A virtual water tank simulation method based on data augmentation and XGBoost

Abstract

The present application belongs to the technical field of engineering experiment, and particularly relates to a virtual water tank simulation method based on data expansion and XGBoost. The method comprises the following steps: step one, collecting scouring test data of a wave-current water tank platform, and establishing a database; step two, data preprocessing; step three, data expansion based on a generative adversarial network (GAN); step four, construction of a scouring depth prediction model based on an extreme gradient boosting (XGB); step five, XGB model parameter optimization; and step six, model performance evaluation. The present application can accurately predict the complex interaction among water flow, bed surface and foundation in a short time, the gap between the prediction result and the actual observation data is very small, and the present application can provide a reliable prediction basis in actual engineering. Meanwhile, the present application can significantly reduce experimental cost and resource consumption, and enhance real-time monitoring and optimization capability. The present application has wide applicability and scalability.

Description

Technical Field

[0001] This invention belongs to the field of engineering experimental technology, specifically relating to a virtual water tank simulation method based on data augmentation and XGBoost. Background Technology

[0002] Localized scour is the process by which the soil surrounding a foundation is eroded and carried away by water flow. It involves complex multi-field relationships in fluid mechanics, geotechnical mechanics, and structural mechanics. Its occurrence and development can reduce the bearing capacity of the foundation, leading to a risk of collapse. Therefore, reliable prediction of localized scour is crucial for the safety design of water-related foundation engineering. In practical engineering, the prediction of localized scour typically relies on wave flume experiments, theoretical empirical formulas, and numerical simulations. While these methods provide some support, these three traditional approaches face technical and application challenges. Theoretical empirical formulas perform inconsistently under different hydrogeological conditions, and the predicted results often deviate significantly from reality. Calculations based solely on theoretical formulas in practical engineering may introduce potential risks. Numerical simulation methods depend on parameter values ​​and modeling methods; different modeling methods or parameter values ​​may yield significantly different results, often requiring verification through physical model experiments before parameter sensitivity analysis can be performed. Wave flume experiments are relatively more reliable, but this method consumes significant manpower, physical resources, and time. It requires experienced testers, thorough experimental preparation, and a large wave flume platform, making this approach costly and time-consuming.

[0003] In recent years, machine learning technology has been gradually introduced into engineering data analysis, especially by using algorithms such as Support Vector Regression (SVR) and Artificial Neural Networks (ANN) to process large amounts of historical data and make predictions under the influence of relevant factors. Machine learning-based prediction methods can extract potential patterns by learning from existing data, thereby making relatively accurate predictions.

[0004] Extreme Gradient Boosting (XGBoost, also abbreviated as XGB) constructs a decision tree system sequentially, where each tree is built to correct the errors made by the previous tree. The algorithm uses gradient descent to minimize the loss function; the core idea is to calculate the loss gradient relative to the model's predictions and use these gradients to guide the creation of new trees that reduce the loss.

[0005] While machine learning predictions are more accurate, they also have many problems: High requirements for training data: Machine learning models rely on a large amount of historical data for training, and the quality and quantity of data directly affect the accuracy of predictions. In the absence of sufficient data, machine learning models may fail to accurately capture the complex changes during the scouring process.

[0006] Limited generalization ability: Although machine learning can learn patterns from data, its model's generalization ability (i.e., the model's performance in new environments or under new conditions) may be weak, especially when faced with different geological conditions or hydrological environments, the model may not be able to adapt well.

[0007] Lack of real-time update capability: Most existing machine learning models are trained on static data and lack the ability to update and optimize data in real time, making them unable to cope with the ever-changing realities on the engineering site. Summary of the Invention

[0008] To address the aforementioned issues, and drawing upon the applicant's extensive accumulated wave flume experimental data and artificial intelligence technology, the purpose of this invention is to propose a virtual flume simulation method based on data augmentation and XGBoost. Based on this method, an intelligent flume is established, overcoming the inherent limitations of traditional physical flume scour tests, such as long testing time, high cost, and poor scalability, thereby achieving efficient and accurate simulation and prediction of the local scour evolution process around underwater foundations.

[0009] Traditional physical tank experiments have limitations; conducting physical model tests typically consumes significant time and resources, resulting in high costs and low efficiency. This invention employs the XGBoost machine learning method, using experimental data accumulated over many years to train a model. This enables rapid processing of large amounts of data and prediction of scour depth and its evolution trends, achieving rapid evaluation and assessment of the scour process. This significantly reduces the human and material costs required for scour testing and saves time, thereby providing safety analysis and optimization suggestions for underwater foundation design and solving the problems of high resource consumption and low efficiency in traditional scour experiments.

[0010] Technical solution A virtual water tank simulation method based on data augmentation and XGBoost includes the following steps: Step 1: Collect scouring test data from the wave flume platform and establish a database; Step 2: Data preprocessing.

[0011] Data preprocessing includes: data cleaning, data standardization, and splitting the data into training and test sets.

[0012] Step 3: Data augmentation based on Generative Adversarial Networks (GANs). Data augmentation is performed using Generative Adversarial Networks (GANs). Through adversarial learning between the generator and discriminator, the database is expanded to provide sufficient data support for subsequent model training.

