Rock-fill dam material grading prediction method based on machine learning

Abstract

The invention discloses a rock-fill dam material gradation prediction method based on machine learning. Firstly, a camera is used for obtaining an image of rock-fill particles on a laboratory scale, and then an image instance segmentation algorithm is used for extracting particle granularity statistical data. A regression prediction model is established according to the relation between percentage distribution of different particle sizes and fractal dimensions. A new optimization algorithm is used for fine adjustment of parameters of the prediction model to predict fractal dimensions, then the percentage of each particle size is subjected to inverse calculation according to the fractal dimensions, and a particle grading curve is generated. Finally, the effectiveness of the method is verified on the scale of a construction site. The method shows high accuracy and excellent generalization ability, and verifies the effectiveness of the method in the aspect of rapid nondestructive testing of rockfill gradation; rapid grading judgment of a construction area can be completed within 5 minutes, and the requirements of modern engineering for high efficiency and high precision are greatly met.

Description

Technical Field

[0001] The present invention relates to the field of rockfill dam material construction, and in particular to a rockfill dam material gradation prediction method based on machine learning. Background Art

[0002] Rockfill dam is the main type of water conservancy engineering structure. Reasonable rockfill gradation can improve the anti-seepage performance and ensure the stability of the dam body. In engineering, the gradation test of the test points is often carried out by artificial screening. Artificial tests have problems such as poor representativeness and low efficiency, and cannot meet the needs of modern construction. Therefore, it is necessary to propose a rapid non-destructive gradation test method.

[0003] With the development of digital technology, machine learning provides a potential solution for the rapid detection of rockfill gradation. The method first collects image samples of the construction area through drones; then segments and extracts the morphological features of each rockfill; finally, the rockfill samples in the area are graded and statistically analyzed and the rockfill gradation parameters are determined by regression. Thanks to the optimization of image recognition algorithms, the analysis of granular materials such as rocks and ores has attracted widespread attention, including particle boundary detection and particle size analysis. Traditional image segmentation technology mainly identifies boundaries and contours by analyzing image gradients or second-order derivatives, but the segmentation accuracy is highly dependent on parameter settings. Compared with traditional computer vision technology, image processing methods based on convolutional neural networks have been studied in many fields, and the application effect has been highly evaluated. Relevant scholars have achieved end-to-end accurate segmentation of granular materials such as gravel and coal through deep learning, and extracted particle size.

[0004] The construction specifications of rockfill dams stipulate that the stone should be leveled by the advance method, and the layer thickness should be controlled between 0.9m and 1m. Due to the large size difference, large-sized stones will roll down first, followed by small-sized stones to fill the gaps. Therefore, the stones show non-uniform distribution characteristics in the vertical direction, and it is difficult to accurately describe the stone gradation of the entire area based only on the morphological characteristics of the rockfill on the plane. Previous studies have not paid enough attention to the differences in the vertical spatial distribution of rockfill. Summary of the invention

[0005] Purpose of the invention: In order to solve the current problems of low efficiency in manual particle screening of earth-rock dam materials and difficulty in accurately characterizing spatial rockfill gradation, the present invention proposes a method for predicting rockfill dam material gradation based on machine learning.

[0006] Technical solution: A method for predicting the material gradation of a rockfill dam based on machine learning, comprising the following steps:

[0007] (1) Conduct indoor tests to obtain surface images corresponding to different rockfill gradations;

[0008] (2) Establishing a rockfill image instance segmentation network, extracting size features and performing classification statistics on all rockfill particles in the apparent image, obtaining a grading parameter data set, and establishing a statistical standard for grading prediction particle size;

[0009] (3) Import the gradation parameter data set into the Gradation-Pre machine learning prediction model to perform global parameter optimization and regression prediction training;

[0010] (4) Collect sample images and take earth-rock dam materials in the natural environment of the construction site as the research object;

[0011] (5) Processing the rockfill image of the construction site using step (2) to obtain the hierarchical statistical results of the rockfill size characteristics, and restoring the scale of the statistical results using the similarity classification theory;

[0012] (6) Using the training results of step (3), regression prediction is performed on the characteristic statistical results after scale restoration obtained in step (5) to achieve rockfill gradation prediction in the target area.

