Marine soil compression parameter prediction model based on Light GBM
Through the marine soil compression parameter prediction model based on Light GBM, the problem of difficult prediction of marine soil compression parameters in marine engineering is solved, and more efficient and reliable soil parameter prediction is achieved, reducing the survey cost and improving efficiency.
Patent Information
- Application Number
- CN202510371982.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-27
- Publication Date
- 2025-06-27
AI Technical Summary
The prior art is difficult to effectively predict the compression parameters of marine soil in marine engineering, which leads to high cost of marine engineering testing and difficulty in obtaining soil parameters at various locations, limiting the development of marine engineering.
A marine soil compression parameter prediction model based on Light GBM is proposed. By collecting soil parameter data, preprocessing data, dividing training sets and test sets, performing k-fold cross-validation, inputting Light-GBM model, debugging hyperparameters, and finding the optimal hyperparameters through Bayesian optimization, the soil parameter prediction model is finally obtained.
This model can accurately predict the compression parameters of marine soil, reduce the cost of marine soil exploration, improve the efficiency of soil parameter exploration, and is suitable for environments with high safety requirements in marine engineering.
Smart Images
Figure CN120217185A_ABST
Abstract
Description
Technical Field:
[0001] The present invention relates to a prediction model for marine soil compression parameters based on Light GBM, aiming to improve the determination of marine geotechnical parameters in engineering projects, improve economic efficiency, especially in the environment of high safety requirements for increasing marine engineering projects in recent years. Background Art:
[0002] With the development of construction technology, various marine engineering constructions have been continuously promoted. The research on the physical and mechanical properties of marine soil has become an important issue in marine engineering practice. At present, the numerical values of the mechanical properties and design parameters of soil almost completely depend on in-situ geotechnical tests (such as sampling, drilling and penetration tests) and subsequent laboratory geotechnical tests (the compressibility and shear strength indexes of each test soil sample directly obtained from consolidation tests and triaxial tests). However, the cost of marine geotechnical engineering tests is high, and it is difficult to carry out a large number of on-site tests with high density for specific projects, and it is difficult to obtain the soil parameters at each location, which restricts the development of marine engineering. Therefore, constructing an efficient and reliable soil parameter prediction model is of great significance to the development of marine geotechnical engineering.
[0003] In response to this special need, existing research has tried to improve the detection performance by improving algorithms, introducing new loss functions or optimizing the feature extraction process. However, these methods mainly focus on stability and strength parameters, and there is a lack of research on deformation parameters. The rise and development of machine learning algorithms have emerged in the field of geotechnical engineering, including random forests, neural networks, etc., which make up for the deficiencies of traditional algorithms in expression ability and low accuracy.
[0004] The prediction model for marine soil deformation parameters based on Light GBM proposed by the present invention will predict the relatively difficult-to-obtain soil compression parameters through relatively easy-to-obtain soil mechanical parameters, and adjust the model through Bayesian parameter tuning and k-fold analysis methods to make the model obtain better prediction performance. This model is beneficial to reducing the cost of marine soil exploration and improving the efficiency of soil parameter exploration. Summary of the Invention:
[0005] In view of the deficiencies of the prior art, the present invention proposes a prediction model for marine soil compression parameters based on Light GBM, which consists of the following steps:
[0006] S1: Collect soil parameter data, including bottom depth, water content, wet density, relative density of soil particles, liquid-plastic limit, compression coefficient and compression modulus, etc.;
[0007] S2: Preprocess the data and supplement the missing parameters;
[0008] S3: Divide the data into a training set and a test set;
[0009] S4: Perform k-fold cross-validation on the training set;
[0010] S5: Input the processed data into the Light-GBM prediction model;
[0011] S6: Set the hyperparameters to be debugged, train the model and predict the target parameters;
[0012] S7: Use Bayesian optimization to find the hyperparameters that optimize the evaluation metrics;
[0013] S8: Obtain the final prediction model for soil parameters.