[0013] Step 4: Construction of a scour depth prediction model based on Extreme Gradient Boosting (XGB). A scour depth prediction model is constructed based on the XGB algorithm. By integrating multiple decision trees and introducing regularization techniques, the complex nonlinear relationship between scour parameters and scour depth is effectively captured.

[0014] Step 5: XGB Model Parameter Optimization. Automatic optimization of key parameters of the XGB model is performed using a strategy combining Bayesian optimization (BO) and 5-fold cross-validation to determine the optimal parameter configuration to balance model complexity and generalization ability.

[0015] Step Six: Model Performance Evaluation. The model performance is systematically evaluated using multiple metrics, including the coefficient of determination (R²), root mean square error (RMSE), mean absolute error (MAE), and probability of prediction interval coverage (PICP).

[0016] Beneficial effects 1. Improve forecasting efficiency and accuracy: By employing the extreme gradient boosting XGB method, this invention enables accurate prediction of complex water flow-bed-foundation interactions in a short time.

[0017] Prediction accuracy: The R² of the model in this invention exceeded 0.88 in the prediction results of other scour tests, indicating that they can maintain highly consistent prediction performance under different datasets.

[0018] Error performance: In the prediction of the maximum scour depth, the root mean square error (RMSE) of the model of this invention is less than 0.7 cm and the mean absolute error (MAE) is less than 0.6 cm. This shows that the model's prediction results are very close to the actual observation data and can provide a reliable prediction basis in practical engineering.

[0019] 2. Significantly reduces experimental costs and resource consumption: Traditional physical water tank experiments require the preparation of a large number of experimental samples and testing equipment, and each test involves complex environmental configurations and multiple trials. This invention avoids these high-cost physical experiments through computer simulation. For example, a traditional physical water tank experiment may take 12 days to complete a full test task (including model making, setup, data acquisition, etc.), while the simulation prediction using this invention only takes a few seconds, which reduces the time and economic cost of each test task by about 1000 times.

[0020] Cost savings: Considering that the experimental cost of each traditional test is approximately $6,850, including the cost of experimental equipment, labor and consumables, the application of this invention can significantly reduce the investment in experimental personnel and equipment, saving more than 90% of the experimental costs.

[0021] Resource optimization: This invention enables experiments to no longer rely on a large amount of physical experimental data, and can make predictions directly using existing data, greatly reducing the consumption of physical resources such as soil samples, equipment debugging, and model making.

[0022] 3. Enhance real-time monitoring and optimization capabilities: In practical engineering applications, this invention introduces a real-time update mechanism, enabling the model to continuously optimize based on real-time field data. This technology helps engineers predict sudden scour risks, such as scour during strong currents and floods. Each real-time update ensures that the model's performance in the actual scenario more closely approximates reality, thereby enhancing its accuracy in predicting future scour depth and process evolution.

[0023] 4. Wide applicability and scalability of the technology: Scalability: Under varying environmental conditions, this invention can expand the applicability of the model based on new data and actual testing, adapting to the design needs of different regions and structural types.

[0024] Applicable to multiple scenarios: Through continuous optimization of machine learning algorithms, this invention can adapt to different soil types, water flow types, foundation shapes and other actual engineering parameters, providing support for scour risk assessment of various structures.

[0025] 5. Supports decision-making and optimized design: This invention is not merely a predictive tool; it also enables engineers to identify potential risks and optimize designs during the design phase by providing real-time scour evolution data and sensitivity analysis. For example, through sensitivity analysis of parameters such as different flow velocities, water depths, and foundation model diameters, this invention helps engineers understand the impact of various factors on scour depth, thereby providing quantitative optimization suggestions for the foundation design.

[0026] Decision support: Based on the prediction results and sensitivity analysis provided by this invention, engineers can reasonably adjust the design parameters such as the size and shape of the underwater foundation to reduce the potential risks caused by scouring.

[0027] Optimized design: Under different operating conditions, the present invention can provide targeted suggestions. For example, in high flow velocity areas, the present invention may recommend increasing the diameter of the foundation or adopting a more stable foundation design, thereby improving the structure's resistance to erosion. Attached Figure Description

[0028] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a schematic diagram of the process of using GAN to enhance the scour depth dataset according to the present invention; Figure 3This is a comparison chart of the predicted depth and the measured scour depth in an embodiment of the present invention; Figure 4 The diagram shows the scouring depth variation under different conditions according to an embodiment of the present invention. Detailed Implementation

[0029] The technical solution provided in this application will be further described below with reference to specific embodiments and accompanying drawings. The advantages and features of this application will become clearer from the following description.

[0030] A virtual water tank simulation method based on data augmentation and XGBoost includes the following steps: (e.g.) Figure 1 ) Step 1: Collect scouring test data from the wave flume platform and establish a database; The scouring test data includes: the evolution of scouring depth under different conditions, the maximum depth at which scouring equilibrium is reached, and the time required to reach scouring equilibrium.

[0031] The experiment used a basic model constructed according to similar proportions, and conducted experiments under different flow velocity conditions, seabed parameters, and foundation types.

[0032] A database was established to collect scour test data, which includes the following key parameters: Median particle size (d) 50 The range is 0.1-5mm; Water depth (hp), ranging from 10-40cm; Flow velocity (V), ranging from 10 to 40 cm / s; The model diameter (R) ranges from 3 to 40 cm; Time (t) records the erosion development at different points in time; scour depth ( ), corresponding to the absolute value of scour depth under different conditions.

