Dynamic compaction construction parameter optimization method based on XGBoost-PSO

Abstract

The present invention relates to the field of foundation treatment technology, and in particular to a method for optimizing dynamic compaction construction parameters based on XGBoost‑PSO, comprising: obtaining multiple sets of tamping data through tamping tests; selecting original features from actual measurement parameters of the test, then using actual measurement parameters to construct interactive features, and selecting the optimal interactive features based on XGboost feature importance ranking, and then analyzing the correlation between features; establishing a tamping prediction model based on the screened feature set; optimizing dynamic compaction parameters of the tamping prediction model using a particle swarm algorithm, and ultimately obtaining the optimal tamping parameter combination. The present invention automatically optimizes tamping parameters under the premise of meeting the stop-tamping standard by constructing a tamping prediction model, thereby achieving the purpose of ensuring construction quality, improving construction efficiency, and reducing construction costs; the present invention is applicable to various foundation construction design stages that require dynamic compaction treatment, and has good engineering adaptability and promotion value.

Description

Technical Field

[0001] The present invention relates to the technical field of foundation treatment, and in particular to a method for optimizing dynamic compaction construction parameters based on XGBoost-PSO. Background Art

[0002] During dynamic compaction construction, tamping volume, as an important indicator of foundation response, is generally considered to directly reflect the effectiveness of soil reinforcement. In practical engineering design, the tamping volume of the last two strokes is often used as the criterion for stopping tamping. Other studies have dynamically assessed the degree of soil reinforcement and saturation by monitoring the changing trend of tamping volume. Therefore, tamping volume is not only a crucial parameter for controlling the construction process but also a crucial basis for parameter optimization and compaction assessment. Current engineering practice often relies on field trials to determine the combination of dynamic compaction construction parameters (such as the number of strokes, hammer weight, lifting height, and hammer diameter). This involves inferring the overall construction parameters from the results of small-scale trials. While this method is practical for engineering, it suffers from its reliance on manual judgment and empirical judgment, its inability to cover the design space for multiple parameters and multiple working conditions, and its high cost and low efficiency, making it difficult to implement on a large scale during the early design phase.

[0003] With the rise of artificial intelligence and machine learning, more and more studies are applying data-driven methods to the fields of engineering prediction and parameter optimization. Among them, some studies have attempted to predict dynamic compaction settlement or reinforcement depth. For example, Zhang Yuchuan et al. constructed a dynamic compaction settlement prediction model based on BP neural network using variables such as unit area compaction energy, number of compaction times, and moisture content; Xu Yongbing et al. compared the performance of BP, SVM, RF, XGBoost and other models in the prediction of effective reinforcement depth, and believed that RF and XGBoost are more suitable for dynamic compaction data modeling under small sample conditions. However, due to factors such as the difficulty in obtaining dynamic compaction engineering data and the high cost of experiments, there is currently a lack of high-quality and diverse dynamic compaction parameter-response data sets, as well as a lack of systematic solutions that combine "prediction model + parameter optimization".

[0004] When it comes to predicting tamping settlement, model performance is highly dependent on the quality of feature engineering. While previous studies have selected input variables based on engineering experience, systematic methods for mining parameter interactions are lacking. Furthermore, the numerous interfering factors of field conditions and high data inconsistency limit further improvements in model performance. To this end, Bian Haiding et al. conducted multiple sets of repeatable indoor model tests using scaled-down models. By controlling variables such as tamping hammer mass, drop distance, and moisture content, they provided a controllable means for exploring the factors influencing tamping settlement.

[0005] In terms of parameter optimization, the particle swarm optimization (PSO) algorithm is widely used in various engineering optimization tasks due to its inherent support for real-number encoding, low parameter count, and fast convergence. Previous studies have attempted to combine it with predictive models. For example, Hu Changming et al. used LSTM-PSO to optimize surface settlement and propulsion parameters during shield tunneling.

[0006] Overall, the following deficiencies still exist in the field of dynamic compaction construction design: the determination of construction parameters is still mainly based on trial compaction and experience, and there is a lack of a generalizable and predictable parameter-response model; existing models mostly focus on prediction and are not coupled with optimization algorithms, making it difficult to achieve automatic optimal design; data sources mainly rely on on-site collection, are greatly affected by external factors, and it is difficult to build a systematic database. Summary of the Invention

[0007] The purpose of this invention is to provide a dynamic compaction parameter optimization method that integrates scaled model testing, a machine learning regression model (XGBoost), and a particle swarm optimization algorithm (PSO). This method uses experimentally acquired system data to construct a high-precision compaction settlement prediction model. This method, combined with an optimization algorithm, searches for the optimal parameter combination while meeting the tamping stop criteria. This method enables automated design of construction parameters, improving construction efficiency and the scientific nature of foundation quality control.

