Intelligent prediction method, device and equipment for impact failure mode of rc component and storage medium

Abstract

The present application relates to the technical field of intelligent detection and safety evaluation of civil engineering structures, and relates to an RC component impact damage mode intelligent prediction method, device, equipment and storage medium. It comprises: S1. obtaining a multi-source data set of RC component impact response, including test data and finite element simulation data; S2. performing data cleaning processing and sample balancing processing on the multi-source data set; S3. training a plurality of candidate machine learning models based on the processed data set, and determining an optimal prediction model through cross-validation; S4. introducing a cost-sensitive mechanism in the decision-making stage of the optimal prediction model, and setting differentiated loss weights for misjudgments of different damage modes; S5. outputting the RC component impact damage mode prediction result generated by the optimal prediction model, and generating a feature explanation result based on explainability analysis. Through data enhancement, model optimization and explanation mechanism, high-precision prediction of RC component impact damage mode is realized when the sample is limited and the classes are imbalanced, and the credibility and practicality of structure safety evaluation are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent detection and safety assessment technology for civil engineering structures, and in particular to intelligent prediction methods, devices, equipment and storage media for impact failure modes of RC components. Background Technology

[0002] In recent years, the frequency of accidents involving vehicle collisions, ship collisions, and rockfalls on civil engineering structures worldwide has been rising, posing a serious threat to the safe operation of critical infrastructure such as bridges and offshore platforms. These impact loads, characterized by high strength, short duration, and high energy, easily induce complex brittle failure modes in reinforced concrete (RC) components, such as punching shear failure and oblique shear failure, with rapid failure processes and severe consequences. Therefore, efficiently, accurately, and reliably predicting the final failure mode of RC components under impact loads has become a core challenge and urgent need for structural safety assessment, protective design optimization, and accident emergency response.

[0003] Currently, mainstream prediction methods mainly rely on three types of traditional technologies, all of which have significant limitations: physical tests (such as drop hammer and pendulum impact tests) provide intuitive and reliable results, but they are costly, time-consuming, have poor repeatability, and pose safety hazards. They also make it difficult to systematically obtain sufficient samples covering key design parameters such as geometric dimensions, material strength, and reinforcement ratio, which severely restricts the universality of the models. Numerical simulations (such as LS-DYNA and ABAQUS explicit dynamic analysis) facilitate parametric research, but they consume a lot of computational resources, are complex in modeling, and are difficult to verify, especially in identifying shear failure and bending-shear coupled failure with insufficient accuracy. Empirical / semi-empirical formulas are simple to calculate, but they are limited by the scale of the original data and complex stress states, especially in the region where strain rate effects are significant and failure modes transition, where the reliability of predictions decreases significantly.

[0004] Machine learning (ML) has shown great potential in structural damage prediction due to its powerful data-driven modeling and nonlinear pattern recognition capabilities. However, most research focuses on seismic load scenarios. When applied to predicting impact failure modes of reinforced concrete (RC) components, relevant research remains relatively scarce and faces three core obstacles: First, due to the cost limitations of physical experiments and detailed simulations, high-quality failure mode labeled samples are scarce, resulting in a severe shortage of training data. Second, the distribution of various failure mode samples is uneven, with the majority class (e.g., bending failure) far exceeding the minority class (e.g., shear failure), causing the model to favor the majority class and reducing its sensitivity to high-risk minority classes. Third, current ML models generally exhibit "black box" characteristics, lacking transparency and interpretability in their decision-making logic, which affects their credibility and practicality in engineering decision-making. Summary of the Invention

[0005] This invention aims to at least solve one of the technical problems existing in the prior art. This invention provides an intelligent prediction method, device, equipment, and storage medium for the impact failure modes of RC components. It can acquire, process, and model multi-source datasets to obtain prediction results of the failure modes of RC components under impact loads. In this method, a training set can be constructed based on experimental data, finite element simulation data, and rule-based augmented data. Data cleaning and sample balancing are performed. Through the training and optimization of candidate machine learning models, a cost-sensitive mechanism is introduced for decision-making. Combined with interpretability analysis, the feature contribution results are output, thereby effectively solving the problems of sample scarcity and class imbalance, reducing the risk of misjudgment of key categories, and achieving high accuracy and interpretability in failure mode prediction.

[0006] In a first aspect, embodiments of the present invention provide an intelligent prediction method for impact failure modes of RC components, the method comprising: S1. Obtain a multi-source dataset related to the impact response of RC components, wherein the multi-source dataset includes experimental data and finite element simulation data; S2. Perform data cleaning and sample balancing on the multi-source dataset; S3. Train multiple candidate machine learning models based on the processed dataset, and determine the optimal prediction model through cross-validation; S4. Introduce a cost-sensitive mechanism in the decision-making stage of the optimal prediction model to set differentiated loss weights for misjudgments of different destruction modes; S5. Output the predicted impact failure mode of the RC component generated by the optimal prediction model, and generate the feature interpretation results based on interpretability analysis.

[0007] The intelligent prediction method for impact failure modes of RC components according to embodiments of the present invention has at least the following beneficial effects: The method of this invention first constructs a multi-source dataset by introducing experimental data and finite element simulation data. Then, it improves the reliability and class distribution balance of the data through data cleaning and sample balancing. Next, it trains candidate machine learning models and selects the best-performing prediction model through cross-validation. Finally, a cost-sensitive mechanism is introduced in the prediction stage to effectively reduce the misjudgment rate of key failure modes, and the importance and operational patterns of features are output through interpretability analysis. This method, through the synergistic effect of data augmentation, model optimization, and interpretation mechanisms, achieves high-precision prediction of impact failure modes of RC components under conditions of limited samples and class imbalance, thereby improving the credibility and practicality of structural safety assessment.

[0008] According to some embodiments of the present invention, the multi-source dataset in S1 includes sample data augmented based on punching failure discrimination rules, wherein the punching failure discrimination rules include: S11. Calculate the punching shear capacity of the sample to be judged; S12. When the peak impact force recorded in the sample is greater than the punching shear bearing capacity, the sample is marked as punching shear failure mode; S13. Add the labeled samples to the multi-source dataset.

[0009] According to some embodiments of the present invention, the data cleaning process in S2 includes: S21. In the decision boundary region, oversampling is performed on minority class samples to generate supplementary samples for the boundary region; S22. Based on the neighborhood consistency rule, determine the neighbor category of each sample. If the majority of the neighbor categories are inconsistent with the sample category, then the sample is removed as a noise sample or a fuzzy sample. S23. After completing the above cleaning, delete the supplementary sample.

[0010] According to some embodiments of the present invention, the data cleaning process sets a credibility level based on the data source, wherein the experimental data is set as the highest level of protected source sample and is not removed during the cleaning process.

[0011] According to some embodiments of the present invention, the finite element simulation data is generated by a parametric simulation method, which covers the following key parameters: Geometric dimensions, including cross-sectional height and cross-sectional width; Material strength, including the compressive strength of concrete and the tensile strength of steel reinforcement; Reinforcement ratio, including parameters for longitudinal reinforcement and stirrups; The parametric simulation method uses a drop hammer impact dynamic analysis model based on planar membrane elements when generating the finite element simulation data.

