A corrugated steel plate shear wall mechanical behavior prediction method and intelligent interaction system

Abstract

A corrugated steel plate shear wall mechanical behavior prediction method and an intelligent interaction system. Based on the verified finite element model of the corrugated steel plate shear wall, parameterized analysis is carried out, the mechanical response data under different parameter combinations are obtained and written into the mechanical behavior database; the mechanical response data are quantified to obtain the mechanical behavior characteristics; a prediction model is constructed with the key design parameters as input and the mechanical behavior characteristics as output, the SHAP contribution degree of each key design parameter to the prediction output is calculated to obtain the explanatory analysis result; the prediction model is deployed in the intelligent interaction system to realize interactive prediction of the optimal solution set. The present application reduces the modeling and solving workload in the engineering design stage by optimizing the prediction process, solves the problem of lack of confidence evaluation of single prediction result, provides reverse design and automatic report generation functions based on performance targets, supports scheme comparison and reliability design.

Description

Technical Field

[0001] This invention relates to the field of seismic resistance of building structures and structural engineering technology, specifically to a method for predicting the mechanical behavior of corrugated steel plate shear walls and an intelligent interactive system with reverse optimization and automatic report generation functions. Background Technology

[0002] Corrugated steel plate shear walls are widely used in multi-story and high-rise buildings and seismic-resistant structural systems due to their lightweight, good load-bearing and energy dissipation capabilities, and ease of construction and assembly. The load-bearing mechanism and failure mode of this type of shear wall are influenced by multiple factors, such as the yield strength of the steel, the thickness of the steel plate, the aspect ratio of the panel, the geometric parameters of the corrugations, and initial defects. Variations in these parameters can significantly affect the peak strength, energy dissipation capacity, and failure mode of the shear wall. Therefore, during the engineering design and scheme comparison stages, it is necessary to evaluate the structural mechanical behavior under different parameter combinations to support structural selection and reliability design.

[0003] In existing technologies, the acquisition of mechanical properties of corrugated steel plate shear walls mainly relies on indoor quasi-static tests and nonlinear finite element numerical analysis. Indoor tests can realistically reflect the stress process of components, but they suffer from problems such as long cycles, high costs, and limited sample sizes, making it difficult to cover the wide range of parameter combinations required in the engineering design stage. Nonlinear finite element analysis can, to some extent, replace tests for parametric studies, but it typically requires model building, mesh generation, setting of material constitutive and geometric defects, setting of loading regimes and contact boundaries, and nonlinear iteration and convergence control during the solution process. For scenarios involving multiple parameters and multiple combinations of solutions, establishing and solving each solution individually is not only labor-intensive but also highly dependent on the experience of the analysts, making it difficult to form a stable, reusable, and rapid evaluation capability.

[0004] Furthermore, existing analysis processes primarily rely on single-test / single-simulation results, lacking a mechanism to systematically consolidate multi-source simulation results into a database and enable interactive access for engineering projects. For engineering designers, in addition to obtaining prediction results, it is crucial to clarify the impact of each design parameter on peak intensity, energy dissipation capacity, and failure modes to facilitate parameter sensitivity analysis and optimization decisions. However, traditional numerical analysis or empirical formulas often struggle to simultaneously output interpretable parameter contribution information, resulting in insufficient traceability and decision support capabilities for the application of results.

[0005] Therefore, there is an urgent need for a method and system that can construct surrogate prediction models based on experimentally verified finite element parameterized data. This system should be able to directly predict the peak strength, energy dissipation capacity, and failure mode of corrugated steel plate shear walls without the need for additional indoor experiments or repeated nonlinear finite element iterative solutions for each set of parameters. It should also be able to output the contribution or importance ranking of key design parameters, thereby reducing the modeling and solving workload in the engineering design stage, improving the efficiency of scheme comparison and selection, and enhancing the interpretability of the results. Summary of the Invention

[0006] To address the problems existing in the prior art, the purpose of this invention is to provide a method and intelligent interactive system for predicting the mechanical behavior of corrugated steel plate shear walls, thereby reducing the workload of modeling and solving in the engineering design stage, solving the problem of lack of confidence assessment for single prediction results, and providing reverse design based on performance targets and automatic report generation functions to support scheme comparison and reliability design.

[0007] The technical solution adopted in this invention is: Firstly, a method for predicting the mechanical behavior of corrugated steel plate shear walls is provided, with the following steps: S1. Parametric analysis is performed on the finite element model of the corrugated steel plate shear wall, which has been verified by physical experiments. Different parameter combinations are formed by changing key design parameters and component configurations. The finite element simulation engine is used to automatically perform parametric modeling and solution of the finite element model corresponding to each parameter combination, obtain mechanical response data, and write it into the mechanical behavior database. The key design parameters include at least the yield strength f. y The steel plate thickness t, aspect ratio α, corrugation angle θ, wave amplitude a, wavelength q, and initial defect δ, wherein the component is configured as stiffening rib geometric parameters or boundary constraint features; S2. Perform feature quantization on the mechanical response data to obtain mechanical behavior features, which include at least peak intensity index, energy dissipation capacity index and failure mode label. S3. Construct a supervised learning prediction model with uncertainty estimation function, take the key design parameters as input, and output the predicted values ​​of mechanical behavior features and their prediction confidence intervals; and calculate the SHAP value or SHAP importance statistic of each key design parameter based on the prediction model. S4. Intelligent Interaction and Predictive Optimization: The predictive model is deployed to the intelligent interaction system, which executes the following sub-processes based on the parameters input by the user: (1) Positive prediction and reliability assessment: Physical consistency logic verification is introduced to check whether the prediction results conform to the monotonicity law of material mechanics; (2) Active learning feedback: When the prediction uncertainty exceeds the preset threshold or the physical consistency verification fails, the system automatically generates the corresponding finite element parameterized modeling script based on the current high-risk parameter combination, and calls the finite element simulation engine described in step S1 to perform automatic supplementary calculation, so as to realize the fully automatic closed-loop update of data and the real-time incremental update of the mechanical behavior database. (3) Reverse optimization design: Discretized manufacturing constraints are introduced, and SHAP heuristics are used to search for Pareto optimal solution set that meets the multi-objective performance requirements in the parameter space.

