郭文燁 張健新

摘 要：神經(jīng)網(wǎng)絡(luò)（ANN）模型作為土木工程領(lǐng)域中一種有效的方法能夠用于解決復(fù)雜的問題。基于試驗(yàn)數(shù)據(jù)采用神經(jīng)網(wǎng)絡(luò)對鋼筋混凝土剪力墻的抗剪承載力進(jìn)行預(yù)測，收集160個(gè)鋼筋混凝土剪力墻在低周往復(fù)荷載下的試驗(yàn)數(shù)據(jù)，建立數(shù)據(jù)庫，選取140個(gè)試驗(yàn)樣本對ANN模型進(jìn)行訓(xùn)練，20個(gè)試驗(yàn)樣本進(jìn)行測試驗(yàn)證。ANN1和ANN2有14個(gè)輸入?yún)?shù)：混凝土抗壓強(qiáng)度、剪跨比、軸壓比、豎向鋼筋強(qiáng)度、橫向鋼筋強(qiáng)度、墻體豎向分布鋼筋配筋率、墻體水平分布鋼筋配筋率、邊緣構(gòu)件縱向鋼筋配筋率、邊緣構(gòu)件橫向鋼筋配筋率、邊緣構(gòu)件與截面面積比、截面高厚比、總截面面積、墻高和截面形狀，輸入數(shù)據(jù)分別被歸一化到區(qū)間[0， 1]和[0.1， 0.9]。兩個(gè)模型的輸出數(shù)據(jù)均為剪力。對比分析ANN模型預(yù)測的鋼筋混凝土剪力墻抗剪承載力與采用規(guī)范GB 50011和ACI 318-14公式計(jì)算的抗剪承載力，結(jié)果表明，神經(jīng)網(wǎng)絡(luò)模型能夠精確地預(yù)測鋼筋混凝土剪力墻的抗剪承載力，具有較好的預(yù)測和泛化能力。

關(guān)鍵詞：神經(jīng)網(wǎng)絡(luò);剪力墻;鋼筋混凝土;模型預(yù)測;抗剪承載力

1 Introduction

Reinforced concrete （RC）shear walls are often used in building structures due to their capacity to resist lateral loads under seismic action[1]. The concrete strength， aspect ratio， axial compression ratio， vertical or horizontal web reinforcement ratio， and vertical or horizontal boundary region reinforcement ratio are critical design parameters that govern the lateral load resistance capacity of RC shear walls[2-3]. The formula used in domestic and foreign codes to calculate the shear bearing capacity is an empirical formula determined by statistical analysis， that reflects the main physical and geometric parameters and considers the factors that influence the reliability. Differences in the calculation model and calculation method are incorporated in current codes， such as GB 50011， ACI 318， and EC2. Furthermore， the strength of the concrete used in the formula for calculating the shear bearing capacity is also different. Generally， most existing methods of calculating the shear bearing capacity of RC shear walls are based on models with limited experimental data， such as shear walls using high-strength steel bars. Therefore， further research on more reliable and effcient structural assessement is needed.

ANNs have been used for simulating engineering problems[4-6]. To predict the axial bearing capacity， Du et al.[7] suggested two ANN models of rectangular concrete-filled steel tubular columns. Kotsovou et al.[8] established an ANN model to predict the load bearing capacity of beam-column joints. However， the ANN models and experimental data are limited. In this study， shear bearing capacity predictions of RC shear walls were developed using artificial neural networks. The developed ANN model provides a reference for prefabricated concrete shear walls， the seismic performance of which are equivalent to cast-in-place RC walls[9-12].

2 Data collection

As shown in Fig.1，test results for 160 RC shear walls with rectangular or barbell sections was found in the literature [2-3， 13-29]. The test information included all parameters that may have an impact on the behavior of the RC shear walls. The test samples exhibited good deformation ability. The parameters for all samples were consistent.

The size parameters of the wall （b， h， and H）， the yield strength of the horizontal reinforcements fy， the concrete compressive strength fc， the aspect ratio λ， the axial compression ratio μ， and the shear force V are included to train and test the ANN models. Finally， 160 test samples were obtained and are summarized in Table 1.

3 Artificial neural networks

3.1 Background information

ANN is an operational model that mimics the neural network of the human brain from the perspective of information processing. ANN is an artificial intelligence technology that can solve complex problems based on input parameters. The effects of these parameters are not explicitly illustrated or quantified. ANNs have the ability to learn， summarize， classify， and predict， and it have been achieved remarkable results in many practical applications over the past years. In this study， ANNs are used to predict the shear bearing capacity of RC shear walls.