[0012] According to some embodiments of the present invention, the training candidate machine learning model in S3 includes: S31. Train multiple candidate machine learning models, including random forest models, support vector machine models, neural network models, and gradient boosting models; S32. Perform hyperparameter search on the candidate machine learning model to obtain training results with different combinations of hyperparameters; S33. The predictive performance of the candidate machine learning models is compared by cross-validation, and the candidate machine learning model with the best predictive performance is selected as the optimal predictive model. According to some embodiments of the present invention, the feature interpretation results in S5 include: S51. Based on the Shapley value method, calculate the global and local contributions of input features to the damage mode prediction results; S52. Generate the importance ranking, dependencies, and interaction effects of input features; S53. Based on the feature interpretation results, explain the role of each input feature in the prediction of bending failure, bending-shear failure, shear failure, and punching failure.

[0013] Secondly, embodiments of the present invention provide an intelligent prediction device for the impact failure mode of RC components, the device comprising: The data acquisition module is used to acquire experimental data, finite element simulation data, and supplementary data; The data processing module is used to perform the data cleaning and sample balancing processes. The model training module is used to train multiple candidate machine learning models and determine the optimal prediction model. The prediction output module is used to introduce a cost-sensitive mechanism during the prediction phase and output the prediction results of the destruction mode. An interpretability analysis module is used to generate feature interpretation results based on the prediction results of the destruction mode.

[0014] The intelligent prediction device for impact failure mode of RC components according to embodiments of the present invention has at least the following beneficial effects: The apparatus of this invention first integrates experimental data, finite element simulation data, and augmented data through a data acquisition module, followed by data cleaning and balancing by a data processing module to improve data quality. Then, a model training module trains and cross-validates candidate machine learning models to determine the best-performing prediction model. Finally, a prediction output module introduces a cost-sensitive mechanism during the prediction phase to reduce the misjudgment rate of key failure modes, and an interpretability analysis module outputs feature importance and dependencies. This apparatus, through modular design, organically combines data augmentation, model optimization, and interpretation mechanisms, achieving efficient, accurate, and interpretable prediction of impact failure modes of RC components, thereby improving the reliability of engineering structural safety assessments and design decisions.

[0015] Thirdly, embodiments of the present invention provide a computer device, the computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the methods described above.

[0016] Fourthly, embodiments of the present invention also provide a computer-readable storage medium storing computer-executable instructions for executing the intelligent prediction method for impact failure modes of RC components as described in the first aspect embodiment. Since the computer-readable storage medium employs all the technical solutions of the intelligent prediction method for impact failure modes of RC components described in the above embodiments, it possesses at least all the beneficial effects brought about by the technical solutions of the above embodiments.

[0017] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the steps of the intelligent prediction method for impact failure mode of RC components in Embodiment 1 of the present invention; Figure 2 This is a flowchart illustrating the intelligent prediction method for impact failure modes of RC components according to Embodiment 3 of the present invention. Figure 3 The improved SMOTE-ENN principle and flowchart of Embodiment 3 of the present invention are shown below; Figure 4 This is a comparison chart of classifier performance under different data processing strategies in Embodiment 3 of the present invention. Figure labels: (a) Peak performance of classifiers under different data strategies, (b) Average performance of classifiers under different databases; Figure 5 This is a confusion matrix diagram of the nine machine learning classification models in Embodiment 3 of the present invention on the test set. The diagram is labeled as follows: (a) Confusion matrix of ANN model, (c) Confusion matrix of SVC model, (d) Confusion matrix of KNN model, (e) Confusion matrix of DT model, (f) Confusion matrix of RF model, (g) Confusion matrix of AdaBoost model, (h) Confusion matrix of CatBoost model, and (i) Confusion matrix of XGBoost model.

[0019] Figure 6 This is a prediction performance graph of the cost-sensitive random forest in Embodiment 3 of the present invention. The figure is labeled as follows: (a) Confusion matrix of prediction results of cost-sensitive random forest on the test set, (b) Comparison of F1 score between cost-sensitive random forest and ordinary random forest; Figure 7 This is a bar chart showing the importance of four categories of input variables in the database of Embodiment 3 of the present invention; Figure 8 This is a beehive graph of the four categories of input variables in the database of Embodiment 3 of the present invention. The figures are labeled as follows: (a) the effect of bending failure on model output; (b) the effect of bending-shear failure on model output; (c) the effect of shear failure on model output; and (d) the effect of punching shear failure on model output. Figure 9 The key feature of the four types of destruction modes in Embodiment 3 of the present invention is the SHAP dependency graph. The diagram shows: (a) key feature dependencies for bending failure; (b) key feature dependencies for bending-shear failure; (c) key feature dependencies for shear failure; and (d) key feature dependencies for punching-shear failure. Figure 10 This is a schematic diagram illustrating the effect of the graphical user interface of the fault mode prediction program in Embodiment 4 of the present invention. Detailed Implementation

[0020] The present invention will be further described in detail below with reference to experimental examples and specific embodiments. However, this should not be construed as limiting the scope of the above-mentioned subject matter of the present invention to the following embodiments. All technologies implemented based on the content of the present invention fall within the scope of protection of the present invention.

[0021] Unless otherwise specified, the terms "upper," "lower," "left," "right," "center," "inner," "outer," and "side" used in the description of specific embodiments of the present invention to indicate orientation or positional relationships are based on the orientation or positional relationships shown in the accompanying drawings, or the orientation or positional relationship in which the product / equipment / device is usually placed during use. These terms are merely for the purpose of facilitating the description of the present invention or simplifying the description in specific embodiments, and for enabling those skilled in the art to quickly understand the solution, and do not indicate or imply that a particular device / component / element must have a specific orientation, or be constructed and operated in a specific positional relationship. Therefore, they should not be construed as limitations on the present invention.

[0022] In the description of the embodiments of this invention, technical terms such as "first" and "second" only distinguish one entity or operation from another, and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary or secondary relationship of the indicated technical features. In the description of the embodiments of this invention, "multiple" means two or more, unless otherwise explicitly defined.

[0023] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0024] Example 1 During the research process, the applicant discovered that when using traditional physical experiments, numerical simulations, or empirical formulas to predict the failure modes of RC components under impact loads, obtaining high-precision and generalizable prediction results requires numerous costly experiments or complex numerical analyses, yielding only limited samples. Furthermore, the prediction accuracy is insufficient in areas of class imbalance and transitional failure modes. In solving practical engineering problems, existing technologies fail to meet the requirements of broad data coverage, high class identification sensitivity, and interpretable results to achieve accurate predictions under limited samples and complex load conditions. Therefore, after researching this issue, the applicant proposed an intelligent prediction method for the impact failure modes of RC components. Addressing the technical problems of scarce samples, class imbalance, and insufficient model interpretability, this method achieves high-precision prediction of RC component failure modes under impact loads through multi-source data construction and expansion, improved data cleaning and balancing strategies, training and optimization based on candidate machine learning models, the introduction of cost-sensitive prediction mechanisms, and the integration of interpretability analysis. This results in reduced misclassification rates for key categories, enhanced interpretability of prediction results, and improved reliability of structural safety assessments.