[0008] As a preferred option, in step S1, the verification of the finite element model includes: comparing the calculated load-displacement response, stiffness degradation characteristics, and buckling mode with the physical test results, and confirming the model's validity when the error meets a preset threshold; after confirming its validity, forming a sample set covering the preset parameter space by adopting one or more strategies such as orthogonal design, Latin hypercube sampling, uniform design, or stratified sampling.

[0009] As a preferred option, in step S1: (1) The aspect ratio α is the ratio of the net height H to the net width L of the panel, α = H / L; (2) The corrugation tilt angle θ is the angle between the main direction of the corrugation and the horizontal boundary line, and the unit is angle or radian; (3) The amplitude a is half of the out-of-plane height between the peak and valley of the ripple, in mm; (4) The wavelength q is the period length between adjacent in-phase points along the main direction of the ripple, in mm; (5) The initial defect δ=kH, where k is a preset ratio in the range of 1 / 200 to 1 / 2000, used to characterize the maximum value of the initial out-of-plane offset of the panel; the initial defect can be applied in the finite element model by buckling mode superposition.

[0010] (6) The geometric parameters of the stiffening rib include at least the width b of the stiffening rib. s With thickness t s .

[0011] As a preferred option, in step S2: The peak strength index is obtained by extracting the maximum shear bearing capacity from the skeleton curve in the mechanical response data; The energy dissipation capacity index is obtained by accumulating the hysteresis loop area under cyclic loading response to obtain the cumulative energy dissipation, and by integrating the load-displacement curve under monotonic loading response to obtain the absorbed energy. As a preferred embodiment, the quantitative discrimination and verification process of the destruction mode label in step S2 is as follows: (1) Scale discrimination: Extract the out-of-plane displacement field w(x,y) from the finite element analysis results and identify the buckling half-wave scale λ; when λ is on the same order of magnitude as the panel size and the proportion of out-of-plane deformation coverage area is not less than the preset threshold η1, the sample is judged to be the overall buckling dominant feature; when λ is on the same order of magnitude as the corrugation geometric parameters, and the local wrinkles are mainly concentrated in the crest or trough area and show a localized distribution, and the proportion of out-of-plane deformation coverage area is not greater than the preset threshold η2, the sample is judged to be the local buckling dominant feature. (2) Consistency check: Extract the equivalent stress field σ eq (x,y) and the equivalent plastic strain field ɛ p (x,y), the above scale discrimination results are verified based on the distribution characteristics of high stress or high plasticity regions; (3) Coupling discrimination and coding: When the scale discrimination results of the same sample simultaneously satisfy the characteristics of overall buckling dominance and local buckling dominance, it is judged as interactive destruction; finally, the judgment results are mapped to discrete category codes to serve as classification output labels for supervised learning.

[0012] As a preferred option, in the sub-process (1) of step S4, the physical consistency logic verification is specifically performed by the system's built-in physical heuristic rule base to perform monotonicity verification on the model output.

[0013] As a preferred embodiment, in sub-process (2) of step S4, the triggering condition for the active learning feedback is: calculating the confidence interval variation coefficient CV = (Y upper -Y lower ) / Y mean ; where Y upper To predict the upper bound, Y lower To predict the lower bound Y mean The predicted value is the point value. If the CV exceeds the preset threshold or the physical consistency verification fails, the system automatically extracts the current high-risk parameter combination and generates a finite element simulation script to be fed back to step S1 for supplementary calculation.

[0014] As a preferred option, in sub-process (2) of step S4, the discretization manufacturing constraint is specifically: during the reverse optimization search process, the continuous design parameters to be optimized are forcibly mapped to a discrete specification variable sequence that conforms to current national standards or preset engineering standards.

[0015] As a preferred solution, to address the technical problem of slow convergence in high-dimensional nonlinear spaces, the reverse optimization design further includes: SHAP heuristic guidance: Extract the global importance ranking of each design parameter using SHAP to guide the search algorithm to prioritize sampling in the feature interval that contributes significantly to performance gain, and use the high-gain interval as the initial population search area of ​​the optimization algorithm, thereby accelerating the convergence of the algorithm by reducing the invalid search space; Multi-objective scheme comparison: With the common optimization objectives of minimizing steel consumption, maximizing load-bearing capacity and maximizing energy consumption capacity, the output is a Pareto optimal solution set including economic and performance-oriented schemes.

[0016] Secondly, the present invention also provides an intelligent interactive system, including a processor, a memory, and a display / input device communicatively connected thereto. The memory stores a computer program, which, when executed by the processor, implements the prediction method described above. When the processor executes the computer program, it at least implements the following functional modules: Database management and feature quantification module: used to store and update the mechanical behavior database, and generate peak strength index, energy dissipation capacity index and failure mode label from mechanical response data; Model training, evaluation, and interpretive analysis module: used to train candidate prediction models, evaluate and determine the final prediction model, and calculate and output SHAP values ​​or SHAP importance statistics; Online prediction and interactive display module: This module receives key design parameters input by the user and outputs prediction results and their confidence intervals. It also provides a human-computer interaction interface that includes parameter input, result display, and interpretive analysis. Active learning engine module: used to monitor and predict risk indicators; when the prediction uncertainty exceeds the preset threshold or the physical consistency verification fails, it automatically generates finite element parametric modeling scripts or simulation input files corresponding to the high-risk parameter combination, triggering data supplementation tasks; and after the supplementation is completed, it writes the new samples into the mechanical behavior database and triggers incremental or periodic updates to the prediction model. Simulation engine module: Used to receive modeling scripts or simulation input files generated by the active learning engine module, perform finite element parametric modeling and solution calculations, output corresponding mechanical response data, and send the mechanical response data back to the database management and feature quantization module; Reverse optimization module: It is used to receive the mechanical performance target value set by the user, and based on the prediction model and combined with discrete manufacturing constraints, it uses heuristic algorithm or Pareto optimization algorithm to reverse search for the optimal parameter combination that satisfies the performance target. Automatic report generation module: Standardized report templates are pre-stored. The system automatically extracts the results based on the prediction results, SHAP contribution ranking, and text of the dominant design parameters, and exports an automated design report containing mechanical performance evaluation and scheme comparison suggestions with one click.

