A multi-strategy fusion RC column earthquake damage mode recognition method

Abstract

The application relates to a multi-strategy fusion RC column earthquake damage mode recognition method, and relates to the technical field of artificial intelligence, and comprises the following steps: (1) establishing a pseudo-static test database; (2) determining input parameters and output parameters, screening the design parameters of the RC column through the standard mutual information and chi-square test method, and finally determining the input parameters and the output parameters of the model; (3) carrying out data preprocessing on the model data set composed of the input and output parameters; and (4) establishing a machine learning model to recognize the earthquake damage mode of the RC column. The application provides a multi-strategy fusion RC column earthquake damage mode recognition method, the BSCFB earthquake damage mode recognition model is established, data oversampling, cost-sensitive learning and model integration are fused, the accuracy of the RC column earthquake damage mode recognition is improved, the machine learning model is explained, and the credibility of the model is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and particularly relates to a multi-strategy fusion RC column earthquake damage mode identification method. BACKGROUND

[0002] Reinforced concrete (RC) column is an important load-bearing component of engineering structure. Due to different design parameters of RC column, the seismic performance of RC column is also different. Under the action of strong earthquake, three damage modes of bending failure, shear failure or flexural-shear failure may occur. Since shear failure and flexural-shear failure belong to brittle failure, they should be avoided as much as possible in the design of actual engineering. In order to better realize performance-based seismic design, it is crucial to determine the influence degree of component design parameters on earthquake damage mode and accurately identify the earthquake damage mode of RC column according to the design parameters.

[0003] The traditional methods for identifying the seismic failure modes of RC columns include the observation method, the methods based on the shear-span ratio, the deformation and the load-carrying capacity, the multi-parameter identification method, etc. The observation method is an important method for identifying the seismic failure modes of RC columns. For example, Zhang et al. conducted the quasi-static test of RC columns, and classified the failure modes of the columns according to the sequence of the yielding of the stirrups and longitudinal reinforcements and the development of the cracks. The observation method is simple and practical, but it is subjective, especially when the failure mode of a column has no obvious characteristics. Moreover, this method is not suitable for identifying the seismic failure modes of RC columns in existing or newly-built structures. Wan et al. and Ghee et al. identified the seismic failure modes of RC columns based on the shear-span ratio and the displacement ductility coefficient, respectively. Some literatures used the ratio of the shear demand to the shear capacity to identify the seismic failure modes of RC columns. The above methods mainly used a single performance index of RC columns to identify their seismic failure modes. Since these methods did not comprehensively consider the main factors affecting the seismic failure modes of RC columns, their identification accuracy was relatively low. Therefore, many scholars began to study the multi-parameter identification method for the seismic failure modes of RC columns. Zhu et al. considered the shear-span ratio, the stirrup ratio and the shear capacity, and classified the seismic failure modes of RC columns into two types, i.e., the flexural failure and the shear failure. Although this method considered the influence of multiple parameters on the seismic failure modes, it did not subdivide the shear and flexural-shear failure modes. Qi et al. analyzed the main factors affecting the seismic failure modes of RC columns based on the Fisher method, and proposed an identification method considering the ratio of the shear demand to the shear capacity and the shear-span ratio. Ma et al. integrated multiple key parameters of RC columns into an empirical index, and proposed a method for identifying the seismic failure modes of RC columns from the perspective of probability. However, the shear failure mechanism of RC columns is complex, and the calculation model of the shear capacity of RC columns has large dispersion and low accuracy. In addition, there are many factors affecting the seismic failure modes of RC columns, including the shear-span ratio, the axial compression ratio, the concrete strength, the stirrup configuration and the longitudinal reinforcement configuration, etc. The existing identification methods cannot consider the comprehensive influence of these factors. Therefore, the accuracy of the traditional multi-parameter identification method is difficult to guarantee.

[0004] With the wide application of machine learning method in civil engineering field, many scholars began to use it to identify the seismic failure mode of RC column. This method can excavate the potential and valuable information from numerous design parameters, and use these information to classify the seismic failure mode of RC column, so as to avoid the identification error caused by inaccurate calculation of shear capacity, incomplete consideration of influencing factors and the like. Li Qiming et al. took the stirrup parameter, longitudinal reinforcement parameter, shear span ratio, axial compression ratio and s / h (stirrup spacing / section height) as characteristic parameters, and established a two-stage RC column seismic failure mode identification method by using support vector machine. The overall accuracy of the model is 95.9%, but the accuracy for flexural-shear failure is 89.2%; Yu Xiaohui et al. proposed a two-stage seismic failure mode identification method by using ET algorithm. Although the overall accuracy of the model on the training set and the test set is 97%, the performance of the model on the test set is the most powerful standard for evaluating the generalization ability and stability of the model. The overall accuracy of the model proposed in the study on the test set is 91%, and the recall rate for shear failure is 74%. SUMMARY

[0005] The present application provides a RC column seismic failure mode identification method, which establishes a BSCFB seismic failure mode identification model, integrates data oversampling, cost-sensitive learning and model integration, improves the accuracy of RC column seismic failure mode identification under unbalanced and relatively balanced data sets, and interprets the proposed machine learning model to improve the credibility of the model.