[0014] 2. A prediction model for marine soil compression parameters based on Light GBM according to claim 1, wherein for the data preprocessing in S2, the "3σ criterion" is used to test the data, and the specific content is as follows:
[0015] S1-1: Construct the original data sequence {y1, y2,... y n};
[0016] S1-2: Calculate the mean value y i , and the variance value σ d ;
[0017] S1-3: Calculate the difference between y i and σ d , and determine whether it is an outlier.
[0018] 3. A prediction model for marine soil compression parameters based on Light GBM according to claim 1, wherein for the Light-GBM model in S5, the model further includes the following features:
[0019] S2-1: Histogram-based decision tree algorithm, discretize the continuous feature values into discrete bins (histograms), and find the optimal split point by statistically analyzing the gradient information (such as mean, variance) of each bin, reducing the computational complexity;
[0020] S2-2: Leaf-wise growth strategy, each time select the node with the largest split gain among all current leaf nodes for splitting (instead of splitting by layer), generating an asymmetric tree;
[0021] S2-3: Direct support for categorical features, no need for one-hot encoding, directly process categorical features through specific algorithms (such as histogram-based categorical feature partitioning);
[0022] S2-4: Automatically handle missing values, and automatically learn the optimal allocation direction (left subtree or right subtree) of missing values during training;
[0023] S2-5: Regularization and overfitting control. Add a regularization term to the loss function to penalize complex models. Limit the tree structure through max_depth and num_leaves, and stop training when the performance on the validation set no longer improves.
[0024] 4. A prediction model for marine soil compression parameters based on Light GBM according to claim 1, wherein the African Vulture Optimization Algorithm (AVOA) is introduced in S7, specifically including the following:
[0025] S3-1: Put the corresponding best feasible solution into the first category, put the corresponding sub-optimal solution in the second category, and classify all the remaining solutions into the third category;
[0026]
[0027] Among them, represents the optimal vulture at the i-th iteration, represents the sub-optimal vulture at the i-th iteration, L is a custom parameter and is between [0, 1]. Calculated according to the roulette wheel strategy.
[0028]
[0029] S3-2: Construct a hunger hierarchy of vultures for the search and development of the algorithm:
[0030]
[0031] Among them, rand1 is a custom constant between [0, 1], and z is a custom constant between [-1, 1]. When |F| is greater than or equal to 1, the vulture enters the search stage. When |F| is less than 1, the vulture enters the development stage.
[0032] S3-3: Search stage:
[0033]
[0034] Among them, the formula represents the position of the n-th vulture at the (i + 1)-th iteration, and each rand is a number randomly and uniformly distributed between [0, 1], represents the gap between the vulture and the current optimal or sub-optimal vulture.
[0035]
[0036] Among them, represents the position of the n-th vulture at the i-th iteration, and C is a random number in the range [0, 2].
[0037] S3-4: When the value of |F| is between 0.5 and 1; a parameter p2 defined within the range [0,1] determines whether the vulture performs foraging competition or hovering flight behavior. Before the vulture acts, a random uniformly distributed number rand2 within the range [0,1] is randomly generated. When rand2 is less than or equal to p2, the foraging competition behavior is executed; otherwise, the hovering flight behavior is executed.
[0038]
[0039] In the hovering flight behavior, the position update formula of the vulture is:
[0040]
[0041] S3-5: When |F| is less than 0.5, all vultures are very hungry:
[0042]
[0043] As a preferred technical solution of the present invention, the k-fold cross-validation in S4 is to further divide the training set in the original dataset into k sets of the same size. Select one of the sets as the validation set, and the remaining k - 1 sets as the training set for training. Repeat this training step k times. Select the hyperparameters that minimize the average error in the k modeling runs as the final hyperparameters and train on the original entire training set.
[0044] The beneficial effects of the present invention are as follows: This method uses the negative gradient of the loss function to approximate the residual value of the current decision tree and uses it to fit a new decision tree. In each iteration, the model remains unchanged, and a new function is added to the model to continuously reduce the difference between the predicted value and the measured value. By retaining the instances with larger gradients and randomly sampling the instances with smaller gradients, a more accurate estimate is obtained.