[0013] Furthermore, the step (1) is to simulate the rockfill paving process at the construction site through indoor tests, and to establish seven rockfill material groups with different grading parameters using the similarity classification principle, so as to obtain the surface images of the rockfill under different gradings.

[0014] Furthermore, in step (2), the method described in the patent "A rockfill dam material segmentation method based on instance segmentation algorithm, CN114399516B" is used to obtain the size data of the rockfill particles and their corresponding area information, and classify them according to the particle size (2.5mm, 5mm, 10mm, 20mm, 40mm, 60mm). For each classification range, the ratio of the total area of ​​particles smaller than the particle size threshold to the area of ​​the entire plane area is calculated, and the area ratio of stones with a size greater than 2.5mm is further calculated. A total of 6 feature parameters are obtained in the data set: Feature-0 (<2.5mm), Feature-1 (<5.0mm), Feature-2 (<10.0mm), Feature-3 (<20.0mm), Feature-4 (<40.0mm), Feature-5 (<60.0mm), Feature-6 (>2.5mm)

[0015] Furthermore, the particle mass fractal distribution function of continuously graded coarse-grained soil is:

[0016]

[0017] When min Approaching 0:

[0018]

[0019] Where d is the particle size, d max is the maximum particle size, D is the fractal dimension, M(d) is the mass of particles with a particle size smaller than d, and M is the total mass of the particles.

[0020] Furthermore, the step (3) comprises:

[0021] (3.1) Construct a regression prediction network. The process of constructing the prediction value is as follows:

[0022]

[0023] where x i is the i-th sample; is the corresponding predicted value; K is the number of trees; f k The weight function corresponding to the independent tree model; J is the set of all trees, and its objective function is described as follows:

[0024]

[0025] in, is the loss function, that is With y i The difference, Ω(f k ) is a regularization term to control the complexity of the model; the second-order Taylor formula is expanded for equation (4), and the constant term is removed. The objective function obtained by integration at the tth iteration is:

[0026]

[0027] Among them, g i is the first-order derivative, h i is the second-order derivative;

[0028] (3.2) Meta-heuristic optimization algorithm is used for parameter optimization, which is divided into two stages: exploration and development;

[0029] (3.3) In the exploration phase, the algorithm first randomly generates candidate solutions globally, with the goal of covering different regions of the solution space, thereby increasing the probability of obtaining the global optimal solution and avoiding falling into the local optimum. The specific implementation method is as follows:

[0030]

[0031] in, is the candidate solution of the search space at iteration t, is the best solution for function evaluation, Refers to the vector between the current solution and the random solution, t1 is a random number based on normal distribution, and t2 is a random value in the interval [0,1];

[0032] Then, a heuristic search is performed to adjust the search direction, so that the algorithm can explore specific areas in the solution space in more detail. The specific implementation method is as follows:

[0033]

[0034] Where r1 and r2 are random integer values ​​between [1, N], and t3 is a random value between [0, 1]. is a binary vector with random value 0 or 1; when When it is 0, it indicates that the current solution has not changed. When it is 1, the current solution will move closer to or farther away from the optimal solution;

[0035] (3.4) In the development phase, the candidate solutions found in the exploration phase are refined to make them approach the optimal solution quickly; first, a local search is performed to find a better solution in the neighborhood of the current solution. The specific implementation method is as follows:

[0036]

[0037] Among them, r3 is a random integer value between [1, N], δ is a value of -1 or +1, which is used to control the search direction; γ t It is a random search factor that gradually decreases with the number of iterations, which helps the algorithm to jump out of the local optimal solution, explore the underexplored areas in the solution space, and ensure the dynamic balance between global and local search. i t It is defined as a diffusion factor between [0.3, 2.6], which dynamically adjusts the search strategy based on the overall performance of the current solution; finally, by using a deterministic local search strategy, combined with the current optimal solution, random factors, convergence speed and dynamic adjustment of the search direction, the algorithm iterates quickly within a specific interval to find the optimal solution; the specific implementation method is as follows:

