A method for predicting the seismic performance of a concrete shear wall based on an XGBoost algorithm technique

Abstract

This invention discloses a method for predicting the seismic performance of concrete shear walls based on the XGBoost algorithm. The method includes: standardizing and preprocessing a database of concrete shear wall seismic performance and selecting features to construct training and testing sets; establishing an interpretable prediction model using machine learning, and obtaining the optimal model through hyperparameter tuning; using the prediction model to derive the predicted values ​​of yield strength, its multiples, peak values, and limit points of samples in the testing set; inputting the predicted skeleton curve points into the idealized bilinear model of Ozcebe, and calculating the complete loading-unloading path of the shear wall using the formula; evaluating the prediction performance through various indicators, and verifying it through OpenSees numerical simulation. This invention solves for the first time the problems of large prediction dispersion and unclear physical mechanisms in machine learning; the prediction model can accurately reflect nonlinear mechanisms, significantly improving the credibility and interpretability of data-driven methods; and through numerical simulation verification, it achieves the integration of data-driven and structural mechanics.

Description

Technical Field

[0001] This invention belongs to the field of seismic analysis of shear wall structural components, specifically relating to a method for predicting the seismic performance of concrete shear walls based on the XGBoost algorithm. Background Technology

[0002] While traditional quasi-static tests can realistically reflect the hysteretic response of concrete shear walls under seismic loads, they are costly, time-consuming, require cumbersome specimen preparation, and involve complex parameter control, making it difficult to meet the high efficiency and scalability requirements of modern high-rise buildings for shear wall performance evaluation.

[0003] In recent years, machine learning methods have emerged as a new approach to predicting the seismic performance of structures, offering high fitting accuracy and modeling efficiency. However, most current research is based on simulation-generated data and suffers from insufficient explanation of physical mechanisms, inadequate visualization, and insufficient construction of restoring force paths, making it difficult to directly apply the prediction results to actual design evaluation and seismic performance optimization. Machine learning technology can efficiently utilize the vast amounts of experimental data accumulated over the past decades from different regions and periods, directly mining the mapping relationships within the data to establish accurate and stable prediction models. Although the application of the aforementioned machine learning methods has demonstrated their unique advantages in predicting the basic performance of concrete structures, most current research consists of "black box" models—they provide accurate predictions based on input but cannot provide explanations for these predictions, which reduces the credibility and application scope of machine learning models.

[0004] Numerical simulation methods such as OpenSees have good modeling capabilities and can visualize and analyze the response process of structures under seismic loading. However, they still rely on complex material constitutive models and detailed structural definitions. Their results can be used to compare and verify the rationality and accuracy of machine learning methods. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a method for predicting the seismic performance of concrete shear walls based on the XGBoost algorithm.

[0006] The objective of this invention is achieved through the following technical solution: a method for predicting the seismic performance of concrete shear walls based on the XGBoost algorithm, comprising the following steps:

[0007] Step (1): Collect the seismic performance database of concrete shear walls, and perform preliminary screening of collected geometric parameters, material parameters, and load parameters to construct the input dataset;

[0008] Step (2): Normalize the database and preprocess it, then filter the features to obtain the input parameters of the model, and confirm that the key prediction points of the skeleton curve are the predicted values ​​of yield and its multiples, peak points, and limit points.

[0009] Step (3): Use machine learning XGBoost to build an interpretable prediction model. The input of the prediction model is the geometric parameters, material parameters and load of the concrete shear wall, and the output is the predicted values ​​of yield and its multiple points, peak points and limit points.

[0010] Step (4): Set the hyperparameter range for the prediction model, and then use the training set to iteratively train the prediction model. During the training process, compare and analyze the predicted values, and automatically adjust the hyperparameters of the prediction model based on the results of the comparison and analysis to obtain the best prediction model. Divide the dataset into training set and test set proportionally, input the samples in the test set into the trained prediction model, and obtain the optimal prediction result, i.e., the horizontal and vertical coordinates of the key prediction points. Use the evaluation index R 2 Perform a precision analysis on it.

[0011] Step (5): Connect the key prediction points obtained by machine learning to obtain the skeleton curve, and quantitatively evaluate the displacement bearing capacity index and the morphological fitting accuracy of the skeleton curve output. Analyze the rationality of the skeleton curve prediction. The skeleton curve fitting accuracy evaluation index is RMNSE.