[0006] To achieve the above object, the technical scheme adopted by the present application is:

[0007] A multi-strategy fusion RC column seismic failure mode identification method, comprising the following steps:

[0008] (1) A pseudo-static test database is established by collecting pseudo-static test data of RC columns;

[0009] (2) The input parameters and output parameters are determined, and the design parameters of the RC column are screened by the standard mutual information and chi-square test method, and finally the input parameters and output parameters of the model are determined;

[0010] (3) The model data set composed of input parameters and output parameters is preprocessed;

[0011] (4) A machine learning model is established to identify the seismic failure mode of the RC column.

[0012] Preferably, in step (2), the design parameters are 17, which are: b, h, l, f c , ρ t , ρ l , f yt , fyl , d t , d l , s, l, n, s h , a t , a l , tc, wherein b is the cross-sectional width, h is the cross-sectional height, l is the column length, f c is the compressive strength of concrete, p t is the stirrup reinforcement ratio, p l is the longitudinal reinforcement ratio, f yt is the yield strength of stirrup, f yl is the yield strength of longitudinal reinforcement, d t is the diameter of stirrup, d l is the diameter of longitudinal reinforcement, s is the stirrup spacing, l is the shear span ratio, n is the axial compression ratio, s h is the ratio of stirrup spacing to cross-sectional height, a t is the stirrup reinforcement parameter, and the calculation formula is a t = f yt / f c · p t , a l is the longitudinal reinforcement parameter, and the calculation formula is a l = f yl / f c · p l , tc is the test configuration form. The output parameters are the seismic failure modes, which are flexural failure, shear failure and flexural-shear failure, respectively, and are represented by 1, 2 and 3, respectively.

[0013] Preferably, in the step (2), the characteristic selection method comprises method one and method two, and the input parameters are screened by the combination of method one and method two; the method one is: the mutual information between each independent variable and the dependent variable is calculated by the mutual information algorithm, and the mutual information is standardized according to formula (1):

[0014]

[0015] In formula (1), MI(A, B) represents the mutual information of variable A and variable B; H(A) and H(B) represent the entropy values of variable A and variable B, respectively; H(A|B) and H(B|A) are conditional entropies, which represent the remaining uncertainty of A or B when B or A is known; H(A, B) represents the joint entropy of variable A and variable B;

[0016] The method two is: the input parameters are screened by using chi-square test, and the cramer's V value between each independent variable and the dependent variable is calculated.

[0017] Preferably, in the step (2), the input parameters are screened according to the standard that the standard mutual information is greater than 0.2 or the Cramer's V value is greater than 0.55, and finally the input parameters of the model are determined as: l, f c , p t , p l , f yt , f yl , s, lambda, n, s h , alpha t , alpha l .

[0018] Preferably, the method of the step (3) is that the data set is randomly divided into a training set and a test set according to a ratio of 7:3, and the sample data is standardized by using a z-score method.

[0019] Preferably, the method of the step (4) is that first, the BoderlineSMOTE algorithm is used to oversample the minority class samples, so that the training set data reaches a balanced state; then, the random forest algorithm is used as a base learner of the AdaBoost algorithm, and cost-sensitive learning is considered, that is, the weights of the samples with different classification results are updated and different cost-sensitive coefficients are assigned; for the multi-classification problem, there are mainly six types of errors in the model, which are: (1) i[0]=1, i[1]=2; (2) i[0]=1, i[1]=3; (3) i[0]=2, i[1]=1; (4) i[0]=2, i[1]=3; (5) i[0]=3, i[1]=1; (6) i[0]=3, i[1]=2, i[0] represents a real class, and i[1] represents a predicted class; since the prediction errors of the shear failure and the flexural-shear failure of the RC column are more costly in actual engineering, and the shear failure and the flexural-shear failure belong to minority class samples, therefore, the cost coefficient of the samples with correct classification is set to 1, the costs of different misclassification types are valued according to the actual engineering experience, and are normalized.