[0033] Step 2: Data preprocessing.

[0034] Based on existing prediction formulas and the experience of professionals, the experimental data are preprocessed, including data cleaning, data standardization, and data segmentation.

[0035] The data cleaning is used to remove outliers and missing values; Data standardization standardizes the flushing process data (flushing depth) to a ratio relative to the maximum flushing depth.

[0036] Data splitting involves dividing the dataset into training and test sets at a ratio of 80% / 20%.

[0037] Step 3: Data augmentation based on Generative Adversarial Networks (GANs). Data augmentation is performed using Generative Adversarial Networks (GANs). Through adversarial learning between the generator and discriminator, the database is expanded to provide sufficient data support for subsequent model training.

[0038] Step 4: Construction of a scour depth prediction model based on Extreme Gradient Boosting (XGB). A scour depth prediction model is constructed based on the XGB algorithm. By integrating multiple decision trees and introducing regularization techniques, the complex nonlinear relationship between scour parameters and scour depth is effectively captured.

[0039] Step 5: XGB Model Parameter Optimization. Automatic optimization of key parameters of the XGB model is performed using a strategy combining Bayesian optimization (BO) and 5-fold cross-validation to determine the optimal parameter configuration to balance model complexity and generalization ability.

[0040] Step Six: Model Performance Evaluation. The model performance is systematically evaluated using multiple metrics, including the coefficient of determination (R²), root mean square error (RMSE), mean absolute error (MAE), and probability of prediction interval coverage (PICP).

[0041] The specific process of step one is as follows: We collected physical tank test data, which came from flushing tests conducted in the tank laboratory.

[0042] Furthermore, the scouring test setup conducted in the water tank laboratory includes: (1) Based on the similarity of the Frod number, determine the similarity scale of the test, calculate the test parameters, determine the pile diameter, water depth, design flow velocity, model soil type and particle size, and ensure that the influence of the sidewall effect is avoided during the test. (2) Fabricate test models to ensure that their dimensions meet the test requirements. Various sizes of pile model foundations made of transparent acrylic material can be selected and placed in sedimentation test areas with different parameters. Ensure that the pile body is perpendicular to the upper surface of the sedimentation tank and flatten the sand surface to reduce the influence of the initial bed surface conditions on the test results. (3) Set up test devices such as current meter, wave height meter, and depth sounder, and adjust them according to actual development; (4) When the scouring depth no longer changes or changes less than 1 mm within 1 hour, it is considered that the scouring development has reached equilibrium and the test can be stopped. During the test, the scouring depth is measured at preset time points (0, 1, 2, 5, 10, 20, 30, 60 minutes, etc.). Depending on the test parameters, the test may last from several hours to tens of hours. (5) Repeat steps (3) to (4) above according to the proposed experimental plan to complete all experiments.

[0043] The specific process of step two is as follows: The collected scouring test data were preprocessed. First, the data was cleaned to remove outliers and missing values ​​with measurement errors exceeding 5% to ensure data quality. The data was then standardized, with the scour depth being standardized as a ratio relative to the maximum scour depth, as follows: , Among them, flushing process parameters This represents the normalized scour depth at time t. This represents the actual scour depth at time t. Indicates the maximum scour depth and scour progress indicators. It is used to characterize the percentage of time t away from the maximum scour depth, and when it reaches 100%, it means that the maximum scour has been reached; this standardization process makes scour depth data under different conditions comparable.

[0044] This invention creatively introduces the concept of a scour process parameter S(t), breaking through the limitations of traditional scour analysis methods. Traditional scour prediction methods typically focus only on the static indicator of the final maximum scour depth, ignoring the dynamic evolution characteristics of scour development. This invention, by constructing a time-dependent scour process parameter S(t), can comprehensively depict the complete time-series development trajectory of scour from its initial occurrence to its equilibrium state. This parameter not only reflects the degree of scour development at any given time t, but more importantly, reveals the dynamic laws of scour evolution, including the segmented characteristics of the initial rapid development stage, the middle slow growth stage, and the later equilibrium stage. Through the S(t) parameter, engineers can predict the scour state at specific points in time, assess the rate of scour development, and determine the time windows required to reach different scour levels, providing a precise scientific basis for the timing selection and strength design of engineering protection measures. The introduction of this innovative parameter transforms scour prediction from traditional "final state prediction" to "full-process monitoring," significantly improving the timeliness and accuracy of scour risk assessment. Finally, the data is split into training and test sets at a ratio of 80% / 20% to ensure that the two sets have similar data distribution characteristics. The training set is used for model training and parameter optimization, while the test set is used to evaluate the model's generalization ability and prediction performance.

[0045] Step three involves augmenting the data using Generative Adversarial Networks (GANs). (e.g., ...) Figure 2 ) To address the specific challenges of data acquisition and insufficient sample size in the field of scour prediction, this invention employs an improved Wasserstein generative adversarial network (WGAN-GP) technique to augment the data. The Wasserstein generative adversarial network includes a generator... and discriminator A multilayer perceptron structure is designed to generate and discriminate scour depth data that conforms to physical characteristics by inputting a random noise vector and conditional parameters. Wasserstein distance and gradient penalty techniques are introduced to improve training stability, with the penalty term coefficient set to 10.