[0012] Step (6): Substitute the predicted key points of the skeleton curve into the Ozcebe bilinear restoring force model to obtain the hysteresis curve.

[0013] Compared with the prior art, the significant advantages of this invention are:

[0014] This invention integrates data-driven and physical modeling to efficiently and accurately predict the hysteretic performance and seismic resistance of steel-concrete composite shear walls. It combines data-driven modeling techniques based on large-scale experimental databases, the high-precision prediction capabilities of the XGBoost machine learning model, the hysteretic path representation capabilities of the Ozcebe bilinear restoring force model, and the mechanical verification capabilities of OpenSees finite element analysis, making the prediction of shear wall seismic performance both physically reliable and engineering-practical. This method not only enables efficient prediction of key performance points (such as yield point, peak value, and ultimate limit point), but also reconstructs the skeleton curve and hysteresis curve, realizing a complete chain from data to performance evaluation. Attached Figure Description

[0015] Figure 1 This is a cross-sectional view of the steel plate and concrete composite shear wall in an embodiment of the present invention.

[0016] Figure 2 This is a schematic diagram of the output parameters in an embodiment of the present invention.

[0017] Figure 3 This is a schematic diagram of XGBoost in an embodiment of the present invention.

[0018] Figure 4 This is a learning curve diagram of the XGBoost training set and validation set in an embodiment of the present invention.

[0019] Figure 5 This is a scatter plot of XGBoost predictions in an embodiment of the present invention.

[0020] Figure 6 This is a diagram of the shear force response hysteresis model in an embodiment of the present invention.

[0021] Figure 7 This is an example diagram comparing the predicted hysteresis curve with the experimental results in an embodiment of the present invention.

[0022] Figure 8 This is a schematic diagram of the hysteresis area selected in an embodiment of the present invention.

[0023] Figure 9 This is a schematic diagram of the OpenSees shear wall analysis model in an embodiment of the present invention.

[0024] Figure 10 This is a comparison chart of machine learning prediction results and numerical simulation results in an embodiment of the present invention.

[0025] Figure 11 This is a flowchart illustrating the implementation of the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0027] The objective of this invention is achieved through the following technical solution: a method for predicting the seismic performance of concrete shear walls based on the XGBoost algorithm, comprising the following steps:

[0028] Step 1: Collect a database of seismic performance of steel plate concrete composite shear walls, with a data volume of at least 300 data points, covering various combinations of geometric, parametric, and load parameters. Perform preliminary screening of the collected geometric, material, and load parameters to construct the input dataset.

[0029] Step 2: Perform Z-score normalization preprocessing on the database. For a certain feature data x, the Z-score normalized data value... The calculation formula is:

[0030]

[0031] In the formula, The mean of the features in the dataset. Let the standard deviation of the dataset features be the standard deviation. After substituting the data into the standardization formula, we get... This is the standardized value.

[0032] Next, feature selection is used to determine the model's input parameters. In feature selection, Lasso regression and RFE (Regression-Free Evaluator) techniques are combined. The objective of Lasso regression is... That is, to find the optimal variable that minimizes the objective function. Where y i Let β be the i-th target variable. j Let be the j-th regression coefficient, the first part This is the sum of squared residuals, which measures the error between predicted and observed values, similar to ordinary linear regression. Part Two For L1 regularization terms, Let the parameter vector be β = (β1, β2, ..., β... p The L1 norm expression of ).

[0033] The core steps of RFE (Regression Forecasting) technology are as follows: First, train the model using the current feature set; in this study, Lasso regression is used as the base model. Second, calculate the importance score w for each feature based on the model's coefficients or feature importance scores. j Next, features that contribute the least to the model are removed, i.e. Find the smallest feature; finally, recursively repeat the above steps until the number of remaining features reaches a preset value. j The calculation formula is:

[0034]

[0035] In the formula, L is the loss function of the model. Let j be the j-th feature.

[0036] After filtering the input parameters, the key prediction points of the skeleton curve are confirmed to be yield points and their multiples, peak points, and limit point prediction values.