[0025] Please refer to Figure 1 , Figure 1 A schematic diagram illustrating the steps of the intelligent prediction method for impact failure modes of RC components provided in an embodiment of the present invention. The method may include: S1. Obtain a multi-source dataset related to the impact response of RC components, wherein the multi-source dataset includes experimental data and finite element simulation data; S2. Perform data cleaning and sample balancing on the multi-source dataset; S3. Train multiple candidate machine learning models based on the processed dataset, and determine the optimal prediction model through cross-validation; S4. Introduce a cost-sensitive mechanism in the decision-making stage of the optimal prediction model to set differentiated loss weights for misjudgments of different destruction modes; S5. Output the predicted impact failure mode of the RC component generated by the optimal prediction model, and generate the feature interpretation results based on interpretability analysis.

[0026] In some specific implementation schemes, including: S1. Obtain a multi-source dataset, which includes experimental data and finite element simulation data. In some embodiments, the multi-source dataset also includes sample data expanded based on punching shear failure discrimination rules, specifically including: calculating the punching shear bearing capacity of the sample to be discriminated; when the peak impact force recorded by the sample is greater than the punching shear bearing capacity, marking the sample as a punching shear failure mode and adding it to the multi-source dataset.

[0027] S2. Perform data cleaning and sample balancing on the multi-source dataset to improve data quality and class distribution balance. In some embodiments, the data cleaning process includes: firstly, oversampling minority class samples in the decision boundary region to generate supplementary samples for the boundary region; then, judging the samples based on the K-nearest neighbor rule, and when the majority neighbor class of a sample is inconsistent with its own class, the sample is judged as a noise or fuzzy sample and removed; after completing the above cleaning, the supplementary samples are deleted to avoid introducing data leakage in the subsequent modeling process. During the data cleaning process, a confidence level can also be set according to the data source, wherein experimental data is set as the highest level of protected source samples and is not removed during the cleaning process.

[0028] S3. Train multiple candidate machine learning models based on the processed dataset, and determine the optimal prediction model through cross-validation. In some embodiments, the candidate machine learning models include random forest models, support vector machine models, neural network models, and gradient boosting models. During training, hyperparameter search is first performed on the candidate models to obtain training results under different hyperparameter combinations; then, their prediction performance is compared through cross-validation, thereby selecting the candidate machine learning model with the best performance as the optimal prediction model.

[0029] S4. A cost-sensitive mechanism is introduced in the decision-making stage of the optimal prediction model to set differentiated loss weights for misclassification of different damage modes. In some embodiments, the cost-sensitive mechanism specifically includes: pre-setting risk cost parameters for misclassification of different categories; obtaining the category probabilities output by each classifier during prediction voting, and calculating the total cost of each category in combination with the risk cost parameters; selecting the category with the minimum total cost as the final prediction result. In this way, the probability of misclassifying a brittle damage mode as a ductile damage mode can be reduced, thereby improving the prediction reliability of the model under key categories.

[0030] S5. Output the failure mode prediction results generated by the optimal prediction model, and generate feature explanation results based on interpretability analysis. The interpretability analysis method can employ the SHAP (Shapley Additive Explanation) algorithm based on Shapley values ​​to quantitatively explain the feature contributions of the model output. In some embodiments, the feature explanation results include: calculating the global and local contributions of input features to the prediction results based on the Shapley value method; generating the importance ranking, dependencies, and interaction effects of input features; and explaining the role of each input feature in the prediction of bending failure, bending-shear failure, shear failure, and punching failure based on the feature explanation results. Specifically, the Shapley value method quantifies the positive or negative effects of different features in different categories of prediction by decomposing the model output into the sum of the baseline prediction probability and the contribution values ​​of each input feature, revealing the nonlinear relationship and interaction effects between features and prediction results.

[0031] The method of this invention first constructs a multi-source dataset by introducing experimental data and finite element simulation data, and then improves data quality through data cleaning and sample balancing. Next, it trains candidate machine learning models and selects the best-performing prediction model through cross-validation. A cost-sensitive mechanism is introduced in the prediction stage to effectively reduce the misclassification rate of key categories. Finally, interpretability analysis is used to output the importance and influence of features. This method, through the synergistic effect of data augmentation, model optimization, and interpretation mechanisms, achieves high-precision prediction of impact failure modes of RC components under conditions of limited samples and class imbalance, thereby improving the credibility and practicality of structural safety assessment.

[0032] The intelligent prediction method for impact failure modes of RC components provided in this invention can be applied to many technical fields, such as structural safety assessment in civil engineering, impact-resistant design of bridges, and protection of tunnels and underground structures. In the above implementation, when predicting the impact failure modes of RC components, the richness of training samples can be improved through the construction and expansion of multi-source data; the model's ability to identify minority failure modes can be enhanced through improved data cleaning and balancing strategies; prediction accuracy can be improved through the selection of candidate machine learning models; the misclassification rate of key categories can be reduced through a cost-sensitive mechanism; and the contribution pattern of input features to the prediction results can be revealed through interpretability analysis. This achieves efficient, accurate, and interpretable prediction of the failure modes of RC components under impact loads.

[0033] Example 2 As a further optimization of the preceding embodiments, this embodiment of the invention provides a specific implementation of an intelligent prediction method for impact failure modes of RC components. It can effectively solve the problems of limited sample size and class imbalance, significantly improve the prediction accuracy of failure modes, and use the SHAP (Shapley Additive Explanation) interpretable model method to quantitatively explain the decision-making mechanism of the machine learning model.

[0034] The specific technical steps are as follows: Step S1: Collect experimental data and finite element analysis results of impact on concrete structures from existing studies, compile detailed sample information (including failure mode labels), and perform characteristic correlation analysis.

[0035] Step S2: Based on the conclusions of the correlation analysis of the impact failure mode laws and characteristics summarized from existing research, select geometric dimensions, impact energy, contact boundary conditions, static bearing capacity, shear span ratio, and axial compression ratio as the input feature set.

[0036] Step S3: Using the Vector2-based impact finite element simulation method, the validated beam-column structure impact conditions are parametrically extended to generate sufficient samples covering key design parameters (geometric dimensions, material strength, reinforcement ratio).

[0037] Step S4: For punching failure, which has the scarcest samples among the four failure modes, apply the punching failure discrimination formula. Implement sample expansion: (1) Calculate the punching shear resistance of test samples in existing studies that only recorded peak impact force but did not specify the failure mode. ; (2) Determining through retrospective analysis: if the maximum impact force recorded in the sample is... If the shearing damage occurs, it can be determined that a punching failure has occurred, and targeted replenishment of this type of sample can be achieved.

[0038] Step S5: Based on the reliability differences among the multi-source datasets (experimental data, external simulation data, Vector2 simulation data, and punching formula extended data) in the expanded dataset (experimental data has the highest reliability, followed by the others), the improved SMOTE-ENN (Synthetic Minority Over-sampling Technique and Edited NearestNeighbours) method is introduced for data cleaning.

[0039] Step S6: Perform standardization preprocessing on the cleaned data and randomly divide the training set and test set.