[0020] Preferably, in the step (4), the model is recorded as a BSCFB earthquake damage pattern recognition model, and in the model, the optimal hyperparameters are determined by using a grid search and a five-fold cross-validation method.

[0021] Preferably, in the step (4), the model is explained by using a kernel explanation in the SHAP method.

[0022] The RC column earthquake damage pattern recognition method with multi-strategy fusion has the following beneficial effects:

[0023] (1) The present application determines the main factors affecting the seismic failure mode of RC columns through standard mutual information and chi-square test, takes them as input parameters of the machine learning model, proposes a BSCFB seismic failure mode recognition model, integrates data oversampling, cost-sensitive learning and model integration three strategies, improves the accuracy of RC column seismic failure mode recognition, and interprets the proposed machine learning model, increases the credibility of the model.

[0024] (2) The method of the present application comprehensively considers the influencing factors, and has high model performance. In the case of class imbalance in the data set, the overall accuracy can reach 97%, the precision and recall of the three different seismic failure modes are all above 91%, and the model has interpretability.

[0025] (3) The method of the present application is simple to apply and has a wide range of applications. It also has good effect on RC column seismic failure mode recognition under relatively balanced data set, and the overall accuracy can reach 94%. BRIEF DESCRIPTION OF DRAWINGS

[0026] Figure 1 : Flowchart of the present application;

[0027] Figure 2 : Prediction result graph of the BSCFB seismic failure mode recognition model of the present application;

[0028] Figure 3 : Analysis result graph of the importance of each influencing factor;

[0029] Figure 1 In the formula, D1 represents the initial sample weight, D2 represents the sample weight for training the second weak classifier, D m represents the sample weight for training the mth weak classifier.

[0030] Figure 2 In the formula, FF, SF and FSF represent flexural failure, shear failure and flexural-shear failure, respectively.

[0031] Figure 3 In the formula, FF, SF and FSF represent flexural failure, shear failure and flexural-shear failure, respectively. DETAILED DESCRIPTION

[0032] The following describes the embodiments of the present application in a step-by-step manner. The description is only a preferred embodiment of the present application and is not intended to limit the protection scope of the present application. Any modification, equivalent replacement and improvement made within the spirit and principles of the present application shall be included in the protection scope of the present application.

[0033] In the description of the present application, it should be noted that the terms "upper", "lower", "left", "right", "top", "bottom", "inner", "outer" and the like indicate the orientation or positional relationship shown in the drawings, and are only for the purpose of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, and a particular orientation configuration and operation, and therefore cannot be understood as a limitation on the present application.

[0034] In the initial embodiment, a multi-strategy fusion RC column earthquake damage mode recognition method is provided, as shown in Figure 1 、 2 , 3, comprising the following steps:

[0035] (1) Establishing a pseudo-static test database by collecting pseudo-static test data of RC columns;

[0036] (2) Determining the input parameters and output parameters, and screening the design parameters of the RC column by the standard mutual information and chi-square test method, and finally determining the input parameters and output parameters of the model;

[0037] (3) Data preprocessing of the data set of the model;

[0038] (4) Establishing a machine learning model to recognize the earthquake damage mode of the RC column.

[0039] In a further embodiment, in step (2), the design parameters are 17, respectively: b, h, l, f c , ρ t , ρ l , f yt , f yl , d t , d l , s, λ, n, s h , α t , α l , tc, wherein b is the cross-sectional width, h is the cross-sectional height, l is the column length, f c is the compressive strength of concrete, ρ t is the stirrup reinforcement ratio, ρ l is the longitudinal reinforcement ratio, f yt is the yield strength of stirrup, f yl is the yield strength of longitudinal reinforcement, d t is the diameter of stirrup, d l is the diameter of longitudinal reinforcement, s is the stirrup spacing, λ is the shear span ratio, n is the axial compression ratio, s h is the ratio of stirrup spacing to cross-sectional height, α t is the stirrup reinforcement parameter, and the calculation formula is α t = f yt / f c · ρt , alpha l is a longitudinal reinforcement parameter, and the calculation formula is alpha l = f yl / f c * rho l , tc is a test configuration form. The output parameters are seismic damage modes, namely flexural failure, shear failure and flexural-shear failure, which are represented by 1, 2 and 3 respectively.