[0046] The core of GAN data augmentation lies in the generator. and discriminator The adversarial learning process. Traditional GAN ​​algorithms generate data solely based on data distribution, which can easily produce false data that violates the physical laws of erosion. This invention introduces erosion mechanics constraints into both the generator G and the discriminator D to ensure the physical validity of the generated data.

[0047] Specifically, we introduce a conditional parameter vector c=[d 50 [,hp,V,R] serves as the scour mechanical constraint condition; Then generator The constraint expression is: , in, For noise, As constraints, This represents the scour depth data output by the generator. To scour mechanical constraint functions, To constrain the allowable threshold, both work together to ensure the physical rationality of the generated data.

[0048] The constraint function specifically includes three key constraints: ① Dimensionless scour depth constraint: ,in Represents the absolute value of the scour depth, model diameter The feature dimensions of the basic structure, For flow rate, It is the acceleration due to gravity. Because of the water depth, This represents a functional relationship established based on hydraulic theory. This constraint ensures a proper match between scour depth and hydraulic conditions; ② The constraint of monotonicity in development: ,in This represents the normalized scour depth at time t. Its first-order partial derivative with respect to time represents the rate of erosion development. The second-order partial derivative represents the change in acceleration, ensuring the physical rationality of the scouring process in the time dimension and guaranteeing that the scouring depth increases monotonically with time and the growth rate decreases. ③ Critical shear stress constraint: For bed shear stress, The critical shear stress is used to ensure that the generated data meets the scour initiation conditions. Fluid density, specifically the density of water, is typically expressed in kg / m³. In bed shear stress calculations, the product of fluid density and the square of the flow velocity determines the dynamic pressure exerted by the water flow on the bed surface, which is a fundamental physical quantity representing the driving force of scouring. The coefficient of friction of the bed surface (dimensionless) reflects the influence of the roughness of the bed surface on the resistance of water flow; its value is related to the particle size of the bed surface, water flow conditions, etc., and is used to convert the flow velocity into the shear stress of the bed surface.

[0049] The discriminator is corrected accordingly. The corrected discriminator Output probability value While determining the authenticity, it also verifies the degree to which physical constraints are satisfied. The scour depth data is input to the decision-maker and can be either a real sample or a generated sample.

[0050] Comprehensive expansion of the time-scale scouring process: Unlike traditional methods that only focus on the final state expansion of the maximum scouring depth, this invention achieves intelligent expansion of the entire scouring process over time.

[0051] By introducing a time-weighted loss function: , in For sample indices (from 1 to n). The total number of samples, The index is the time index (from 1 to T), where T is the total length of the time series. For time weighting factors, where The time decay parameter emphasizes the importance of the initial scour development. Here is the loss metric function used to calculate the first... One sample in Real-time normalized erosion depth With generated value The differences between them.

[0052] The generation process employs a conditional time series generation strategy: ,in For noise, As a constraint, ensure that the generated flushing process sequence at each time node All satisfy: S'(t) i+1 )≥S'(t i And S'( → 100%. Simultaneously, a flushing process filtering mechanism is established: retaining those that meet the requirements... The generated samples, where =5%, =3% represents the allowable deviation at 10 minutes and 30 minutes, respectively. and These are the reference normalized flush depths at 10 minutes and 30 minutes, respectively.

[0053] The training process of GAN adopts an alternating optimization strategy. In each iteration, the generator G is fixed first, and the discriminator D parameters are updated 5 times. Then, the discriminator D is fixed, and the generator G parameters are updated once.

[0054] This invention achieves three key innovations based on traditional GANs: First, a conditional constraint mechanism is introduced, which, based on the physical characteristics of scour, introduces a conditional parameter vector c=[d 50 The first constraint is to use [, hp, V, R] as generation constraints to ensure that the generated data conforms to the physical laws of scour mechanics; the second is to integrate prior knowledge of scour mechanics, embedding the physical relationship between scour depth and hydraulic parameters into the generator network structure to avoid generating data that does not conform to physical laws; the third is to extract features at multiple scales (mainly time scales), designing a multilayer perceptron structure for scour time history data to effectively capture the scour development characteristics at different time scales.

[0055] This embodiment successfully expanded 48 sets of original data to 500 sets through data augmentation technology. The statistical distribution characteristics of the generated data are highly consistent with the original data, providing sufficient and high-quality data support for the subsequent training of machine learning models.

[0056] The specific process of step four is as follows: This invention employs the Extreme Gradient Boosting (XGBoost) algorithm to construct a scour depth prediction model. This algorithm introduces regularization techniques on top of the traditional gradient boosting decision tree, effectively avoiding overfitting and improving model performance. The key innovation of this invention lies in constructing a feature system oriented towards scour mechanics, deeply integrating the physical mechanism of scour into the XGBoost algorithm framework.

[0057] The prediction function of the XGBoost model is as follows: , in, This represents the predicted scour depth value for the i-th sample. The function represents the k-th decision tree, where K is the total number of decision trees. The input feature vector [d] for the i-th sample 50 ,hp,V,R,t].