[0040] Step S7: Use the Borderline-SMOTE algorithm to synthesize and amplify minority class destruction pattern samples (such as shearing destruction) in the training set to achieve class balance in the training set.

[0041] Step S8: Based on the augmented dataset, select 9 machine learning algorithms to build a prediction model; (1) Perform hyperparameter grid search optimization on the training set; (2) The model was trained using the optimal parameters, and its performance was evaluated through cross-validation; (3) Select the Random Forest model with the best prediction performance as the optimal predictor.

[0042] Step S9: Based on the selected best-performing Random Forest model, a parameterized cost-sensitive matrix is ​​introduced. Cost-sensitive voting is performed during the model voting phase to learn the difference in misclassification loss, thereby reducing the misclassification rate of key categories.

[0043] Step S10: Using the SHAP interpretable method, perform global and local interpretation analysis on the optimal failure mode prediction model established in Step 8 to obtain the feature importance ranking of different failure modes; quantitatively analyze the contribution and influence of input features on the prediction results.

[0044] In step S1, based on the crack development of the specimen in the experiment or the distribution of effective plastic strain in the structure in the finite element results, it is determined whether the sample has bending failure, bending-shear failure, shear failure or punching failure; the characteristic correlation analysis is performed using the Pearson correlation coefficient.

[0045] In step S2, the specific model input parameters include the cross-sectional height. h Cross-sectional width b structural radius R s Structural span s , shear span ratio λ Static bending bearing capacity B r Static shear bearing capacity S r Contact stiffness M c Contact area A c Radius of the impacting object (hammer head) R h Concrete compressive strength f c axial compression ratioF a Mass of the impacting object (hammer) M Velocity of the impacting object (hammer) V The static bending and shear bearing capacities are calculated with reference to the Design Code for Reinforced Concrete and Prestressed Concrete Bridges and Culverts of Highways (JTG 3362-2018) and the Design Code for Collision Resistance of Highway Bridges (JTG / T 3360-02-2020). The specific calculation formulas are shown below: (1) Bending capacity: (1) (2) In the formula The design value of bending moment ( ); Design value of axial compressive strength of concrete ( ); The design values ​​of tensile and compressive strength of longitudinal ordinary steel bars ( ); Effective height of the cross section ( ); The height of the compression zone of the cross section ( ), The cross-sectional areas of the longitudinal ordinary steel bars in the tension zone and compression zone are respectively ( For specific methods of determining material parameters, please refer to Section 5.2.2 of the Design Specification for Highway Reinforced Concrete and Prestressed Concrete Bridges and Culverts (JTG 3362-2018).

[0046] (2) Shear bearing capacity: (3) (4) (5) (6) In the formula This represents the design value of the shear capacity of the component. This represents the design value of the shear capacity of the concrete portion. This represents the design value of the shear capacity of the stirrup section. The design value of the shear bearing capacity of the steel frame is given, and the unit of the design bearing capacity is kN. The shear impact reduction factor is taken as 0.7; The shear strength (kPa) of concrete can be found in Appendix A.1.1 of the Code for Collision Resistance Design of Highway Bridges (JTG / T 3360-02—2020). The gross cross-sectional area of ​​the component ( ) ; The number of independent stirrup rings; The total cross-sectional area of ​​the spiral or annular stirrups on the same cross section ( ); The design value of the tensile strength of the stirrup ( ); The diameter of the stirrup ring measured from the centerline of the stirrup ring ( ); s Stirrup spacing ( ); The total cross-sectional area of ​​the stirrups on the same section in the loading direction ( ); The effective shear height of the cross section in the loading direction is the distance (m) from the edge of the compression zone to the point of application of the resultant force of the tensile reinforcement. The design value of the yield strength of the steel frame ( ); The web height of the steel frame ( ); The web height of the steel frame ( ). In step S3, rectangular elements are used to model the concrete, drop hammer, and support plate. Longitudinal reinforcement is simulated using truss elements, and stirrups are embedded into the concrete matrix in a dispersed manner. The element mass is distributed using the lumped mass method at nodes; drop hammer nodes are only subject to mass constraints in the y-direction, while other nodes are considered for bidirectional mass distribution in both x and y directions.

[0047] The drop hammer consists of three layers of rigid rectangular elements with extremely high material stiffness to simulate rigid body behavior. The drop hammer and the specimen are connected by compression-only truss elements, enabling automatic separation after impact. The steel plate-concrete interface is modeled using uniaxial stiffness bearing elements to more realistically reflect the stress distribution in the loading region. While ensuring clear representation of the structural stress characteristics, the model was appropriately simplified; the boundaries of beams and columns are simulated using compression-only truss members and steel plate supports, respectively, improving computational efficiency and numerical stability.

[0048] For material constitutive modeling, the Hognestad parabolic model (applicable to compressive strength ≤ 40 MPa) was used for ordinary concrete; the Montoya / Ottosen (1979) model was used for confined concrete. The material properties and mechanical behavior of the reinforcing steel were set using the default parameters of the VecTor2 software.

[0049] In step S4, the formula... k T It is a dimensionless coefficient that includes inertial force, shear strength of steel reinforcement, etc., and is taken as 6.5 according to existing research; f t The tensile strength of concrete, D andW These represent the depth and width of the component's cross-section, respectively.

[0050] Step S5 first involves oversampling the minority class samples by using linear interpolation to generate synthetic samples among the minority class samples. This balances the class distribution and enhances the model's ability to distinguish the minority class. After oversampling, the dataset is cleaned using an undersampling method based on the K-nearest neighbor rule—Edited Nearest Neighbors (ENN). Specifically, ENN examines each sample; if its majority nearest neighbors belong to different classes, the sample is considered to be noise or a point with ambiguous boundaries and is removed from the training set. Based on the standard SMOTE-ENN method, we have made the following key improvements: (1) The standard SMOTE method in the first step is changed to Borderline-SMOTE to more effectively handle minority class samples located at the decision boundary. The algorithm is only performed on minority class samples near the boundary to generate new samples. While enriching the samples in the boundary region, it also improves the ability of the subsequent ENN algorithm to identify and remove noise and fuzzy boundary points.

[0051] (2) Protection mechanism. We introduced the “protected source” mechanism. Based on the different sources of data mentioned above, we believe that the experimental data has the highest accuracy and reliability. We marked the samples from the experimental source as “protected source” to prevent accidental deletion, retain reliable samples, and avoid the loss of important information.

[0052] (3) Oversampling data deletion mechanism. The first step of the above method is to use Borderline-SMOTE to synthesize new samples in the boundary region to enhance the sample distribution in the boundary region and improve the ability of ENN to remove noise and ambiguity. However, if these synthesized samples are retained in subsequent modeling, they may lead to data leakage due to implicit test set information, thereby overestimating the model performance. Therefore, after Borderline SMOTE-ENN processing is completed, all synthesized samples should be deleted to avoid the risk of information leakage.

[0053] In step S6, the data is processed using the Z-score standardization method, and the training set and test set are randomly divided, with 80% used as the training set and 20% as the test set.

[0054] In step S7, only minority class oversampling is performed on the training set to prevent data leakage.