This study uses a back-propagation （BP） algorithm， as shown in Fig.2. A typical artificial neuron is shown in Fig.3. Three layers are included in the ANNs： input layer， hidden layer， and output layer. Each layer comprises k neurons， three neurons， and two neurons， respectively.

The connections between interrelated neurons with a set specific weight are multiplied by the input data produced by the neuron. The values obtained in a particular layer are passed through the link and summed up with the bias （refer to Fig.2）[8]. A predefined activation is used to represent the relationship between the inputs and the outputs， as shown in the followingwhere yi is the output of the ANN， wij is the weight coefficients of the jth neuron， xj is the input data， θi is the bias of the neuron， and g（·） is the activation function. In this study， input and hidden layers used sigmoid activation functions， and the output layer used the tan-sigmoid activation function.

3.2 Input and output data

The input parameters were selected based on the dominant effect of the parameters on the behavior of the RC shear wall， and included the concrete compressive strength （fc）， the aspect ratio （λ）， the axial compression ratio （μ）， the vertical reinforcement yield strength （fy，vw）， the horizontal reinforcement yield strength （fh，vw）， the vertical reinforcement web ratio （ρvw）， the horizontal reinforcement web ratio （ρhw）， the vertical reinforcement boundary region ratio （ρvc）， the horizontal reinforcement boundary region ratio （ρhc）， the sectional area ratio of the boundary region to the total cross-section area （Ab/Ag）， the sectional height thickness ratio （lw/tw）， the total section area （Ag）， the wall height （H）， and the section shape （the rectangular section is “0” and the barbell section is “1”）.

Since the performance of the RC shear walls specified in the code is determined by the limit of the shear load capacity， we take the maximum shear （Vmax） as the target parameter. The maximum and minimum values of the input and output data are listed in Table 2. Table 3 shows the correlation between the input parameters used for the prediction of the shear bearing capacity of the RC shear walls. Some parameters are weakly correlated while others are strongly correlated. For example， the correlation coefficient between the lw/tw and Ab/Ag was -0.763， which indicates a strong negative relationship. The correlation coefficient between the H and λ was 0.449， which indicates a weak positive relationship. The sequence of the correlation for the input parameters from strong to weak were Ag， ρvc， ρhc， fc，Ab/Ag， section shape， fy，vw，fh，vw，H，ρhw，λ，ρvw， μ， and lw/tw.

To minimize the deviation of the ANN and low convergence rates， the values of the input and output data are normalized using Eq.（2）.

3.3 Training and testing of the ANNs

In this study， the network was built using the ANN toolbox in MATLAB. The BP network with 15 hidden layers was used to build the model of RC shear walls. The 160 experimental samples were randomly divided into two groups， 140 samples for training， and 20 samples for testing. In order to verify the effect of normalization equation on the ANNs prediction， the two control groups ANN1 and ANN2 were normalized in the range [0， 1] and [0.1， 0.9].

The training process of the neural network involves adjusting the network's weights and deviations （initially randomly assigned） to optimize the network's performance in the iterative process. The error performance index of the forward network is MSE， which is the mean square error between the network output and the target. The neural network would modify the network node weight， according to MSE. At the same time， in order to reduce the error in each iteration， ANN used the back-error propagation algorithm. After the error was calculated， the weights and bias were readjusted.

The calibration procedure of the ANN model is shown in Fig.4. This was repeated until one of the following conditions was met： 1） After 500 training sessions， the algorithm will stop the training process. 2） The error-index reaches 10-5. 3） The validation check occurs 10 times.

The ANN values （ANN-output） and test values （targets） are illustrated in Fig.5. The ANNs predicted values were close to the experimental values with good deformation ability， indicating that the ANN1 and ANN2 models successfully learned the relationship between input and output data. In addition， the predicted values and test values that were closer to each other in different normalized ranges were in the range [0， 1] rather than the range [0.1， 0.9].

The ratio of output to target OTR， mean value MV， and standard deviation SD are used to evaluate the behavior of the model.

Where Oi and Ti are the prediction values of the ANN models and the maximum shear of the experimental samples， respectively. n is the total sample number.

Curves of OTR and sample number for ANN1， ANN2， GB 50011， and ACI 318 are presented in Fig.6. Two predicted values in the ANN1 model exhibited the errors of 8.1% and 8.7%， which were overestimated. Two predicted values in the ANN2 model exceeded the error of 8.0%. One was underestimated and the other was overestimated. The SD was 0.036 1 in ANN1 and 0.041 2 in ANN2 （refer to Table 4）. Therefore， the ANN1 model was superior to the ANN2 model in calculating the shear bearing capacity of RC shear walls.