[0008] To achieve the above objectives, the present invention provides a method for optimizing dynamic compaction construction parameters based on XGBoost-PSO, comprising the following steps:

[0009] S1. Dynamic compaction model test: Design and manufacture a dynamic compaction model, and conduct compaction tests using test rammers of various specifications and the dynamic compaction model to obtain compaction settlement data under several different parameter combinations;

[0010] S2. Feature selection: First, select the original features from the actual measured parameters of the dynamic compaction model test; then use the actual measured parameters to construct interactive features, and select the optimal interactive features based on the XGboost feature importance ranking; then perform correlation analysis on the original features and the optimal interactive features;

[0011] S3. Based on the feature set obtained by screening in step S2, an XGBoost tamping settlement prediction model is established;

[0012] S4. Use the particle swarm algorithm to optimize the compaction parameters of the XGBoost compaction settlement prediction model, and finally obtain the optimal compaction parameter combination.

[0013] Furthermore, step S1 specifically includes:

[0014] S1.1. Determine Similarity Criteria and Scale Ratio: Based on the three similarity criteria of geometric similarity, physical similarity, and mechanical similarity, and in combination with the actual engineering context, select model control variables and determine the similarity coefficient system required for the impact test using the second similarity theorem and dimensional analysis. The model control variables include rammer mass M, rammer lifting height H, rammer diameter D, number of impacts N, impact energy E, ramming sinkage s, moisture content w, and soil dry density d.

[0015] S1.2. Prepare a dynamic compaction model box, configure a triangular bracket for the dynamic compaction model box, and install an electromagnetic adsorption device on the top of the triangular bracket to simulate the free fall of a heavy object during the dynamic compaction process;

[0016] S1.3. Test rammers of different specifications are provided; some of the hammers have holes on their surfaces to reduce air resistance and to meet actual engineering conditions.

[0017] S1.4. Arrange the measuring device and perform the tamping operation: Design different tamping parameters to conduct multiple groups of tamping tests. Set the number of tamping times for each group of tests to 8 to 18 times. After each tamping, read the tamping amount data and record it in the database.

[0018] Furthermore, in step S1.4, after each set of tamping is completed, the soil is disturbed again and loosely spread and leveled to maintain the consistency of the initial state of the soil.

[0019] Furthermore, step S2 specifically includes:

[0020] S2.1. Original feature selection: Based on the parameters actually measured in the model test data, rammer mass, lifting height, rammer diameter, and number of ramming strokes were selected as original feature variables. Grey correlation analysis was used to calculate the correlation between each original feature variable and the target variable. Original feature variables with a correlation greater than the correlation threshold were considered key influencing features and included in subsequent modeling. The correlation threshold was 0.6.

[0021] S2.2, Optimal interactive feature selection: The specific steps are:

[0022] (1) Determine the combination form of interactive features: Select the ram mass M and the lifting height H as the basic variables for constructing interactive features, and set three types of combination forms: MH a 、M b H and M c H d ; where a, b, c and d are power parameters used to adjust the nonlinear shape;

[0023] (2) Set the parameter search range and step size: To ensure search accuracy and computational efficiency, set a, b, c, d ∈ [0.1, 2, 0.1];

[0024] (3) Construct interaction features and perform K-fold cross validation: Use K-fold cross validation to train the XGBoost model for the three types of combined interaction features, and calculate the feature importance index of each interaction feature in each round of training, and use weighted average to obtain the stable importance of each interaction feature; where K=5;

[0025] (4) Calculate the total importance of features and select the optimal power combination: MH a 、M b H and M c H d The average importance values of the three interactive features are added together to define the total importance of the current combination. After traversing all power combinations, the power combination with the largest total importance is selected as the final interactive feature construction strategy.

[0026] (5) Construct the optimal interaction features and incorporate them into model training: Based on the evaluation results, the interaction features corresponding to the optimal power parameters are formally incorporated into the model input feature set, and the original features are combined to participate in the XGBoost model training;

[0027] S2.3. Combine the original features selected in step S2.1 and the optimal interaction features selected in step S2.2 as variables for Spearman calculation analysis. If the correlation coefficient ρ between variables is greater than 0.8, the collinearity is considered too high and features are removed. Then, the variance inflation factor method is used to calculate vif to ensure that the remaining feature vif value is less than 10.