[0058] The prediction function integrates multiple decision trees to evaluate the key parameters [d] of the scour process. 50Nonlinear combinatorial learning of [hp,V,R,t] enables accurate prediction of the temporal evolution of scour depth. Each decision tree... During training, the system automatically learns the optimal split points and weight combinations between different scouring parameters, such as high flow velocity V and small particle size d. 50 Predicting larger scour depths under certain conditions, and considering the enhancing effect of flow around the foundation under conditions of large water depth hp and large foundation size R.

[0059] The objective function of the XGBoost model is: , The objective function consists of two parts: the first part As the loss function, this embodiment uses the squared loss function to measure the difference between the predicted value and the true value; Part Two This is a regularization term used to control model complexity and prevent overfitting. Through this design, the model learns the complex relationships between scour parameters while avoiding excessive memorization of specific scour conditions from the training data, thus improving its generalization ability to predict scour under different engineering environments. The true scour depth value (label value) of the i-th training sample, i.e., the measured scour depth used during model training, is compared with the predicted value. Together, they form the basis for calculating the loss function.

[0060] Among them, the regularization term: , The regular expression consists of two parts: Control the number of leaf nodes, where Let be the number of leaf nodes in the k-th tree. (Hyperparameter gamma) is the penalty parameter for the number of leaf nodes; For L2 regularization terms, where The leaf node weight vector, (Hyperparameter lambda) is the L2 regularization parameter, which prevents leaf node weights from becoming too large.

[0061] Unlike traditional machine learning methods that simply treat these parameters as numerical inputs, this invention introduces input feature vectors. = [d 50 The scouring mechanics-guided design of [hp, V, R, t] fully explores the scouring physics of each parameter: median particle size d 50The flow velocity (hp) directly determines the scour resistance and initiation characteristics of sediment, and is the fundamental source of scour resistance. Water depth (hp) affects velocity distribution and bed shear stress, controlling the driving force of scour. Flow velocity (V) is the main dynamic factor in the scour process, exhibiting a strongly nonlinear relationship with scour depth. The radius (R) of the basic model determines the flow characteristics and eddy current intensity, influencing the spatial development pattern of scour. The time parameter (t) captures the temporal evolution of scour from its initiation to equilibrium. This characteristic construction based on scour mechanics enables the XGBoost model to better understand and learn the complex interactions between water flow, sediment, and structure.

[0062] The algorithm training process is as follows: First, the mean of the training samples is used as the initial prediction. The initial value is obtained by minimizing the loss function, providing a starting point for subsequent iterations.

[0063] Then comes the iterative training phase, where each iteration revolves around optimizing the objective function: calculating the residual gradient between the current model's predicted value and the true value, based on the first and second partial derivatives of the loss function with respect to the predicted value; constructing a new decision tree to fit the residual gradient, with the split nodes of the decision tree preferentially selecting parameter combinations that have the most significant impact on scour depth, such as V² and d. 50 The ratio relationship between R and hp, and the geometric ratio relationship between R and hp are calculated; the optimal weight of each leaf node is calculated, which is determined by the residual gradient, the second gradient and the regularization parameter, to ensure that the model complexity is controlled while fitting the residual; the learning rate is applied to scale the contribution of the new tree, and then the prediction function of the overall model is updated.

[0064] Repeat the iterative process until the preset upper limit of the number of trees is reached or the verification error no longer decreases significantly. At this point, the model is considered to have converged to the optimal state.

[0065] Through this progressive model building approach, the XGBoost algorithm can gradually correct prediction biases, with each newly added decision tree focusing on improving the shortcomings of the preceding model. The introduction of regularization terms effectively prevents the model from overfitting the training data, improving the model's generalization ability. This algorithm is particularly suitable for handling the complex nonlinear relationship between hydraulic parameters and scour depth in scour prediction. By incorporating the scour mechanics mechanism into this progressive model building approach, the XGBoost algorithm can gradually reveal and learn the interaction laws of various parameters during the scour process, with each newly added decision tree focusing on improving the prediction accuracy under specific scour conditions. This algorithm is particularly suitable for handling hydraulic parameters [V, hp], geometric parameters [R], and material parameters [d] in scour prediction. 50 The strong coupling nonlinear relationship between the time parameter [t] and the time parameter [t] provides a high-precision intelligent solution for predicting the time history of scour depth in engineering practice.

[0066] The specific process of step five is as follows: The hyperparameters of the XGBoost algorithm directly affect its performance. In this embodiment, we use a strategy that combines Bayesian optimization (BO) with 5-fold cross-validation to automatically optimize the nine key hyperparameters of the XGBoost algorithm.

[0067] The search space for the nine hyperparameters is set as follows: gamma parameter [0.1, 5] logarithmic distribution, learning rate [0.01, 0.3] uniform distribution, maximum depth [3, 10] integer distribution, number of trees [100, 1500] integer distribution, subsampling rate [0.6, 1.0] uniform distribution, column sampling rate [0.6, 1.0] uniform distribution, minimum subweight [1, 10] integer distribution, alpha parameter [0, 1.0] uniform distribution, and lambda parameter [0, 2.0] uniform distribution.