[0038]

[0039] Among them, α is the convergence speed factor, t4 and t5 are random values ​​between [0,1], It represents a dynamic and guiding force, which adjusts the position of individuals according to the current state of the algorithm and the information of the objective function, thereby helping the algorithm to effectively search and find the optimal solution in a complex solution space. Through this comprehensive search strategy, the algorithm can significantly improve the search efficiency and quickly converge to the optimal solution while ensuring the quality of the solution.

[0040] Furthermore, the step (4) is to use a drone to take images of the surface of the earth-rock dam construction site, and the shooting time is after the surface is paved and the site is leveled, and before vibration rolling; considering the image acquisition quality and to ensure the stability of the model, the image acquisition environment is divided into five types according to weather conditions, including sunny days, cloudy days, evenings, rainy days, and haze weather, and the rock types of different material yards are collected according to the degree of weathering of the rocks, including slightly weathered and unweathered.

[0041] Furthermore, in step (5), due to the different particle size ranges of the construction site and the basic experiment, the particle size thresholds corresponding to the input features of the regression model will be different. In order to extract the features related to the target task in a hierarchical manner, this study calculated the change particle size threshold of the rockfill material based on the similarity fractal principle:

[0042]

[0043] Where, d i Enter the particle size threshold corresponding to feature-i in the target task, is the particle size threshold corresponding to feature-i input in the basic experiment, d max is the maximum particle size of the filling stone in the target task, It is the maximum particle size of the filling stone in the basic experiment.

[0044] Furthermore, the step (6) is: inputting the characteristic statistical result after scale restoration obtained in step (5) into the regression prediction model constructed in step (3) to obtain the corresponding fractal dimension; and transforming the fractal dimension using the similarity classification principle to obtain the predicted rockfill gradation curve of the image area.

[0045] An electronic device comprises a processor and a memory; the memory is used to store a computer program, and the computer program is run by the processor to execute a method for predicting the material gradation of a rockfill dam based on machine learning.

[0046] A computer storage medium, characterized in that the computer storage medium stores a computer program, which is run by a processor to execute a method for predicting the gradation of rockfill dam materials based on machine learning.

[0047] Beneficial effects: 1. Laboratory tests and field applications show that the method exhibits high accuracy and excellent generalization ability, confirming its effectiveness in achieving rapid non-destructive detection of rockfill gradation; 2. It can quickly determine the gradation of the construction area within 5 minutes, greatly meeting the requirements of modern engineering for high efficiency and high precision. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 is a flow chart of the present invention;

[0049] Figure 2 It is the simulation test method of the present invention;

[0050] Figure 3 This is the comparison of prediction results of 100 test data;

[0051] Figure 4 It is a comparison between the grading determination result and the actual value. DETAILED DESCRIPTION

[0052] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0053] The present invention first uses a camera to acquire images of rockfill particles at a laboratory scale, and then uses an image instance segmentation algorithm to extract particle size statistics. A regression prediction model is established based on the relationship between the percentage distribution of different particle sizes and the fractal dimension. A new optimization algorithm is used to fine-tune the prediction model parameters to predict the fractal dimension, and then the percentage of each particle size is inversely deduced based on the fractal dimension to generate a particle grading curve. Finally, the effectiveness of the proposed method is verified at the construction site scale.