[0028] Furthermore, step S3 specifically includes:

[0029] Step S3.1, Data Partitioning and Standardization: The original dataset is divided into training and test sets according to a set ratio; the model input features are normalized and preprocessed to ensure data consistency across different dimensions;

[0030] S3.2, XGBoost model initialization: Use the XGBoost regression model to construct a nonlinear mapping relationship f: X→y, where X is the feature vector and y is the tamping amount;

[0031] S3.3. Hyperparameter Grid Search and Tuning: Grid search is used to optimize the key hyperparameters of the XGBoost model, with the goal of minimizing the cross-validation error. Five-fold cross-validation is used to calculate the negative mean squared error for each set of hyperparameters, and the model combination with the best performance is selected.

[0032] S3.4. Model Prediction and Performance Evaluation: The optimized model is used to make predictions on the training and test sets. The following metrics are used to comprehensively evaluate the model performance: RMSE (root mean square error), MAE (mean absolute error), and R² (coefficient of determination).

[0033] S3.5. Output the optimal model and parameters: Finally, retain and output the best model after parameter adjustment and its corresponding optimal hyperparameter combination, providing a core support tool for subsequent parameter optimization based on the prediction model.

[0034] Furthermore, step S4 specifically includes:

[0035] S4.1. Start the optimization system and prepare for dynamic compaction parameter search and prediction evaluation: Load the XGBoost compaction settlement prediction model, which has been trained based on historical compaction test data, to predict the compaction settlement response under the input parameter combination;

[0036] S4.2. Set the initial parameters of PSO, including particle swarm size, maximum number of iterations, inertia factor, individual / global learning factor, and search space boundary;

[0037] S4.3. Initialize the position and velocity of the particles, and randomly generate the position (i.e., parameter combination) and initial velocity of the individual particles in the parameter space;

[0038] S4.4. Predicting the tamping response using the XGBoost model. Input the parameter combination of each particle into the XGBoost model to obtain the predicted tamping response.

[0039] S4.5. Construct an objective function that integrates construction cost and tamping response, in the form of P = α × N + β × M × H + penalty. α and β are weights, representing the importance of the number of tamping blows and construction cost, respectively, and should be set based on engineering experience and actual conditions. Penalty is the penalty term for violating the tamping standard; if the tamping amount exceeds the corresponding energy level threshold, a positive penalty is imposed. N, M, and H represent the number of tamping blows, the mass of the tamping hammer, and the lifting height, respectively.

[0040] S4.6. Update the individual optimal value P_best, record the current historical optimal fitness value of each particle and its corresponding position; update the global optimal value G_best, select the particle with the smallest fitness value among all current particles as the global optimal solution, record its position, and determine whether it meets the stop tamping criteria. If the current predicted tamping amount meets the stop tamping criteria under the corresponding tamping energy, continue; otherwise, update the particle position and velocity and continue iteration;

[0041] S4.7. Output the optimal compaction parameter combination: The final output is the compaction parameter combination that meets the stop compaction standard and has the minimum fitness, and the optimization process is completed.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] The present invention first constructs a systematic test data set through a scaled model test, and collects the tamping response under different parameter combinations on the basis of controlling multiple variables such as rammer mass, lifting height, number of tamping times, and hammer diameter. The model test data provides a reliable physical basis and data support for subsequent machine learning modeling and intelligent optimization. Subsequently, a dynamic compaction parameter optimization method based on the combination of XGBoost regression prediction and PSO particle swarm optimization algorithm is proposed. By constructing a tamping prediction model, the tamping parameters are automatically optimized under the premise of meeting the stop-tamping standard, thereby achieving the purpose of ensuring construction quality, improving construction efficiency and reducing construction costs. The method of the present invention is applicable to the design stage of various foundation constructions that require dynamic compaction treatment, and has good engineering adaptability and promotion value.