[0017] The beneficial effects of this invention are: Compared with the prior art, the present invention has the following significant advantages: Firstly, it significantly reduces the computational cost and barrier to entry in the design phase. This invention constructs a parametric database based on a finite element model verified by physical experiments and trains a surrogate model. In the engineering application phase, no additional indoor experiments or time-consuming nonlinear finite element iterative solutions are required; the mechanical properties of corrugated steel plate shear walls can be obtained in real time through the system. Furthermore, this invention incorporates stiffening rib parameters and boundary constraint characteristics into the prediction range, enabling the prediction method to adapt to more complex engineering structural requirements.

[0018] Secondly, it improves the reliability and security of the prediction results. The prediction model constructed in this invention has an uncertainty estimation function, providing prediction confidence intervals along with the output point prediction values. This enables engineers to identify potential biases of the model in sparse data regions (such as extreme parameter combinations), avoiding blindly accepting low-confidence prediction results, thereby effectively reducing the safety risks of structural design.

[0019] Third, it enhances the interpretability of the results and its ability to support decision-making. By calculating the SHAP value, this invention can quantitatively provide the direction and extent of the contribution of parameters such as yield strength, geometric dimensions, and initial imperfections to bearing capacity and failure mode (for example, clearly indicating whether plate thickness or corrugation angle dominates the improvement in bearing capacity of the current scheme). This "white-box" analysis provides a direct basis for parameter sensitivity analysis.

[0020] Fourth, it achieves a functional closed loop from "forward prediction" to "reverse design". The reverse optimization module integrated in this invention changes the traditional "trial and error" design process, and can automatically deduce the optimal parameter combination based on the design goal (such as the target bearing capacity). Combined with the automatic report generation module, the system can generate standardized technical reports containing professional suggestions, cloud map simulations and data charts with one click, which significantly improves the automation level and commercial value of engineering bidding, scheme presentation and design delivery. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0022] Figure 1 This is a schematic diagram of the overall process of the method for predicting the mechanical behavior of corrugated steel plate shear walls according to the present invention. Figure 2 This is a schematic diagram illustrating the relationship between the construction of the mechanical behavior database and feature quantization in this invention; Figure 3 This is a structural block diagram of the intelligent interactive system of the present invention; Figure 4 A schematic diagram of the corrugated steel plate shear wall structure and its key geometric parameters; Figure 5 A schematic diagram of the user interface of an intelligent interactive system; Figure 6 A schematic diagram illustrating the SHAP (Self-Programmable Array) for a corrugated steel plate shear wall sample. Figure 7 This is a schematic diagram of the reverse optimization algorithm based on the prediction model of the present invention. Detailed Implementation

[0023] The present invention will now be described in detail through exemplary embodiments. However, it should be understood that, without further description, elements, structures, and features in one embodiment may be advantageously incorporated into other embodiments.

[0024] The present invention will be described below with reference to embodiments. It should be understood that the embodiments are only used to explain the present invention and are not intended to limit the scope of protection of the present invention. Without departing from the concept of the present invention, those skilled in the art can make equivalent substitutions or modifications to the step sequence, parameter value range, model type, and interface form in the embodiments, all of which should fall within the scope of protection of the present invention.

[0025] In this specification, the "mechanical behavior database" refers to a data set consisting of multiple sets of key design parameter samples, their corresponding mechanical response data, and mechanical behavior features / feature labels obtained by quantifying the mechanical response data, i.e., the corresponding set of "key design parameters - mechanical response data - feature labels"; the "mechanical behavior features" refer to the characterization indicators and category labels obtained by quantifying the mechanical response data, including at least peak strength indicators, energy dissipation capacity indicators, and failure mode labels.

[0026] The following is in conjunction with the appendix Figure 1-7 The detailed process of a method for predicting the mechanical behavior of corrugated steel plate shear walls and the module composition of its intelligent interactive system are described below: A method for predicting the mechanical behavior of corrugated steel plate shear walls, the specific steps of which are as follows: Step S1: Construction of the mechanical behavior database S11. First, establish a finite element model of the corrugated steel plate shear wall and verify the finite element model using existing physical test results. The verification content may include, but is not limited to: load-displacement response curves (including skeleton curves), stiffness degradation characteristics, and buckling modes. By comparing the finite element calculation results with the experimental results, when the difference meets the preset threshold requirements, it is confirmed that the finite element model can be used for parametric analysis and database construction.

[0027] The preset threshold can be characterized by a quantifiable error index, preferably including: peak shear capacity V. max The relative error between the energy consumption capacity E and the energy dissipation capacity E, wherein the relative error can be expressed by formula e x =|X FE -Xtest | / X test Calculate by multiplying by 100% (X is V) max (or E), with a threshold requirement that the error of each key indicator does not exceed 15%.

[0028] Furthermore, to enhance the consistency verification of the overall shape and stiffness degradation characteristics of the curve, the similarity range of the lateral force-displacement curve can be quantified by the overall error index of the curve, preferably using the root mean square error (RMSE) or normalized root mean square error (NRMSE); where RMSE can be calculated according to formula (1), and NRMSE can be calculated according to formula (2): in, , The lateral force values ​​can be obtained from the skeleton curve (or envelope curve) at a unified displacement node (which can be obtained through interpolation or resampling) using finite element and experimental data. The peak lateral force is used in the experiment. Preferably, the normalized root mean square error (NRMSE) is used to characterize the overall error of the curve, and NRMSE ≤ 15% is used as the threshold requirement.

