A method for predicting marine soil compression parameters

Abstract

The application provides a marine soil compression parameter prediction method, comprising the following steps: S1, collecting target site soil body parameter data; S2, pre-processing the target site soil body parameter data collected in step S1 to fill in missing data; S3, dividing the pre-processed data set into a training set and a test set, and performing k-fold cross-validation on the training set; S4, inputting the processed data into a LightGBM model, setting the hyperparameters that need to be debugged for the model, training the model and predicting the target parameter; S5, finding the hyperparameters that make the evaluation index optimal through Bayesian optimization, and evaluating whether the prediction result meets the requirements through the model accuracy evaluation index; S6, obtaining a final soil body parameter prediction model. The application can predict difficult-to-obtain deformation parameters by using some easily-obtained soil body physical property parameters, thereby reducing the marine soil body survey cost and improving the soil body parameter survey efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of soil parameter testing, and particularly relates to a prediction method of marine soil compression parameters. BACKGROUND

[0002] With the continuous development of construction technology, various marine engineering constructions are continuously promoted. The exploration of the physical and mechanical properties of marine soil has become an important problem in marine engineering practice.

[0003] At present, the mechanical properties and design parameters of soil are almost completely dependent on in-situ geotechnical tests (such as sampling, drilling and penetration testing) and subsequent indoor geotechnical tests (strength and deformation indexes of each test soil sample directly obtained from consolidation tests and triaxial tests). However, marine geotechnical engineering tests are costly, and for specific projects, it is difficult to carry out a large number of site tests in densely located positions, and it is difficult to obtain soil parameters at each position, which limits the development of marine engineering construction.

[0004] Therefore, it is of great significance to develop a high-efficiency and reliable soil parameter prediction system for marine geotechnical engineering. If the soil compression parameters that are difficult to obtain can be predicted from the basic physical properties of the soil that are easy to obtain, it will help to reduce the cost of marine soil exploration and improve the efficiency of soil parameter determination. SUMMARY

[0005] The present application aims at the above problems, and provides a marine soil compression parameter prediction method based on Bayesian optimization and LightGBM algorithm.

[0006] To this end, the above-mentioned purpose of the present application is realized by the following technical scheme:

[0007] A prediction method of marine soil compression parameters, comprising the following steps:

[0008] S1, collecting target site soil parameter data;

[0009] S2, preprocessing the target site soil parameter data collected in step S1 to fill in the missing values of the data;

[0010] S3, dividing the preprocessed data set into a training set and a test set, and performing k-fold cross-validation on the training set;

[0011] S4, inputting the processed data into a LightGBM model, setting the hyperparameters to be debugged for the model, training the model and predicting the target parameters;

[0012] S5, finding the hyperparameters that make the evaluation index optimal through Bayesian optimization, and evaluating whether the prediction result meets the requirements through the model accuracy evaluation index;

[0013] If the requirement is not met, steps S2, S3 and S4 are performed again, specifically, the data preprocessing mode is adjusted, the data set division area is changed, the cross-validation number is changed, the hyperparameters to be debugged and the hyperparameter range are re-set, the model is trained and the target parameter is predicted until the accuracy evaluation index requirement is met;

[0014] S6, a final soil parameter prediction model is obtained.

[0015] While the above technical solutions are adopted, the application can also adopt or combine the following technical solutions:

[0016] As a preferred technical solution of the application: in step S1, the target site soil parameter data includes sample bottom depth, water content, wet density, soil particle relative density, liquid plastic limit, compression coefficient and compression modulus.

[0017] As a preferred technical solution of the application: in step S2, the data missing value filling method includes deleting missing rows and random forest filling method.

[0018] As a preferred technical solution of the application: in step S3, the preprocessed data set is randomly divided into a training set and a test set according to a ratio of 4:1, wherein the training set is used to build a model, and the test set is used to verify the actual prediction ability of the model.

[0019] The k-fold cross-validation is to further divide the training set in the original data set into k sets of the same size, select one set as the validation set, and use the remaining k-1 sets as the training set and train them, repeat the training step k times, and finally select the hyperparameters that make the average error of k times modeling reach the minimum as the final hyperparameters for training on the original whole training set.

[0020] As a preferred technical solution of the application: in step S4, the hyperparameters include max_depth, num_leaves, learning_rate, bagging_fraction and feature_fraction.