[0040] In further embodiments,

[0041] In step (2), the feature selection method includes method one and method two, and the input parameters are screened by the combination of method one and method two; the method one is: the mutual information between each independent variable and the dependent variable is calculated by the mutual information algorithm, and the mutual information is standardized according to formula (1):

[0042]

[0043] In formula (1), MI(A, B) represents the mutual information of variable A and variable B; H(A) and H(B) represent the entropy values of variable A and variable B respectively; H(A|B) and H(B|A) are conditional entropies, which represent the remaining uncertainty of A or B when B or A is known; H(A, B) represents the joint entropy of variable A and variable B;

[0044] The method two is: in order to ensure that the selected parameters can comprehensively and effectively reflect the data information, the input parameters are screened by using chi-square test, and the cramer's V value between each independent variable and the dependent variable is calculated. The calculation results of the two feature selection methods are shown in Table 1.

[0045] Table 1 NMI and cramer's V value of each parameter of RC column

[0046] b h l f c ]]> t ]]> ​ l ]]> ​ f yt ]]> f yl ]]> d t ]]> NMI 0.19 0.14 0.29 0.27 0.24 0.31 0.34 0.35 0.12 Cramer's V 0.55 0.47 0.80 0.89 0.78 0.85 0.92 0.91 0.44 d l ]]> s Lambda n s h ]]> t ]]> ​ l ]]> ​ <![CDATA[t c ]]> NMI 0.18 0.16 0.28 0.22 0.26 0.30 0.30 0.06 Cramer's V 0.54 0.58 0.75 0.74 0.82 0.98 0.97 0.26

[0047] In further embodiments, as shown in Table 1, in step (2), the input parameters are screened according to the standard that the standard mutual information is greater than 0.2 or the cramer's V value is greater than 0.55, and the input parameters of the model are finally determined as: l, f c , rho t , rho l , f yt , f yl , s, lambda, n, s h , alpha t , alpha l .

[0048] In a further embodiment, the method of step (3) is to randomly divide the data set into a training set and a test set in a ratio of 7:3. To improve the accuracy and stability of the machine learning model, after dividing the training set and the test set, the sample data is standardized by using the z-score method.

[0049] In a further embodiment, the method of step (4) is: first, use the BoderlineSMOTE algorithm to oversample the minority class samples, so that the training set data reaches a balanced state; then use the random forest algorithm as the base learner of the AdaBoost algorithm, and consider cost-sensitive learning, that is, update the weights of samples with different classification results and assign different cost-sensitive coefficients; for this multi-classification problem, the model mainly exists six types of errors, which are: i[0]=1,i[1]=2;(2)i[0]=1,i[1]=3;(3)i[0]=2,i[1]=1;(4)i[0]=2,i[1]=3;(5)i[0]=3,i[1]=1;(6)i[0]=3,i[1]=2,i[0] represents the true class, and i[1] represents the predicted class; Since in actual engineering, the prediction error of RC column shear failure and bending shear failure has a greater cost, and the shear failure and bending shear failure in the data belong to minority class samples, therefore, the cost coefficient of the sample with correct classification is set to 1, the cost of different misclassification types is valued according to the actual engineering experience, and is normalized. The final assignment result is shown in Table 2.

[0050] Table 2 Cost-sensitive coefficients of different classification results

[0051] i[1,1] i[1,2] i[1,3] i[2,1] i[2,2] i[2,3] i[3,1] i[3,2] i[3,3] Un-normalized 1 2 2 4 1 3 3 2 1 Normalized 1 / 19 2 / 19 2 / 19 4 / 19 1 / 19 3 / 19 3 / 19 2 / 19 1 / 19

[0052] In a further embodiment, in step (4), the model is denoted as the BSCFB seismic damage pattern recognition model, and in the model, the optimal hyperparameters are determined by grid search and five-fold cross-validation.

[0053] In a further embodiment, to enhance the reliability of the model, in step (4), the kernel explanation in the SHAP method is used to explain the model. The influence of each input parameter on the model is shown in Figure 3 .

[0054] As can be seen from Figure 3 (a), from the overall model, l and λ are the most important factors affecting the seismic damage pattern of RC columns, followed by α l ,f c ,f yt ,f yl ,ρ l ,ρ t , and finally α t ,s,n,sh . Figure 3 (b) shows the influence of each characteristic parameter on bending failure, with the importance ranked as l>λ>α. l >ρ t >f c >ρ l >f yl >f yt >α t >s>s h >n. Figure 3 (c) shows the degree of influence of each characteristic parameter on shear failure, and their importance is ranked as λ>l>ρ. t >α l >f yl >ρ l >α t >f yt >s>f c >n>s h . Figure 3 (d) Show the degree of influence of each characteristic parameter on bending-shear failure, and their importance is ranked as l>f c >f yt >α l >λ>f yl >ρ l >ρ t >α t >s h >n>s. And l, λ, ρ t f c f yl f yt α t The larger the relative value, the greater the α value. l ρ l ,s,s h The smaller the relative values ​​of and n, the more likely bending failure will occur. Conversely, the more likely non-bending failure will occur. In practical engineering design, a comprehensive design should be carried out based on the degree of influence and positive / negative correlation of each influencing factor on the seismic failure mode of RC columns, in order to avoid shear failure or bending-shear failure as much as possible.