[0037] Step 3: Utilize machine learning XGBoost to build an interpretable prediction model. The input to the prediction model is the geometric parameters, material parameters, and load of the concrete shear wall. The output is the predicted values ​​of yield and its multiples, peak points, and limit points. The mathematical model of XGBoost can be regarded as an additive model composed of K regression decision trees, as shown in the following formula.

[0038]

[0039] Where is the predicted value of the i-th target; n is the number of trees; f k Let F be a function in the function space F; F is the set of all possible decision trees.

[0040] The objective function of the XGBoost regression model consists of two parts: a loss function and a regularization term. The loss function represents the difference between the predicted and true values, indicating the degree of model fit. The regularization term controls the model's complexity, prevents overfitting, and ensures both accuracy and generalization ability. The objective function formula and the objective function of the model after iterative training to step t are shown below:

[0041]

[0042]

[0043] in, The objective function is... The loss function; For regularization terms, For the prediction model of the first t-1 rounds, This refers to the prediction model newly added in round t.

[0044] Step 4: Set the hyperparameter range for the prediction model, then iteratively train the prediction model using the training set. During training, compare and analyze the predicted values, and automatically adjust the hyperparameters of the prediction model based on the results of this analysis to obtain the best prediction model. Divide the dataset into training and test sets proportionally, and input the samples from the test set into the trained prediction model to obtain the optimal prediction results, i.e., the x and y coordinates of the key prediction points. Utilize the evaluation metric R... 2 The precision analysis of it is expressed as follows:

[0045]

[0046] in, The actual value is the corresponding model prediction value. , The average value of the dataset. It is the sum of squared residuals. This is the total sum of squares.

[0047] Step 5: Connect the key prediction points obtained from machine learning to obtain the skeleton curve. Quantitatively evaluate the displacement bearing capacity indicators and the shape fitting accuracy of the skeleton curve output, and analyze the rationality of the skeleton curve prediction. The skeleton curve fitting accuracy evaluation index is RMNSE, and its calculation formula is:

[0048]

[0049] Among them, ypred,i It is the value of the predicted curve, y true,i is the actual value of the curve, and n is the number of data points.

[0050] Step Six: Input the predicted skeleton curve key points into the Ozcebe bilinear restoring force model to obtain the hysteresis curve. If it exceeds in one direction... At least once, and in the quadrant where unloading occurred, the yield load had not previously exceeded. Then the unloading occurs along a straight line upwards to the zero-load axis. (Greater than) The slope is shown in the following formula (e.g., AB, KL segment):

[0051]

[0052] in, The slope from the origin to the cracking point. This refers to the distance from the yield point in one quadrant to the cracking point in another quadrant. Less than The slope is taken as k1 (e.g., segment CD). When the yield load is exceeded... When the unloading stiffness is higher than the cracking load (such as segments GH, MN, RS, VW, and YZ), the calculation is as follows:

[0053]

[0054] The unloading stiffness below the cracking load (such as segments HI, NO, ST, and WX) is calculated as follows:

[0055]

[0056] If the loading direction exceeds Then reload to When, along the straight line passing through point ( , (e.g., DE, OP, TU); then load more than The reloading will travel upwards along a straight line to the main curve, passing through point ( , (e.g., EF, PQ, UY); Finally, outside the intersection of the reload branch and the primary curve, the loading follows the primary curve (e.g., FG, QR). If unloading is completed before reaching the zero-load axis, reloading in the same quadrant will follow a straight line pointing to the immediately preceding peak point (e.g., XV). , , , The definition is shown in the following formula:

[0057]

[0058]

[0059]

[0060]

[0061]

[0062] in, It is the axial compressive force. Where n is the nominal axial load capacity, and n is the number of counter tracking cycles at constant displacement. , , It is a mathematical coefficient. This represents the peak displacement. For the peak load, For the maximum displacement, This represents the principal tangential stress corresponding to the maximum displacement.

[0063] Next, the model performance is quantitatively evaluated and its rationality is analyzed by the accuracy of curve shape fitting, hysteresis area error and equivalent viscous damping coefficient error.

[0064] Step 7: Build a numerical model of the shear wall based on OpenSees for comparison and verification. Judge the rationality of the numerical model based on the ultimate bearing capacity, ductility coefficient, and equivalent viscous damping coefficient.

[0065] (7.1) Divide the fiber model section into partitions, and then further discretize each region into several fiber elements. Select fiber elements based on the stiffness method to subdivide the structural members into multiple segments.