[0029] Furthermore, the stiffness degradation characteristic can be achieved at several preset displacement levels {Δ} j Extracting the secant stiffness K j =V j / △ j Among them, V j It can be taken as the corresponding displacement amplitude Δ j The peak lateral force or the lateral force at the corresponding point on the skeleton curve, {△ j The selected point can be either the test loading displacement level or a preset displacement control point. The mean absolute percentage error (MAPE) is used. K Characterizing its consistency: Preferably, when MAPE K When the stiffness degradation rate is ≤15%, it is considered that the stiffness degradation characteristics are consistent between the finite element results and the experimental results.

[0030] Specifically, this embodiment is based on the cyclic loading test of trapezoidal corrugated steel plate shear walls by Emami et al., and a finite element model for simulating its seismic performance is established using ABAQUS software. To verify the accuracy of the model, the vertical and horizontal corrugation test data provided by the test are selected as the benchmark. The shear wall specimens are subjected to the same preset horizontal cyclic load as in the test to conduct pushover tests, and the nonlinear response behavior of the shear wall under cyclic shear force is recorded. It should be understood that the finite element model can be evaluated by comparing one dimension—the shear wall failure mode, yield zone and plastic strain development distribution, or lateral force-displacement curve—or by combining multiple dimensions. That is, based on the actual cracking damage distribution and failure mode of the shear wall, the material constitutive model parameters, contact setting parameters, and mesh parameters are adjusted. When the relative error and / or NRMSE and MAPE of the key indicators are... K Once the threshold requirement of no more than 15% is met, stop adjusting the parameters and confirm that the model is effective.

[0031] Table 1 shows an example of comparing simulation results with experimental data. The energy dissipation capacity E can be obtained by accumulating the hysteretic torus of the hysteretic load-displacement curve, with the accumulation range consistent with the experimental loading level. Table 1 shows that the errors of all key indicators are controlled within 15%, indicating that the model has high accuracy. S12. To facilitate consistency management of engineering input and database, this embodiment adopts common definition methods for some geometric parameters: the aspect ratio α can be characterized by the ratio of the panel's net height H to its net width L, i.e., α = H / L; the corrugation tilt angle θ can be defined as the angle between the main corrugation direction and the horizontal boundary line; the wavelength q can characterize the period length between adjacent in-phase points along the main corrugation direction; the amplitude a can be used to characterize the characteristic quantity of the out-of-plane height between corrugation peaks and valleys, and maintains a consistent definition in the database; the initial defect δ can be given according to the defect amplitude related to the panel's net height, for example, taking δ = kH (k takes 1 / 200~1 / 2000), used to characterize the maximum value of the panel's initial out-of-plane offset; in finite element modeling, the initial defect can be applied by buckling mode superposition, i.e., first perform linear buckling analysis to obtain the modal shape, then scale the mode according to the preset amplitude and superimpose it into the initial geometry; stiffening rib geometric parameters (if applicable): define the width b of the stiffening rib. s Thickness t s The parameters include their spacing and arrangement, or the definition of the bending stiffness ratio of edge members. The definition and application of these parameters ensure consistency in the database sample size, facilitating subsequent feature quantification and model training.

[0032] S13. Based on confirming the validity of the finite element model, the value range and combination method of key design parameters are set to form a sample set covering the preset parameter space. Specifically, this embodiment, based on the finite element model established in S11, establishes 576 corresponding finite element models by adjusting the shear wall parameters in the finite element model. The parameter ranges are derived from engineering experience and experimental coverage. The key design parameters include at least the yield strength f. y Steel plate thickness t, aspect ratio α, corrugation tilt angle θ, wave amplitude a, wavelength q, initial defect δ.

[0033] To ensure the sample set has good representativeness of the parameter space, this embodiment employs one or more strategies among orthogonal design, Latin hypercube sampling, uniform design, or stratified sampling to construct parameter combinations. Orthogonal design is suitable for scenarios where a small number of samples cover multiple factors and levels; Latin hypercube sampling or uniform design is suitable for uniform coverage of continuous parameter spaces; and stratified sampling is suitable for scenarios where key parameter intervals are segmented and the sample ratio of different intervals is controlled. In practical applications, the appropriate strategy can be selected based on the parameter dimensions, value range, and database size requirements, and the generation criteria should be kept consistent during database updates to ensure the repeatability of model training and comparative evaluation. Table 2 shows the value range and examples of key design parameters. S14. After generating the parameter combinations, perform finite element analysis on each set of parameter samples to obtain the corresponding mechanical response data and write it into the mechanical behavior database. The mechanical response data may include, but is not limited to: load-displacement response curves, hysteresis curves, stress / strain fields (e.g., equivalent stress, equivalent plastic strain), displacement fields (e.g., out-of-plane displacement fields), and buckling morphology information, thereby establishing a correspondence between key design parameters and mechanical response data. Further, the mechanical response data is quantified to generate mechanical behavior features / feature tags, and stored together with the corresponding key design parameters and mechanical response data, thus constituting the mechanical behavior database.

[0034] Step S2: Quantitative extraction of mechanical behavior features In this embodiment, the mechanical behavior features are obtained by quantifying the mechanical response data obtained in step S1, in order to construct the "true value" labels required for the supervised learning prediction model. The mechanical behavior features include at least peak intensity indicators, energy dissipation capacity indicators, and failure mode labels.

[0035] S21. Peak Strength Index Extraction: The peak strength index can be extracted based on the skeleton curve (or envelope curve) obtained from finite element analysis. Specifically, the maximum shear capacity V corresponding to the skeleton curve is taken. maxAs a peak intensity indicator, V max This represents the absolute value of the maximum shear force on the skeleton curve, expressed in kN. This index will be one of the main physical quantities used in subsequent regression model predictions.

[0036] S22. Energy Dissipation Capacity Index Extraction: The energy dissipation capacity index is obtained according to different loading regimes as follows: When the mechanical response data is a cyclic loading response, the hysteresis loop area of ​​each loading cycle is calculated based on the hysteretic load-displacement curve and accumulated to obtain the cumulative energy dissipation E as the energy dissipation capacity index; when the mechanical response data is a monotonic loading response, the absorbed energy is obtained by integrating the displacement interval based on the load-displacement curve as the energy dissipation capacity index. The unit of cumulative energy dissipation or absorbed energy is kN·mm.