[0044] In addition to the above-described objects, features and advantages, the present invention has other objects, features and advantages. The present invention will be further described in detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] The accompanying drawings are used to provide a further understanding of the embodiments of the present invention and constitute a part of the specification. Together with the following detailed description, they are used to explain the embodiments of the present invention, but do not constitute a limitation of the embodiments of the present invention. In the accompanying drawings:

[0046] Figure 1 This is the overall process of the dynamic compaction construction parameter optimization method based on XGBoost-PSO in the present invention;

[0047] Figure 2 It is a structural schematic diagram of the model test device of the present invention;

[0048] Figure 3 It is a schematic diagram of the rammer model in the present invention;

[0049] Figure 4 This is a flow chart of the dynamic compaction parameter optimization using the PSO algorithm in the present invention. DETAILED DESCRIPTION

[0050] The present invention will be described in detail below with reference to the various embodiments shown in the accompanying drawings, but it should be noted that these embodiments are not limitations of the present invention, and any equivalent transformations or substitutions in functions, methods, or structures made by ordinary technicians in this field based on these embodiments are all within the scope of protection of the present invention.

[0051] See Figure 1 This embodiment provides a method for optimizing dynamic compaction construction parameters based on XGBoost-PSO, comprising the following steps:

[0052] S1. Dynamic compaction model test: Design and manufacture a dynamic compaction model, and conduct compaction tests using test rammers of various specifications and the dynamic compaction model to obtain compaction sinking data under several different parameter combinations. Specifically, this includes:

[0053] S1.1. Determine similarity criteria and scale ratios. Based on the three similarity criteria of geometric similarity, physical similarity, and mechanical similarity, and in combination with the actual engineering context, parameters such as rammer mass M, rammer lifting height H, rammer diameter D, number of ramming blows N, ramming energy E, ramming sinking amount s, moisture content w, and soil dry density d are selected as model control variables. The second similarity theorem and dimensional analysis method are used to determine the similarity coefficient system required for the ramming test to guide the design of the scaled model.

[0054] Step S1.2, design the dynamic compaction model. Make a dynamic compaction model box 1 of rectangular structure with dimensions of 1000mm×1000mm×800mm. The box is made of wood to meet the boundary requirement that "the action boundary of the dynamic compaction model box is at least 4 times the diameter of the rammer" to prevent the influence of boundary effects on the test; equip the dynamic compaction model box with a triangular bracket 2, and install an electromagnetic adsorption device 3 on the top of the triangular bracket. The dynamic compaction model can achieve a free-falling hammer height of up to 1.5m to simulate the free-falling behavior of heavy objects during the dynamic compaction process. The dynamic compaction model, the triangular bracket and the electromagnetic adsorption device together constitute a test device. The test device is as follows: Figure 2 shown.

[0055] S1.3, configure test rammers of different specifications. Specifically, five types of round iron rammers with masses of 5kg, 8kg, 11kg, 17kg, and 22kg are available. Each type of rammer has three specifications with diameters of 120mm, 150mm, and 180mm. Some of the hammer surfaces have holes to reduce air resistance and to meet actual engineering conditions. The rammer is shown in the figure below. Figure 3 shown.

[0056] S1.4. Set up the measuring device and perform the tamping operation. Design different tamping parameters (such as tamping hammer weight M, lifting height H, tamping hammer diameter D and number of tamping times N, etc.) to conduct multiple groups of tamping tests. Use a millimeter ruler to record the vertical distance from the geometric center of the tamping hammer to the soil surface as the measurement value of the tamping amount; set the number of tamping times for each group of tests to 8 to 18 times; preferably, set the number of tamping times for each group of tests to 12 times. During the test, an electromagnetic adsorption device is used to control the free fall of the tamping hammer. After each tamping, the tamping amount data is read and recorded in the database. After each tamping, the settlement of the soil needs to be measured with a ruler. To ensure the independence of the test, the soil is disturbed again and loosely spread and leveled after each group of tamping is completed to maintain the consistency of the initial state of the soil in multiple groups of tamping tests.

[0057] S2. Feature selection: First, select the original features from the actual measured parameters of the dynamic compaction model test; then use the actual measured parameters to construct interactive features, and select the optimal interactive features based on the XGboost feature importance ranking; then perform correlation analysis on the original features and the optimal interactive features, and finally obtain the input features of the subsequent XGboost prediction model. The specific steps are:

[0058] S2.1. Original feature selection: Based on the parameters actually measured in the model test data, the rammer mass M, lifting height H, rammer diameter D, and number of ramming strokes N were selected as original feature variables. The Grey Relational Analysis (GRA) method was used to calculate the correlation between each original feature variable and the target variable (ramming amount). Original feature variables with a correlation greater than a threshold of 0.6 were considered key influencing features and included in subsequent modeling.