[0068] Bayesian optimization employs the RBF kernel selection method, automatically learning kernel parameters through maximum likelihood estimation. Performance evaluation utilizes 5-fold cross-validation, with the root mean square error (RMSE) as the objective function. The average of five validation results is used for each parameter configuration. The optimization process consists of two phases: the initialization phase uses Latin hypercube sampling to generate 10 initial parameter configurations, ensuring uniform coverage of the search space; the iterative optimization phase executes 50 iterations, updating the Gaussian process model based on historical data in each iteration to obtain new parameter configurations, and then performing 5-fold cross-validation evaluation. The entire optimization process is repeated three times using different random seeds, selecting the optimal configuration through 10-fold cross-validation comparison, while an early stopping mechanism is introduced to prevent overfitting.

[0069] Through this system optimization process, the optimal parameter configuration for predicting scour depth was finally determined as follows: gamma value 0.01807, learning rate 0.29361, maximum depth 7, number of trees 866, subsampling rate 0.8, column sampling rate 0.7, minimum subweight 3, alpha parameter 0.1, and lambda parameter 1.0. This optimized parameter configuration effectively balances the model's complexity and generalization ability, ensuring the accuracy and stability of the prediction system.

[0070] Step six involves systematically evaluating the performance of each model using multiple complementary statistical metrics. These metrics include the coefficient of determination (R²), root mean square error (RMSE), mean absolute error (MAE), and the probability of coverage of the predicted interval.

[0071] The coefficient of determination is calculated as follows: The root mean square error (RMSE) measures the model's ability to explain data variation; it is calculated as follows: The mean absolute error (MAE) measures the absolute deviation between the predicted and actual values; it is calculated as follows: It provides an intuitive measure of prediction error; the prediction interval coverage probability calculates the proportion of actual values ​​falling within the prediction interval, assessing the reliability of the model. It is the measured scour depth value of the i-th test sample. It is the predicted scour depth of the i-th test sample; Meaning and Step Four Same, but step six has been changed. This corresponds more closely to the symbol system for scouring depth in step one. It is the average of the measured scour depths of all test samples.

[0072] These evaluation metrics comprehensively reflect the model's predictive accuracy and stability from different perspectives. Through performance evaluation on the test set, the constructed XGB model has been verified to accurately predict the scour depth development process under different conditions, providing reliable technical support for scour risk assessment and protection design in engineering practice.

[0073] Example 1 The generalization ability of the model is evaluated based on a virtual water tank simulation method. This embodiment uses additional flushing data from the water tank experiments that were not included in the training set to evaluate the generalization ability of the proposed virtual water tank simulation method. The model's gamma value is 0.01807, the learning rate is 0.29361, the maximum depth is 7, and the number of trees is 866. The results are as follows... Figure 3 As shown, the prediction results of the method of the present invention are... Larger values ​​for RMSE and MAE indicate that the predicted results are very close to the actual results, meaning that the model has high prediction accuracy and good generalization ability, and also performs well in predicting exogenous data.

[0074] Example 2 Based on the method of this invention, an AI-powered water tank was designed, and parameter sensitivity analysis was performed. This embodiment utilizes a trained intelligent model to systematically analyze various factors affecting scour depth. As shown in Table 1, by designing 160 sets of input conditions with different parameter combinations, the AI-powered water tank predicted the maximum scour depth and scour evolution process under various conditions. Figure 4 As shown, the analysis of the prediction results reveals the influence of different parameters on the scour depth: (1) Particle size (d) 50(1) Effects of flow velocity (V): Under steady flow conditions, smaller particles lead to a faster scour depth reaching a steady state, indicating that finer sediments are more easily transported by water flow and settle rapidly. (2) Effects of water depth (hp): Studies have shown that as water depth increases, scour depth also increases accordingly. (3) Effects of flow velocity (V): Flow velocity and scour depth are significantly positively correlated. Higher flow velocity leads to greater scour depth, which may be due to the stronger erosive force and enhanced turbulence generated by higher flow velocity, resulting in more severe erosion and greater scour depth around the foundation. (4) Effects of model diameter (R): The influence of pile radius on scour depth is relatively limited, which may be due to the small variation in pile diameter in the training data and the insensitivity of the training model to changes in pile diameter. By analyzing the influence of various parameters on scour depth, it was found that flow velocity has the most significant effect, with higher flow velocity leading to greater scour depth; the influence of soil particle size on scour depth decreases with increasing particle size. The parameter analysis of the AI ​​flume illustrates the key role of hydrodynamics and soil particle properties in determining the scour depth around hydraulic structures, and can provide guidance for the design and construction of hydraulic structures.

[0075] Table 1 Sensitivity Analysis Parameter Settings Conventional scouring tests often require physical model testing for verification before parameter sensitivity analysis can be performed; wave flume testing requires a lot of manpower, material resources and time, requires testers to have rich testing experience, to make sufficient test preparations, and to be equipped with a large wave flume platform, making this method costly and time-consuming.

[0076] The above description is merely a description of preferred embodiments of this application and is not intended to limit the scope of this application in any way. Any changes or modifications made by those skilled in the art based on the above-disclosed technical content should be considered as equivalent and valid embodiments and fall within the scope of protection of the technical solution of this application.