[0037] S23. As a preferred embodiment, the quantitative discrimination and verification process of the destruction mode label in step S2 is as follows: (1) Scale discrimination: Extract the out-of-plane displacement field w(x,y) from the finite element analysis results and identify the buckling half-wave scale λ; when λ is on the same order of magnitude as the net height H or net width L, and the proportion of the out-of-plane deformation coverage area is not less than the preset threshold η1, the sample is determined to be a dominant feature of overall buckling; when λ is on the same order of magnitude as the wave geometric parameter wavelength q or wave amplitude a, and the local wrinkles are mainly concentrated in the wave crest or wave trough region and show a localized distribution, and the proportion of the out-of-plane deformation coverage area is not greater than the preset threshold η2, the sample is determined to be a dominant feature of local buckling; wherein, the proportion of the deformation coverage area can be defined as satisfying the absolute value |w(x,y)|≥w th The proportion of the area of ​​the region to the total area of ​​the panel, w th The preset threshold; (2) Consistency check: Extract the equivalent stress field σ eq (x,y) and the equivalent plastic strain field ɛ p (x,y), identify σ ep ≥σ th or ε p ≥ε th High stress or high plasticity areas; if the area mainly covers a large continuous area of ​​the panel, the check is overall buckling dominant; if the area is mainly concentrated in a local area near the corrugated lines or crests and troughs, the check is local buckling dominant. (3) Coupling discrimination and coding: When the scale discrimination results of the same sample simultaneously satisfy the characteristics of overall buckling dominance and local buckling dominance, it is judged as interactive destruction; finally, the judgment results are mapped to discrete category codes to serve as classification output labels for supervised learning.

[0038] Step S3: Predictive model training, evaluation, and interpretive analysis S31. Dataset Partitioning and Feature Construction: The database samples constructed in step S1 and the mechanical behavior features extracted in step S2 are combined to form a dataset. The input feature vector X consists of key design parameters, including at least the yield strength f. y The parameters include: steel plate thickness t, aspect ratio α, corrugation tilt angle θ, wave amplitude a, wavelength q, initial defect δ, and stiffening rib geometric parameters or boundary constraint characteristics; the output label Y includes at least: peak strength index, energy dissipation capacity index (regression target), and failure mode label (classification target).

[0039] The dataset is divided into training and test sets (e.g., 70%:30% of the sample size). On the training set, k-fold cross-validation combined with grid search or Bayesian optimization is used to optimize hyperparameters to determine the optimal parameter combination for the model.

[0040] S32. Model performance evaluation: In this embodiment, the trained model is independently evaluated on the test set to determine the final prediction model.

[0041] (1) Regression Model Evaluation (Peak Intensity and Energy Dissipation Capacity) The model performance is evaluated using a pre-defined regression error index, including the coefficient of determination (R²). 2 The relative error level is reflected by indicators such as mean absolute error (MAE), mean absolute percentage error (MAPE), root mean square error (RMSE), and mean square error (MSE); their expressions are as follows: In the formula, t i t represents the measured value (finite element true value); t represents the measured average value; p i is the predicted value; n is the number of samples.

[0042] Table 3 shows examples of machine learning model performance evaluation results. As shown in Table 3, XgBoost and SVR models exhibit better generalization ability on the test set (higher R-values). 2 (With a lower MAPE), it can be given priority as a candidate model.

[0043] (2) Classification Model Evaluation (Damage Mode): Multi-class classification evaluation indicators are used to evaluate the model performance. Accuracy: The percentage of samples in the test set that correctly predict the target data; Macro-F1 (Macro-average F1): The arithmetic mean of the F1 scores calculated for each damage mode category is used to objectively evaluate the model when the number of samples in each category is unbalanced. Confusion Matrix: A statistical representation of the correspondence between true and predicted classes in a test set.

[0044] Table 4 shows examples of the evaluation results of the destruction mode classification model on the test set. Table 5 shows an example of a confusion matrix for damage mode classification (test set). Explanation: Mode I = Overall damage (code 0); Mode II = Local damage (code 1); Mode III = Interactive damage (code 2). Tables 4-5 show that the classification model has high accuracy and macro-average F1 score on the test set. Furthermore, the confusion matrix indicates that misclassifications among the damage modes are mainly concentrated in adjacent or similar damage mechanism categories, suggesting that the prediction has good engineering applicability.

[0045] S33. Construction and Interpretation of Uncertainty Prediction Models (1) Constructing a prediction model with uncertainty estimation function: In response to the requirement in the claims regarding the output of "predicted confidence intervals", this embodiment introduces an uncertainty quantification mechanism based on the above algorithm. The prediction model preferably adopts a gradient boosting decision tree (such as XGBoost / LightGBM) combined with a quantile loss function, or adopts Gaussian process regression, or adopts a Bayesian neural network. Taking the gradient boosting decision tree as an example, the model is configured as a multi-output or multi-model structure, respectively predicting the target variable's: point prediction value (usually the conditional mean or 50th quantile); prediction lower bound (e.g., 5th quantile, corresponding to Q). 0.05 Predict the upper bound (e.g., 95th percentile, corresponding to Q). 0.95 The resulting 90% prediction confidence interval [Q] 0.05 Q 0.95 This is used to quantify the model's certainty regarding the current input parameters. When the interval width is too large, it prompts the user that the current design parameters are in a sparse region of the training data and should be used with caution.

[0046] (2) SHAP Interpretive Analysis: After determining the final prediction model, the SHAP values ​​of each key design parameter are calculated based on the model (usually based on a point prediction model). For the regression output of peak intensity and energy dissipation capacity, and the classification output of failure modes, the global importance ranking and sample-level contribution decomposition are calculated respectively.

[0047] Global importance: The average contribution of each parameter to the model output is quantified by calculating the average absolute value of the SHAP of all samples in the test set.

[0048] Directional analysis: Combining the SHAP dependency plot, this analysis examines whether increasing parameter values ​​positively enhances or negatively weakens the bearing capacity. Through these steps, a comprehensive prediction model is constructed that combines high-precision point prediction, uncertainty quantification alerts, and white-box interpretation capabilities, laying the foundation for the subsequent deployment of intelligent interactive systems.

[0049] S4, Intelligent Interactive System Deployment and Online Prediction S41. System Deployment and Interface Architecture This invention deploys the final prediction model (integrating the point prediction model and the uncertainty estimation model) determined in step S3 into an intelligent interactive system. The system architecture adopts a modular design and features a human-computer interaction interface, specifically including: a parameter input interface, a prediction result display interface, an interpretive analysis interface, a reverse engineering interface, and a report generation interface. The user inputs f through the parameter input interface. y After obtaining the parameters t, α, θ, a, q, δ and stiffening ribs, the system calls the background prediction model through the online inference engine and outputs the predicted mechanical properties and their confidence intervals.

[0050] S42. System Module Function Definition: Database management module: used to store the mechanical behavior database, model version iteration information, and user historical prediction records; Feature quantization and training module: Enables feature extraction from mechanical response data and supports periodic model updates and training; Interpretive Analysis Module: Calls the SHAP algorithm to calculate the contribution of each parameter and generates intuitive visualization charts; Online prediction and reverse optimization module: Supports forward prediction from "design parameters → mechanical performance" and reverse intelligent search from "target performance → optimal parameters"; Automatic report generation module: Based on preset standardized templates, it extracts predicted data, explanatory conclusions and comparison suggestions, and generates technical reports with one click.

[0051] S43. Positive Forecasting and Uncertainty Assessment Process When a user performs a positive prediction, the system sequentially executes the following multi-level safeguard mechanism to output the result: (1) Parameter space validity and distribution density verification: The system first checks whether the input vector X is within the preset value range of the mechanical behavior database. In addition to boundary checks, the system introduces an outlier detection algorithm (such as Euclidean distance analysis) to calculate the distribution density of the input point and the existing sample set; if the input parameter is in the "extrapolation region" where the data is extremely sparse, the system will automatically increase the uncertainty weight of the sample.

[0052] (2) Multi-task prediction and confidence assessment: Call the supervised learning model constructed in step S3 and output the peak intensity V simultaneously. max Point prediction of energy consumption capacity E and confidence interval based on 90% quantile [Y] lower ,Y upper ].

[0053] (3) Physical consistency logic check: The system has a built-in physical heuristic rule base to perform monotonicity checks on the model output. For example, when other geometric parameters are fixed, the V... max Does the predicted value vary with the steel plate thickness t or yield strength f? y The prediction shows a positive growth trend as the value increases. If the prediction results show fluctuations that violate the basic laws of materials mechanics, the system will mark the prediction as "physical logic failure" and force a jump to the feedback mechanism in step 4.

[0054] (4) Active learning feedback loop: If the confidence interval coefficient of variation CV = (Y upper -Y lower ) / Y mean If the risk exceeds a preset threshold (e.g., 15%), or if the physical consistency check determines the failure, the system will trigger an automatic recalculation suggestion. The system's internal logic automatically extracts the current high-risk parameter combinations, generates corresponding finite element modeling scripts (e.g., Python or TCL scripts) using preset parametric templates, and automatically drives the simulation engine (e.g., ABAQUS or ANSYS) in step S1 to perform unattended, precise recalculation. After the calculation is complete, the system automatically extracts response features and feeds them back to the mechanical behavior database to update the proxy model, thereby achieving self-evolution of prediction accuracy and fully automatic closed-loop data updates.

[0055] (5) Visualization: While displaying the predicted value on the interface, the confidence interval is marked by the shaded area, and the relative position of the predicted point in the global parameter space is displayed simultaneously.

[0056] S44 Reverse Optimization Design Example To address the trade-off between "performance compliance" and "cost control" in practical engineering, the reverse optimization module executes the following logic: (1) Construction of multi-objective function: Allows users to set multi-dimensional design objectives, including but not limited to: target bearing capacity range V target Energy consumption threshold value E min and the target of minimizing steel consumption Min Weight .

[0057] (2) Introducing Discrete Manufacturing Constraints: In the optimization calculation, the system forces specific continuous parameters to be transformed into discrete specification variables. For example, the steel plate thickness t is only taken from steel plate specifications that conform to the current national standards, such as {6,8,10,12,14,16}, and the wavelength q must meet the modulus requirements of the processing machinery to ensure that the optimization results meet the actual construction needs.

[0058] (3) SHAP Heuristic Initial Value Guidance: To address the slow convergence of the optimization search caused by the numerous design parameters and highly nonlinear mechanical behavior of corrugated steel plate shear walls, this embodiment utilizes the SHAP global importance ranking calculated in step S3 for intelligent acceleration during the initialization of the search population. By analyzing the contribution of each feature to the performance objective, the system guides the search algorithm to prioritize intensive sampling within the "high-sensitivity" feature interval that significantly contributes to performance gains. For example, if SHAP analysis shows that the plate thickness t is the dominant factor, the algorithm concentrates the initial solution space within the optimal theoretical interval of t, thereby effectively avoiding the algorithm from getting trapped in local optima and significantly shortening the iteration convergence time of genetic algorithms or particle swarm algorithms in complex high-dimensional spaces.

[0059] (4) Pareto front selection: The optimization module outputs a set of Pareto optimal solutions for the user to weigh: Option A (Economy): Just meets the strength requirements, with the thinnest plate thickness and the least amount of steel used.

[0060] Option B (Performance-based): Significantly improves seismic redundancy and energy dissipation capacity while moderately increasing plate thickness.

[0061] (5) Interactive optimization decision: The system displays the scores of each scheme in four dimensions, namely strength, energy consumption, weight and construction difficulty, through a parallel coordinate graph, which helps designers select the final implementation scheme with one click and import it into the automatic report generation module.

[0062] S45 Automatic Report Generation Example The system executes the following automated generation logic based on the prediction or optimization results: Data extraction: Automatically retrieves the input parameters, predicted mean, confidence interval, and physical verification status of the current scheme.

[0063] Text synthesis: Professional evaluations are synthesized based on a rule base. For example: "The predicted bearing capacity of this scheme is 801.32kN. SHAP analysis shows that yield strength contributes the most to the improvement of bearing capacity (SHAP value +49.10), while initial defects are the main weakening factor." File export: Export the above text, mechanical performance curves, and visualization charts (such as...). Figure 6 Automatically format and generate a PDF "Design Performance Evaluation Report for Corrugated Steel Shear Wall".

[0064] S46, Example of positive forecast data To verify the model's accuracy, typical parameter combinations from the database were selected for inference, as shown in Table 6 below.

[0065] S47, Example of SHAP Interpretive Analysis The interpretive analysis interface provides a decomposition result of "baseline value + contribution of each feature = predicted value" for a single sample based on SHAP.

[0066] Sample a: V max =760.96kN. Among them, the initial defect δ=2.40mm, aspect ratio α=0.90 and wavelength q=300mm have a positive contribution to the predicted value, while the plate thickness t=6.00mm and small amplitude a=15.00mm have a negative contribution (e.g. -126.41kN and -58.22kN respectively). Sample b: V max =801.32kN. Among them, the material yield strength fy=307MPa and the amplitude a=20.00mm contribute positively (e.g. +49.10kN and +46.76kN), while the plate thickness t=6.00mm and the initial defect δ=13.50mm contribute negatively (e.g. -82.14kN and -70.42kN). Sample c (stiffening measures): After introducing longitudinal stiffening ribs (width b) s =60mm, thickness t s =6mm), V max The load capacity was increased to 876.26 kN. SHAP decomposition showed that the introduction of stiffening rib parameters played a significant positive and dominant role in improving the load-bearing capacity, among which the stiffening rib width b... s The SHAP contribution is +75.40 kN, and the thickness t s The stiffening rib parameter contributed +39.90 kN, effectively compensating for the reduced load-bearing capacity caused by the thinner plate. Furthermore, in the classification output interpretation of failure modes, the stiffening rib parameter had the highest positive SHAP value for the probability output of the "overall failure" category, indicating that this feature was accurately identified by the model as a key structure altering the structural failure mechanism.

[0067] Table 6. Input parameters and prediction results for typical samples (V max Examples of destruction modes (E and destruction modes) are as follows: This invention also relates to an intelligent interactive system, including a processor, a memory, and a display / input device communicatively connected thereto. The memory stores a computer program, which, when executed by the processor, implements the prediction method described above. When the processor executes the computer program, it at least implements the following functional modules: Database management and feature quantification module: used to store and update the mechanical behavior database, and generate peak strength index, energy dissipation capacity index and failure mode label from mechanical response data; Model training, evaluation, and interpretive analysis module: used to train candidate prediction models, evaluate and determine the final prediction model, and calculate and output SHAP values ​​or SHAP importance statistics; Online prediction and interactive display module: This module receives key design parameters input by the user and outputs prediction results and their confidence intervals. It also provides a human-computer interaction interface that includes parameter input, result display, and interpretive analysis. Active learning engine module: used to monitor and predict risk indicators; when the prediction uncertainty exceeds the preset threshold or the physical consistency verification fails, it automatically generates finite element parametric modeling scripts or simulation input files corresponding to the high-risk parameter combination, triggering data supplementation tasks; and after the supplementation is completed, it writes the new samples into the mechanical behavior database and triggers incremental or periodic updates to the prediction model. Simulation engine module: Used to receive modeling scripts or simulation input files generated by the active learning engine module, perform finite element parametric modeling and solution calculations, output corresponding mechanical response data, and send the mechanical response data back to the database management and feature quantization module; Reverse optimization module: It is used to receive the mechanical performance target value set by the user, and based on the prediction model and combined with discrete manufacturing constraints, it uses heuristic algorithm or Pareto optimization algorithm to reverse search for the optimal parameter combination that satisfies the performance target. Automatic report generation module: Standardized report templates are pre-stored. The system automatically extracts the results based on the prediction results, SHAP contribution ranking, and text of the dominant design parameters, and exports an automated design report containing mechanical performance evaluation and scheme comparison suggestions with one click.

[0068] The parts not described in detail in this embodiment are existing technologies.

Claims

1. A method for predicting the mechanical behavior of corrugated steel plate shear walls, characterized in that: The steps are as follows: S1. Parametric analysis is performed based on the finite element model of the corrugated steel plate shear wall verified by physical experiments. Different parameter combinations are formed by changing key design parameters and component configurations. The finite element simulation engine is used to automatically parametrically model and solve the finite element model corresponding to each parameter combination, obtain mechanical response data, and write it into the mechanical behavior database. S2. Perform feature quantization on the mechanical response data to obtain mechanical behavior features, which include at least peak intensity index, energy dissipation capacity index and failure mode label. S3. Construct a supervised learning prediction model with uncertainty estimation function, taking the key design parameters as input, and outputting the predicted values ​​of mechanical behavior features and their prediction confidence intervals. And based on the prediction model, calculate the SHAP value or SHAP importance statistic for each key design parameter; S4. Intelligent Interaction and Predictive Optimization: The predictive model is deployed to the intelligent interaction system, which executes the following sub-processes based on the parameters input by the user: (1) Positive prediction and reliability assessment: Physical consistency logic verification is introduced to check whether the prediction results conform to the monotonicity law of material mechanics; (2) Active learning feedback: When the prediction uncertainty exceeds the preset threshold or the physical consistency verification fails, the system automatically generates the corresponding finite element parameterized modeling script based on the current high-risk parameter combination, and calls the finite element simulation engine described in step S1 to perform automatic supplementary calculation, so as to realize the fully automatic closed-loop update of data and the real-time incremental update of the mechanical behavior database. (3) Reverse optimization design: Discretized manufacturing constraints are introduced, and SHAP heuristics are used to search for Pareto optimal solution set that meets the multi-objective performance requirements in the parameter space.