[0059] S2.2, Optimal interactive feature selection: The specific steps are:

[0060] (1) Determine the combination form of interactive features: Select the ram mass M and the lifting height H as the basic variables for constructing interactive features, and set three types of combination forms: MH a 、M b H and M c H d Where a, b, c and d are power parameters used to adjust the nonlinear shape;

[0061] (2) Set the parameter search range and step size: To ensure search accuracy and computational efficiency, set a, b, c, d ∈ [0.1, 2, 0.1].

[0062] (3) Construct interaction features and perform K-fold cross-validation: Use K-fold cross-validation (K=5) to train the XGBoost model for the three types of combined interaction features, and calculate the "feature importance" index (feature importance) of each interaction feature in each round of training, and use weighted average to obtain the stable importance of each interaction feature.

[0063] (4) Calculate the total importance of features and select the optimal power combination: add the average importance values of the three interactive features and define it as the "total importance" of the current combination; after traversing all power combinations, select the power combination with the largest total importance as the final interactive feature construction strategy.

[0064] (5) Construct the optimal interaction features and incorporate them into model training: Based on the evaluation results, the interaction features corresponding to the optimal power parameters are formally incorporated into the model input feature set, and the original features are combined to participate in the XGBoost model training.

[0065] S2.3. Correlation analysis between features:

[0066] To avoid the effects of multicollinearity between features, the Spearman rank correlation coefficient was used to analyze the correlation between variables. The original features selected in step S2.1 and the optimal interaction features selected in step S2.2 were combined as variables for Spearman's calculation analysis. If the correlation coefficient ρ between variables was greater than 0.8, the collinearity was considered excessive and features were removed. The variance inflation factor (vif) method was then used to calculate the remaining features to ensure that the vif value was less than 10, thereby ensuring the stability and generalization ability of model training. The variance inflation factor (vif) is an important indicator for detecting the presence of collinearity.

[0067] S3. Based on the feature set obtained by screening in step S2, an XGBoost tamping settlement prediction model is established. The specific steps are as follows:

[0068] S3.1. Data division and standardization: The original data set (i.e., the tamping settlement data set under different parameter combinations collected during the model test in step s1) is divided into a training set and a test set in a ratio of 8:2; StandardScaler is used to normalize and preprocess the input features to ensure data consistency between different dimensions.

[0069] S3.2. XGBoost model initialization: The XGBoost regression model is used to construct a nonlinear mapping relationship f: X→y, where X is the feature vector and y is the tamping amount.

[0070] S3.3. Hyperparameter grid search and tuning: Grid search (GridSearchCV) is used to perform combinatorial optimization of the key hyperparameters of the XGBoost model (such as n_estimators, max_depth, learning_rate, subsample, colsample_bytree, etc.), with the optimization goal of minimizing the cross-validation error. Five-fold cross-validation is used to calculate the negative mean squared error under each set of hyperparameters to select the model combination with the best performance.

[0071] S3.4. Model prediction and performance evaluation: The optimized model is used to make predictions on the training set and test set, and the following indicators are used to comprehensively evaluate the model performance: RMSE (root mean square error), MAE (mean absolute error), and R² (coefficient of determination).

[0072] S3.5. Output the optimal model and parameters: Finally, retain and output the best model after parameter adjustment and its corresponding optimal hyperparameter combination, providing a core support tool for subsequent parameter optimization based on the prediction model.

[0073] S4. Use particle swarm optimization to optimize the compaction parameters of the XGBoost compaction settlement prediction model, and finally obtain the optimal compaction parameter combination; the flow chart of the dynamic compaction parameter optimization using the PSO algorithm is as follows: Figure 4 The specific steps are as follows:

[0074] S4.1. Start the optimization system and prepare for dynamic compaction parameter search and prediction evaluation. Load the XGBoost compaction settlement prediction model, which has been trained based on historical compaction test data. Load the XGBoost model to predict the compaction settlement response under the input parameter combination.

[0075] S4.2. Set the initial parameters of PSO, including particle swarm size, maximum number of iterations, inertia factor, individual / global learning factor, and search space boundary.

[0076] S4.3. Initialize the position and velocity of the particles, and randomly generate the position (i.e., parameter combination) and initial velocity of the individual particle swarms in the parameter space.