[0045] Adopt the mutually exclusive feature merging technology to merge the mutually exclusive features within a certain conflict ratio, achieving the effect of dimensionality reduction from the perspective of reducing features without causing information loss. The prediction model using this method is original in the field of predicting marine soil parameters. Description of the drawings:
[0046] Figure 1 is the flow chart of the present invention;
[0047] Figure 2 is the overall structure module diagram of the present invention. Detailed implementation manners:
[0048] The present invention will be further described below in conjunction with the accompanying drawings and embodiments. However, the present invention can be implemented in many different ways and should not be construed as limited to the embodiments shown; on the contrary, these embodiments provide implementations that meet the applicable legal requirements for those skilled in the art.
[0049] Embodiment: Taking a total of 482 sets of geotechnical test data obtained from a certain offshore wind farm project as the research object, according to the physical and mechanical parameters of the soil, a Light-GBM prediction model is established to predict the deformation-related parameters of the soil, namely the compression factor and the compression modulus. And it is compared and verified with a certain measured value, the optimal hyperparameters are obtained through Bayesian hyperparameter tuning, and the actual prediction effect of the constructed model is evaluated by combining multiple indicators.
[0050] In this embodiment, the evaluation indicators we adopt include the mean absolute error, the root mean square error, the mean absolute percentage error, and the coefficient of determination.
[0051] The mean absolute error is the average of the absolute values of the differences between the predicted values and the true values. It reflects the accuracy of the model in prediction. The calculation formula for the mean absolute error is as follows:
[0052]
[0053] In this formula, represents the model prediction value; while y i represents the actual true value.
[0054] The root mean square error is the average of the squared differences between the predicted values and the true values. The closer the predicted value is to the true value, the more accurate the model is, and the lower the RMSE value. The calculation formula for the root mean square error is as follows:
[0055]
[0056] In this formula, still represents the model prediction value; while y i represents the actual true value.
[0057] The mean absolute percentage error is the average of the percentages of the errors between the predicted values and the measured values to the measured values. The closer the predicted value is to the true value, the more accurate the model is, and the lower the MAPE value. The calculation formula for MAPE is as follows:
[0058]
[0059] The coefficient of determination R 2 is calculated as:
[0060]
[0061] These metrics together provide a comprehensive evaluation of the model's performance. These metrics are crucial for understanding the advantages and disadvantages of the model and guiding further optimization.
[0062] The experimental environment of this embodiment is as follows: Processor: Intel Core i9-13900H, Memory is 32GB DDR5 (5200MHz), Graphics card used is NVIDIA GeForce RTX 4060 Laptop GPU, Storage hard disk is 1024GB SSD, Operating system used is Windows 11. Development environment: Python version is 3.11, PyTorch version is 1.8.0, torchvision version is 0.9.0
[0063] The prediction results are shown in Table 1-3 below:
[0064]
[0065] Table 1 Light-GBM Bayesian hyperparameter tuning results for compression coefficient a v
[0066]
[0067] Table 2 Light-GBM Bayesian hyperparameter tuning results for compression modulus E s
[0068]
[0069] Table 3 Accuracy evaluation metrics for different models
[0070] In summary, the present invention proposes a prediction model for marine soil compression parameters based on Light-GBM, aiming to improve the determination of marine geotechnical parameters in engineering projects and enhance economic efficiency. Past research has mainly attempted to improve detection performance by improving algorithms, introducing new loss functions, or optimizing the feature extraction process. These methods mainly target stability and strength parameters, and there is a lack of research on deformation parameters. To make up for this lack, the present invention introduces a prediction model for marine soil compression parameters, which can not only predict the relatively difficult-to-obtain soil compression parameters through relatively easy-to-obtain soil mechanical parameters, but also promote the development of the Light-GBM algorithm in the field of geotechnical engineering to a certain extent. Through the above technologies, the model shows excellent performance in predicting marine soil compression parameters.