[0021] The max_depth represents the maximum depth of the tree model; the num_leaves represents the number of nodes on a tree; the learning_rate represents the learning rate; the bagging_fraction represents the proportion of the number of models in the whole model; and the feature_fraction is used to control the split point in other parameters.

[0022] As a preferred technical solution of the application: in step S4, the objective function of model training is:

[0023]

[0024] In the above formula: Obj K is the objective function of model training; L(x, y) is a square loss function, i.e. (x-y) 2 ; y i is the true value of the predicted quantity; is the result of the Kth learning; Ω(f K ) is the regularization term of the Kth tree; c K-1 is the regularization term sum of the first K-1 trees; f K is the function of the Kth tree. The meaning of the objective function is to find a tree that minimizes the value of the objective function.

[0025] As a preferred technical solution of the present application:

[0026] Let g i denote the first-order derivative of the loss function of the ith sample, h i denote the second-order derivative of the loss function of the ith sample, and the result of the second-order Taylor expansion of the objective function of model training is:

[0027]

[0028] In the above formula: Obj K is the objective function of model training; L(x, y) is a square loss function, i.e. (x-y) 2 ; y i is the true value of the predicted quantity; is the result of the Kth learning; Ω(f K ) is the regularization term of the Kth tree; c K-1 is the regularization term sum of the first K-1 trees; f K is the function of the Kth tree; g i is the first-order derivative of the loss function of the ith sample; h i is the second-order derivative of the loss function of the ith sample.

[0029] As a preferred technical solution of the present application: in step S5, the model accuracy evaluation indicators include: mean absolute error MAE (Mean absolute error), root mean square error RMSE (Root mean square error), mean absolute percentage error MAPE (Mean absolute percentage error), and determination coefficient R 2 (R-square).

[0030] As a preferred technical solution of the present invention: In step S5, Bayesian optimization provides a region for each hyperparameter value, that is, searches for the values ​​of each hyperparameter that make the objective function reach the optimal value.

[0031] As a preferred technical solution of the present invention: the model takes the sample depth, water content, wet density, relative density of soil particles, and liquid limit as inputs, and the compression coefficient and compression modulus as outputs.

[0032] This invention provides a method for predicting the compression parameters of marine soil, which has the following beneficial effects:

[0033] 1) Compared with traditional test methods for obtaining compression parameters, this method for predicting marine soil compression parameters can predict deformation parameters that are difficult to obtain by using some easily obtainable soil physical property parameters, which helps to reduce the cost of marine soil exploration and improve the efficiency of soil parameter exploration.

[0034] 2) Compared with traditional prediction models, this model improves prediction performance and accuracy by using Bayesian k-fold cross-validation and Bayesian optimization. Attached Figure Description

[0035] Figure 1 The flowchart illustrates the steps of the method for predicting marine soil compression parameters provided by this invention.

[0036] Figure 2 This is a frequency histogram of soil parameter data for the target site in an embodiment of the present invention.

[0037] Figure 3 This is a comparison chart of model predictions and actual values ​​in an embodiment of the present invention. Detailed Implementation

[0038] The present invention will be described in further detail with reference to the accompanying drawings and specific embodiments.

[0039] Reference Figure 1 As shown, a method for predicting marine soil compression parameters based on Bayesian optimization and the LightGBM algorithm includes the following steps:

[0040] S1. Collect soil parameter data for the target site, including sample depth, moisture content, wet density, relative density of soil particles, liquid limit, plastic limit, compression coefficient, and compression modulus.

[0041] S2. Data preprocessing, mainly involving handling missing values; methods for imputing missing values ​​include deleting missing rows and random forest imputation.

[0042] S3. Randomly divide the preprocessed dataset into training and test sets in a 4:1 ratio. The training set is used to build the model, and the test set is used to validate the model's actual predictive ability. The data is divided into training and test sets; k-fold cross-validation further divides the original training set into k sets of equal size. One set is selected as the validation set, and the remaining k-1 sets are used as the training set for training. This training step is repeated k times. Finally, the hyperparameters that minimize the average error in the k modeling iterations are selected as the final hyperparameters. Training is then performed on the original full training set, followed by k-fold cross-validation on the training set.