[0077] S4.4. Use the XGBoost model to predict the tamping response. Input the parameter combination of each particle into the XGBoost model to obtain its predicted tamping amount (especially the average tamping amount of the last two hits).

[0078] S4.5. Construct an objective function that integrates construction cost and tamping response, using the form P = α × N + β × M × H + penalty. α and β are weights, representing the importance of the number of tamping strokes and construction cost, respectively, and should be set based on project experience and actual conditions. Penalty is a penalty term for violating the tamping standard; if the tamping amount exceeds the corresponding energy level threshold, a positive penalty is applied. N, M, and H represent the number of tamping strokes, tamping hammer mass, and lifting height, respectively. The feasible ranges for each parameter are shown in Table 1.

[0079] Table 1 Feasible range of each parameter setting

[0080]

[0081] S4.6. Update the individual optimal value P_best, recording the current historically optimal fitness value and corresponding position of each particle. Update the global optimal value G_best, select the particle with the smallest fitness value among all current particles as the global optimal solution, record its position, and determine whether it meets the stop-ramming standard. If the current predicted ramming settlement meets the stop-ramming condition for the corresponding ramming energy, continue iteration. Otherwise, update the particle position and velocity and continue iteration. Specifically, the stop-ramming standard refers to the "Technical Specifications for Building Foundation Treatment": the average ramming settlement of the last two impacts for ramming energies less than 4000 kN·m, 4000-6000 kN·m, and greater than 6000 kN·m should not exceed 50 mm, 100 mm, and 200 mm, respectively.

[0082] S4.7. Output the optimal compaction parameter combination: The final output is the dynamic compaction parameter combination that meets the stop-compaction standard and has the minimum adaptability (including the number of compaction times N, the rammer mass M, the rammer lifting height H, and the rammer diameter D). The optimization process is completed and the program ends.

[0083] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. A dynamic compaction construction parameter optimization method based on XGBoost-PSO, characterized in that: The following steps are involved: S1. Dynamic compaction model test: Design and manufacture a dynamic compaction model, and conduct compaction tests using test rammers of various specifications and the dynamic compaction model to obtain compaction settlement data under several different parameter combinations; S2. Feature selection: First, select the original features from the actual measured parameters of the dynamic compaction model test; Then, the actual measurement parameters are used to construct the interactive features, and the optimal interactive features are selected based on the XGboost feature importance ranking; then the correlation analysis is performed between the original features and the optimal interactive features; Step S2 specifically includes: S2.

1. Original feature selection: Based on the parameters actually measured in the model test data, the rammer mass M, lifting height H, rammer diameter D, and number of ramming strokes N were selected as original feature variables. The grey correlation analysis method was used to calculate the correlation between each original feature variable and the target variable. Original feature variables with a correlation greater than the correlation threshold were considered key influencing features and included in subsequent modeling. S2.2, optimal interactive feature selection, the specific steps are: (1) Determine the combination form of interactive features: select the ram mass M and the lifting height H as the basic variables for constructing interactive features, and set three types of combination forms MH a 、M b H and M c H d ; where a, b, c and d are power parameters used to adjust the nonlinear shape; (2) Set the parameter search range and step size: set a, b, c, d ∈ [0.1, 2, 0.1]; (3) Construct interaction features and perform K-fold cross validation: Use K-fold cross validation to train the XGBoost model for the three types of combined interaction features, and calculate the feature importance index of each interaction feature in each round of training, and use weighted average to obtain the stable importance of each interaction feature; where K=5; (4) Calculate the total importance of features and select the optimal power combination: MH a 、M b H and M c H d The average importance values of the three interactive features are added together to define the total importance of the current combination. After traversing all power combinations, the power combination with the largest total importance is selected as the final interactive feature construction strategy. (5) Construct the optimal interaction features and incorporate them into model training: Based on the evaluation results, the interaction features corresponding to the optimal power parameters are formally incorporated into the model input feature set, and the original features are combined to participate in the XGBoost model training; S2.

3. Combine the original features selected in step S2.1 and the optimal interaction features selected in step S2.2 as variables for Spearman's equation analysis. If the correlation coefficient ρ between variables is greater than 0.8, the collinearity is considered too high and features are removed. Variance inflation factor (vif) is then used to calculate the remaining features to ensure that the vif value is less than 10. S3. Based on the feature set obtained by screening in step S2, an XGBoost tamping settlement prediction model is established; S4. Use the particle swarm algorithm to optimize the compaction parameters of the XGBoost compaction settlement prediction model, and finally obtain the optimal compaction parameter combination.