[0055] In step S8, hyperparameter optimization employs a strategy combining randomized search and grid search. First, randomized search (RandomizedSearchCV) efficiently samples within a predefined parameter space (the number of samples is controlled by setting n_iter) to quickly locate globally optimal regions. Then, based on the preliminary results, grid search (GridSearchCV) refines the search for local parameter combinations to achieve a balance between efficiency and accuracy. Throughout the process, 10-fold cross-validation is used to evaluate model stability and generalization ability. The dataset is divided into 10 subsets, which alternately serve as the validation set, while the remainder serves as the training set. The final average evaluation result is taken to reduce sample splitting bias and ensure robust evaluation.

[0056] The selected evaluation metrics include accuracy, precision, recall, and F1 score, with the specific expressions as follows: (7) (8) (9) (10) In the formula, TP A true positive is defined as a case where both the actual and predicted classes are positive. TN A true negative example is one in which both the true class and the predicted class are negative. FP A false positive is a case where the true class is negative but the predicted class is positive. FN A false negative is a case where the true class is positive but the predicted class is negative.

[0057] In step S9, the parameterized loss matrix is ​​set as follows: (11) In the formula, the adjustable parameters reflect the engineering risk cost, and the parameters satisfy the constraints. This is to quantify the relative severity of different types of misjudgments.

[0058] Parameter values ​​in the loss matrix The optimal parameter combination was determined using 10-fold cross-validation, which directly affects the effectiveness of the model. Based on the established optimal cost matrix, the following operations were performed during the voting phase of the random forest model: (1) Obtain the predicted category for each tree: (12) (2) Calculate the sample x Categorize into Category j Total cost: (13) (3) Select the class with the lowest total cost as the final prediction result: (14) In the formula, For input parameters, K The number of trees in the random forest model. For the random forest model k Trees were used to quantify the relative severity of different types of misjudgments. The categories in the loss matrix Predicted as j The loss.

[0059] In step S10, SHAP global and local interpretability analysis was performed by category to clearly reveal the specific contribution patterns of different features to the prediction results of each category: (1) First, feature importance analysis was performed. Bar charts of feature importance ranking based on average SHAP value and beehive plots of SHAP value for each damage mode were drawn for different categories of samples. In the bar charts, features with higher average SHAP value represent higher feature importance. There are significant differences in feature importance ranking among different classes. In the beehive plots, each point represents a sample, showing the correlation and distribution of feature values ​​and SHAP values ​​of each parameter for each sample.

[0060] (2) Subsequently, SHAP dependency graphs of the four features with the highest global contribution for each of the four failure modes (bending failure, bending-shear failure, shear failure, and punching failure) were constructed to deeply analyze the nonlinear dependency relationship and mechanism between feature values ​​and corresponding SHAP values.

[0061] In the main effects dimension, the vertical axis SHAP value quantifies the direction and strength of the contribution of features to the predicted probability of a specific damage mode—when the eigenvalue > 0, the eigenvalue drives the model to increase the predicted probability of that mode; when the SHAP value < 0, it has an inhibitory effect. The functional relationship between eigenvalues ​​and SHAP values ​​(such as exhibiting monotonicity, threshold abruptness, or nonlinear oscillation) essentially reflects the mathematical representation of the physical mechanism.

[0062] In the interaction effect dimension, color coding is performed based on the automatically recognized strong interaction features (interaction_index), and the color gradient changes reveal the coupling relationship between the horizontal axis features and the associated parameters.

[0063] The intelligent prediction method for impact failure modes of RC components provided in this invention effectively solves the problems of limited sample size, imbalanced category distribution, and insufficient model interpretability commonly found in current research, thus significantly improving the accuracy of predicting RC structure failure modes under impact. Through the improved SMOTE-ENN data cleaning method, it effectively handles the confidence differences between different data sources, enhances the ability to identify and remove noisy and ambiguous samples, and lays a data foundation for high-precision model prediction. Furthermore, it adopts a full-chain technical paradigm of "multi-source data augmentation - data cleaning - model optimization - mechanism interpretation," and through synergistic innovation in data augmentation and model interpretability, accurately predicts the failure modes of RC structures under impact. This provides a general framework for machine learning applications in small-sample scenarios and has direct guiding value for fields such as impact-resistant design and safety assessment in civil engineering.

[0064] Example 3 As a further optimization of the preceding embodiments, this embodiment of the invention provides a specific implementation of a method for intelligent prediction of impact failure modes of RC components, including the following: Step S1: Collect experimental data and finite element analysis results of impact on concrete structures from existing studies, and statistically analyze detailed sample information. Based on the crack development of the specimens in the tests or the distribution of effective plastic strain in the structure in the finite element results, determine whether the samples experienced bending failure, bending-shear failure, shear failure, or punching shear failure. In the end, a total of 417 impact test and finite element simulation cases from 46 studies were collected, including 311 impact tests and 106 finite element simulations, covering concrete beam and column structures with cross-sectional shapes including square and circular. The number of samples corresponding to the four failure modes were 153, 91, 132, and 41, respectively. The Pearson correlation coefficient was used for characteristic correlation analysis.

[0065] Step S2: Based on the correlation analysis conclusions of impact failure mode laws and characteristics summarized from existing research, select geometric dimensions, impact energy, contact boundary conditions, static bearing capacity, shear span ratio, and axial compression ratio as the input feature set. Specific model input parameters include section height. h Cross-sectional width b structural radius R s Structural span s , shear span ratio λ Static bending bearing capacity B r Static shear bearing capacity S r Contact stiffness M cContact area A c Radius of the impacting object (hammer head) R h Concrete compressive strength f c axial compression ratio F a Mass of the impacting object (hammer) M Velocity of the impacting object (hammer) V The static bending and shear bearing capacities are calculated with reference to the Design Code for Reinforced Concrete and Prestressed Concrete Bridges and Culverts of Highways (JTG 3362-2018) and the Design Code for Collision Resistance of Highway Bridges (JTG / T 3360-02-2020). The specific calculation formulas are shown below: (1) Bending capacity: (16) (17) In the formula The design value of bending moment ( ); Design value of axial compressive strength of concrete ( ); The design values ​​of tensile and compressive strength of longitudinal ordinary steel bars ( ); Effective height of the cross section ( ); The height of the compression zone of the cross section ( ), The cross-sectional areas of the longitudinal ordinary steel bars in the tension zone and compression zone are respectively ( For specific methods of determining material parameters, please refer to Section 5.2.2 of the Design Specification for Highway Reinforced Concrete and Prestressed Concrete Bridges and Culverts (JTG 3362-2018).

