Rockfill dry density prediction model training method and rockfill dry density prediction system

Abstract

The invention relates to rock-fill dry density prediction, particularly provides a training method of a rock-fill dry density prediction model and a rock-fill dry density prediction system, and aims to solve the problem that a method for interpreting the rock-fill dry density by an additional mass method is poor in scene adaptability due to model simplification and high dependence on specific parameter and local data fitting. And therefore, the problems of low prediction precision, poor robustness and poor universality in an engineering scene can be solved. In order to achieve the purpose, the training method of the rock-fill body dry density prediction model comprises the steps that the corresponding rock-fill body dry density prediction model is established according to scene categories; and under the scene category, forming an input feature vector by the first frequency, the maximum frequency difference, the vibration mass and the dynamic stiffness of the single measuring point of the rockfill, and training the rockfill dry density prediction model by taking the dry density of the measuring point as an output target. According to the method, the rock-fill body dry density prediction model based on the scene category is constructed, so that the nondestructive testing precision and robustness of the rock-fill body dry density are remarkably improved.

Description

Technical Field

[0001] This invention relates to the prediction of dry density of rockfill bodies, specifically providing a training method for a dry density prediction model of rockfill bodies and a dry density prediction system for rockfill bodies. Background Technology

[0002] One of the core quality control indicators for the construction of rockfill dams, roadbeds, and other engineering projects is the compacted dry density of the rockfill mass. Although the traditional pit test method is recognized as the benchmark method for dry density testing, it has inherent drawbacks such as being destructive, inefficient, costly, and having limited representativeness, and cannot meet the needs of modern engineering projects for large-scale, rapid, and non-destructive testing.

[0003] Therefore, non-destructive testing techniques, represented by the mass-added method, have been widely applied. This method involves applying an additional mass to the surface of the rockfill and collecting vibration signals to obtain its dynamic parameters, thereby indirectly deriving physical parameters such as dry density. Its core interpretation methods can be summarized into the following four categories, but all have significant limitations: 1. Correlation Method, also known as Unit Regression Method. This method is based on statistical principles and establishes a univariate or multivariate regression model between dynamic parameters and pit-measured dry density. Its fundamental drawbacks are: it usually assumes a simple linear relationship between dynamic parameters and density, ignoring the nonlinear and multi-factor coupling characteristics of rockfill as a complex granular material, leading to oversimplification of the model; the established regression model is highly dependent on data from specific projects or material sources, and once the project scenario, material gradation, and moisture content change, the model's prediction accuracy will drop sharply, lacking robustness and transferability, resulting in poor model universality; it usually only uses a few parameters for fitting, failing to fully explore the rich information contained in the vibration signal, resulting in limited model representation ability and insufficient utilization of model information.

[0004] 2. The calibration coefficient method, based on the mass-elastic model theory, introduces a calibration coefficient N related to the properties of the rockfill and calibrates it using known pit-measured densities. Its key drawbacks are: a fragile theoretical foundation, as the formula derivation relies on the seismic wave velocity Vp0, and accurately obtaining Vp0 is a significant technical challenge. Rockfill exhibits strong anisotropy; the wave velocity difference between different test directions at the same location can exceed 30%, leading to the physical paradox of "higher wave velocity, lower inverted density" when calculating density values ​​based on a single-direction wave velocity, severely compromising the reliability and accuracy of the method. The parameter's physical meaning is ambiguous: the calibration coefficient N is a lumped parameter lacking a clear physical meaning; its value fluctuates drastically with changes in the material area and operating conditions, essentially still "calibrating a measurement point with a measurement point," making it difficult to predict uncalibrated areas.