[0066] (7.2) The Concrete02 constitutive model was used for concrete material, and the Steel02 constitutive model was used to calculate the material properties of steel bars;

[0067] (7.3) The simulated hysteresis curve was obtained by performing Pushover analysis using the displacement control mode;

[0068] (7.4) The mechanical rationality and practicality of the prediction model are finally verified by comparing the error of ultimate bearing capacity, the error of displacement ductility coefficient and the error of equivalent viscous damping with the prediction results.

[0069] The present invention will now be further described in conjunction with the specification and accompanying drawings.

[0070] Furthermore, combined Figure 11In step one, based on a systematic literature review method, 299 sets of test data for steel-concrete composite shear walls from multiple countries and regions around the world, including China, the United States, Canada, Australia, Germany, and the United Kingdom, were collected and organized to construct a relatively complete test database. The data sources mainly include academic papers published in authoritative domestic and international journals, standard experimental reports, and typical engineering cases, covering research results from different geographical regions and design code systems, and possessing significant regional representativeness and engineering reference value.

[0071] like Figure 1 As shown, in addition to manually screening the features to exclude obviously repetitive and substitutable parameters, this paper also refers to the calculation formulas for the eccentric compression bearing capacity of the positive section of steel-concrete composite shear walls and steel plate concrete shear walls in the specifications. Finally, for a specific steel plate concrete composite shear wall, it is defined as being represented by 21 different features, which are usually divided into three categories: geometric dimensions, material parameters and loads.

[0072] like Figure 2 As shown, different feature points form skeleton curves, and the output is divided into two main categories: displacement and bearing capacity.

[0073] Furthermore, in step two, the input parameters are Z-score standardized. After the process, the offset of each feature has been effectively eliminated, and the standardized mean of all parameters is between -0.01 and +0.02, and the standard deviation is between 1.00 and 1.09.

[0074] Next, feature selection is performed. First, the Lasso regression model is initialized with a regularization parameter α=0.01. Second, the RFE model is initialized with Lasso regression as the base model and the number of target features is set. Finally, the input feature matrix and the target variable are fitted to obtain the feature matrix after feature selection. After feature selection, the following 15 input parameters are confirmed, forming the optimal feature subset that combines statistical independence and engineering interpretability of geometric parameters, material properties and loads.

[0075] like Figure 3As shown, further, in step three, extreme gradient boosting (XGBoost) is an efficient and scalable gradient boosting tree algorithm proposed by Chen et al. This algorithm uses decision trees as base learners and predicts the class or regression value of samples by integrating multiple base learners. Unlike random forests, XGBoost uses gradient boosting for training; each decision tree is built based on the prediction residuals of the previous tree, gradually reducing the prediction error through iterative optimization. In addition, XGBoost introduces L1 and L2 regularization terms to constrain the weight allocation of leaf nodes in the decision tree, effectively preventing model overfitting. XGBoost significantly improves training speed through parallel computation and optimization algorithms, making it particularly suitable for large-scale datasets and highly flexible, with a certain tolerance for outliers, but it consumes significant computational resources.

[0076] Furthermore, in step four, the hyperparameters of the algorithm are set as follows: the number of weak learners ranges from [100, 1000], the maximum tree depth ranges from [3, 10], the learning rate ranges from [0.001, 0.3], the minimum sample weight of leaf nodes ranges from [1, 10], the node splitting criterion ranges from [0, 0.5], the random sampling ratio ranges from [0.6, 1], and the feature sampling ratio for each tree ranges from [0.6, 1].

[0077] To automate the search for optimal hyperparameter combinations, a tree-based Bayesian optimization method is employed to quickly locate the optimal parameter combination. Specifically, the Optuna library in Python is used to optimize the hyperparameters of machine learning algorithms. Compared to traditional grid search and random search, it dynamically adjusts the search space by constructing a probabilistic model of the objective function, thus avoiding exploration of invalid regions and effectively reducing the risk of overfitting. This method not only significantly shortens training time while maintaining model performance but also better adapts to high-dimensional hyperparameter spaces. In the Optuna optimization framework, the three core concepts are the objective function, trial, and study. This study uses Optuna's TPESampler as a sampler to ensure experimental reproducibility. TPESampler, based on a tree-structured Parzen Estimator, intelligently samples hyperparameter combinations by modeling the probability distribution of the objective function. The optimization objective is set to maximize the objective function value (direction='maximize'), and the optimization process is initiated using the study.optimize method. The number of trials (n_trials) was set to 100 to strike a balance between optimization performance and computational cost. Finally, the model was retrained using the optimal hyperparameters obtained from the optimization, and its performance was evaluated.