[0054] like Figure 1 As shown, a method for predicting material gradation of rockfill dam based on machine learning includes the following steps:

[0055] (1) Conduct indoor tests to obtain surface images corresponding to different rockfill gradations;

[0056] (2) Establishing a rockfill image instance segmentation network, extracting size features and performing classification statistics on all rockfill particles in the apparent image, obtaining a grading parameter data set, and establishing a statistical standard for grading prediction particle size;

[0057] (3) Import the gradation parameter data set into the Gradation-Pre machine learning prediction model to perform global parameter optimization and regression prediction training;

[0058] (4) Collect sample images and take earth-rock dam materials in the natural environment of the construction site as the research object;

[0059] (5) Processing the rockfill image of the construction site using step (2) to obtain the hierarchical statistical results of the rockfill size characteristics, and restoring the scale of the statistical results using the similarity classification theory;

[0060] (6) Using the training results of step (3), regression prediction is performed on the characteristic statistical results after scale restoration obtained in step (5) to achieve rockfill gradation prediction in the target area.

[0061] The step (1) simulates the stone paving process in the laboratory by using a proportional scaling method, with the purpose of obtaining the surface image of the stone pile under different gradations. Figure 2 As shown, the stone paving process of the simulation experiment includes: unloading, leveling, and compacting. In the construction organization design of rockfill dam, the maximum size of the rockfill is between 600mm and 800mm, and the fractal dimension of the dam material is limited to between 2.55 and 2.65. Fractal dimension D is a common method to characterize the grading curve of continuously graded earth-rock materials. The larger the fractal dimension, the more fine particles there are. According to the fractal theory of grading of earth-rock dam materials, seven kinds of sample data with fractal dimensions of 2.45, 2.50, 2.55, 2.60, 2.65, 2.70, and 2.75 are produced. The particle mass fractal distribution function of continuously graded coarse-grained soil is:

[0062]

[0063] When min Approaching 0:

[0064]

[0065] Where d is the particle size, d max is the maximum particle size, D is the fractal dimension, M(d) is the mass of particles with a particle size smaller than d, and M is the total mass of particles. Through the proportional scaling method, the experimental particle size scale is controlled to be 0.1-60mm. Table 1 shows the detailed particle size and the mass proportion of particles smaller than this size under different fractal dimensions.

[0066] Table 1 The proportion of particle size intervals with different fractal dimensions

[0067]

[0068]

[0069] In step (2), the method described in the patent "A rockfill dam material segmentation method based on instance segmentation algorithm, CN114399516B" is used to obtain the size data of the rockfill particles and their corresponding area information, and classify them according to the particle size (2.5mm, 5mm, 10mm, 20mm, 40mm, 60mm). For each classification range, the ratio of the total area of ​​particles smaller than the particle size threshold to the area of ​​the entire plane area is calculated, and the area ratio of stones larger than 2.5mm is further calculated. A total of 6 feature parameters are obtained in the data set: Feature-0 (<2.5mm), Feature-1 (<5.0mm), Feature-2 (<10.0mm), Feature-3 (<20.0mm), Feature-4 (<40.0mm), Feature-5 (<60.0mm), Feature-6 (>2.5mm)

[0070] The step (3) comprises:

[0071] The characteristic parameters obtained in step (2) and the corresponding gradation characteristics are used as training data to construct the Gradation-Pre regression prediction model, and the global optimal solution is found by adopting a meta-heuristic optimization strategy. The regression prediction algorithm consists of a series of decision trees. The core goal is to accurately fit the residuals by learning a new function f(x) so that the predicted values ​​of all trees are as close to the actual observed values ​​as possible. The process of constructing the predicted value is as follows:

[0072]

[0073] where x i is the i-th sample; is the corresponding predicted value; K is the number of trees; f k The weight function corresponding to the independent tree model; J is the set of all trees, and its objective function is described as follows:

[0074]

[0075] in, is the loss function, that is With y i The difference, Ω(f k ) is a regularization term to control the complexity of the model; the second-order Taylor formula is expanded for equation (4), and the constant term is removed. The objective function obtained by integration at the tth iteration is:

[0076]

[0077] Among them, g i is the first-order derivative, h i is the second-order derivative;