[0005] 3. Plate Method / Isoline Method: This method calculates density by establishing a two-dimensional interpolation map of the parameters of the added mass method and the results of the pit measurement method. The parameters of the added mass method include dynamic stiffness and vibrating mass, while the results of the pit measurement method include wet density and moisture content. The two-dimensional interpolation map includes a plate or isolines. Its core issues include: sensitivity to data distribution; constructing the plate requires that the measurement points be uniformly distributed in the parameter space. Actual engineering data is often discrete and non-uniform, leading to local distortions and abnormal density gradients in sparse data regions. It is highly dependent on the chosen algorithm; different algorithms perform differently in terms of smoothness, detail restoration, and trend prediction, lacking a unified and optimal selection criterion, thus introducing uncertainty. Significant boundary effects occur; in data boundary regions, due to the lack of neighboring data support, the interpolation extrapolation results often have large errors and low reliability.

[0006] In summary, all existing interpretation techniques for the augmented mass method share common bottlenecks: weak model adaptability, strong scene dependence, and insufficient accuracy and stability. The root cause lies in the failure to systematically address the following issues: (1) How to establish a high-precision density prediction model that can integrate multi-dimensional dynamic characteristics (such as first frequency, frequency difference, vibrating mass, dynamic stiffness, etc.) with complex engineering scenario attributes; (2) How to mine and solidify transferable "knowledge" from a large amount of cross-engineering and cross-scenario detection data to support high-precision prediction in new scenarios, rather than relying solely on shallow fitting of specific datasets.

[0007] Therefore, there is an urgent need in this field for a training method and a system for predicting the dry density of rockfill bodies to solve the above problems. Summary of the Invention

[0008] The present invention aims to solve the above-mentioned technical problems, namely, to solve the problems that the current method of interpreting the dry density of rockfill is due to its simplified model, poor scene adaptability and high dependence on specific parameters and local data fitting, resulting in low prediction accuracy, insufficient robustness and universality in complex and ever-changing engineering scenarios.

[0009] In a first aspect, the present invention provides a training method for a dry density prediction model of a rockfill mass, the training method comprising the following steps: Establish a corresponding dry density prediction model for riprap based on the scene category; Under the aforementioned scenario category, the initial frequency, maximum frequency difference, vibrating mass, and dynamic stiffness of a single measuring point of the rockfill body are used as the input feature vector, and the dry density of the measuring point is used as the output target to train the dry density prediction model of the rockfill body.

[0010] In a specific implementation of the above training method, the scene category includes multi-level scene categories; and / or The input feature vector and the corresponding output target constitute a knowledge unit, and the number of knowledge units used to train the dry density prediction model of the rockfill mass is at least 200; and / or The dry density prediction model for the rockfill mass includes an input layer, a hidden layer, and an output layer. The number of hidden layers is determined based on the number of knowledge units used to train the dry density prediction model; and / or Training the dry density prediction model for the rockfill mass also includes the following steps: Based on the actual dry density value and the predicted dry density value output by the rockfill dry density prediction model, determine whether the rockfill dry density prediction model is qualified.

[0011] In a specific implementation of the above training method, "training the dry density prediction model of the rockfill body" includes the following steps: The data of the knowledge units are preprocessed to obtain standardized training data; The dry density prediction model for the rockfill is trained using the standardized training data.

[0012] In a specific implementation of the above training method, "determining whether the dry density prediction model of the rockfill body is qualified based on the predicted dry density value and the actual dry density value output by the rockfill dry density prediction model" includes the following steps: Calculate the absolute error between the predicted dry density value and the actual value. The formula for calculating the absolute error is as follows: ; The dry density prediction model is deemed qualified based on whether the absolute error is greater than a first preset value; or Calculate the relative error between the predicted dry density value and the actual value. The formula for calculating the relative error is as follows: ; Whether the dry density prediction model is qualified is determined based on whether the relative error is greater than a second preset value.

[0013] In a second aspect, the present invention provides a rockfill dry density prediction system, which includes multiple rockfill dry density prediction models corresponding one-to-one with multiple scene categories, and the rockfill dry density prediction models are trained by the training method described above.