[0078] Furthermore, the model performance was evaluated using a ten-fold cross-validation method. The dataset was divided into ten subsets, with nine subsets used for training and one subset for validation in each iteration. Finally, samples from the test set were input into the trained prediction model to obtain the optimal prediction results, i.e., the x and y coordinates of the key prediction points. The learning curves of the training and validation sets were evaluated using the R² score, represented by the y-axis of the learning curve graph. The red curve consistently remained close to 1.0, exhibiting minimal fluctuation, indicating that the model fit the training data very well. Furthermore, the fitting ability remained stable as the training set size increased, demonstrating effective learning of data features during training and maintaining high prediction accuracy on the validation set, exhibiting good robustness. Figure 4 As shown.

[0079] Furthermore, the prediction model outputs the value of a single point on the skeleton curve to determine the coefficients. As a standard for evaluating the prediction results of the model. Key displacement prediction The result is =0.97, =0.93, =0.97, =0.92, =0.68, =0.96, =0.87, =0.97, =0.94, =0.91,2 =0.95, 1.5 =0.92, 0.5 =0.89, 0.5 =0.88, 1.5 =0.92, 2 =0.96, which shows that most R values ​​are 0.96. 2 The values ​​are high and stable, indicating a very good fit. Key load prediction. The result is =0.93, =0.93, =0.98, =0.98, =0.95, =0.97, =0.95, =0.99, =0.99, =0.94, =0.98, =0.97, =0.96, =0.97, =0.98, =0.98, indicating that all XGBoost predictions are at or above 0.93. It excels at capturing the strong logical relationship between material strength, dimensional parameters, axial compression ratio, and the ultimate bearing capacity of a structure. Therefore, it performs exceptionally well with physically clear and decisive indicators such as maximum load, peak point, and yield point. A scatter plot is output using a positive peak load as an example. =0.97, as Figure 5 As shown.

[0080] Further, in step five, the coordinates of the predicted points are connected to obtain the skeleton curve. The skeleton curve is then compared and analyzed based on the mechanical characteristics of displacement and the mechanical characteristics of bearing capacity. The mechanical characteristics of displacement are analyzed using the displacement ductility coefficient and the ultimate elastic-plastic displacement angle. The displacement ductility coefficient is the ultimate displacement Δ... u For yield displacement Δ y The ratio of the ultimate elastic-plastic displacement angle to the ultimate displacement Δ. u and wall height h wThe mechanical characteristics of bearing capacity are analyzed using the bearing capacity reserve coefficient and initial stiffness. The bearing capacity reserve coefficient is the maximum load divided by the yield load minus 1, and the initial stiffness is the ratio of the yield load to the yield displacement. The results show that R² exceeds 0.9 in multiple indicators such as initial stiffness and ultimate displacement angle, and reaches 0.71 in ductility coefficient, demonstrating excellent predictive performance.

[0081] Furthermore, the NRMSE was used to quantitatively evaluate the fitting accuracy of the skeleton curve morphology. Since the predicted minimum and maximum displacements differ from the actual displacements, the abscissas of the two curves are not completely consistent. Therefore, the abscissas of the two curves need to be aligned first. Thus, interpolation was first performed on the other curve within a shorter abscissa range, divided into 100 equal parts. Then, the corresponding ordinates were found, and the NRMSE of the overlapping x-region was calculated. Finally, the NRMSE values ​​of all valid rows were summed and averaged. The final NRMSE of XGBoost was 0.038, further validating the rationality and accuracy of XGBoost's predictions.

[0082] Furthermore, in step six, the Ozcebe bilinear restoring force model diagram is as follows: Figure 6 As shown,

[0083] Furthermore, when incorporating the predicted key points, not only were the key feature points of the skeleton curve predicted using the aforementioned machine learning model, including the yield point, peak point, failure point, and 85% peak load drop point, but load-displacement data points at different displacement multiples were also extracted. Specifically, load values ​​corresponding to 0.5, 1.5, 2, 2.5, 3, 3.5, and 4 times the yield displacement were supplemented to more accurately reconstruct the hysteresis curve. An example comparison of the predicted and experimental hysteresis curves is shown in the figure below. Figure 7 As shown.

[0084] Furthermore, the model performance is quantitatively evaluated and its rationality analyzed by comparing the hysteresis area error with the equivalent viscous damping coefficient error. The maximum bearing capacity cycle and the cycle before failure are selected as the hysteresis area evaluation objects. The first attainment of the maximum bearing capacity, i.e., the ultimate limit state, reflects the performance of the limit state, while the last complete cycle before failure reflects the structural collapse warning state. Each hysteresis loop is calculated in sub-loops, with one loop defined as the point from the positive intersection of the vertical axis to the point where the loop ends and intersects the vertical axis. Figure 8 As shown, the circle corresponding to the maximum load is red, intersecting the y-axis at the green point; the circle before failure is orange, intersecting the y-axis at the yellow point. Then, Simpson's integral method is used to calculate the area of ​​each circle. Simpson's integral method is a numerical integration method used to approximate the definite integral of a function. For a function f(x) on the interval [a, b], the approximate integral value of Simpson's formula is:

[0085]

[0086] In the formula, a and b are the endpoints of the interval. The midpoint of the interval is given by a coefficient ratio of 1:4:1, with each of the two endpoints having a weight of 1 and the midpoint having a weight of 4.

[0087] The calculation results show that the maximum bearing capacity area error is 3.2%, the area error before failure is 4.8%, and the overall error is 4.0%, which is excellent. This indicates that XGBoost can accurately fit the stiffness degradation, strength evolution, and energy dissipation process of the component throughout the hysteresis process, especially demonstrating outstanding performance in path continuity in the pinch-out section and large deformation region. It is suitable for high-precision nonlinear hysteresis behavior modeling. The error of the equivalent viscous damping coefficient in the maximum bearing capacity circle is 4.5%, the error of the equivalent viscous damping coefficient before failure is 6.1%, and the overall error is 5.3%, indicating that XGBoost prediction is very good and can effectively reproduce the energy dissipation characteristics of shear walls under strongly nonlinear conditions.

[0088] Furthermore, in step seven, a numerical model of the shear wall is established based on OpenSees for comparative verification. Numerical simulation is performed on the composite shear wall with built-in steel plates and reinforced concrete, using an analysis model combining a cyclic softening model and a fiber model. The analysis unit is as follows: Figure 9 As shown. The shear wall model was divided into 8 segments along its height and 3 segments along its width for numerical simulation. The vertical load was divided into 4 equal parts and applied to the top of the model. Fiber elements based on the stiffness method were used in the analysis. The concrete material of the edge members was selected using the Concrete02 constitutive model, while the reinforcing bars and I-beams used the Steel02 constitutive model. The shear wall panels used quadrilateral elements J2 Plasticity Material, with the mesh size based on the plastic hinge length of the wall, using CSMM material. The internal steel plates used quadrilateral elements, with the material being Plane Stress Concrete Materials. Pushover analysis was performed using the displacement control mode to obtain the simulated hysteresis curves. The machine learning-predicted hysteresis curves were compared with the numerical simulation curves, as shown in the figure. Figure 10 As shown.

[0089] Furthermore, the average error in ultimate bearing capacity was 5.8%, the average error in ultimate displacement was 33.9 mm, the average error in ductility coefficient was 10.6%, the average linear elastic-plastic displacement angle was 0, the plastic zone rotation angle error ratio was 3.6%, the error in the hysteresis loop area of ​​the maximum bearing capacity circle was 16.5%, and the error in the equivalent viscous damping coefficient of the maximum bearing capacity circle was 17.3%. These are all within acceptable ranges, but worse than the hysteresis curve predicted by machine learning. By cross-comparing the two methods, not only can the response characteristics of each model at different stages be revealed, but the bias caused by a single modeling path can also be effectively avoided, thus more systematically and comprehensively demonstrating the effectiveness and reliability of machine learning methods in predicting the hysteretic performance of steel plate concrete composite shear walls.