Claims

1. A virtual water tank simulation method based on data augmentation and XGBoost, characterized in that, Includes the following steps: Step 1: Collect scouring test data from the wave flume platform and establish a database; Step 2: Data preprocessing; Data preprocessing includes: data cleaning, data standardization, and splitting the data into training and testing sets; Step 3: Data augmentation based on Generative Adversarial Networks (GANs); Generative Adversarial Networks (GANs) are used to augment the data. Through adversarial learning between the generator and the discriminator, the database is expanded to provide sufficient data support for subsequent model training. Step 4: Constructing a scour depth prediction model based on extreme gradient boosting XGB; A scour depth prediction model is constructed based on the XGB algorithm. By integrating multiple decision trees and introducing regularization techniques, the complex nonlinear relationship between scour parameters and scour depth is effectively captured. Step 5: Optimize XGB model parameters; Automatic optimization of key parameters of the XGB model is performed by using a strategy that combines Bayesian optimization and 5-fold cross-validation to determine the optimal parameter configuration to balance model complexity and generalization ability. Step Six: Model Performance Evaluation; The model performance is systematically evaluated using multiple evaluation metrics, including the coefficient of determination R², root mean square error RMSE, mean absolute error MAE, and prediction interval coverage probability PICP.

2. The virtual water tank simulation method based on data augmentation and XGBoost according to claim 1, characterized in that, In step one, The scour test data includes: the evolution of scour depth under different conditions, the maximum depth at which scour equilibrium is reached, and the time required to reach scour equilibrium; the test uses a basic model made according to a similar scale relationship, and conducts the test using different flow velocity conditions, seabeds with different parameters, and different types of foundations; A database was established to collect scour test data, which included the following parameters: Median particle size d 50 The range is 0.1-5mm; Water depth hp, ranging from 10-40cm; Flow velocity V, ranging from 10 to 40 cm / s; The model diameter R ranges from 3 to 40 cm. Time t records the erosion and development at different points in time; scour depth This corresponds to the absolute value of the scour depth under different conditions.

3. The virtual water tank simulation method based on data augmentation and XGBoost according to claim 1, characterized in that, The specific process of step two is as follows: The collected scouring test data were preprocessed. First, the data was cleaned to remove outliers and missing values ​​with measurement errors exceeding 5% to ensure data quality. The data was then standardized, with the scour depth being standardized as a ratio relative to the maximum scour depth, as follows: ， Among them, flushing process parameters This represents the normalized scour depth at time t. This represents the actual scour depth at time t. Indicates the maximum scour depth and scour progress indicators. It is used to characterize the percentage of the distance from the maximum scour depth at time t, and when it reaches 100%, it means that the maximum scour has been reached; Finally, the data is split into training and test sets at a ratio of 80% / 20% to ensure that the two sets have similar data distribution characteristics. The training set is used for model training and parameter optimization, while the test set is used to evaluate the model's generalization ability and prediction performance.

4. The virtual water tank simulation method based on data augmentation and XGBoost according to claim 1, characterized in that, The specific process of step three is as follows: The data is augmented using an improved Wasserstein generative adversarial network (WGAN-GP) technique; the Wasserstein generative adversarial network includes a generator. and discriminator ; Introduce conditional parameter vector c=[d 50 [,hp,V,R] serves as the scour mechanical constraint condition; Then generator The constraint expression is: ， in, For noise, As constraints, This represents the scour depth data output by the generator. To scour mechanical constraint functions, To constrain the allowable threshold, both factors work together to ensure the physical reasonableness of the generated data; The constraint function specifically includes three key constraints: ① Dimensionless scour depth constraint: ,in Represents the absolute value of the scour depth, model diameter The feature dimensions of the basic structure, For flow rate, It is the acceleration due to gravity. Because of the water depth, This represents a functional relationship established based on hydraulic theory; this constraint ensures a reasonable match between scour depth and hydraulic conditions. ② The constraint of monotonicity in development: ,in This represents the normalized scour depth at time t. Its first-order partial derivative with respect to time represents the rate of erosion development. The second-order partial derivative represents the change in acceleration, ensuring the physical rationality of the scouring process in the time dimension and guaranteeing that the scouring depth increases monotonically with time and the growth rate decreases. ③ Critical shear stress constraint: For bed shear stress, The critical shear stress is used to ensure that the generated data meets the scour initiation conditions. For fluid density; Bed surface friction coefficient; The discriminator is corrected accordingly. The corrected discriminator Output probability value While determining the authenticity, it also verifies the degree to which physical constraints are satisfied. The scour depth data is input to the decision-maker; A comprehensive expansion of the timescale erosion process: introducing a time-weighted loss function: ， in For the sample index, from 1 to n, The total number of samples, Here, T is the time index, and T is the total length of the time series. For time weighting factors, where The time decay parameter emphasizes the importance of the initial scour development. Here is the loss metric function used to calculate the first... One sample in Real-time normalized erosion depth With generated value The differences between them; The generation process employs a conditional time series generation strategy: ,in For noise, As a constraint, ensure that the generated flushing process sequence at each time node All satisfy: S'(t) i+1 )≥S'(t i And S'( →100%; Simultaneously, establish a flushing process filtering mechanism: retain those meeting the requirements. The generated samples, where =5%, =3% represents the allowable deviation at 10 minutes and 30 minutes, respectively. and These are the reference normalized flush depths at 10 minutes and 30 minutes, respectively. The training process of GAN adopts an alternating optimization strategy. In each iteration, the generator G is fixed first, and the discriminator D parameters are updated 5 times. Then, the discriminator D is fixed, and the generator G parameters are updated once.