[0014] In a third aspect, the present invention provides a method for predicting the dry density of rockfill bodies, which is applied to the aforementioned rockfill dry density prediction system. The prediction method includes the following steps: Select the corresponding dry density prediction model for the rockfill based on the scene category of the rockfill; The initial frequency, maximum frequency difference, vibrating mass, and dynamic stiffness of a single measuring point of the rockfill are input into the dry density prediction model of the rockfill, and the dry density prediction model of the rockfill outputs the predicted dry density value of the measuring point.

[0015] In a third aspect, the present invention provides a dry density prediction device for rockfill bodies, which is applied to the aforementioned dry density prediction system for rockfill bodies.

[0016] In a fourth aspect, the present invention provides a computer device comprising one or more processors and a memory; wherein the memory stores computer-readable instructions; and the one or more processors read the computer-readable instructions to cause the computer device to perform the method described above.

[0017] In a fifth aspect, the present invention provides a computer-readable storage medium comprising computer-readable instructions that, when executed on a computer, cause the computer to perform the methods described above.

[0018] In a sixth aspect, the present invention provides a computer program product comprising computer-readable instructions that, when executed on a computer, cause the computer to perform the methods described above.

[0019] By employing the above technical solutions, this invention significantly improves the accuracy and robustness of non-destructive testing of rockfill dry density by constructing a scene-category-based dry density prediction model. Specifically: 1) Multi-factor integrated modeling fully integrates multi-dimensional dynamic characteristics such as initial frequency, frequency difference, vibrating mass, and dynamic stiffness, avoiding the simplification and distortion of traditional single-factor or linear models; 2) Completely abandoning the dependence on wave velocity Vp0, fundamentally eliminating the "velocity paradox" caused by anisotropy, ensuring parameter stability and result comparability; 3) Based on the nonlinear mapping capability of deep learning, it automatically captures the complex characteristics of rockfill, with inversion errors significantly lower than traditional methods such as the calibration coefficient method; 4) Relying on deep feature extraction, it improves the adaptability to heterogeneous materials containing boulders and interlayers, effectively suppressing the "pseudo-anomaly zone" phenomenon in the contour line method. Overall, this invention forms a closed-loop technology chain of "scene classification, database training, and precise matching," providing a high-precision, high-reliability, and universally applicable intelligent solution for rockfill compaction quality testing.

[0020] Furthermore, this invention achieves systematic optimization of training data quality and consistency by performing data preprocessing for each sub-knowledge base, thus laying a solid foundation for accurate training and stable prediction of the subsequent model. Specifically, data cleaning and outlier removal effectively filter out noise and invalid data introduced by equipment, environment, or human operation, ensuring the authenticity and reliability of knowledge units from the source and significantly reducing the risk of interference from poor-quality data to model training. Feature normalization unifies the order of magnitude and scale of each feature parameter under different engineering scenarios, eliminating the weight imbalance problem caused by the difference in the units of features. This not only greatly improves the training efficiency and convergence speed of the deep learning model, but also provides a unified data benchmark for the model to perform stable and reliable generalization predictions in different scenarios. This preprocessing step is an indispensable key technical link for this method to achieve high-precision and high-robust dry density prediction. Detailed Implementation

[0021] Preferred embodiments of the present invention will now be described. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0022] It should be noted that in the description of this invention, the terms "upper," "lower," "left," "right," "inner," and "outer," which indicate directional or positional relationships, are merely for ease of description and do not indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0023] Furthermore, it should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "setting," and "connection" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a direct connection, an indirect connection through an intermediate medium, or a connection within two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0024] To address the problems of low prediction accuracy, insufficient robustness, and inadequate universality in complex and variable engineering scenarios caused by the current method of interpreting dry density of rockfill due to model simplification, poor scene adaptability, and high dependence on fitting specific parameters and local data, this embodiment discloses a training method for a dry density prediction model of rockfill. The method includes the following steps: S101. Construct a knowledge base based on the original data. The knowledge base includes multiple knowledge units. Each knowledge unit includes a corresponding scene category, parameters, and dry density. The scene categories are multi-level. In this embodiment, the scene categories include four levels. Specifically, the first-level scene category is the project name; the second-level scene category is the unit project, such as upper reservoir and lower reservoir; the third-level scene category is the sub-project, such as main dam, secondary dam, reservoir area, and drainage body; the fourth-level scene category is the sub-item project, such as cushion material, main stockpile, secondary stockpile, and transition material. Parameters include initial frequency, maximum frequency difference, vibrating mass, dynamic stiffness, and dry density. The initial frequency is denoted by f0 and its unit is Hz. The maximum frequency difference is denoted by Δf0 and its unit is Hz. The vibrating mass is denoted by M0 and its unit is kg. The dynamic stiffness is denoted by K and its unit is kN / mm. The dry density is obtained through pit testing and its symbol is ρ, with a unit of g / cm³.

[0025] S102, classify all knowledge units according to scenario categories, and knowledge units with the same scenario category form sub-knowledge bases. When the number of knowledge units in a sub-knowledge base is less than 200, supplement the number of knowledge units using a sample label expansion method to ensure that the number of knowledge units in each sub-knowledge base is no less than 200. The sample label expansion method is the same as the sample label expansion method in the prior art, and its specific steps will not be repeated here.

[0026] The following is an example of the knowledge base in this embodiment:

[0027] S103. Preprocessing is performed on the data of all knowledge units to obtain standardized training data. Preprocessing specifically includes data cleaning, outlier removal, and feature normalization. Data cleaning and outlier removal remove invalid data generated during the collection process due to equipment errors, environmental interference, or improper operation, preventing poor-quality data from causing model training bias and ensuring the authenticity and reliability of the input data. Feature normalization eliminates the scale influence of different scene categories on the measurement point data, ensuring that all features are of the same order of magnitude, avoiding interference with model parameter optimization due to feature weight imbalance, and improving training efficiency and convergence speed. Through the above preprocessing, a corresponding standardized training dataset is generated for each sub-knowledge base.

[0028] S104 takes the first frequency, maximum frequency difference, vibrating mass, and dynamic stiffness as input features, and dry density as the output target. Specifically, the vector of input features is... The output target is .in, , , , These are the first frequency, maximum frequency difference, vibrating mass, and dynamic stiffness of the i-th knowledge unit in the sub-knowledge base, respectively. Let n be the dry density of the i-th knowledge unit in the sub-knowledge base, and n be the total number of knowledge units in the sub-knowledge base.

[0029] S105. For each sub-knowledge base, establish a corresponding dry density prediction model for the rockfill. Specifically, based on the sample size and parameter characteristics of the sub-knowledge base, select a dry density prediction model such as RNN, CNN, or improved extreme learning machine as the basic architecture, preferentially using a fully connected or convolutional structure of input layer-hidden layer-output layer. To ensure that the model complexity matches the data volume of the sub-knowledge base, the number of hidden layers is determined according to the number of knowledge units used to train the dry density prediction model. Specifically, if the number of knowledge units is 200-500, use 2-3 hidden layers; if the number of knowledge units is more than 500, use 3-5 hidden layers.

[0030] S106. The corresponding dry density prediction model of the rockfill body is pre-trained using standardized training data from the sub-knowledge base, so that the dry density prediction model can form a precise correlation mapping mechanism and dry density prediction capability between the input features of the rockfill body measurement points and the dry density under the corresponding four-level scene category.

[0031] Furthermore, after the dry density prediction model for the rockfill mass is trained, an identifier is added to the model based on the four-level scene category, training completion time, and performance indicators. For example, XX Project - Upper Reservoir - Main Dam - Subbase Material_20251111_MAE0.023. The complete model parameter file is saved, including the weights, biases, hyperparameter configurations, activation functions, etc., for subsequent use and fine-tuning. The model's training set, validation set, and test set loss curves, key performance indicators, etc., are recorded to generate a model performance report, serving as the basis for subsequent cross-scene fusion or transfer learning.