[0043] S4. Input the processed data into the LightGBM model, set the hyperparameters that need to be tuned, train the model and predict the target parameters. The selected hyperparameters include: max_depth, num_leaves, learning_rate, bagging_fraction and feature_fraction; max_depth represents the maximum depth of the tree model; num_leaves represents the number of nodes in a tree; learning_rate represents the learning rate; bagging_fraction represents the proportion of data used in the entire model; feature_fraction is used to control the split point among other parameters.

[0044] S4. Use Bayesian optimization to find the hyperparameters that optimize the evaluation metrics. These hyperparameters include max_depth, num_leaves, learning_rate, bagging_fraction, and feature_fraction. The optimal hyperparameters are calculated using mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination (R²). 2 Four model accuracy evaluation metrics are used to assess whether the prediction results meet the requirements;

[0045] S5. The final soil parameter prediction model is obtained, with the sample depth, water content, wet density, relative density of soil particles, and liquid and plastic limits as inputs, and the compression coefficient and compression modulus as outputs. This model enables the prediction of difficult-to-obtain deformation parameters using some readily available soil physical property parameters.

[0046] The following is an illustration using a specific example:

[0047] This embodiment uses 482 sets of geotechnical test data obtained from an offshore wind farm project as the research object, and selects the following characteristics: sample depth D, water content W0, wet density ρ, and relative density of soil particles G. s Liquid limit W L , plastic limit W P Compression factor av Compression modulus E s Parameter distribution as follows Figure 2 As shown.

[0048] A marine soil compression parameter prediction model based on the LightGBM algorithm was established using Python. The `info` function was used to check the dataset, which contains 4520 data points. Among them, ρ, G... s a v E s Two sets of data were missing. After observing the dataset, it was found that the missing data originated from the same two soil samples; therefore, these two soil samples were directly deleted. As for the liquid limit W... L , plastic limit W P The missing value rate was 22.3%, and direct deletion would be a significant waste of data. The LightGBM algorithm can make predictions even with missing values. By comparing the prediction results of no processing and imputing the data with average values ​​of 39.96 and 24.08, the accuracy of the two methods was found to be similar. Therefore, it was decided not to use the liquid limit W. L , plastic limit W P The missing soil parameters and preprocessing methods are shown in Table 1. Table 1 shows the missing soil parameters and preprocessing methods in the embodiments of the present invention.

[0049] Table 1

[0050]

[0051] The preprocessed dataset was randomly divided into a training set and a test set in a 4:1 ratio. The training set was used to build the model, and the test set was used to validate the model's actual predictive ability. The training set consisted of 384 samples, and the test set consisted of 96 samples.

[0052] To obtain better calculation results, 10-fold cross-validation was used to validate the established model.

[0053] Five parameters were set with tuning ranges, and RMSE was selected as the evaluation metric. Bayesian optimization was used to iteratively obtain lower RMSE accuracy values. The hyperparameters obtained from Bayesian tuning were then used to build the model with the highest accuracy. The final tuning results are shown in Tables 2 and 3. Table 2 shows the compression coefficient 'a' in the embodiment of this invention. v The LightGBM Bayesian hyperparameter tuning results are shown in Table 3. Table 3 presents the results for the compression modulus E in the embodiments of this invention. S Results of LightGBM Bayesian hyperparameter tuning.

[0054] Table 2

[0055]

[0056] Table 3

[0057]

[0058] The predictive ability of the established model was further evaluated by comparing it with a model based on the XGBoost algorithm, a model without k-fold cross-validation, and a model with Bayesian parameter tuning. The accuracy evaluation metrics for different models are shown in Table 4. Table 4 compares the accuracy evaluation results of different models in the embodiments of this invention. It can be seen that the established model shows some improvement compared to the model without k-fold cross-validation and Bayesian parameter tuning. Compared to the model with added manual features, the predictive accuracy of the established model is not significantly different.

[0059] Table 4

[0060]

[0061] Comparison of predicted and actual values ​​obtained in embodiments of the present invention: Figure 3 As shown. From Figure 3 As can be seen, when the soil parameter values ​​are small, the difference between the predicted and actual values ​​is small. This indicates that when there is a large amount of small data, the model can learn sufficiently and improve the simulation accuracy. This model has high accuracy in predicting the compression parameters of the two types of soil, and the predicted data can replace the actual values.

