high-quality panel rockfill dam partition design images can be generated.

[0157] Specifically, the process of augmenting and expanding the upstream dam slope ratio, the downstream dam slope ratio, and the area of each region by parametric adjustment according to the distribution law to construct multi-source heterogeneous training datasets and test datasets includes:

[0158] Determine the parameter thresholds of the upstream dam slope ratio, the downstream dam slope ratio, and the area of each region;

[0159] Define the adjustment step lengths of the upstream dam slope ratio, the downstream dam slope ratio, and the areas of each region;

[0160] Establish a combination rule based on the parameter thresholds and the adjustment step lengths;

[0161] Generate a series of parameter value combinations according to the combination rule;

[0162] Perform geometric transformation and simulation on the input image according to the generated parameter value combinations to generate a target profile image;

[0163] Assign the corresponding parameter values to each of the target images as label information;

[0164] Construct the training dataset and the test dataset based on the target input image and the label information.

[0165] Specifically, determine the minimum value (U_min) and the maximum value (U_max) of the upstream dam slope ratio (U), for example, U_min = 1:1.5, U_max = 1:2.5. Determine the minimum value (D_min) and the maximum value (D_max) of the downstream dam slope ratio (D), for example, D_min = 1:1.5, D_max = 1:3. Determine the minimum value (A_min) and the maximum value (A_max) of the area of each region (A), for example, A_min = 1000 square meters, A_max = 5000 square meters. Set the adjustment step length (U_step) of the upstream dam slope ratio, for example, U_step = 0.1. Set the adjustment step length (D_step) of the downstream dam slope ratio, for example, D_step = 0.1. Set the adjustment step length (A_step) of the area of the region, for example, A_step = 100 square meters. Create a rule to generate all possible parameter combinations according to the parameter thresholds and step lengths of the upstream dam slope ratio, the downstream dam slope ratio, and the area of the region.

[0166] Example of the rule: For each U value, from U_min to U_max, with U_step as the step length, traverse all D values, from D_min to D_max, with D_step as the step length, and then for each U-D combination, traverse all A values, from A_min to A_max, with A_step as the step length. Write a script using a programming language (such as Python) to generate a list of parameter value combinations according to the combination rule. For each input image, perform geometric transformation according to the parameter value combinations to simulate different dam slope ratios and areas of the region. Use image processing software or a custom algorithm to implement these transformations, such as adjusting the shape and size of the image. For each generated target profile image, store its corresponding parameter values (U, D, A) as label information.

[0167] Specifically, the automated parametric adjustment and simulation process significantly reduces the time for manual design, enabling designers to quickly explore multiple design options and optimize the design process. The multi-source heterogeneous dataset contains various possible profile designs, which helps avoid overfitting during model training and ensures that the model can handle diverse design requirements. By simulating different parameter combinations, the optimal or most unfavorable design parameters can be identified, thus guiding the actual engineering design and improving the quality and safety of the final design.

[0168] Specifically, the process of validating the second initial training model with the validation dataset to obtain the second target training model includes:

[0169] Create a binary mask from the profile contour parameters of the input image, where the profile contour area is marked as 1 and the remaining contour areas are marked as 0;

[0170] Filter the output features according to the mask to obtain a filtered result;

[0171] Determine that the generated image is only valid within the panel profile contour according to the filtered result.

[0172] Specifically, the mask formula is as follows:

[0173]

[0174]

[0175] In the formula, is the contour mask matrix, is the value of each point in the matrix; are different feature matrices, which are the main heap, secondary heap, drainage area, mold-increasing area, and scaling ratio respectively, are the normalized scalar values of different features.

[0176] Specifically, for each input image, first identify the parameters of the profile contour, which define the shape and position of the panel rockfill dam profile. Using these parameters, generate a binary mask of the same size as the input image. In this mask, the pixels within the profile contour are set to 1, while the pixels outside the contour are set to 0. This can be achieved through image processing techniques such as edge detection and region filling. After the model processes the input image, a series of feature maps will be output, which represent the model's different levels of understanding of the input image. Filter these output feature maps using the created binary mask. The specific operation is to multiply the mask with each feature map pixel by pixel, so that only the features corresponding to the pixels with a value of 1 in the mask are retained, and the features corresponding to the pixels with a value of 0 are set to 0.

[0177] Specifically, by only focusing on the area within the profile outline, the model can more accurately learn the features related to the design, reducing the interference of external noise or irrelevant information. By excluding the areas outside the outline, unnecessary computational amount is reduced, computational efficiency is improved, and resource consumption is decreased. Ensuring that the generated panel rockfill dam zoning design image only contains the valid information within the profile outline helps to generate a design scheme that better meets the actual engineering requirements.

[0178] Specifically, the process of filtering the output features according to the mask to obtain a filtering result includes:

[0179]

[0180] Wherein, are the original output height H, width W, and channels C of the generation network, represents the element-wise multiplication Hadamard product, is the valid generated image after being filtered by the mask, is the binary mask matrix, satisfying:

[0181] .

[0182] Specifically, the original output (height H, width W, channels C) of the generation network is subjected to an element-wise multiplication (Hadamard product) operation with the binary mask matrix. The result of the multiplication operation will generate a new image, in which only the area of the profile outline retains the original output features, and the remaining areas are set to 0. This new image is the valid generated image after being filtered by the mask. Analyze the valid generated image after being filtered by the mask to ensure that the generated image is only valid within the panel profile outline. According to the analysis result, adjust the parameters of the generative adversarial network model to optimize the quality and accuracy of the generated image, and ensure that the generated image meets the design requirements.

[0183] Specifically, through mask filtering, it can be ensured that the generated image is only valid within the panel profile outline, avoiding the interference of generating irrelevant areas, thereby improving the accuracy and reliability of the generated image. Mask filtering can reduce the amount of image information that the generation network needs to process, thereby reducing the computational amount and computational time. The area of the generated image can be controlled by adjusting the type and parameters of the mask, for example, generating areas of specific shapes or sizes.

[0184] As Figure 5 shown, in this embodiment, a case study is carried out on a certain concrete face rockfill dam in China. The seismic basic intensity of this project is degree VII, the design earthquake is 0.305g, and the ultimate seismic resistance capacity is about 0.6g. The main rockfill materials of the project include basalt, tuffaceous basalt, etc. According to the design requirements of the rockfill dam, seismic design requirements, and the mechanical parameters and volume of the original rock and other parameters.

[0185] The intelligent design process of a concrete face rockfill dam is specifically as follows:

[0186] 1. Intelligent prediction of dam body profile parameters:

[0187] Machine learning algorithms are used for predicting the dam body profile parameters, and the key parameters for the surface zoning design are generated through intelligent mapping. Among them, the input features include 5 features: dam height of 240 m, saturated compressive strength of the main rockfill of 80 MPa, saturated compressive strength of the secondary rockfill of 50 MPa, quantity of the main rockfill source of 11.9 million m³, and quantity of the secondary rockfill source of 6.4 million m³; the output features are the upstream and downstream slope ratios of 1.4, and the area ratio of secondary rockfill:drainage:main rockfill:incremental model = 0.24:0.12:0.57:0.07.

[0188] 2. Intelligent generation design of the rockfill dam profile zoning:

[0189] (1) Input data construction: According to the data feature expression methods in Specific Embodiment 1 and 2, text-image feature fusion is realized. In the design of this concrete face rockfill dam, the input image feature is the outer contour of the rockfill dam profile determined according to the design requirements, and the input text feature is the normalized areas of the main rockfill, secondary rockfill, drainage, and incremental model areas of this project (secondary rockfill:drainage:main rockfill:incremental model = 0.24:0.12:0.57:0.07).

[0190] (2) Intelligent design: A generative adversarial network model is constructed according to the method in Specific Embodiment 2, and the intelligent generation model is trained and tested. After receiving the image-text input data features, the trained model quickly generates an intelligent design result in pixel expression, and this result includes the layout positions and dimensions of the main rockfill, secondary rockfill, drainage, and incremental model areas in the rockfill dam profile.

[0191] So far, the technical solutions of the present invention have been described in combination with the preferred embodiments shown in the accompanying drawings. However, it is easy for those skilled in the art to understand that the protection scope of the present invention is obviously not limited to these specific embodiments. Without departing from the principle of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the protection scope of the present invention.

[0192] The above are only the preferred embodiments of the present invention and are not used to limit the present invention; for those skilled in the art, the present invention can have various changes and modifications. Any modification, equivalent substitution, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. An intelligent design method for a concrete face rockfill dam profile based on data feature fusion, characterized in that: include: Step S1, collecting initial data and foundation surface parameters required for the design of a concrete face rockfill dam, wherein the initial data includes characteristic water level, seismic properties and stone material parameters, and also collecting historical data corresponding to the initial data; Step S2, constructing a learning prediction model according to the initial data and the historical data to predict the dam body height, upstream dam slope ratio and downstream dam slope ratio of the rockfill dam; Step S3, calculating according to the dam body height, the upstream dam slope ratio, the downstream dam slope ratio and the foundation surface parameters to obtain dam body contour parameters, and predicting the target partition area ratios of the main stacking area, the secondary stacking area, the drainage area and the mold increasing area according to the dam body contour parameters; Step S4, drawing an input image according to the dam body contour parameters, and fusing the input image, the characteristic water level, the seismic properties and the area ratios of each target partition to obtain a multi-dimensional feature vector; Step S5, using the multi-dimensional feature vector as a model input to construct a generative adversarial network model, and generating a panel rockfill dam zoning design image according to the generative adversarial network model.

2. The method for intelligent design of a concrete face rockfill dam profile based on data feature fusion according to claim 1 is characterized in that: The process of step S2 includes: Normalizing the characteristic water level, the earthquake resistance property, the stone material parameter and the historical data to obtain a first processing result; Performing feature extraction on the first processing result to obtain a first feature result; Characterizing the first characteristic result to obtain a characterization result; The characterization results are used as inputs of the learning prediction model to predict the dam body height, the upstream dam slope ratio and the downstream dam slope ratio.

3. The method for intelligent design of a concrete face rockfill dam profile based on data feature fusion according to claim 2 is characterized in that: The process of using the characterization result as the input of the learning prediction model to predict the dam body height, the upstream dam slope ratio and the downstream dam slope ratio includes: Performing correlation analysis on the characterization results to obtain analysis results; The dam body height and the slope ratio corresponding to the historical data are used as target variables; Using a part of the analysis results and the target variable as a training set, and another part of the analysis results and the target variable as a validation set; Training the learning prediction model according to the training set to obtain a first initial training model; Verifying the initial training model according to the verification set to obtain a first target training model; The dam body height, the upstream dam slope ratio and the downstream dam slope ratio are predicted according to the target training model.

4. The method for intelligent design of a concrete face rockfill dam profile based on data feature fusion according to claim 3 is characterized in that: The process of calculating according to the dam body height, the upstream dam slope ratio, the downstream dam slope ratio and the foundation surface parameters to obtain the dam body contour parameters comprises: The ratio of the dam body height to the upstream dam slope length is taken as the upstream dam slope ratio, the ratio of the dam body height to the downstream dam slope length is taken as the downstream dam slope ratio, and the width of the dam body bottom in contact with the foundation is taken as the foundation surface width; The sum of the upstream dam slope length, the downstream dam slope length and the foundation surface width is taken as the dam crest width; The dam body profile parameters are determined according to the dam body height, the dam top width, the upstream dam slope ratio and the downstream dam slope ratio using a geometric method.

5. The method for intelligent design of a concrete face rockfill dam profile based on data feature fusion according to claim 4 is characterized in that: The process of predicting the target partition area ratios of the main heap area, the secondary heap area, the drainage area and the mold-increasing area according to the dam body contour parameters includes: The dam body is divided into a main pile area, a secondary pile area, a drainage area and an increased mold area according to the dam body contour parameters; Calculate the area of ​​the main pile area, the secondary pile area, the drainage area and the mold increase area according to a geometric method; Comparing the area of ​​each of the regions with the total area of ​​the dam to obtain the area ratio of each sub-region; Collecting the partition area ratio data of the historical data, and analyzing the partition area ratio data to obtain a second analysis result; The second analysis result is used as an input of a prediction model, and the prediction model is trained using the partition area ratio data to obtain the area ratio of each target partition.

6. The method for intelligent design of a concrete face rockfill dam profile based on data feature fusion according to claim 5 is characterized in that: The process of step S4 includes: Converting the characteristic water level, the seismic resistance attribute and the partition area ratio into numerical features to obtain a conversion result; Fusing the conversion result and the input image to obtain a fusion result; Extracting multimodal features from the fusion result to obtain a second feature result; Performing normalization and feature alignment processing on the second feature result to obtain a second processing result; The second processing results are stacked along the data channel to obtain the multi-dimensional feature vector.

7. The method for intelligent design of a concrete face rockfill dam profile based on data feature fusion according to claim 6 is characterized in that: The process of step S5 includes: Counting the cross-sectional profile parameters of the input image, the initial data, the coordinate parameters of each of the subareas and the cross-sectional area ratio parameters to obtain the distribution rules of different parameters; According to the distribution law, the upstream dam slope ratio, the downstream dam slope ratio, and the area of ​​each region are augmented and expanded by parameterized adjustment to construct a multi-source heterogeneous training data set and a test data set; Performing feature data fusion on the training data set and the test data set to construct a corresponding input feature data set and an output feature data set; Taking the input feature data set as an input parameter, and using the training data set to train the generative adversarial network model to obtain a second initial training model; Verifying the second initial training model according to the verification data set to obtain a second target training model; The model is trained according to the second objective to generate the zoning design image of the concrete face rockfill dam.

8. The method for intelligent design of a concrete face rockfill dam profile based on data feature fusion according to claim 7 is characterized in that: The process of augmenting and expanding the upstream dam slope ratio, the downstream dam slope ratio, and the area of ​​each region by using parameterized adjustment according to the distribution law to construct a multi-source heterogeneous training data set and a test data set includes: Determining parameter thresholds of the upstream dam slope ratio, the downstream dam slope ratio and the area of ​​each region; defining the adjustment step length of the upstream dam slope ratio, the downstream dam slope ratio and the area of ​​each region; Establishing a combination rule according to the parameter threshold and the adjustment step size; Generate a series of parameter value combinations according to the combination rule; Performing geometric transformation and simulation on the input image according to the generated parameter value combination to generate a target cross-sectional image; Assigning corresponding parameter values ​​to each of the target profile images as label information; The training data set and the test data set are constructed according to the target cross-sectional image and the label information.

9. The method for intelligent design of a concrete face rockfill dam profile based on data feature fusion according to claim 8 is characterized in that: The process of verifying the second initial training model according to the verification data set to obtain the second target training model includes: Creating a binary mask using the profile parameters of the input image, wherein the profile area is marked as 1 and the remaining profile areas are marked as 0; Filtering the output feature according to the mask to obtain a filtering result; The generated image is determined according to the filtering result to be valid only within the panel cross-section contour.

10. The method for intelligent design of a concrete face rockfill dam profile based on data feature fusion according to claim 9, characterized in that: The process of filtering the output feature according to the mask to obtain a filtering result includes: ;in, is the original output height H, width W, and channel C of the generative network, represents the element-wise multiplication Hadamard product, is the effective generated image after mask filtering, is a binary mask matrix that satisfies: 。

Images