[0066] (2) Shear bearing capacity: (18) (19) (20) (twenty one) In the formula This represents the design value of the shear capacity of the component. This represents the design value of the shear capacity of the concrete portion. This represents the design value of the shear capacity of the stirrup section. The design value of the shear bearing capacity of the steel frame is given, and the unit of the design bearing capacity is kN. The shear impact reduction factor is taken as 0.7; The shear strength (kPa) of concrete can be found in Appendix A.1.1 of the Code for Collision Resistance Design of Highway Bridges (JTG / T 3360-02—2020). The gross cross-sectional area of ​​the component ( ) ; The number of independent stirrup rings; The total cross-sectional area of ​​the spiral or annular stirrups on the same cross section ( ); The design value of the tensile strength of the stirrup ( ); The diameter of the stirrup ring measured from the centerline of the stirrup ring ( ); s Stirrup spacing ( ); The total cross-sectional area of ​​the stirrups on the same section in the loading direction ( ); The effective shear height of the cross section in the loading direction is denoted as , and the distance (m) from the edge of the compression zone to the point of application of the resultant force of the tensile reinforcement is denoted as . The design value of the yield strength of the steel frame ( ); The web height of the steel frame ( ); The web height of the steel frame ( ). Step S3: Using the Vector2-based impact finite element simulation method, the mechanical response of the verified beam-column structure under impact conditions is parametrically extended to generate sufficient samples covering key design parameters (including geometric dimensions, material strength, and reinforcement ratio). The specific simulation method is as follows.

[0067] The simulation uses rectangular elements to model the concrete, drop hammer, and support plate. Longitudinal reinforcement is simulated using truss elements, with stirrups embedded in the concrete matrix in a dispersed manner. Element mass is distributed using the lumped mass method at nodes; drop hammer nodes are constrained only in the y-direction, while other nodes are considered for bidirectional mass distribution in both x and y directions.

[0068] The drop hammer consists of three layers of rigid rectangular elements with extremely high material stiffness to simulate rigid body behavior. The drop hammer and the specimen are connected by compression-only truss elements, enabling automatic separation after impact. The steel plate-concrete interface is modeled using uniaxial stiffness bearing elements to more realistically reflect the stress distribution in the loading region. While ensuring clear representation of the structural stress characteristics, the model was appropriately simplified; the boundaries of beams and columns are simulated using compression-only truss members and steel plate supports, respectively, improving computational efficiency and numerical stability.

[0069] For material constitutive modeling, the Hognestad parabolic model (applicable to compressive strength ≤ 40 MPa) was used for ordinary concrete; the Montoya / Ottosen (1979) model was used for confined concrete. The material properties and mechanical behavior of the reinforcing steel were set using the default parameters of the VecTor2 software.

[0070] Step S4: For punching failure, which has the scarcest samples among the four failure modes, apply the punching failure discrimination formula. Implement sample expansion: (1) Calculate the punching shear resistance of test samples in existing studies that only recorded peak impact force but did not specify the failure mode. ; (2) Determining through retrospective analysis: if the maximum impact force recorded in the sample is... If the shearing damage occurs, it can be determined that a punching failure has occurred, and targeted replenishment of this type of sample can be achieved.

[0071] In the formula k T It is a dimensionless coefficient that includes inertial force, shear strength of steel reinforcement, etc., and is taken as 6.5 according to existing research; f t The tensile strength of concrete, D and W These represent the depth and width of the component's cross-section, respectively.

[0072] Step S5: Given the varying reliability of the multi-source datasets (experimental data, external simulation data, Vector2 simulation data, and punching formula-extended data) in the expanded dataset (experimental data has the highest reliability, followed by the others), an improved SMOTE-ENN (Synthetic Minority Over-sampling Technique and Edited Nearest Neighbours) method is introduced for data cleaning. The principle and process of this method are as follows: Figure 3 As shown. Figure 3 An improved data processing flow is shown, which is as follows: First, based on the distribution of majority and minority class samples, the Borderline-SMOTE method is used to oversample the minority class samples in the decision boundary region to generate new boundary minority class samples; then, the ENN (Edited Nearest Neighbour) method is used to clean the dataset, removing noise points and blurred boundary samples, thus obtaining the ENN-cleaned sample set; finally, the aforementioned synthetic samples are deleted through a clean mechanism to avoid potential data leakage and ensure the reliability and generalization of the training set.

[0073] Specifically, this method first oversamples the minority class samples, using linear interpolation to generate synthetic samples among the minority class samples, thereby balancing the class distribution and enhancing the model's ability to distinguish the minority class. After oversampling, an undersampling method based on the K-nearest neighbor rule—Edited Nearest Neighbours (ENN)—is used to clean the dataset. Specifically, ENN examines each sample; if its majority nearest neighbors belong to different classes, the sample is considered to be noise or a point with blurred boundaries, and it is removed from the training set. Based on the standard SMOTE-ENN method, we have made the following key improvements: (1) The standard SMOTE method in the first step is changed to Borderline-SMOTE to more effectively handle minority class samples located at the decision boundary. The algorithm is only performed on minority class samples near the boundary to generate new samples. While enriching the samples in the boundary region, it also improves the ability of the subsequent ENN algorithm to identify and remove noise and fuzzy boundary points.

[0074] (2) Protection mechanism. We introduced the “protected source” mechanism. Based on the different sources of data mentioned above, we believe that the experimental data has the highest accuracy and reliability. We marked the samples from the experimental source as “protected source” to prevent accidental deletion, retain reliable samples, and avoid the loss of important information.

[0075] (3) Oversampling data deletion mechanism. The first step of the above method is to use Borderline-SMOTE to synthesize new samples in the boundary region to enhance the sample distribution in the boundary region and improve the ability of ENN to remove noise and ambiguity. However, if these synthesized samples are retained in subsequent modeling, they may lead to data leakage due to implicit test set information, thereby overestimating the model performance. Therefore, after Borderline SMOTE-ENN processing is completed, all synthesized samples should be deleted to avoid the risk of information leakage.

[0076] The effectiveness of the strategy, such as Figure 4 As shown, Figure 4 (a) represents the peak performance of the classifier under different data strategies. Figure 4 (b) represents the average performance of the classifier across different databases. Specifically, Figure 4(a) The vertical axis uses a two-character naming convention to identify different strategy combinations. The first character indicates the application of the data augmentation strategy: "N" indicates the original dataset without data augmentation and cleaning, "M" indicates the dataset with data augmentation but not cleaning, and "I" indicates the augmented dataset with both augmentation and cleaning. The second character indicates the use of the training set balancing strategy: "N" indicates no training set balancing strategy, "S" indicates the SMOTE balancing strategy, and "B" indicates the Borderline-SMOTE balancing strategy. Experimental results show that after processing the augmented database using the proposed improved SMOTE-ENN data cleaning strategy, the model trained on this data improves performance by 7 to 11 percentage points on key evaluation metrics such as Accuracy, Recall, and Precision. Figure 4 As shown in (b), the best performance was achieved when the augmented database was combined with the Borderline SMOTE balance, with accuracy, recall, and precision improved by 12.3%, 19.2%, and 16.5% respectively compared to the baseline, fully validating the effectiveness of this method in improving the model's generalization ability and classification performance.

[0077] Step S6: The cleaned data is processed using the Z-score standardization method, and the training set and test set are randomly divided in an 8:2 ratio.

[0078] Step S7: Use the Borderline-SMOTE algorithm to synthesize and amplify only the minority of destruction pattern samples (such as shearing destruction) in the training set to achieve class balance in the training set and prevent data leakage.

[0079] Step S8: Based on the augmented dataset, select 9 machine learning algorithms to build a prediction model; (1) Perform hyperparameter grid search optimization on the training set; (2) The model was trained using the optimal parameters, and its performance was evaluated through cross-validation; (3) Select the random forest model with the best prediction performance as the final predictor.