[0062] The above specific embodiments are used to explain and illustrate the present invention, and are only preferred embodiments of the present invention, not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made to the present invention within the spirit and scope of the claims shall fall within the protection scope of the present invention.

Claims

1. A method for predicting marine soil compression parameters, characterized in that: The method includes the following steps: S1. Collect soil parameter data for the target site; S2. Preprocess the soil parameter data of the target site collected in step S1 to fill in the missing data values. S3. Divide the preprocessed dataset into a training set and a test set, and perform k-fold cross-validation on the training set. S4. Input the processed data into the LightGBM model, set the hyperparameters that need to be adjusted for the model, train the model and predict the target parameters; S5. Find the hyperparameters that make the evaluation index optimal through Bayesian optimization, and evaluate whether the prediction results meet the requirements through the model accuracy evaluation index. If the requirements are not met, repeat steps S2, S3, and S4. Specifically, adjust the data preprocessing method, change the dataset partitioning region, change the cross-validation fold number, reset the hyperparameters that need to be debugged and the range of hyperparameters, train the model and predict the target parameters until the accuracy evaluation index requirements are met. S6. Obtain the final soil parameter prediction model; In step S4, the objective function for model training is: In the above formula: Obj K Let L(x,y) be the objective function for model training; L(x,y) is the squared loss function, i.e., (x,y) 2 ;y i The true value of the predicted quantity; This represents the result of the Kth learning iteration; Ω(f) K ) represents the regularization term for the Kth tree; c K-1 f is the sum of the regularization terms for the first K-1 trees; K The function for the Kth tree; Use g i Let h be the first derivative of the loss function for the i-th sample. i Let be the second derivative of the loss function for the i-th sample. The result of the second Taylor expansion of the objective function for model training is: In the above formula: Obj K Let L(x,y) be the objective function for model training; L(x,y) is the squared loss function, i.e., (x,y) 2 ;y i The true value of the predicted quantity; This is the result of the (K-1)th learning iteration; Ω(f K ) represents the regularization term for the Kth tree; c K-1 f is the sum of the regularization terms for the first K-1 trees; K The function for the Kth tree; g i h is the first derivative of the loss function for the i-th sample; i Let be the second derivative of the loss function for the i-th sample.

2. The method for predicting marine soil compression parameters according to claim 1, characterized in that: In step S1, the soil parameters of the target site include: sample depth, moisture content, wet density, relative density of soil particles, liquid limit, plastic limit, compression coefficient, and compression modulus.

3. The method for predicting marine soil compression parameters according to claim 1, characterized in that: In step S2, the methods for imputing missing data values ​​include: deleting missing rows and random forest imputation.

4. The method for predicting marine soil compression parameters according to claim 1, characterized in that: In step S3, the preprocessed dataset is randomly divided into a training set and a test set in a 4:1 ratio. The training set is used to build the model, and the test set is used to verify the actual predictive ability of the model. k-fold cross-validation involves further dividing the training set in the original dataset into k sets of equal size, selecting one set as the validation set, and using the remaining k-1 sets as the training set for training. This training step is repeated k times, and finally, the hyperparameters that minimize the average error in the k modeling iterations are selected as the final hyperparameters. Training is then performed on the original entire training set.

5. The method for predicting marine soil compression parameters according to claim 1, characterized in that: In step S4, the hyperparameters include max_depth, num_leaves, learning_rate, bagging_fraction, and feature_fraction; The `max_depth` represents the maximum depth of the tree model; the `num_leaves` represents the number of nodes in a tree; the `learning_rate` represents the learning rate; the `bagging_fraction` represents the proportion of the data used in the entire model; and the `feature_fraction` is used to control the split point among other parameters.

6. The method for predicting marine soil compression parameters according to claim 1, characterized in that: In step S5, the model accuracy evaluation metrics include: Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), and Coefficient of Determination (R²). 2 .

7. The method for predicting marine soil compression parameters according to claim 1, characterized in that: In step S5, Bayesian optimization provides a region for each hyperparameter value, that is, it searches for the values ​​of each hyperparameter that make the objective function optimal.

8. The method for predicting marine soil compression parameters according to claim 1, characterized in that: The model takes the sample depth, moisture content, wet density, relative density of soil particles, and liquid and plastic limits as inputs, and the compressibility coefficient and compressibility modulus as outputs.

Images