[0078] The meta-heuristic optimization algorithm is used for parameter optimization, which is mainly divided into two stages: exploration and development. In the development stage, the algorithm first randomly generates candidate solutions globally, with the aim of covering different areas of the solution space, thereby increasing the probability of obtaining the global optimal solution and avoiding falling into the local optimum. The specific implementation method is as follows:

[0079]

[0080] in, is the candidate solution of the search space at iteration t, is the best solution for function evaluation, Refers to the vector between the current solution and the random solution, t1 is a random number based on normal distribution, and t2 is a random value in the interval [0,1]; then a heuristic search is performed to adjust the search direction so that the algorithm can explore specific areas in the solution space in more detail. The specific implementation method is as follows:

[0081]

[0082] Where r1 and r2 are random integer values ​​between [1, N], and t3 is a random value between [0, 1]. is a binary vector with random value 0 or 1; when When it is 0, it indicates that the current solution has not changed. When it is 1, the current solution will move closer to or farther away from the optimal solution;

[0083] In the development phase, the candidate solutions found in the exploration phase are refined to quickly approach the optimal solution; first, a local search is performed to find a better solution in the neighborhood of the current solution. The specific implementation method is as follows:

[0084]

[0085] Among them, r3 is a random integer value between [1, N], δ is a value of -1 or +1, which is used to control the search direction; γ t It is a random search factor that gradually decreases with the number of iterations, which helps the algorithm to jump out of the local optimal solution, explore the underexplored areas in the solution space, and ensure the dynamic balance between global and local search. i t It is defined as a diffusion factor between [0.3, 2.6], which dynamically adjusts the search strategy based on the overall performance of the current solution; finally, by using a deterministic local search strategy, combined with the current optimal solution, random factors, convergence speed and dynamic adjustment of the search direction, the algorithm iterates quickly within a specific interval to find the optimal solution; the specific implementation method is as follows:

[0086]

[0087] Among them, α is the convergence speed factor, t4 and t5 are random values ​​between [0,1], It represents a dynamic and guiding force, which adjusts the position of individuals according to the current state of the algorithm and the information of the objective function, thereby helping the algorithm to effectively search and find the optimal solution in a complex solution space. Through this comprehensive search strategy, the algorithm can significantly improve the search efficiency and quickly converge to the optimal solution while ensuring the quality of the solution.

[0088] Furthermore, the step (4) is to use a drone to take images of the surface of the earth-rock dam construction site, the shooting time is after the surface is paved and the site is leveled, and before vibration rolling; the particle size range of the rockfill material is 0-800 mm; the drone is set to fly at a net altitude of 3 meters, and after calibration, a single image corresponds to an actual construction site size of 4*3 meters; considering the image acquisition quality and to ensure the stability of the model, the image acquisition environment is divided into five types according to weather conditions, including sunny days, cloudy days, evenings, rainy days, and haze weather, and the rock types of different material yards are collected according to the degree of weathering of the rocks, including slightly weathered and unweathered.

[0089] Furthermore, in step (5), due to the different particle size ranges of the construction site and the basic experiment, the particle size thresholds corresponding to the input features of the regression model will be different. In order to extract the features related to the target task in a hierarchical manner, this study calculated the change particle size threshold of the rockfill material based on the similarity fractal principle, following the following equation:

[0090]

[0091] Where, d i Enter the particle size threshold corresponding to feature-i in the target task, is the particle size threshold corresponding to feature-i input in the basic experiment, d max is the maximum particle size of the filling stone in the target task, It is the maximum particle size of the filling stone in the basic experiment.

[0092] Furthermore, the step (6) is: inputting the characteristic statistical result after scale restoration obtained in step (5) into the regression prediction model constructed in step (3) to obtain the corresponding fractal dimension; and transforming the fractal dimension using the similarity classification principle to obtain the predicted rockfill gradation curve of the image area.