[0080] Hyperparameter optimization employs a strategy combining randomized search and grid search. First, randomized search (RandomizedSearchCV) efficiently samples within a predefined parameter space (the number of samples is controlled by setting `n_iter`) to quickly locate globally optimal regions. Then, based on the initial results, grid search (GridSearchCV) refines the search for local parameter combinations to achieve a balance between efficiency and accuracy. Throughout the process, 10-fold cross-validation is used to evaluate model stability and generalization ability. The dataset is divided into 10 subsets, which alternately serve as the validation set, while the remainder serves as the training set. The final evaluation result is the average of the results to reduce sample splitting bias and ensure robust evaluation.

[0081] The selected evaluation metrics include accuracy, precision, recall, and F1 score, with the specific expressions as follows: (twenty two) (twenty three) (twenty four) (25) In the formula, TP A true positive is defined as a case where both the actual and predicted classes are positive. TN A true negative example is one in which both the true class and the predicted class are negative. FP A false positive is a case where the true class is negative but the predicted class is positive. FN A false negative is a case where the true class is positive but the predicted class is negative.

[0082] Based on the processed dataset, nine machine learning models, including Artificial Neural Network (ANN), Support Vector Classification (SVC), K-Nearest Neighbors (KNN), Decision Tree (DT), Random Forest (RF), Adaptive Boosting (AdaBoost), Light Gradient Boosting Machine (LightGBM), Categorical Gradient Boosting (CatBoost), and Extreme Gradient Boosting (XGBoost), were trained and evaluated. The evaluation results on the test set are as follows: Figure 5 As shown, the results are presented in the form of a confusion matrix. The random forest model achieved a prediction accuracy of 91.4%, a mean precision of 92.0%, and a recall of 92.8% on the test set.

[0083] Step S9: Based on the established best-performing random forest model, a parameterized cost-sensitive matrix is ​​introduced. Specifically, 10-fold cross-validation is used to determine the optimal parameter combination of the cost-sensitive matrix. Cost-sensitive voting is performed during the model voting phase to learn the difference in misclassification loss, thereby reducing the misclassification rate of key categories. The established optimal cost matrix is ​​shown below: (26) Based on the optimal cost matrix established through grid search cross-validation, during the voting phase of the random forest model, the class prediction probability of each tree is obtained, and the sample is calculated. x Categorize into Category j The total cost is calculated, and the class with the lowest total cost is selected as the final prediction result. Figure 6 (a) shows the prediction results of cost-sensitive random forest on the test set. Cost-sensitive voting reduces the second diagonal elements in the confusion matrix, effectively reducing the critical misclassification of brittle failure as ductile failure, and improving the prediction accuracy from 91.4% to 93.1%. Figure 6 (b) Further, it is shown that, except for punching (which already has 100% accuracy), the F1-scores of all failure modes are improved (1-4%), with the largest improvement observed in the prediction of bending-shear failure. This demonstrates the effectiveness of the developed method in identifying complex failure modes with both ductile and brittle characteristics.

[0084] Step S10: Using the SHAP interpretable method, the optimal damage mode prediction model established in Step S8 was subjected to global and local interpretive analysis by category to clearly reveal the specific contribution patterns of different features to the prediction results of each category. This yielded the ranking of feature importance for different damage modes and the quantitative analysis results and influence patterns of the contribution of input features to the prediction results. (1) First, feature importance analysis was performed, and bar charts ranking feature importance based on average SHAP values ​​for different categories of samples and beehive plots of SHAP values ​​for each disruption mode were drawn, as shown below. Figure 7 and Figure 8 As shown. Figure 7 In the bar chart shown, features with higher average SHAP values ​​represent greater feature importance. Figure 8 (a) The effect of bending failure on model output. Figure 8 (b) The impact of bending and shear failure on the model output. Figure 8 (c) The effect of shear failure on model output. Figure 8 The impact of (d) punching failure on the model output is shown in the bee colony diagram, where each point represents a sample, illustrating the correlation and distribution of the eigenvalues ​​and SHAP values ​​of each sample's parameters. Further analysis of feature importance under different failure modes reveals significant differences in feature importance ranking across different classes: In both bending and flexural-shear failure, impact velocity, flexural capacity, and shear capacity are the three most important characteristics, indicating that these three factors have a common influence on determining these failure modes. However, in flexural-shear failure, the impact mass, span, and shear span ratio are significantly more important than in bending failure.

[0085] In shear failure, the flexural capacity, shear capacity, and shear span ratio of a member have the most significant impact. Notably, the characteristic importance and SAP value of the shear span ratio in this type of failure are much higher than in other failure types.

[0086] In punching failure, the impact velocity SHAP value and importance are significantly amplified compared to other failure types, and it plays a decisive role in this category, with its value far exceeding the contribution value of other characteristics.

[0087] The above phenomena are consistent with the results and conclusions of relevant studies.

[0088] (2) The feature with the highest global contribution was plotted—impact velocity. V The SHAP dependency graphs for the four failure modes (bending failure, bending-shear failure, shear failure, and punching failure) are as follows: Figure 9 As shown, where Figure 9 (a) Key feature dependencies of bending failure, Figure 9 (b) Key characteristic dependencies of bending-shear failure, Figure 9 (c) Key feature dependencies of shear failure, Figure 9 The key feature dependency relationship of (d) punching failure is used to deeply analyze the nonlinear dependency relationship and mechanism between eigenvalues ​​and corresponding SHAP values.

[0089] In the main effects dimension, the vertical axis SHAP value quantifies the direction and strength of the contribution of features to the predicted probability of a specific damage mode: when the eigenvalue > 0, the eigenvalue drives the model to increase the predicted probability of that mode; when the SHAP value < 0, it has an inhibitory effect. The functional relationship between eigenvalues ​​and SHAP values ​​(such as exhibiting monotonicity, threshold abruptness, or nonlinear oscillation) essentially reflects the mathematical representation of the physical mechanism.

[0090] In the interaction effect dimension, color coding is performed based on automatically identified strong interaction features (interaction_index), and the color gradient changes reveal the coupling relationship between the horizontal axis features and the associated parameters. Dependency graph analysis shows that the same feature has significantly different effects on the SHAP value under different failure modes. In bending failure, the SHAP value increases with... V The increasing and continuously decreasing trend indicates that the higher the impact velocity, the lower the probability of bending failure; in bending-shear failure, the SHAP value is related to lg( V） There is a threshold relationship, lg( V When ) < 0.8, it increases, lg( V The SAP value decreases when it exceeds 0.8, reflecting the transition from ductile bending to brittle shear failure; in shear failure, the SAP value is around 0.5. <lg( V The concentration of SHAP values ​​is relatively high in the range of <0.9; however, in punching failure, the SHAP value increases with... V The continued rise indicates that high-speed impacts are more likely to cause this type of damage.

[0091] Further analysis revealed that other key characteristics also exhibited failure mode-dependent response patterns: such as flexural bearing capacity. B r The failure exhibits a distinct threshold characteristic in punching shear failure, a negative correlation in bending failure, and a positive correlation in shear failure; shear capacity is positively correlated in bending and negatively correlated in shear; the shear span ratio is linearly positively correlated with the SHAP value in shear. These results reveal the complexity of the failure mechanism and emphasize the importance of modal destructive characteristics in contributing to the failure.