The outputs of the RC shear walls are the results calculated by the formulas Eqs.（6） to （9）.

Fig.6 shows the OTRs calculated by ANN1， ANN2， GB 50011， and ACI 318. Table 4 lists MVs and SDs using the testing data for ANN1， ANN2， GB 50011， and ACI 318.

Results predicted by the ANN1 and ANN2 models matched those calculated by GB 50011 and ACI 318 very well.

The results predicted by the ANN1 and ANN2 models matched those calculated by GB 50011 and ACI 318 very well. There were two outputs with an error of over 8% for both ANN1 and ANN2， but they did not exceed 10%. Two out of twenty in ANN1 were overestimated. One was overestimated in ANN2， and the other was underestimated. These results show that the ANN model exhibited a significant improvement compared to the standard GB 50011 and ACI 318. Compared with the experimental data， fourteen results predicted by GB 50011 exceed 10% difference based on the OTRs. There were sixteen predicted results with errors exceeding 10% in ACI 318. The SDs of ANN1 and ANN2 were 0.036 1 and 0.041 2， much lower than those of GB 50011 and ACI 318 （refer to Table 3）. Compared with GB 50011 and ACI 318， the ANNs exhibited better performance on predicting the shear bearing capacity of RC shear walls.

There were thirteen results with errors exceeding 10% in GB 50011 and three in ACI 318 were underestimated. The MVs of the results predicted by GB 50011 and ACI 318 were 0.954 4 and 0.825 6， respectively. ANN models exhibited higher MVs than GB 50011 and ACI 318， indicating that the formulas were conservative in GB 50011 and ACI 318 due to the usage of high strength materials. The SD of GB 50011 and ACI 318 reached 0.189 7 and 0.223 6， which were larger than the ANN models.

The ANN1 and ANN2 models had the two largest MVs， while ANN1 and ANN2 exhibited smaller SDs. Thus， ANN models can accurately predict the shear bearing capacity of RC shear walls. Compared with the design codes， ANN models may be safer.

5 Conclusions

Two ANN models with fourteen input parameters were developed， based on experimental data. An efficient learning model based on ANNs was proposed to evaluate the load bearing capacity of RC shear walls. The prediction results show that ANN models predict the load bearing capacity favorably using parameters such as the aspect ratio， the axial compression ratio， the concrete and reinforcement strength， the boundary region and web reinforcement ratio， and the sectional ratio and size， thus accurate predictions can be provided.

The ANN1 and ANN2 models exhibit a better correlation with the experimental results than the codes GB 50011 and ACI 318. The ANN models exhibit better accuracy in prediction and generalization capacity. The BP algorithm can be effectively adopted in the shear strength prediction of RC shear walls.

Application of developed ANNs can be extended by further experimental tests including other shaped sections as input data. More studies on RC shear walls including high strength concrete and high strength reinforcements are valuable for the structures adopting RC shear walls.

Acknowledgements

The authors would like to acknowledge the financial support from the Natural Science Foundation of Hebei Province （No. E2018202290）.

References：

[1] NI X Y， CAO S Y， LIANG S， et al. High-strength bar reinforced concrete walls： Cyclic loading test and strength prediction [J]. Engineering Structures， 2019， 198： 109508.

[2] PARK H G， BAEK J W， LEE J H， et al. Cyclic loading tests for shear strength of low-rise reinforced concrete walls with grade 550 MPa bars [J]. ACI Structural Journal， 2015， 112（3）： 299-310.

[3] CHENG M Y， HUNG S C， LEQUESNE R D， et al. Earthquake-resistant squat walls reinforced with high-strength steel [J]. ACI Structural Journal， 2016， 113（5）： 1065-1076.

[4] SIAM A， EZZELDIN M， EL-DAKHAKHNI W. Machine learning algorithms for structural performance classifications and predictions： Application to reinforced masonry shear walls [J]. Structures， 2019， 22： 252-265.

[5] TRUNG N T， SHAHGOLI A F， ZANDI Y， et al. Moment-rotation prediction of precast beam-to-column connections using extreme learning machine [J]. Structural Engineering and Mechanics， 2019， 70（5）： 639-647.

[6] GEYER P， SINGARAVEL S. Component-based machine learning for performance prediction in building design [J]. Applied Energy， 2018， 228： 1439-1453.

[7] DU Y S， CHEN Z H， ZHANG C Q， et al. Research on axial bearing capacity of rectangular concrete-filled steel tubular columns based on artificial neural networks [J]. Frontiers of Computer Science， 2017， 11（5）： 863-873.

[8] KOTSOVOU G M， COTSOVOS D M， LAGAROS N D. Assessment of RC exterior beam-column joints based on artificial neural networks and other methods [J]. Engineering Structures， 2017， 144： 1-18.

[9] SOUDKI K， WEST J S， RIZKALLA S H， et al. Horizontal connections for precast concrete shear wall panels under cyclic shear loading [J].PCI Journal， 1996， 41（3）： 64-80.

[10] SMITH B J， KURAMA Y C， MCGINNIS M J. Behavior of precast concrete shear walls for seismic regions： comparison of hybrid and emulative specimens [J]. Journal of Structural Engineering， 2013， 139（11）： 1917-1927.

[11] CHEN Y， ZHANG Q， FENG J， et al. Experimental study on shear resistance of precast RC shear walls with novel bundled connections [J]. Journal of Earthquake and Tsunami， 2019， 13： 1940002.

[12] LI J B， WANG L， LU Z， et al. Experimental study of L-shaped precast RC shear walls with middle cast-in-situ joint [J]. The Structural Design of Tall and Special Buildings， 2018， 27（6）： e1457.

[13] OESTERLE R G， FIORATO A E， JOHAL L S， et al. Earthquake resistant structural walls tests of isolated walls [R]//Report to National Science Foundation， Skokie （IL， USA）： PCA Construction Technology Laboratories， 1976.

[14] LEFAS I D， KOTSOVOS M D， AMBRASEYS N N. Behavior of reinforced concrete structural walls： strength， deformation characteristics， and failure mechanism [J].ACI Structural Journal， 1990， 87（1）： 23-31.

[15] PILAKOUTAS K， ELNASHAI A. Cyclic behavior of RC cantilever walls， part I： Experimental results [J]. ACI Structural Journal， 1995， 92（3）： 271-281.

[16] SALONIKIOS T N， KAPPOS A J， TEGOS I A， et al. Cyclic load behavior of low-slenderness reinforced concrete walls： Design basis and test results [J]. ACI Structural Journal， 1999， 96（4）： 649-660.

[17] RIVA P， FRANCHI A. Behavior of reinforced concrete walls with welded wire mesh subjected to cyclic loading [J].ACI Structural Journal， 2001， 98（3）： 324-334.

[18] HIDALGO P， LEDEZMA C， JORDAN R. Seismic behavior of squat reinforced concrete shear walls [J]. Earthquake Spectra， 2002， 18（2）： 287-308.

[19] DAZIO A， BEYER K， BACHMANN H. Quasi-static cyclic tests and plastic hinge analysis of RC structural walls [J]. Engineering Structures， 2009， 31（7）： 1556-1571.

[20] LIU X， BURGUEO R， EGLESTON E， et al. Inelastic web crushing performance limits of high-strength-concrete structural wall： Single wall test program： Report No. CEE-RR-2009/03 [R]. Michigan State University， East Lansing， MI， 2009， 281.

[21] TRAN T. Experimental and analytical studies of moderate aspect ratio reinforced concrete structural walls [D]. University of California， 2012.

[22] ALARCON C， HUBE M A， DE LA LLERA J C. Effect of axial loads in the seismic behavior of reinforced concrete walls with unconfined wall boundaries [J]. Engineering Structures， 2014， 73： 13-23.

[23] HUBE M A， MARIHUN A， DE LA LLERA J C， et al. Seismic behavior of slender reinforced concrete walls [J]. Engineering Structures， 2014， 80： 377-388.

[24] WANG Y. Effects of web reinforcement discontinuities on the seismic response of structural walls [D]. West Lafayette， Indiana： Purdue University. 2014.

[25] TENG S， CHANDRA J. Cyclic shear behavior of high-strength concrete structural walls [J].ACI Structural Journal， 2016， 113（6）： 1335-1345.

[26] GUO R，ZHU K. Experimental research on seismic behavior of RC shear wall with 600 megapascal high-strength reinforced bars [J]. Industrial Construction， 2017， 47（6）： 34-39.

[27] HUBE M， SANTA-MARIA H， LOPEZ M. Experimental campaign of thin reinforced concrete shear walls for low-rise constructions [C]//Proceedings of Sixteenth World Conference on Earthquake， Santiago， 2017.

[28] CHEN X L. Study on seismic performance of RC shear wall structure reinforced with HRB600 high strength reinforcement [D]. Chongqing： Chongqing University， 2018.

[29] LIU Y. Experimental study on seismic performance of concrete shear walls with 600 MPa steel bars [D]. Nanjing： Southeast University， 2019.

（編輯 章潤紅）