2. The prediction method according to claim 1, characterized in that: In step S1, the verification of the finite element model includes: comparing the calculated load-displacement response, stiffness degradation characteristics, and buckling mode with the physical test results, and confirming the model's validity when the error meets a preset threshold; after confirming its validity, forming a sample set covering the preset parameter space by adopting one or more strategies such as orthogonal design, Latin hypercube sampling, uniform design, or stratified sampling.

3. The prediction method according to claim 1, characterized in that, In step S1: the key design parameters include at least the yield strength f. y The steel plate thickness t, aspect ratio α, corrugation angle θ, corrugation amplitude a, wavelength q, and initial defect δ, wherein the component is configured as stiffening rib geometric parameters or boundary constraint features, and the stiffening rib geometric parameters include at least the stiffening rib width b. s With thickness t s .

4. The prediction method according to claim 1, characterized in that, In step S2: The peak strength index is obtained by extracting the maximum shear bearing capacity from the skeleton curve in the mechanical response data; The energy dissipation capacity index is obtained by accumulating the hysteresis loop area under cyclic loading response to obtain the cumulative energy dissipation, and by integrating the load-displacement curve under monotonic loading response to obtain the absorbed energy.

5. The prediction method according to claim 1, characterized in that, In step S2, the specific process for quantifying and verifying the destruction mode label is as follows: (1) Scale discrimination: Extract the out-of-plane displacement field w(x,y) from the finite element analysis results and identify the buckling half-wave scale λ; when λ is on the same order of magnitude as the panel size and the proportion of out-of-plane deformation coverage area is not less than the preset threshold η1, the sample is determined to be the overall buckling dominant feature. When λ is on the same order of magnitude as the corrugation geometry parameter, and the local wrinkles are mainly concentrated in the crest or trough region and exhibit a localized distribution, and the proportion of their out-of-plane deformation coverage area is not greater than the preset threshold η2, the sample is determined to have a local buckling dominant feature. (2) Consistency check: Extract the equivalent stress field σ eq (x,y) and the equivalent plastic strain field ɛ p (x,y), the above scale discrimination results are verified based on the distribution characteristics of high stress or high plasticity regions; (3) Coupling discrimination and coding: When the scale discrimination results of the same sample simultaneously satisfy the characteristics of overall buckling dominance and local buckling dominance, it is judged as interactive destruction; finally, the judgment results are mapped to discrete category codes to serve as classification output labels for supervised learning.

6. The prediction method according to claim 1, characterized in that, In sub-process (1) of step S4, the physical consistency logic verification specifically involves the system's built-in physical heuristic rule base to perform monotonicity verification on the model output.

7. The prediction method according to claim 1, characterized in that, In sub-process (2) of step S4, the triggering condition for the active learning feedback is: calculating the confidence interval variation coefficient CV = (Y upper -Y lower ) / Y mean ; where Y upper To predict the upper bound, Y lower To predict the lower bound Y mean The predicted value is the point value. If the CV exceeds the preset threshold or the physical consistency verification fails, the system automatically extracts the current high-risk parameter combination and generates a finite element simulation script to be fed back to step S1 for supplementary calculation.

8. The prediction method according to claim 1, characterized in that, In subprocess (2) of step S4, the discretization manufacturing constraint is specifically: during the reverse optimization search process, the continuous design parameters to be optimized are forcibly mapped to a discrete specification variable sequence that conforms to the current national standards or the preset engineering standards.

9. The prediction method according to claim 1 or 8, characterized in that, To address the technical problem of slow convergence in high-dimensional nonlinear spaces, the reverse optimization design further includes: SHAP heuristic guidance: Extract the global importance ranking of each design parameter using SHAP to guide the search algorithm to prioritize sampling in the feature interval that contributes significantly to performance gain, and use the high-gain interval as the initial population search area of ​​the optimization algorithm, thereby accelerating the convergence of the algorithm by reducing the invalid search space; Multi-objective scheme comparison: With the common optimization objectives of minimizing steel consumption, maximizing load-bearing capacity, and maximizing energy consumption capacity, the output is a Pareto optimal solution set that includes economical and performance-oriented schemes.

10. An intelligent interactive system, characterized in that, The system includes a processor, a memory, and a display / input device communicatively connected thereto. The memory stores a computer program, which, when executed by the processor, implements the prediction method according to any one of claims 1 to 9. When the processor executes the computer program, it at least implements the following functional modules: Database management and feature quantification module: used to store and update the mechanical behavior database, and generate peak strength index, energy dissipation capacity index and failure mode label from mechanical response data; Model training, evaluation, and interpretive analysis module: used to train candidate prediction models, evaluate and determine the final prediction model, and calculate and output SHAP values ​​or SHAP importance statistics; Online prediction and interactive display module: This module receives key design parameters input by the user and outputs prediction results and their confidence intervals. It also provides a human-computer interaction interface that includes parameter input, result display, and interpretive analysis. Active learning engine module: used to monitor and predict risk indicators; when the prediction uncertainty exceeds the preset threshold or the physical consistency verification fails, it automatically generates finite element parametric modeling scripts or simulation input files corresponding to the high-risk parameter combination and triggers data supplementation tasks. After the calculation is completed, the new samples will be written into the mechanical behavior database, and the incremental or periodic updates to the prediction model will be triggered. Simulation engine module: Used to receive modeling scripts or simulation input files generated by the active learning engine module, perform finite element parametric modeling and solution calculations, output corresponding mechanical response data, and send the mechanical response data back to the database management and feature quantization module; Reverse optimization module: It is used to receive the mechanical performance target value set by the user, and based on the prediction model and combined with discrete manufacturing constraints, it uses heuristic algorithm or Pareto optimization algorithm to reverse search for the optimal parameter combination that satisfies the performance target. Automatic report generation module: Standardized report templates are pre-stored. The system automatically extracts the results based on the prediction results, SHAP contribution ranking, and text of the dominant design parameters, and exports an automated design report containing mechanical performance evaluation and scheme comparison suggestions with one click.

Images