[0092] Experimental results show that the prediction method proposed in this invention can accurately predict the failure mode of RC components under impact load, and has good interpretability, which is consistent with the conclusions of various studies.

[0093] Example 4 As a further optimization of the preceding embodiments, this application provides a graphical user interface (GUI) for intelligent prediction of impact-induced failure modes, such as... Figure 10 As shown. This GUI is developed based on the Python language and built using the PyQt5 library, enabling visualization of the impact failure mode prediction model for reinforced concrete (RC) components.

[0094] In the GUI, users can input structural and load parameters, including but not limited to geometric dimensions, material strength, and impact conditions. The input range is consistent with the model parameters in the previous embodiment. After receiving the input, the system can call the trained machine learning model in real time to perform predictions and output the impact failure mode of the RC component and the corresponding feature interpretation results in the interface.

[0095] This GUI, as an accessible tool, simplifies the deployment and use of models, providing non-professional users with an intuitive interactive approach and real-time analysis feedback, thereby improving the practicality and engineering scalability of the intelligent prediction method for impact failure modes.

[0096] Example 5 As a further optimization of the preceding embodiments, this embodiment of the invention provides an intelligent prediction device for the impact failure mode of RC components. The device includes: The data acquisition module is used to acquire experimental data, finite element simulation data, and supplementary data. For example, experimental data can come from publicly available literature on impact tests of RC components, finite element simulation data can be generated through parametric simulation using finite element software such as Vector2, and supplementary data can be used to supplement samples with missing labels based on punching shear failure discrimination rules.

[0097] The data processing module performs the data cleaning and sample balancing processes to improve data reliability and class distribution balance. For example, an improved SMOTE-ENN method can be used to oversample minority class samples in the decision boundary region, then use the K-nearest neighbor rule to remove noisy samples, and delete the generated synthetic samples after processing. At the same time, the experimental data is set as a protected source to prevent accidental deletion.

[0098] The model training module is used to train multiple candidate machine learning models and determine the optimal prediction model through cross-validation. For example, random forests, support vector machines, neural networks, and gradient boosting models can be selected as candidate models. Hyperparameters are optimized through grid search or random search, and 10-fold cross-validation is used to evaluate model performance.

[0099] The prediction output module introduces a cost-sensitive mechanism during the prediction phase, assigning differentiated weights to misclassifications of different damage modes and outputting the predicted damage modes. For example, a cost-sensitive matrix can be set up to add risk weights to different misclassification types during classification voting, ultimately selecting the category with the lowest total cost as the output, thereby reducing the probability of brittle damage modes being misclassified.

[0100] The interpretability analysis module is used to generate feature interpretation results based on the failure mode prediction results, thereby revealing the contribution patterns of input features to the prediction results. For example, the Shapley value method can be used to rank the importance of each input feature, generating dependency graphs and interaction effect graphs to illustrate the different mechanisms of action of key features such as impact velocity, span, and flexural bearing capacity in bending failure, flexural-shear failure, shear failure, and punching shear failure.

[0101] It should be understood that the modules provided in the above embodiments are only illustrative examples of the functional division described above. In practical applications, the functions can be performed by different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above.

[0102] The functional modules in the above embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of the embodiments of the present invention.

[0103] Based on the same application concept, embodiments of the present invention also provide a computer device, which may include a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the intelligent prediction method for impact failure modes of RC components as described above.

[0104] Based on the same concept, embodiments of the present invention also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the intelligent prediction method for impact failure modes of RC components as described above.

[0105] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects.

[0106] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0107] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. An intelligent prediction method for impact failure modes of RC components, characterized in that, include: S1. Obtain a multi-source dataset related to the impact response of RC components, wherein the multi-source dataset includes experimental data, finite element simulation data, and sample data expanded based on the punching shear failure discrimination rule; The finite element simulation data is generated using a parametric simulation method, which covers the following key parameters: Geometric dimensions, including cross-sectional height and cross-sectional width; Material strength, including the compressive strength of concrete and the tensile strength of steel reinforcement; Reinforcement ratio, including parameters for longitudinal reinforcement and stirrups; The parametric simulation method uses a drop hammer impact dynamic analysis model based on planar membrane elements when generating the finite element simulation data. The punching failure discrimination rules include: S11. Calculate the punching shear capacity of the sample to be judged; S12. When the peak impact force recorded in the sample is greater than the punching shear bearing capacity, the sample is marked as punching shear failure mode; S13. Add the labeled samples to the multi-source dataset; S2. Perform data cleaning and sample balancing on the multi-source dataset; The data cleaning process includes: S21. In the decision boundary region, oversampling is performed on minority class samples to generate supplementary samples for the boundary region; S22. Based on the neighborhood consistency rule, determine the neighbor category of each sample. If the majority of the neighbor categories are inconsistent with the sample category, then the sample is removed as a noise sample or a fuzzy sample. S23. After completing the above cleaning, delete the supplementary sample; The data cleaning process sets a credibility level based on the data source, with experimental data being set as the highest level of protected source sample and not being removed during the cleaning process; S3. Train multiple candidate machine learning models based on the processed dataset, and determine the optimal prediction model through cross-validation; S4. Introduce a cost-sensitive mechanism in the decision-making stage of the optimal prediction model to set differentiated loss weights for misjudgments of different destruction modes; S5. Output the predicted impact failure mode of the RC component generated by the optimal prediction model, and generate the feature interpretation results based on interpretability analysis.

2. The method according to claim 1, characterized in that, The candidate machine learning models trained in S3 include: S31. Train multiple candidate machine learning models, including random forest models, support vector machine models, neural network models, and gradient boosting models; S32. Perform hyperparameter search on the candidate machine learning model to obtain training results with different combinations of hyperparameters; S33. The predictive performance of the candidate machine learning models is compared by cross-validation, and the candidate machine learning model with the best predictive performance is selected as the optimal predictive model.

3. The method according to claim 1, characterized in that, The feature interpretation results described in S5 include: S51. Based on the Shapley value method, calculate the global and local contributions of input features to the damage mode prediction results; S52. Generate the importance ranking, dependencies, and interaction effects of input features; S53. Based on the feature interpretation results, explain the role of each input feature in the prediction of bending failure, bending-shear failure, shear failure, and punching failure.

4. An intelligent prediction device for impact failure modes of RC components, characterized in that, For carrying out the method of claim 1, comprising: The data acquisition module is used to acquire experimental data, finite element simulation data, and sample data expanded based on the punching failure discrimination rule; The data processing module is used to perform data cleaning and sample balancing. The model training module is used to train multiple candidate machine learning models and determine the optimal prediction model. The prediction output module is used to introduce a cost-sensitive mechanism during the prediction phase and output the prediction results of the destruction mode. An interpretability analysis module is used to generate feature interpretation results based on the prediction results of the destruction mode.

5. A computer device, characterized in that, The computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the method as described in any one of claims 1 to 3.

6. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method as described in any one of claims 1 to 3.

Images