[0093] The total amount of data sets produced in the study is 500, of which the training set has 400 samples and the test set has 100 samples. The performance of the present invention and several other machine learning models on regression tasks is compared. As can be seen from Table 2, the model performance of Gradation-Pre is excellent, with the coefficient of determination (R^2), mean square error (MSE) and mean absolute error (MAE) being 0.967, 3.7e-4 and 1.328e-2 respectively. These results show that the Gradation-Pre model surpasses some traditional classical models in accuracy, thus fully proving the effectiveness of the method proposed in this study.

[0094] Table 2 Machine learning model evaluation

[0095]

[0096] Figure 3 The verification results of 100 test data are shown. The closer the data point is to the function line y=x, the smaller the difference between the predicted value and the true value. It can be observed from the figure that most of the fractal dimension prediction results are concentrated near the true value, and the deviation of most of the prediction points is within plus or minus 0.025. This shows that the effective planar characteristics of the rockfill are strongly correlated with the overall fractal dimension. Although there is a large deviation between some predicted values ​​and the true value, this is difficult to avoid in actual situations. Because stones are non-viscous discrete particles, in experiments and actual engineering operations, it is impossible to ensure that the particles can be completely evenly distributed in the box before dumping. In addition, the irregular shape of the particles will also cause some interference to the final measurement results. Overall, the method used still shows high accuracy and reliability in predicting fractal dimensions.

[0097] Two sample images obtained during construction were randomly selected for prediction. The predicted fractal dimension was plotted as a gradation curve, with the minimum particle size set to 0.1 mm and the maximum particle size set to the upper limit of the actual detection particle size range of 600 mm. Figure 4 The comparison between the gradation determination results and the actual values ​​is shown. Among them, the solid line represents the predicted value, the dotted line represents the true value, and the black curve is the upper and lower envelope. It can be seen from the curve that the predicted gradation curve is smoother than the actual curve, while the actual curve fluctuates around the predicted curve. This phenomenon reflects that since the rockfill material itself is a loose particle body, the upper and lower layers are unevenly distributed, and it is difficult to accurately predict the mass ratio of particles in each particle size range by relying solely on the prediction method. Despite this, the similarity between the predicted gradation curve and the actual gradation curve still shows that the fractal dimension method used has high effectiveness and reliability in gradation prediction.

Claims

1. A method for predicting the gradation of rockfill dam materials based on machine learning, characterized in that: The following steps are involved: (1) Conduct indoor tests to obtain surface images corresponding to different rockfill gradations; (2) Establishing a rockfill image instance segmentation network, extracting size features and performing classification statistics on all rockfill particles in the apparent image, and obtaining a grading parameter dataset; (3) Import the gradation parameter data set into the Gradation-Pre machine learning prediction model to perform global parameter optimization and regression prediction training; (4) Collect sample images and take earth-rock dam materials in the natural environment of the construction site as the research object; (5) Processing the rockfill image of the construction site using step (2) to obtain the hierarchical statistical results of the rockfill size characteristics, and restoring the scale of the statistical results using the similarity classification theory; (6) Using the training results of step (3), regression prediction is performed on the characteristic statistical results after scale restoration obtained in step (5) to achieve rockfill gradation prediction in the target area.

2. The method for predicting the material gradation of rockfill dam based on machine learning according to claim 1, characterized in that: The step (1) simulates the stone paving process in the laboratory by using a proportional scaling method to obtain the surface images of the stone pile under different gradations.

3. The method for predicting material gradation of rockfill dam based on machine learning according to claim 1, characterized in that: The step (2) is to obtain the size data of the rockfill particles and their corresponding area information, and classify them according to the particle size; for each classification range, calculate the ratio of the total area of ​​particles smaller than the particle size threshold to the area of ​​the entire plane area.

4. The method for predicting material gradation of rockfill dam based on machine learning according to claim 1, characterized in that: The particle mass fractal distribution function of continuously graded coarse-grained soil is: When min Approaching 0: Where d is the particle size, d max is the maximum particle size, D is the fractal dimension, M(d) is the mass of particles with a particle size smaller than d, and M is the total mass of the particles.