5. The virtual water tank simulation method based on data augmentation and XGBoost according to claim 1, characterized in that, Step four involves the following steps: An extreme gradient boosting (XGBoost) algorithm is used to construct a scour depth prediction model. Based on the traditional gradient boosting decision tree, regularization technology is introduced to construct a feature system oriented towards scour mechanics, and the physical mechanism of scour is deeply integrated into the XGBoost algorithm framework. The prediction function of the XGBoost model is as follows: ， in, This represents the predicted scour depth value for the i-th sample. The function represents the k-th decision tree, where K is the total number of decision trees. The input feature vector [d] for the i-th sample 50 ,hp,V,R,t]; The prediction function integrates multiple decision trees to evaluate the key parameters [d] of the scour process. 50 Nonlinear combinatorial learning of [hp,V,R,t] enables accurate prediction of the temporal evolution of scour depth; each decision tree During training, the system automatically learns the optimal segmentation point and weight combination between different scouring parameters. The objective function of the XGBoost model is: ， The objective function consists of two parts: the first part Loss function; Part Two This is a regularization term used to control model complexity and prevent overfitting. Let be the true scour depth value of the i-th training sample; Among them, the regularization term is: ， The regular expression consists of two parts: Control the number of leaf nodes, where Let be the number of leaf nodes in the k-th tree. The penalty parameter for the number of leaf nodes; For L2 regularization terms, where The leaf node weight vector, This is the L2 regularization parameter, which prevents leaf node weights from becoming too large. The algorithm training process is as follows: First, the mean of the training samples is used as the initial prediction. The initial values ​​are obtained by minimizing the loss function, providing a starting point for subsequent iterations; Then, the iterative training phase begins, with each iteration focusing on optimizing the objective function: calculating the residual gradient between the current model's predicted value and the true value, based on the first and second partial derivatives of the loss function with respect to the predicted value; constructing a new decision tree to fit the residual gradient, with the split nodes of the decision tree prioritizing the parameter combinations that have the most significant impact on the scour depth; calculating the optimal weight for each leaf node, which is determined by the residual gradient, the second gradient, and the regularization parameter, ensuring that the model complexity is controlled while fitting the residual; applying the learning rate to scale the contribution of the new tree, and then updating the overall model's prediction function; Repeat the iterative process until the preset upper limit of the number of trees is reached or the verification error no longer decreases significantly. At this point, the model is considered to have converged to the optimal state.

6. The virtual water tank simulation method based on data augmentation and XGBoost according to claim 1, characterized in that, Step five involves the following steps: For the nine key hyperparameters of the XGBoost algorithm, an automatic optimization strategy combining Bayesian optimization and 5-fold cross-validation is adopted. The search space for the nine hyperparameters is set as follows: gamma parameter [0.1, 5] logarithmic distribution, learning rate [0.01, 0.3] uniform distribution, maximum depth [3, 10] integer distribution, number of trees [100, 1500] integer distribution, subsampling rate [0.6, 1.0] uniform distribution, column sampling rate [0.6, 1.0] uniform distribution, minimum subweight [1, 10] integer distribution, alpha parameter [0, 1.0] uniform distribution, and lambda parameter [0, 2.0] uniform distribution. Bayesian optimization employs the RBF kernel selection method, automatically learning kernel parameters through maximum likelihood estimation. Performance evaluation utilizes 5-fold cross-validation, with the root mean square error (RMSE) as the objective function. The average of five validation results is used for each parameter configuration. The optimization process consists of two phases: the initialization phase uses Latin hypercube sampling to generate 10 initial parameter configurations, ensuring uniform coverage of the search space; the iterative optimization phase executes 50 iterations, updating the Gaussian process model based on historical data in each iteration to obtain new parameter configurations, and then performing 5-fold cross-validation evaluation; the entire optimization process is repeated three times using different random seeds, selecting the optimal configuration through 10-fold cross-validation comparison, while introducing an early stopping mechanism to prevent overfitting; finally, the optimal parameter configuration for predicting scour depth is determined.

7. The virtual water tank simulation method based on data augmentation and XGBoost according to claim 1, characterized in that, The optimal parameter configuration for predicting scour depth is: gamma value 0.01807, learning rate 0.29361, maximum depth 7, number of trees 866, subsampling rate 0.8, column sampling rate 0.7, minimum subweight 3, alpha parameter 0.1, and lambda parameter 1.

0.

8. The virtual water tank simulation method based on data augmentation and XGBoost according to claim 1, characterized in that, The specific process of step six is: to systematically evaluate the performance of each model through multiple complementary statistical indicators; Evaluation metrics include the coefficient of determination (R²), root mean square error (RMSE), mean absolute error (MAE), and probability of coverage of the predicted interval. The coefficient of determination is calculated as follows: It measures the model's ability to explain data variation; The root mean square error is calculated as follows: It measures the absolute deviation between the predicted value and the actual value; The mean absolute error is calculated as follows: It provides an intuitive measure of prediction error; It is the measured scour depth value of the i-th test sample. It is the predicted scour depth of the i-th test sample; It is the average of the measured scour depths of all test samples; The predicted interval coverage probability is calculated to determine the proportion of actual values ​​falling within the predicted interval, thus assessing the reliability of the model.

Images