2. The method for optimizing dynamic compaction construction parameters according to claim 1, wherein: Step S1 specifically includes: S1.

1. Determine Similarity Criteria and Scale Ratio: Based on the three similarity criteria of geometric similarity, physical similarity, and mechanical similarity, and in combination with the actual engineering context, select model control variables and determine the similarity coefficient system required for the impact test using the second similarity theorem and dimensional analysis. The model control variables include rammer mass M, rammer lifting height H, rammer diameter D, number of impacts N, impact energy E, ramming sinkage s, moisture content w, and soil dry density d. S1.

2. Prepare a dynamic compaction model box, configure a triangular bracket for the dynamic compaction model box, and install an electromagnetic adsorption device on the top of the triangular bracket to simulate the free fall of a heavy object during the dynamic compaction process; S1.

3. Test rammers of different specifications are provided; some of the hammers have holes on their surfaces to reduce air resistance and to meet actual engineering conditions. S1.

4. Arrange the measuring device and perform the tamping operation: Design different tamping parameters to conduct multiple groups of tamping tests. Set the number of tamping times for each group of tests to 8 to 18 times. After each tamping, read the tamping amount data and record it in the database.

3. The method for optimizing dynamic compaction construction parameters according to claim 2, wherein: In step S1.4, the soil is disturbed again and loosely spread and leveled after each set of tamping is completed to maintain the consistency of the initial state of the soil in multiple sets of tamping tests.

4. The method for optimizing dynamic compaction construction parameters according to claim 1, wherein: Step S3 specifically includes: S3.

1. Data Partitioning and Standardization: Divide the original dataset into training and test sets according to a set ratio; perform normalization preprocessing on the model input features to ensure data consistency across different dimensions. S3.2, XGBoost model initialization: Use the XGBoost regression model to construct a nonlinear mapping relationship f: X→y, where X is the feature vector and y is the tamping amount; S3.

3. Hyperparameter Grid Search and Tuning: Grid search is used to optimize the key hyperparameters of the XGBoost model, with the goal of minimizing the cross-validation error. Five-fold cross-validation is used to calculate the negative mean squared error for each set of hyperparameters, and the model combination with the best performance is selected. S3.

4. Model Prediction and Performance Evaluation: The optimized model is used to make predictions on the training and test sets, and the root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²) are used to comprehensively evaluate the model performance. S3.

5. Output the optimal model and parameters: Finally, retain and output the best model after parameter adjustment and its corresponding optimal hyperparameter combination, providing a core support tool for subsequent parameter optimization based on the prediction model.

5. The method for optimizing dynamic compaction construction parameters according to claim 2, wherein: Step S4 specifically includes: S4.

1. Start the optimization system and prepare for dynamic compaction parameter search and prediction evaluation: Load the XGBoost compaction settlement prediction model, which has been trained based on historical compaction test data, to predict the compaction settlement response under the input parameter combination; S4.

2. Set the initial parameters of PSO, including particle swarm size, maximum number of iterations, inertia factor, individual / global learning factor, and search space boundary; S4.

3. Initialize the position and velocity of the particles, and randomly generate the position and initial velocity of individual particles in the parameter space; S4.

4. Predicting the tamping response using the XGBoost model. Input the parameter combination of each particle into the XGBoost model to obtain the predicted tamping response. S4.

5. Construct an objective function that integrates construction cost and tamping response, in the form of P = α × N + β × M × H + penalty. α and β are weights, representing the importance of the number of tamping blows and construction cost, respectively, and should be set based on engineering experience and actual conditions. Penalty is the penalty term for violating the tamping standard; if the tamping amount exceeds the corresponding energy level threshold, a positive penalty is imposed. N, M, and H represent the number of tamping blows, the mass of the tamping hammer, and the lifting height, respectively. S4.

6. Update the individual optimal value P_best, record the current historical optimal fitness value of each particle and its corresponding position; update the global optimal value G_best, select the particle with the smallest fitness value among all current particles as the global optimal solution, record its position, and determine whether it meets the stop tamping criteria. If the current predicted tamping amount meets the stop tamping criteria under the corresponding tamping energy, continue; otherwise, update the particle position and velocity and continue iteration; S4.

7. Output the optimal compaction parameter combination: The final output is the compaction parameter combination that meets the stop compaction standard and has the minimum fitness, and the optimization process is completed.

Images