5. The method for predicting the material gradation of rockfill dam based on machine learning according to claim 1, characterized in that: The step (3) comprises: (3.1) Construct a regression prediction network. The process of constructing the prediction value is as follows: where x i is the i-th sample; is the corresponding predicted value; K is the number of trees; f k The weight function corresponding to the independent tree model; J is the set of all trees, and its objective function is described as follows: in, is the loss function, that is With y i The difference, Ω(f k ) is a regularization term to control the complexity of the model; the second-order Taylor formula is expanded for equation (4), and the constant term is removed. The objective function obtained by integration at the tth iteration is: Among them, g i is the first-order derivative, h i is the second-order derivative; (3.2) Meta-heuristic optimization algorithm is used for parameter optimization, which is divided into two stages: exploration and development; (3.3) In the exploration phase, the algorithm first randomly generates candidate solutions globally. The specific implementation method is as follows: in, is the candidate solution of the search space at iteration t, is the best solution for function evaluation, Refers to the vector between the current solution and the random solution, t1 is a random number based on normal distribution, and t2 is a random value in the interval [0,1]; Then, a heuristic search is performed to adjust the search direction. The specific implementation method is as follows: Where r1 and r2 are random integer values ​​between [1, N], and t3 is a random value between [0, 1]. is a binary vector with random value 0 or 1; when When it is 0, it indicates that the current solution has not changed. When it is 1, the current solution will move closer to or farther away from the optimal solution; (3.4) In the development phase, the candidate solutions found in the exploration phase are refined to make them approach the optimal solution quickly; first, a local search is performed to find a better solution in the neighborhood of the current solution. The specific implementation method is as follows: Among them, r3 is a random integer value between [1, N], δ is a value of -1 or +1, which is used to control the search direction; γ t is a random search factor that decreases as the number of iterations increases. It is defined as a diffusion factor between [0.3, 2.6], which dynamically adjusts the search strategy based on the overall performance of the current solution; finally, by using a deterministic local search strategy, combined with the current optimal solution, random factors, convergence speed and dynamic adjustment of the search direction, the algorithm iterates quickly within a specific interval to find the optimal solution; the specific implementation method is as follows: Among them, α is the convergence speed factor, t4 and t5 are random values ​​between [0,1], It represents a dynamic and guiding force, which adjusts the position of individuals according to the current state of the algorithm and the information of the objective function, thereby helping the algorithm to effectively search and find the optimal solution in a complex solution space.

6. The method for predicting the material gradation of rockfill dam based on machine learning according to claim 1, characterized in that: The step (4) is to use a drone to take images of the surface of the earth-rock dam construction site. The shooting time is after the surface is paved and the site is leveled, and before vibration rolling. The image acquisition environment is divided into five types according to weather conditions, including sunny days, cloudy days, evenings, rainy days, and haze weather. Rock types from different material yards are collected according to the degree of weathering of the rocks, including slightly weathered and unweathered.

7. The method for predicting the material gradation of rockfill dam based on machine learning according to claim 1, characterized in that: In step (5), the threshold value of the change in particle size of the rockfill material is calculated based on the similarity fractal principle: Where, d i Enter the particle size threshold corresponding to feature-i in the target task, is the particle size threshold corresponding to feature-i input in the basic experiment, d max is the maximum particle size of the filling stone in the target task, It is the maximum particle size of the filling stone in the basic experiment.

8. The method for predicting material gradation of rockfill dam based on machine learning according to claim 1, characterized in that: The step (6) is: inputting the characteristic statistical result after scale restoration obtained in step (5) into the regression prediction model constructed in step (3) to obtain the corresponding fractal dimension; and transforming the fractal dimension using the similarity classification principle to obtain the predicted rockfill gradation curve of the image area.

9. An electronic device, characterized in that: The method comprises a processor and a memory; the memory is used to store a computer program, and the computer program is run by the processor to execute the method according to any one of claims 1 to 8.

10. A computer storage medium, characterized in that: The computer storage medium stores a computer program, which is executed by a processor to perform the method according to any one of claims 1 to 8.

Images