Claims

1. A multi-strategy fusion RC column earthquake damage mode recognition method, characterized by: It comprises the following steps: (1) Establish a pseudo-static test database by collecting RC column pseudo-static test data; (2) Determine the input parameters and output parameters, and screen the design parameters of the RC column by the standard mutual information and chi-square test method to finally determine the input parameters and output parameters of the model; (3) Data preprocessing is performed on the model data set composed of input parameters and output parameters; (4) Establish a machine learning model to identify the seismic failure mode of the RC column; The design parameters in the step (2) are 17, respectively: b, h, l, f c , p t , p l , f yt , f yl , d t , d l , s, l, n, s h , a t , a l , tc, wherein b is the cross-sectional width, h is the cross-sectional height, l is the column length, f c is the compressive strength of concrete, p t is the stirrup reinforcement ratio, p l is the longitudinal reinforcement ratio, f yt is the yield strength of stirrup, f yl is the yield strength of longitudinal reinforcement, d t is the diameter of stirrup, d l is the diameter of longitudinal reinforcement, s is the stirrup spacing, l is the shear span ratio, n is the axial compression ratio, s h is the ratio of stirrup spacing to cross-sectional height, a t is the stirrup reinforcement parameter, and the calculation formula is a t = f yt / f c · p t , a l is the longitudinal reinforcement parameter, and the calculation formula is a l = f yl / f c · p l , and tc is the test configuration form; the output parameter is the seismic damage mode, which is flexural failure, shear failure and flexural-shear failure, respectively represented by 1, 2 and 3. In step (2), the feature selection method includes method one and method two, which are combined to screen the input parameters; method one is to calculate the mutual information between each independent variable and the dependent variable by the mutual information algorithm, and to standardize the mutual information according to formula (1): In formula (1), MI(A,B) represents the mutual information of variables A and B; H(A) and H(B) represent the entropy values of variables A and B; H(A|B) and H(B|A) are conditional entropies, which represent the remaining uncertainty of A or B when B or A is known; H(A,B) represents the joint entropy of variables A and B; Method two is to screen the input parameters by chi-square test and calculate the cramer's V value between each independent variable and the dependent variable; In step (2), the input parameters are screened according to the standard that the standard mutual information is greater than 0.2 or the Cramer's V value is greater than 0.55, and finally the input parameters of the model are determined as: l, f c , p t , p l , f yt , f yl , s, l, n, s h , a t , a l .

2. The multi-strategy fused RC column seismic damage pattern recognition method of claim 1, characterized in that: The method of step (3) is to randomly divide the data set into a training set and a test set in a ratio of 7:3, and to standardize the sample data by z-score.

3. The multi-strategy fused RC column seismic damage pattern recognition method of claim 2, characterized in that: The method of step (4) is: first, use the BoderlineSMOTE algorithm to oversample the minority class samples to balance the training set data; then use the random forest algorithm as the base learner of the AdaBoost algorithm, and consider the cost-sensitive learning, that is, update the weights of samples with different classification results and assign different cost-sensitive coefficients; for this multi-classification problem, the model mainly has six types of errors, which are: i[0]=1,i[1]=2;(2)i[0]=1,i[1]=3;(3)i[0]=2,i[1]=1;(4)i[0]=2,i[1]=3;(5)i[0]=3,i[1]=1;(6)i[0]=3,i[1]=2, i[0] represents the true class, and i[1] represents the predicted class; Since the prediction error of RC column shear failure and bending shear failure has a greater cost in actual engineering, and shear failure and bending shear failure belong to minority class samples, the cost coefficient of correctly classified samples is set to 1, and the cost of different misclassification types is valued according to the actual engineering experience and normalized.

4. The multi-strategy fused RC column seismic damage pattern recognition method of claim 3, characterized in that: In step (4), the model is denoted as BSCFB seismic failure mode identification model, and the optimal hyperparameters are determined by grid search and five-fold cross-validation.

5. The multi-strategy fused RC column seismic damage pattern recognition method of claim 4, characterized in that: In step (4), the kernel explanation in the SHAP method is used to explain the model.

Images