[0032] This embodiment also discloses a rockfill dry density prediction system, which includes multiple rockfill dry density prediction models that correspond one-to-one with multiple scene categories.

[0033] This embodiment also discloses a method for predicting the dry density of rockfill bodies. This method utilizes the aforementioned rockfill dry density prediction system and includes the following steps: S201, Select the corresponding dry density prediction model for the rockfill based on the scene category of the rockfill.

[0034] S202 involves preprocessing parameters from multiple measuring points on the rockfill to obtain standardized parameters. These parameters include the initial frequency, maximum frequency difference, vibrating mass, and dynamic stiffness. This preprocessing method is consistent with the preprocessing method used when pre-training the dry density prediction model.

[0035] S203, construct the input feature vector based on the standardized parameters. The input feature vector is arranged in the order of first frequency, maximum frequency difference, vibrating mass, and dynamic stiffness.

[0036] S204, input the input feature vector into the dry density prediction model, and the dry density prediction model outputs the predicted value of dry density.

[0037] S205. Select a portion of the measuring points from multiple measuring points in the rockfill body, perform actual measurements on the selected measuring points, and obtain the true value of the dry density.

[0038] S206. Accuracy verification is performed based on the predicted and actual values ​​of the dry density at the selected measurement points. Accuracy verification includes the following two implementation methods.

[0039] The first embodiment of accuracy verification is as follows: calculate the absolute error between the predicted and actual values ​​of dry density. The absolute error is equal to the absolute value of the difference between the predicted and actual values. If the absolute value of the error is not greater than 0.05 g / cm³, it means that the dry density prediction model meets the engineering testing accuracy standard.

[0040] The second embodiment of accuracy verification is as follows: calculate the relative error between the predicted and actual values ​​of dry density. The relative error is equal to the ratio of the absolute value of the difference between the predicted and actual values ​​to the actual value. If the relative error is not greater than 2.5%, it means that the dry density prediction model meets the engineering testing accuracy standard.

[0041] If the dry density prediction model does not meet the engineering inspection accuracy standards, the target prediction model will be fine-tuned and trained based on the actual values ​​of the sampled samples and the corresponding input parameters until the prediction error meets the requirements. If the dry density prediction model meets the engineering inspection accuracy standards, a prediction report will be output.

[0042] S207, the prediction report includes the predicted dry density values ​​at multiple measurement points, the corresponding prediction accuracy index, the four-level scene category of the rockfill, and the dry density prediction model identifier.

[0043] The following example illustrates the specific process of this method for predicting the dry density of rockfill: There are 20 measurement points to be predicted, with their fourth-level scene category being XX Project - Upper Reservoir - Main Dam - Main Material Stockpile. A pre-trained network model that perfectly matches the traceability attributes of this four-dimensional scene was retrieved and matched. The multi-parameters of the 20 measurement points were preprocessed according to the standard process consistent with the sub-knowledge base data preprocessing to form standardized input feature vectors. These vectors were then input into the pre-trained network model to predict the dry density of the measurement points, as detailed in the table below. Measurement points 5 and 15 were sampled, and their pit-based dry densities were measured, as shown in Table 2. The absolute error of measurement point 5 was 0.03 g / cm³, and the absolute error of measurement point 15 was 0.04 g / cm³, both satisfying the error ≤ 0.05 g / cm³. The relative error of measurement point 5 was 1.29%, and the relative error of measurement point 15 was 1.75%, both satisfying the error ≤ 2.5%, both meeting the engineering inspection accuracy standard, demonstrating the reliability of the pre-trained network model.

[0044] This embodiment also discloses a dry density prediction device for rockfill bodies, which is applied to the above-mentioned dry density prediction method for rockfill bodies.

[0045] This embodiment also discloses a computer device, which includes one or more processors and a memory; wherein the memory stores computer-readable instructions; the one or more processors read the computer-readable instructions to enable the computer device to implement the above-described method for predicting the dry density of rockfill.

[0046] This embodiment also discloses a computer-readable storage medium including computer-readable instructions that, when executed on a computer, cause the computer to perform the above-described method for predicting the dry density of rockfill.

[0047] This embodiment also discloses a computer program product, which includes computer-readable instructions that, when executed on a computer, cause the computer to perform the above-described method for predicting the dry density of rockfill.

[0048] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles 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 such changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A training method for a dry density prediction model of riprap, characterized in that, Includes the following steps: Establish a corresponding dry density prediction model for riprap based on the scene category; Under the aforementioned scenario category, the initial frequency, maximum frequency difference, vibrating mass, and dynamic stiffness of a single measuring point of the rockfill body are used as the input feature vector, and the dry density of the measuring point is used as the output target to train the dry density prediction model of the rockfill body.

2. The training method according to claim 1, characterized in that, The scene categories include multi-level scene categories; and / or The input feature vector and the corresponding output target constitute a knowledge unit, and the number of knowledge units used to train the dry density prediction model of the rockfill mass is at least 200; and / or The dry density prediction model for the rockfill mass includes an input layer, a hidden layer, and an output layer. The number of hidden layers is determined based on the number of knowledge units used to train the dry density prediction model; and / or Training the dry density prediction model for the rockfill mass also includes the following steps: Based on the actual dry density value and the predicted dry density value output by the rockfill dry density prediction model, determine whether the rockfill dry density prediction model is qualified.

3. The training method according to claim 2, characterized in that, Training the dry density prediction model for the rockfill mass includes the following steps: The data of the knowledge units are preprocessed to obtain standardized training data; The dry density prediction model for the rockfill is trained using the standardized training data.

4. The training method according to claim 2, characterized in that, "Determining whether the dry density prediction model of the rockfill body is qualified based on the predicted dry density value and the actual dry density value output by the dry density prediction model" includes the following steps: Calculate the absolute error between the predicted dry density value and the actual value. The formula for calculating the absolute error is as follows: Absolute error = |predicted value - actual value|; The dry density prediction model is deemed qualified based on whether the absolute error is greater than a first preset value; or Calculate the relative error between the predicted dry density value and the actual value. The formula for calculating the relative error is as follows: Whether the dry density prediction model is qualified is determined based on whether the relative error is greater than a second preset value.

5. A dry density prediction system for riprap, characterized in that, It includes multiple dry density prediction models for rockfill bodies that correspond one-to-one with multiple scene categories, and the dry density prediction models for rockfill bodies are trained by the training method described in any one of claims 1-4.

6. A method for predicting the dry density of a rockfill mass, characterized in that, Applied to the dry density prediction system for rockfill bodies according to claim 5, the prediction method includes the following steps: Select the corresponding dry density prediction model for the rockfill based on the scene category of the rockfill; The initial frequency, maximum frequency difference, vibrating mass, and dynamic stiffness of a single measuring point of the rockfill are input into the dry density prediction model of the rockfill, and the dry density prediction model of the rockfill outputs the predicted dry density value of the measuring point.

7. A device for predicting the dry density of a rockfill mass, characterized in that, It is applied to the dry density prediction system for rockfill as described in claim 5.

8. A computer device, characterized in that, The device includes one or more processors and a memory; wherein the memory stores computer-readable instructions; the one or more processors read the computer-readable instructions to cause the computer device to perform the method of claim 6.

9. A computer-readable storage medium, characterized in that, It includes computer-readable instructions that, when executed on a computer, cause the computer to perform the method of claim 6.

10. A computer program product, characterized in that, It includes computer-readable instructions that, when executed on a computer, cause the computer to perform the method of claim 6.

Images