[0090] In summary, this demonstrates that the method proposed in this application is effective.

[0091] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for predicting the seismic performance of a concrete shear wall based on the XGBoost algorithm technique, characterized in that, Includes the following steps: Step (1): Collect the seismic performance database of concrete shear walls and perform preliminary screening of the collected parameters to construct the input dataset; Step (2): Perform normalization preprocessing and feature filtering on the database to obtain the input parameters of the model, and confirm the category of key prediction points of the skeleton curve; Step (3): Use the XGBoost algorithm of machine learning to establish an interpretable prediction model. The input of the prediction model is the geometric parameters, material parameters and load of the concrete shear wall, and the output is the predicted values ​​of yield and its multiple points, peak points and limit points. Step (4): Set the hyperparameter range for the prediction model, and then use the training set to iteratively train the prediction model. During the training process, compare and analyze the predicted values, and automatically adjust the hyperparameters of the prediction model according to the comparison and analysis results to obtain the best prediction model. Divide the dataset into training set and test set according to the proportion, input the samples in the test set into the trained prediction model, and obtain the optimal prediction result, i.e., the horizontal and vertical coordinates of the key prediction points. Step (5): Connect the key prediction points obtained by the machine learning XGBoost algorithm to obtain the skeleton curve, and quantitatively evaluate the single point value, displacement bearing capacity index, and skeleton curve shape fitting accuracy of the skeleton curve output, and analyze the rationality of the skeleton curve prediction. Step (6): Substitute the predicted key points of the skeleton curve into the Ozcebe bilinear restoring force model to obtain the hysteresis curve.

2. The method of claim 1, wherein, Step (1) specifically involves collecting a database of seismic performance of concrete shear walls. The database contains three types of parameters: geometric parameters, material parameters, and load parameters. The specific parameters are: wall width, wall height, wall thickness, edge column length, edge column width, yield strength of longitudinal reinforcement in edge columns, yield strength of distributed reinforcement in the wall, yield strength of stirrups, concrete compressive strength, vertical reinforcement ratio of columns, horizontal reinforcement ratio of columns, vertical reinforcement ratio of walls, horizontal reinforcement ratio of walls, yield stress of boundary steel profiles, axial load, test axial compression ratio, and design axial compression ratio. Then, the distribution functions are fitted to these parameters, which are the Lorentz function, Gaussian function, and Lognormal function, which are three typical distribution functions.

3. The method of claim 1, wherein: The normalization preprocessing method mentioned above is the Z-score normalization method.

4. The method of claim 1, wherein: The feature selection method employed is the Lasso regression model, and the Recursive Feature Elimination (RFE) technique is used for data extraction.

5. The method of claim 1, wherein, Step (2) specifically includes the following sub-steps: (2.1) The Z-score standardization method is used to convert the dataset into a standard normal distribution with a mean of 0 and a standard deviation of 1, thereby eliminating the influence of the data's dimensions; (2.2) Initialize the Lasso regression model and set the regularization parameter α; (2.3) Initialize the RFE model, using Lasso regression as the base model, and set the number of target features; (2.4) Finally, the input feature matrix and the target variable are fitted to obtain the feature matrix after feature selection, which is the final input model training feature.

6. The method of claim 1, wherein, Step (4) specifically includes the following sub-steps: (4.1) Set the search space for the seven hyperparameters of the XGBoost model: learning rate, maximum tree depth, number of weak learners, minimum sum of weights of leaf nodes, node splitting criteria, proportion of random sampling, and feature sampling proportion of each tree. (4.2) The optimization process adopts the Bayesian optimization method in the Optuna library. The optimal hyperparameter combination is searched within the above parameter range through multiple trials and iterations. The R² score is used as the evaluation index for ten-fold cross-validation training. (4.3) The optimization algorithm automatically selects the hyperparameter combination that maximizes the cross-validation R² score. After determining the optimal hyperparameter combination, the trained model is used to predict the test set and outputs the load and displacement values ​​corresponding to each prediction point.

7. The method according to claim 1, characterized in that: The overall fitting accuracy evaluation index of the skeleton curve is the standardized root mean square error (NRMSE).

8. The method according to claim 1, characterized in that: The single-point value output by the skeleton curve and the evaluation index of displacement bearing capacity are determined by the coefficient of determination (R²).

Images