[0071] The above embodiments merely represent several implementation manners of the present invention. The description thereof is relatively specific and detailed, but it should not be construed as a limitation on the scope of the invention patent. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all fall within the protection scope of the present invention.
Claims
1. A marine soil compression parameter prediction model based on Light GBM, the specific steps are as follows: S1: Collect soil parameter data, including sample bottom depth, moisture content, wet density, soil particle relative density, liquid and plastic limits, compression coefficient and compression modulus, etc.; S2: preprocess the data and supplement the missing parameters; S3: Divide the data into training set and test set; S4: Perform k-fold cross validation on the training set; S5: Input the processed data into the Light-GBM prediction model; S6: Set the hyperparameters that need to be debugged, train the model and predict the target parameters; S7: Use Bayesian optimization to find the hyperparameters that optimize the evaluation index; S8: Obtain the final soil parameter prediction model.
2. A marine soil compression parameter prediction model based on Light GBM according to claim 1, characterized in that: The data preprocessing in S2 uses the "3σ criterion" to test the data, which specifically includes the following contents: S1-1: Construct the original data sequence {y1, y2, ...y n }; S1-2: Calculate the average value y i , variance value σ d ; S1-3: Calculate y i With σ d Difference, determine whether it is an outlier.
3. The marine soil compression parameter prediction model based on Light GBM according to claim 1 is characterized in that: The Light-GBM model in S5 further includes the following features: S2-1: The histogram-based decision tree algorithm discretizes the continuous feature values into discrete bins (histograms), and finds the optimal split point by counting the gradient information (such as mean and variance) of each bin, thus reducing the computational complexity; S2-2: Leaf-wise growth strategy, each time the node with the largest splitting gain among all current leaf nodes is selected for splitting (rather than splitting by layer), generating an asymmetric tree; S2-3: Direct support for categorical features, without one-hot encoding, and directly processing categorical features through specific algorithms (such as histogram-based categorical feature partitioning); S2-4: Automatically handle missing values and automatically learn the optimal allocation direction of missing values (left subtree or right subtree) during training; S2-5: Regularization and overfitting control, add regularization terms to the loss function, penalize complex models, limit the tree structure through max_depth and num_leaves, and stop training when the performance of the validation set no longer improves.
4. A marine soil compression parameter prediction model based on Light GBM according to claim 1, characterized in that The African Vulture Optimization Algorithm (AVOA) was introduced in S7, which includes the following contents: S3-1: The best feasible solution is placed in the first category, the suboptimal solution is placed in the second category, and the remaining solutions are all placed in the third category; in, represents the optimal vulture at the i-th iteration, represents the suboptimal vulture at the i-th iteration, L is a custom parameter and is between [0,1]. Calculated according to the Roulette strategy. S3-2: Constructing a vulture hunger hierarchy for algorithm search and development: Among them, rand1 is a custom constant between [0, 1], z is a custom constant between [-1, 1]. When |F| is greater than or equal to 1, the vulture enters the search phase, and when |F| is less than 1, the vulture enters the development phase. S3-3: Search phase: The formula represents the position of the nth vulture at the i+1th iteration, and each rand is a number randomly and uniformly distributed between [0,1]. Indicates the gap between the vulture and the current best or second best vulture. in, represents the position of the nth vulture in the ith iteration, and C is a random number in the range [0, 2]. S3-4: When the value of |F| is between 0.5 and 1; the parameter p2 defined in the range of [0,1] determines whether the vulture performs foraging competition or hovering flight behavior. Before the vulture takes action, a random uniformly distributed number rand2 in the range of [0,1] is randomly generated. When rand2 is less than or equal to p2, the foraging competition behavior is performed, otherwise the hovering flight behavior is performed. In the hovering flight behavior, the vulture's position update formula is: S3-5: When |F| is less than 0.5, all vultures are very hungry: