• <tr id="yyy80"></tr>
  • <sup id="yyy80"></sup>
  • <tfoot id="yyy80"><noscript id="yyy80"></noscript></tfoot>
  • 99热精品在线国产_美女午夜性视频免费_国产精品国产高清国产av_av欧美777_自拍偷自拍亚洲精品老妇_亚洲熟女精品中文字幕_www日本黄色视频网_国产精品野战在线观看 ?

    基于啟發(fā)式方法和支持向量機(jī)方法預(yù)測有機(jī)物的熱導(dǎo)率

    2012-12-11 09:10:14時靜潔陳利平陳網(wǎng)樺
    物理化學(xué)學(xué)報 2012年12期
    關(guān)鍵詞:熱導(dǎo)率南京向量

    時靜潔 陳利平 陳網(wǎng)樺,* 石 寧 楊 惠 徐 偉

    (1南京理工大學(xué)化工學(xué)院安全工程系,南京210094;2化學(xué)品安全控制國家重點(diǎn)實(shí)驗室,山東青島266071)

    基于啟發(fā)式方法和支持向量機(jī)方法預(yù)測有機(jī)物的熱導(dǎo)率

    時靜潔1,2陳利平1陳網(wǎng)樺1,*石 寧2楊 惠1徐 偉2

    (1南京理工大學(xué)化工學(xué)院安全工程系,南京210094;2化學(xué)品安全控制國家重點(diǎn)實(shí)驗室,山東青島266071)

    構(gòu)建147個有機(jī)物分子結(jié)構(gòu)與其熱導(dǎo)率值之間的定量結(jié)構(gòu)-性質(zhì)關(guān)系(QSPR)模型,探討影響有機(jī)物熱導(dǎo)率的結(jié)構(gòu)因素.以147個化合物作為樣本集,隨機(jī)選擇118個作為訓(xùn)練集,29個作為測試集.應(yīng)用CODESSA軟件計算了組成、拓?fù)?、幾何、靜電和量子化學(xué)等描述符,通過啟發(fā)式方法(HM)篩選得到5個結(jié)構(gòu)參數(shù)并建立線性回歸模型;用所選5個結(jié)構(gòu)參數(shù)作為支持向量機(jī)(SVM)的輸入,建立非線性的支持向量機(jī)回歸模型.預(yù)測結(jié)果表明:支持向量機(jī)回歸模型的性能(復(fù)相關(guān)系數(shù)R2=0.9240)雖略低于啟發(fā)式回歸模型的性能(R2=0.9267),但是支持向量機(jī)方法預(yù)測性能(R2=0.9682)高于啟發(fā)式方法的預(yù)測性能(R2=0.9574),對于QSPR模型來說,預(yù)測性能更重要.因此,總體來說支持向量機(jī)方法優(yōu)于啟發(fā)式方法.支持向量機(jī)方法和啟發(fā)式方法的提出為工程上提供了一種根據(jù)分子結(jié)構(gòu)預(yù)測有機(jī)物熱導(dǎo)率的新方法.

    啟發(fā)式方法;支持向量機(jī);熱導(dǎo)率;預(yù)測;定量結(jié)構(gòu)-性質(zhì)關(guān)系

    1 Introduction

    Thermal conductivity,also known as the coefficient of thermal conductivity,reflects the capacity of heat transmission.1Thermal conductivity data are most important for the engineering design of any thermal processes.2Thermal conductivities of organics at 20°C in a liquid state were studied in this research. Usually thermal conductivity is experimentally measured.Several different approaches could potentially be used to study thermal conductivity:(i)heat flux meter techniques,3(ii)flash technique,4(iii)transient plane-source method,5(iv)transient hotwire method.6Yet such measuring methods are time-consuming and technique-limited.Furthermore,an accurate determination of this property is very difficult because of convection and radiation accompanying the heat losses during experiment.7Thus,the prediction of the thermal conductivity with a theoretical method becomes important and necessary.

    Quantitative structure-property relationship(QSPR)models are obtained through analyzing and calculating the correlation between the property and a variety of structural information.8-11The purpose of constructing a QSPR model is to find factors that determine the property.Meanwhile,such a model can also predict the property of compounds including those not yet synthesized,since the property is determined by the molecular structure which is translated into the so-called molecular descriptors.Recently,HM and SVM have been developed to predict nematic transition temperatures in themotropic liquid crystal,12binding affinities of adenosine A2Areceptor antagonists,13the activities of imidazothienopyrazines14and so on.Bini and his research team15also built an HM model on the thermal conductivity,but their study had some disadvantages.There were only 33 data in their sample dataset,although the squared correlation coefficient(R2)was a bit higher.Furthermore,they did not divide the sample dataset into the training set and the test one.As a result,it is impossible to know the predictive ability of the model.What is more,they did not study the non-linear model.Hence,we cannot figure out the non-linear relationship between the thermal conductivity and the molecular structure.

    The objective of our work is to find the structural factors which play important roles in the thermal conductivity for organics and to establish the quantitative linear and non-linear relationships(HM and SVM)between the thermal conductivity and molecular descriptors.

    2 Materials and methodology

    2.1 Data

    The accuracy of the prediction model can be affected directly by the reliability of experimental values.Thus it is important to select a reliable dataset.The data source in this work is from Data Manual for the Physical Property of Organic Compounds which is carefully compiled by Qingdao Institute of Chemical.

    The whole dataset is composed of 147 compounds containing C,H,O,N,Cl,S,and F,and covers hydrocarbons,halogen compounds,ethers,alcohols,esters,aldehydes,ketones, amines,and amino compounds.These compounds with wide chemical diversities lay the foundation for a robust and effective prediction model.In order to build a QSPR model,the dataset was randomly divided into two subsets,the training set and the test one,consisting of 118(80%)and 29(20%)compounds,respectively.The training set was used to select variables and to construct the models,and the performance of the models was evaluated by the test one.In addition,for the purpose of comparison,the two subsets which were employed to build an HM model,and the two ones employed to build an SVM model,were exactly the same.

    2.2 Descriptors

    Descriptors are defined as numerical characteristics associated with chemical structures.They are derived from the chemical constitution,topology,geometry,wave function,potential energy surface,and some combinations of these items of a given chemical structure.The value of a particular descriptor can be set by the user or calculated automatically by software,such as CODESSA.Each descriptor value must be associated with a previously defined structure.In this paper,CODESSA was employed to compute values of these descriptors.The software, designed by Katritzky team,can be used to calculate 804 descriptors covering constitutional,geometrical,topological,electrostatic,quantum-chemical,and thermodynamic descriptors. The classification of these descriptors is listed in Table 1.

    The selection of suitable molecule descriptors is very critical throughout the QSPR research.An efficient descriptor must be capable of providing as much structural information as possible,and the more precise the better.Recently,there are a variety of variable selection methods such as the genetic algorithm (GA)method and the HM.Herein,HM was chosen in this paper to select variables due to the two advantages of HM.First, with this approach,the global optimal solution could be found.16When selecting variables from large amounts of data, we are in need of a method to go beyond the local best solution so as to find a global best solution.But owing to their own limitations,traditional methods could be only used to search the partial solution,not the global optimal solution.The second advantage is that HM is very simple and convenient compared with GA.Although GA can also obtain the global optimal solution,people need to compile complicated codes in MATLAB environment when using this method.

    HM is employed to remove some descriptors by pretreat-ment according to the following four criteria.(1)Not all compounds share the same parameters;(2)to all compounds,numerical value of descriptors changes within a small range;(3) in the equation related to parameters,HM eliminates the descriptors which do not match the following criterion:the F-test F<0.1;(4)HM deletes the descriptors whose t-test value t is less than the defined one.After the pre-selection of descriptor, multiple linear regression method is developed in a stepwise procedure.First,starting with the top descriptor from the pre-selected list of descriptors,the two-parameter correlation is calculated using the following pairs:the first descriptor with each of the remaining descriptors and the second descriptor with each of the remaining descriptors,etc.The best pairs,as evidenced by the highest F-values in the two-parameter correlations,are chosen and used for further inclusion of descriptors in a similar manner.Then,new descriptors are added one-byone until the pre-selected number of descriptors in the model is achieved.Astepwise addition of further descriptor scales is performed to find the best multi-parameter regression models with the optimum values of statistical criteria(highest values of R2, the F-test,and the standard deviation(S)).From the above processes,five descriptors are selected from descriptors pool and the linear model is produced by the HM.17

    Table 1 Classification of descriptors

    2.3 Computational methods

    SVM was introduced by Vapnik18and has been applied to classification as well as regression tasks.For a given regression problem,the main goal of using SVM is to find the optimal hyperplane with the largest margin separating classes of data.The schematic diagram is shown in Fig.1.

    For the linear regression problem,a data set is considered, and each input{(xi,yi),i=1,2,…,n}is mapped into the corresponding output.The approximate value can be obtained by the linear function:

    In order to ensure the flatness of the function(1),finding out the minimum w is essential.A hypothesis is suggested:all the data points can fit with the linear function in the accuracy of ε. Then minimizing w transforms into the problem of reducing model complexity,namely,the following quadratic programming problem:min(1/2||w||2)

    Fig.1 Principle description of SVM for regression problems

    In consideration of the fitting error,the slack variablesand C are introduced.The constant C>0 is a regularization constant determining the trade-off between the training error and the model flatness.Then the problem correspondingly changes into the following optimization one:

    The solution is a linear regression function of the optimal hyperplane as follows:

    where,αi,and b are the parameters which play the performance of determining the optimal hyperplane and can be achieved by means of working out the above constraint conditions(3).

    For the non-linear regression problem,with the help of the nonlinear mapping ψ,all data points are mapped into the high dimensional feature space.Then the problem can be perfectly solved just by using the above linear regression method.With the transformation of the nuclear function,the points are successfully mapped into the high dimensional feature space by the support vector machine.The kernel function accords with the following constraint:K(x,xi)=<ψ(x)·ψ(xi)>.

    Once the coefficients are determined,the regression estimate is given by Eq.(5):

    In the present study,the kernel function mainly contains four forms:linear nuclear,polynomial nuclear,radial basis function (RBF),and sigmoid.Herein,the widely used technique RBF was adopted for research.

    2.4 Model validation

    Model validation is proved to be crucial to QSPR modelling. It is acknowledged that the three aspects of fitting ability,robust performance,and predictive power are all very important. If one of them is ignored,we would never reach any comprehensive evaluation.Hence,according to OECD principles,the QSPR models,which have been built,must be comprehensively validated from the above mentioned three indices.The quality of fitting ability of the models is judged by the squared correlation coefficient R2,the average absolute error(AAE),and the root mean square error(RMSE).R2is an indicator that measures linear correlation degree between one variable and another.RMSE indicates dispersion degree of random error.The larger R2is,the smaller RMSE will be,and the model will have more fitting ability.However,good fitness does not stand for good robustness and predictive ability,thus internal validation is considered to be necessary for model validation.The internal predictive capability of a model is evaluated by leave-one-out cross-validationon the training set,which is defined as the following equation(6):

    Excellent Q2can illustrate robustness as well as excellent internal predictive ability of the QSPR models;yet,it cannot guarantee the true predictive ability of the models.Roy20and Pinheiro21et al.pointed out that the external validation was a crucial and indispensable validation method used to determine the true predictive ability of the QSPR models for new chemicals.The predictive ability of a model on external validation set can be expressed byby the following equation(7):

    where,yiandare respectively the experimental,predicted values of the test set,andis the mean experimental lnTC values of the samples in the training set.

    3 Results and discussion

    3.1 Interpretation of the selected descriptors

    The descriptors of each compound were calculated by CODESSA.HM can give one-to six-parameter models.When adding another descriptor cannot improve significantly the statistics of a model,it shows that the optimum subset size has been achieved.To avoid over-parametrization of the model,an increase in the value which is less than 0.02 is chosen as the breakpoint criterion.22The influences of the numbers of descriptors on R2are shown in Fig.2.

    Fig.2 shows that 5-parameter correlation is an ideal choice and adopted for the model input.Therefore,five descriptors were proposed as the model input.

    3.2 Results of HM

    After the heuristic reduction,a linear model was built shown in the following equation(8):

    Fig.2 R2versus descriptor number

    where,lnTC is the logarithm value of the thermal conductivity, F is Fish criterion,S is corrected mean square error,and n is the number of the sample.The type and the definition of the five descriptors and statistical parameters are described in Table 2.The comparison of the predicted and the experimental values are presented in Fig.3.

    The aim of this study is to seek the structural factors that influence the thermal conductivity by the analysis of the descriptors.There are five descriptors including two electrostatic descriptors,two constitutional descriptors,and one quantumchemical descriptor.These descriptors reflect different characters of the molecular structure.It is observed that constitutional descriptors and electrostatic descriptors play a main role.HA dependent HDSA-1/TMSA among electrostatic descriptors represents solvent-accessible surface area of H-bonding donor H atoms which is affected by the size of the hydrogen bonding interaction.The definition of FPSA3 fractional PPSA(PPSA3/ TMSA)[Zefirov?s PC]is FPSA3=PPSA3/TMSA,PPSA3 indicates total charge weighted partial positively charged molecular surface area,TMSA refers to total molecular surface area, and FPSA3 is expressed by fractional atomic charge weighted partial positive surface area.Relative molecular weight and relative number of C atoms are the constitutional descriptors. Quantum-chemical descriptor of the model is minimun(Min>0.1)bond order of an F atom.How much influence each descriptor has on the thermal conductivity is judged by comparing the coefficients before descriptors.If the coefficient is positive,the correlation between the descriptors and the thermal conductivity is positive,otherwise,negative.The larger the absolute value of the coefficient is,the more influence the descriptor is.Accordingly the permutation order is X3>X1>X2>X4>X5.

    3.3 Results of SVM

    For further research on the non-linear relationship betweenthe thermal conductivity and the molecular structure,SVM was proposed for the non-linear model.Descriptors were selected by HM as input and the thermal conductivity as output. It is difficult to choose related parameters when SVM is developed to predict.Inappropriate parameter selection can make a serious impact on the precision or accuracy of the prediction.

    Table 2 Selected descriptors and statistical parameters

    Fig.3 Comparison between the predicted and experimental lnTC by HM for test set

    RBF was chosen as a kernel function for appropriate related parameters which is most widely used at present.23-25Grid point search(GS)was applied to choose the best parameter combination.The search range of C and γ named the width of RBF, were both from 2-8to 28,and the step length was 1.Then the best coefficient was determined on the basis of Q2loocalculated by leave-one-out cross-validation for the training set.The optimal parameters are displayed as follows:C=256,γ=0.2500,ε= 0.1.The main performance parameters are presented in Table 3 and the comparison between the predicted and the experimental values are shown in Fig.4.

    3.4 Comparison and analysis of the results

    From Fig.3 and Fig.4,one can see that the calculated conversion values of HM and SVM are in good agreement with the experimental ones(Table 4),and the prediction accuracy is satisfying.As shown in Table 3,RMSE and AAE of HM and SVM are both small for the test set,and compared the training set with the test one,the prediction error is close to each other. This depicts that the constructed models have not only higher prediction ability but also better generalization performance than previous ones.

    In addition,the specific calculation results of relative errorfor HM and SVM are shown in Fig.5.In terms of HM,the mean relative error is 0.7961%and the maximum relative error is 2.986%.The relative errors of 52 compounds are smaller than 0.5%,whose number accounts for about 35%of the whole samples.For SVM,the mean relative error is 0.7435% and the maximum relative error is 2.965%.The relative errors of 70 compounds are smaller than 0.5%,whose number accounts for more than 50%of the whole samples.Compared with HM,the number of larger prediction error in SVM is significantly reduced.The data in Table 3 reveals that for the training set,R2of SVM is a bit smaller than that of HM,but for the test set,R2of SVM is higher than that of HM;however,prediction ability is regarded more important,therefore,SVM model for the thermal conductivity is better than HM model.

    Table 3 Performance comparison between the results of two models

    Fig.4 Comparison between the predicted and experimental lnTC by SVM for test set

    Although prediction effect of the two models in this study is satisfying,abnormal values still exist.Abnormal values(Table 5)severely influence the prediction performance of the models.If abnormal values of HM and SVM are screened,the performance of the models is greatly improved.

    Further residual analysis on the test set of the two models was performed(Fig.6).The residuals of HM model and SVM model are both randomly distributed in both sides of the line. Thus it can be concluded that the system error is not produced in the built process for HM model and SVM model,and the built models are very robust.

    Fig.5 Number of compounds for each interval and the percent errors obtained by HM and SVM

    Table 4 Experimental and predicted lnTC by HM and SVM

    Table 5 Abnormal values of predicted ln(TC/(mW·m-1·K-1))

    Fig.6 Comparative residuals vs experimental lnTC of test set for the HM and SVM models

    4 Conclusions

    (1)HM in software CODESSA not only screens molecular descriptors but also builds a linear model in this paper.Then SVM constructs a non-linear model with the screened descriptors.The two models are satisfying;especially SVM has stronger ability to predict.

    (2)Analysis of the screened abnormal values and the residual figure can improve the built model and make the model robust.

    (3)The interpretation of the models indicates that the influential factors are solvent-accessible surface area of H-bonding donor H atoms,fractional atomic charge,weighted partial positive surface area,relative number of C atoms,and relative molecular weight.Of all the factors,solvent-accessible surface area of H-bonding donor H atoms,fractional atomic charge,and weighted partial positive surface area are primary ones.

    (1) Gao,S.;Cao,C.Z.Acta Phys.-Chim.Sin.2006,22,1478. [高 碩,曹晨忠.物理化學(xué)學(xué)報,2006,22,1478.]doi:10.3866/ PKU.WHXB20061209

    (2) Khajeh,A.;Modarress,H.Struct.Chem.2011,22,1315.doi: 10.1007/s11224-011-9828-6

    (3) Rides,M.;Morikawa,J.;Halldahl,L.;Hay,B.;Lobo,H.; Dawson,A.;Allen,C.Polymer Testing 2009,28,480.doi: 10.1016/j.polymertesting.2009.03.002

    (4) Coquard,R.;Panel,B.Int.J.Therm.Sci.2009,48,747.doi: 10.1016/j.ijthermalsci.2008.06.005

    (5) Huang,L.H.;Liu,L.S.J.Food Eng.2009,95,179.doi: 10.1016/j.jfoodeng.2009.04.024

    (6) Nagasaka,Y.;Nagashima,A.Rev.Sci.Instrum.1981,52,229. doi:10.1063/1.1136577

    (7) Sastri,S.R.S.;Rao,K.K.Chem.Eng.J.1999,74,161.

    (8)Toropov,A.A.;Toropova,A.P.;Benfenati,E.J.Math.Chem. 2009,46,1060.doi:10.1007/s10910-008-9491-3

    (9) Shi,J.J.;Chen,L.P.;Shi,N.;Xu,W.;Yang,H.;Chen,W.H. China Safety Science Journal 2011,21,125.[時靜潔,陳利平,石 寧,徐 偉,楊 惠,陳網(wǎng)樺.中國安全科學(xué)學(xué)報,2011, 21,125.]

    (10)Tamm,K.;Burk,P.J.Mol.Model.2006,12,417.doi:10.1007/ s00894-005-0062-2

    (11) Gharagheizi,F.Comput.Mater.Sci.2007,40,159.doi:10.1016/ j.commatsci.2006.11.010

    (12) Gong,Z.G.;Zhang,R.S.;Xia,B.B.;Hu,R.J.;Fan,B.T. QSAR Comb.Sci.2008,27,1282.doi:10.1002/qsar.200860027 (13)Lu,P.;Wei,X.;Zhang,R.S.;Yuan,Y.G.;Gong,Z.G.Med. Chem.Res.2011,20,1220.doi:10.1007/s00044-010-9431-1

    (14)Long,W.;Liu,P.X.;Li,X.R.;Xu,Y.;Yu,J.;Ma,S.T.;Yu,L. L.;Zou,Z.M.J.Chemometrics 2009,23,304.doi:10.1002/ cem.1235

    (15) Bini,R.;Malvaldi,M.;Pitner,W.R.;Chiappe,C.J.Phys.Org. Chem.2008,21,622.doi:10.1002/poc.1337

    (16) Pan,Y.Research on Prediction Model and Quantitative Relationship between the Structures and Flammability Characteristics of Organic Compounds.Ph.D.Dissertation, Nanjing University of Technology,Nanjing,2009. [潘 勇.有機(jī)物定量結(jié)構(gòu)-燃爆特性相關(guān)性及預(yù)測模型研究[D].南京:南京工業(yè)大學(xué),2009.]

    (17)Katritzky,A.R.;Lobanov,V.S.;Karelson,M.CODESSA Version2.0 Reference Manual;University of Florida:Florida, 1995-1997.

    (18) Vapnik,V.N.The Nature of Statistical Learning Theory;Wiley: New York,1998.

    (19) Ojha,P.K.;Mitra,I.;Das,R.N.;Roy,K.Chemomet.Intell. Lab.Syst.2011,107,194.doi:10.1016/j.chemolab.2011.03.011

    (20) Roy,K.;Mitra,I.;Kar,S.J.Chem.Inf.Model.2012,52,396. doi:10.1021/ci200520g

    (21)Pinheiro,L.M.V.;Ventura,M.C.M.M.;Moita,M.L.C.J.J. Mol.Liq.2010,154,102.doi:10.1016/j.molliq.2010.04.013

    (22) Strouf,O.Chemical Pattern Recognition;Wilely:New York 1986.

    (23) Lin,S.L.;Liu,Z.Journal of Zhejiang University of Technology 2007,35,163.[林升梁,劉 志.浙江工業(yè)大學(xué)學(xué)報,2007, 35,163.]

    (24) Pan,Y.;Jiang,J.C.;Wang,R.;Cao,H.Y.;Cui,Y.J.Hazard. Mater.2009,164,1242.doi:10.1016/j.jhazmat.2008.09.031

    (25)Yang,H.;Chen,L.P.;Xie,C.X.;Shi,N.;Chen,W.H.Fire Safety Science 2011,20,62.[楊 惠,陳利平,謝傳欣,石 寧,陳網(wǎng)樺.火災(zāi)科學(xué),2011,20,62.]

    July 16,2012;Revised:September 10,2012;Published on Web:September 27,2012..

    Prediction of the Thermal Conductivity of Organic Compounds Using Heuristic and Support Vector Machine Methods

    SHI Jing-Jie1,2CHEN Li-Ping1CHEN Wang-Hua1,*SHI Ning2YANG Hui1XU Wei2
    (1Department of Safety Engineering,School of Chemical Engineering,Nanjing University of Science&Technology,Nanjing 210094, P.R.China;2State Key Laboratory of Chemical Safety and Control,Qingdao 266071,Shandong Province,P.R.China)

    To build the quantitative structure-property relationship(QSPR)between the molecular structures and the thermal conductivities of 147 organic compounds and investigate which structural factors influence the thermal conductivity of organic molecules,the topological,constitutional,geometrical, electrostatic,quantum-chemical,and thermodynamic descriptors of the compounds were calculated using the CODESSA software package,where these descriptors were pre-selected by the heuristic method (HM).The dataset of 147 organic compounds was randomly divided into a training set(118),and a test set (29).As a result,a five-descriptor linear model was constructed to describe the relationship between the molecular structures and the thermal conductivities.In addition,a non-linear regression model was built based on the support vector machine(SVM)with the same five descriptors.It was concluded that,although the fitting performance of the SVM model(squared correlation coefficient,R2=0.9240)was slightly worse than that of the HM model(R2=0.9267),the predictive performance of the SVM model(R2=0.9682)was better than that of the HM model(R2=0.9574).As the predictive parameter is more important than the fitting parameter,it can be seen that the SVM model is superior to the HM model.The proposed methods(SVM and HM)can be successfully used to predict the thermal conductivity of organic compounds with pre-selected theoretical descriptors,which can be directly calculated solely from the molecular structure.

    Heuristic method; Support vector machine; Thermal conductivity;Prediction;QSPR

    10.3866/PKU.WHXB201209273

    O641

    ?Corresponding author.Email:chenwh_nust@163.com;Tel:+86-25-84315526.

    The project was supported by the National Key Basic Research Program of China(973)(2010CB735510).

    國家重點(diǎn)基礎(chǔ)研究發(fā)展規(guī)劃項目(973)(2010CB735510)資助

    猜你喜歡
    熱導(dǎo)率南京向量
    南京比鄰
    “南京不會忘記”
    空位缺陷對單層石墨烯導(dǎo)熱特性影響的分子動力學(xué)
    向量的分解
    聚焦“向量與三角”創(chuàng)新題
    連續(xù)碳纖維鋁基復(fù)合材料橫向等效熱導(dǎo)率的模擬分析
    Si3N4/BN復(fù)合陶瓷熱導(dǎo)率及其有限元分析
    南京·九間堂
    金色年華(2017年8期)2017-06-21 09:35:27
    又是磷復(fù)會 又在大南京
    向量垂直在解析幾何中的應(yīng)用
    极品教师在线视频| 男人狂女人下面高潮的视频| 亚洲av免费在线观看| 国产激情偷乱视频一区二区| 精品久久久久久成人av| 久久九九热精品免费| 好男人视频免费观看在线| 日韩成人av中文字幕在线观看| 夜夜看夜夜爽夜夜摸| 国产成人91sexporn| 亚洲真实伦在线观看| av福利片在线观看| 麻豆久久精品国产亚洲av| 在线观看66精品国产| 一个人看视频在线观看www免费| 成人一区二区视频在线观看| 如何舔出高潮| 日韩三级伦理在线观看| 亚洲精品成人久久久久久| 麻豆国产97在线/欧美| 成人漫画全彩无遮挡| 在线观看美女被高潮喷水网站| 狂野欧美激情性xxxx在线观看| 国产精品爽爽va在线观看网站| 国产成人影院久久av| 1024手机看黄色片| 青春草国产在线视频 | 最近最新中文字幕大全电影3| 99热这里只有是精品50| 久久久久久久久中文| 亚洲最大成人中文| 欧美日韩综合久久久久久| 天堂中文最新版在线下载 | 日本成人三级电影网站| 成人毛片60女人毛片免费| 欧美三级亚洲精品| 日日啪夜夜撸| 男人舔女人下体高潮全视频| 国产精品,欧美在线| 舔av片在线| 日韩欧美精品免费久久| 午夜爱爱视频在线播放| 青春草国产在线视频 | 久久久久久久久中文| 精品国内亚洲2022精品成人| 国产日本99.免费观看| а√天堂www在线а√下载| 欧美zozozo另类| 国产精品久久久久久av不卡| 亚洲四区av| 国产真实乱freesex| 在线免费观看不下载黄p国产| 如何舔出高潮| 观看免费一级毛片| 人人妻人人看人人澡| 成人高潮视频无遮挡免费网站| 在线国产一区二区在线| 人妻制服诱惑在线中文字幕| 天天一区二区日本电影三级| 你懂的网址亚洲精品在线观看 | 欧美精品国产亚洲| 免费观看的影片在线观看| 九九爱精品视频在线观看| 国产毛片a区久久久久| 国产一区二区在线观看日韩| 午夜福利在线观看吧| 一卡2卡三卡四卡精品乱码亚洲| 亚洲国产精品合色在线| 3wmmmm亚洲av在线观看| 美女内射精品一级片tv| 真实男女啪啪啪动态图| 1000部很黄的大片| 国产精品日韩av在线免费观看| 好男人在线观看高清免费视频| 久久久久久久亚洲中文字幕| 男女啪啪激烈高潮av片| 亚洲最大成人av| 国产日本99.免费观看| 日韩在线高清观看一区二区三区| 国产精品无大码| 欧美性猛交╳xxx乱大交人| 久久精品国产亚洲网站| 亚洲人成网站在线观看播放| 哪个播放器可以免费观看大片| 日本免费a在线| 成人二区视频| 精品不卡国产一区二区三区| 天堂√8在线中文| 午夜福利视频1000在线观看| 国产精品电影一区二区三区| 日本五十路高清| 少妇人妻一区二区三区视频| 18禁在线播放成人免费| 国产v大片淫在线免费观看| 伦精品一区二区三区| 少妇被粗大猛烈的视频| 国产男人的电影天堂91| 麻豆成人午夜福利视频| 久久人妻av系列| 日本-黄色视频高清免费观看| 国产精品一及| av女优亚洲男人天堂| 午夜福利成人在线免费观看| 老司机福利观看| 又爽又黄无遮挡网站| 麻豆成人午夜福利视频| 欧美高清性xxxxhd video| 国产av麻豆久久久久久久| 你懂的网址亚洲精品在线观看 | 波多野结衣巨乳人妻| 亚洲在线观看片| 国产精品不卡视频一区二区| 看十八女毛片水多多多| 天堂中文最新版在线下载 | 亚洲精品日韩在线中文字幕 | 成人亚洲精品av一区二区| 亚洲在线自拍视频| 欧美人与善性xxx| 人体艺术视频欧美日本| 亚洲国产精品sss在线观看| 国产精品一区二区三区四区久久| 卡戴珊不雅视频在线播放| 99热这里只有是精品在线观看| 男女那种视频在线观看| 少妇的逼水好多| 伊人久久精品亚洲午夜| 国产乱人视频| 国内揄拍国产精品人妻在线| 97在线视频观看| av又黄又爽大尺度在线免费看 | 欧美+亚洲+日韩+国产| 久久这里只有精品中国| av在线天堂中文字幕| 99在线视频只有这里精品首页| 午夜福利成人在线免费观看| 亚洲三级黄色毛片| 人妻制服诱惑在线中文字幕| 亚洲av中文av极速乱| 久久久成人免费电影| 亚洲最大成人av| 国产精品久久视频播放| 成年av动漫网址| 白带黄色成豆腐渣| 看非洲黑人一级黄片| 成人午夜高清在线视频| 欧美xxxx性猛交bbbb| av专区在线播放| 18+在线观看网站| 亚洲在线观看片| 内射极品少妇av片p| 国产 一区 欧美 日韩| 一级av片app| 男人的好看免费观看在线视频| 99国产精品一区二区蜜桃av| 亚洲七黄色美女视频| 午夜免费激情av| 色综合站精品国产| 国产高清激情床上av| 美女国产视频在线观看| 一进一出抽搐gif免费好疼| av天堂中文字幕网| 99热精品在线国产| 特大巨黑吊av在线直播| 久久久久性生活片| 麻豆国产av国片精品| 国产中年淑女户外野战色| 欧美xxxx性猛交bbbb| 男插女下体视频免费在线播放| 国产精品精品国产色婷婷| 青春草亚洲视频在线观看| 99久久精品热视频| 日本爱情动作片www.在线观看| 午夜福利在线在线| 少妇被粗大猛烈的视频| 国产国拍精品亚洲av在线观看| 美女黄网站色视频| 国产一区二区激情短视频| av在线观看视频网站免费| 九九在线视频观看精品| 国内精品美女久久久久久| 一个人看视频在线观看www免费| 国产91av在线免费观看| 尤物成人国产欧美一区二区三区| 日本三级黄在线观看| 欧美一级a爱片免费观看看| 蜜桃久久精品国产亚洲av| 成人二区视频| 99久久精品国产国产毛片| av又黄又爽大尺度在线免费看 | 欧美高清成人免费视频www| 国产 一区 欧美 日韩| 一级av片app| 国产精品野战在线观看| 日本爱情动作片www.在线观看| 免费人成视频x8x8入口观看| 欧美成人a在线观看| 日本av手机在线免费观看| 床上黄色一级片| 国产精品无大码| kizo精华| 欧美日韩一区二区视频在线观看视频在线 | 久久这里只有精品中国| 国产亚洲91精品色在线| 插阴视频在线观看视频| 欧美成人免费av一区二区三区| 在线观看66精品国产| 丰满乱子伦码专区| 国产大屁股一区二区在线视频| 亚洲激情五月婷婷啪啪| 插阴视频在线观看视频| 日本熟妇午夜| АⅤ资源中文在线天堂| 国产v大片淫在线免费观看| av.在线天堂| 嫩草影院精品99| 国产av一区在线观看免费| 22中文网久久字幕| 成人午夜高清在线视频| 日韩在线高清观看一区二区三区| 爱豆传媒免费全集在线观看| 午夜精品国产一区二区电影 | 69av精品久久久久久| 精品久久久噜噜| 精品人妻视频免费看| 美女cb高潮喷水在线观看| 日韩在线高清观看一区二区三区| 97超视频在线观看视频| 校园春色视频在线观看| 波多野结衣高清无吗| 欧美zozozo另类| 欧美潮喷喷水| 免费av毛片视频| 久久精品夜色国产| 国产一级毛片七仙女欲春2| 青春草亚洲视频在线观看| 日韩欧美国产在线观看| 日韩中字成人| 亚洲成人久久爱视频| 看十八女毛片水多多多| or卡值多少钱| 国产日韩欧美在线精品| 久久精品久久久久久噜噜老黄 | 成人欧美大片| 99热网站在线观看| 婷婷色av中文字幕| 国产精品久久久久久亚洲av鲁大| 一本久久精品| 少妇高潮的动态图| 超碰av人人做人人爽久久| 亚洲人成网站高清观看| 久久精品夜夜夜夜夜久久蜜豆| 国产精品美女特级片免费视频播放器| 国产熟女欧美一区二区| 亚洲av熟女| 看片在线看免费视频| 日韩制服骚丝袜av| 极品教师在线视频| 亚洲av中文av极速乱| 一级毛片我不卡| 伦精品一区二区三区| 免费不卡的大黄色大毛片视频在线观看 | 国产精品免费一区二区三区在线| 婷婷色av中文字幕| 久久久午夜欧美精品| 99riav亚洲国产免费| 伦精品一区二区三区| 亚洲国产欧洲综合997久久,| 级片在线观看| 男女下面进入的视频免费午夜| 久久精品夜色国产| 麻豆久久精品国产亚洲av| 日韩av在线大香蕉| 搞女人的毛片| 舔av片在线| 亚洲欧美日韩高清在线视频| 99久久精品热视频| 中文字幕制服av| 日本免费一区二区三区高清不卡| 国产亚洲精品久久久com| 国产成人freesex在线| 国国产精品蜜臀av免费| 中文字幕久久专区| 99热只有精品国产| 校园春色视频在线观看| 亚洲精品日韩av片在线观看| 日韩一区二区三区影片| 日韩欧美 国产精品| 一区二区三区免费毛片| 国产毛片a区久久久久| 中文字幕熟女人妻在线| 国产激情偷乱视频一区二区| 国产精品一区二区三区四区久久| 免费黄网站久久成人精品| 午夜福利在线在线| 三级毛片av免费| 久久精品91蜜桃| 亚洲欧美日韩卡通动漫| 国产成人freesex在线| 淫秽高清视频在线观看| 国产色婷婷99| 久久久久久久久久黄片| 国产亚洲av片在线观看秒播厂 | 亚洲在久久综合| 色综合站精品国产| 亚洲丝袜综合中文字幕| 又粗又硬又长又爽又黄的视频 | 亚洲自拍偷在线| 国产老妇伦熟女老妇高清| 在线观看66精品国产| 搡老妇女老女人老熟妇| 亚洲性久久影院| 亚洲国产欧美人成| 成人欧美大片| 麻豆国产97在线/欧美| 国产综合懂色| 91麻豆精品激情在线观看国产| 成人亚洲精品av一区二区| 国产亚洲5aaaaa淫片| 日本欧美国产在线视频| 美女脱内裤让男人舔精品视频 | 欧美一级a爱片免费观看看| 日韩制服骚丝袜av| 别揉我奶头 嗯啊视频| 97人妻精品一区二区三区麻豆| 午夜激情欧美在线| 亚洲激情五月婷婷啪啪| 国产真实乱freesex| 亚洲中文字幕日韩| videossex国产| 久久久久国产网址| 最近中文字幕高清免费大全6| 中文字幕av在线有码专区| 国产成人精品婷婷| 成人综合一区亚洲| 免费不卡的大黄色大毛片视频在线观看 | a级毛片免费高清观看在线播放| 成人美女网站在线观看视频| 国产精品久久久久久精品电影| 国产成人精品久久久久久| 亚洲精品国产成人久久av| 大又大粗又爽又黄少妇毛片口| 国产精品一及| 在线播放国产精品三级| 免费av毛片视频| 欧洲精品卡2卡3卡4卡5卡区| eeuss影院久久| 婷婷色av中文字幕| 哪个播放器可以免费观看大片| 99热这里只有是精品在线观看| 91午夜精品亚洲一区二区三区| 好男人在线观看高清免费视频| 国产熟女欧美一区二区| 性色avwww在线观看| h日本视频在线播放| 久久6这里有精品| 特级一级黄色大片| 免费看美女性在线毛片视频| 特级一级黄色大片| а√天堂www在线а√下载| 亚洲欧美成人精品一区二区| 亚洲18禁久久av| 最近最新中文字幕大全电影3| 男女做爰动态图高潮gif福利片| 午夜精品国产一区二区电影 | 亚洲欧美日韩高清专用| 亚洲第一区二区三区不卡| 在线观看午夜福利视频| 中文亚洲av片在线观看爽| 一级毛片我不卡| 三级男女做爰猛烈吃奶摸视频| 岛国在线免费视频观看| 久久久久国产网址| av在线播放精品| 不卡视频在线观看欧美| 亚洲aⅴ乱码一区二区在线播放| 伊人久久精品亚洲午夜| 亚洲aⅴ乱码一区二区在线播放| 看免费成人av毛片| 国产成人a∨麻豆精品| 白带黄色成豆腐渣| 午夜福利在线观看吧| 亚洲经典国产精华液单| 亚洲高清免费不卡视频| 国产老妇女一区| 美女内射精品一级片tv| 九九在线视频观看精品| 国产精华一区二区三区| 国产女主播在线喷水免费视频网站 | www.色视频.com| 亚洲图色成人| 国产精品久久久久久精品电影| 国产乱人偷精品视频| 菩萨蛮人人尽说江南好唐韦庄 | 精品久久久久久久久久久久久| 国产成人91sexporn| 亚洲第一电影网av| 久久精品国产清高在天天线| 亚洲欧美日韩高清在线视频| 国产精品三级大全| 赤兔流量卡办理| 91久久精品国产一区二区三区| 久久这里有精品视频免费| 久久精品综合一区二区三区| 日韩欧美在线乱码| 午夜亚洲福利在线播放| 欧美不卡视频在线免费观看| 性色avwww在线观看| 国产白丝娇喘喷水9色精品| 一个人观看的视频www高清免费观看| 人人妻人人澡人人爽人人夜夜 | 一区福利在线观看| 国产91av在线免费观看| 丰满的人妻完整版| 精品久久久久久久久久免费视频| 一边亲一边摸免费视频| 人人妻人人看人人澡| 国产精品精品国产色婷婷| 色5月婷婷丁香| 日本av手机在线免费观看| 一个人看视频在线观看www免费| 国产黄片视频在线免费观看| 国产精品.久久久| 一区二区三区免费毛片| 亚洲欧美精品自产自拍| 人妻少妇偷人精品九色| 人人妻人人澡欧美一区二区| 成人亚洲精品av一区二区| 久久精品夜夜夜夜夜久久蜜豆| 亚洲国产精品久久男人天堂| 国产私拍福利视频在线观看| 联通29元200g的流量卡| 久久久久久大精品| 国产探花极品一区二区| 午夜福利在线观看吧| 你懂的网址亚洲精品在线观看 | 青春草视频在线免费观看| 国产精品三级大全| 欧美日韩国产亚洲二区| 一个人观看的视频www高清免费观看| 一级黄色大片毛片| 国产成人影院久久av| 国产成人精品久久久久久| 能在线免费观看的黄片| 亚洲成人精品中文字幕电影| 九九爱精品视频在线观看| 国产精品,欧美在线| 日韩大尺度精品在线看网址| 欧美一区二区国产精品久久精品| 久久这里只有精品中国| 久久精品国产亚洲av天美| 热99在线观看视频| 国产成人精品一,二区 | 日韩国内少妇激情av| 69av精品久久久久久| 欧美日韩综合久久久久久| 给我免费播放毛片高清在线观看| 亚洲精品亚洲一区二区| 蜜桃亚洲精品一区二区三区| 国产精品久久久久久精品电影| 亚洲在线自拍视频| 亚洲最大成人av| 国产精品乱码一区二三区的特点| 男女视频在线观看网站免费| 午夜福利在线观看免费完整高清在 | 国产精品乱码一区二三区的特点| 亚洲精品影视一区二区三区av| 免费观看在线日韩| 三级毛片av免费| 少妇人妻一区二区三区视频| 国产av一区在线观看免费| 亚洲自偷自拍三级| 午夜免费男女啪啪视频观看| 亚洲七黄色美女视频| 女同久久另类99精品国产91| 在线免费十八禁| 最好的美女福利视频网| 欧美另类亚洲清纯唯美| 亚州av有码| 午夜免费男女啪啪视频观看| а√天堂www在线а√下载| 亚洲成人av在线免费| 国产精品嫩草影院av在线观看| 久久久久久九九精品二区国产| 色视频www国产| 日韩精品有码人妻一区| 精品久久久噜噜| 国产一区二区三区在线臀色熟女| 国产精品国产高清国产av| 春色校园在线视频观看| 你懂的网址亚洲精品在线观看 | 亚洲五月天丁香| 国产免费男女视频| 久久精品综合一区二区三区| 久久人妻av系列| 久久精品夜夜夜夜夜久久蜜豆| 嘟嘟电影网在线观看| 神马国产精品三级电影在线观看| 国产精品一区www在线观看| 久久精品91蜜桃| 性色avwww在线观看| 国产精品一区二区在线观看99 | 国产伦精品一区二区三区四那| 亚洲国产欧美在线一区| 十八禁国产超污无遮挡网站| 亚洲av男天堂| 亚洲天堂国产精品一区在线| 日韩制服骚丝袜av| 欧洲精品卡2卡3卡4卡5卡区| 黄片wwwwww| 成人亚洲精品av一区二区| 亚洲精品成人久久久久久| 久久草成人影院| 美女大奶头视频| 免费看日本二区| 国产午夜精品论理片| 中国国产av一级| 色综合色国产| 成人av在线播放网站| 国产 一区精品| 十八禁国产超污无遮挡网站| 青春草亚洲视频在线观看| 亚洲av二区三区四区| 亚洲国产精品成人久久小说 | 内地一区二区视频在线| 亚洲内射少妇av| 日韩,欧美,国产一区二区三区 | 99久国产av精品| eeuss影院久久| 精品日产1卡2卡| 国产 一区精品| 欧美成人免费av一区二区三区| 婷婷色综合大香蕉| 九九爱精品视频在线观看| 婷婷六月久久综合丁香| 尤物成人国产欧美一区二区三区| 99久久中文字幕三级久久日本| 国产亚洲av片在线观看秒播厂 | 秋霞在线观看毛片| 日产精品乱码卡一卡2卡三| 青春草国产在线视频 | 最近中文字幕高清免费大全6| 亚洲在久久综合| 日日摸夜夜添夜夜爱| 亚洲色图av天堂| 亚洲精品久久国产高清桃花| 女同久久另类99精品国产91| 久久久久九九精品影院| 午夜福利在线观看免费完整高清在 | 精品99又大又爽又粗少妇毛片| 国产不卡一卡二| 精品久久久久久成人av| 国产伦精品一区二区三区视频9| 国产精品永久免费网站| 免费av毛片视频| 久久久a久久爽久久v久久| 美女黄网站色视频| 天堂网av新在线| 亚洲国产欧美人成| 97超碰精品成人国产| 少妇裸体淫交视频免费看高清| 欧美激情国产日韩精品一区| 亚洲欧美日韩卡通动漫| 伦理电影大哥的女人| 成人无遮挡网站| 国产午夜福利久久久久久| 国产老妇伦熟女老妇高清| 男女啪啪激烈高潮av片| 国产精品一二三区在线看| 亚洲成人中文字幕在线播放| 亚洲久久久久久中文字幕| 亚洲av.av天堂| 亚洲欧美日韩高清在线视频| av在线播放精品| 边亲边吃奶的免费视频| 内地一区二区视频在线| 欧美性猛交黑人性爽| 一级黄色大片毛片| 欧美色欧美亚洲另类二区| 日韩欧美三级三区| 日日摸夜夜添夜夜添av毛片| 日韩高清综合在线| 国产极品天堂在线| 观看免费一级毛片| 观看美女的网站| 我的老师免费观看完整版| 老司机福利观看| 国产精品日韩av在线免费观看| 婷婷亚洲欧美| 99热6这里只有精品| 国产精品久久久久久精品电影| 日日摸夜夜添夜夜添av毛片| 春色校园在线视频观看| 免费观看a级毛片全部| 联通29元200g的流量卡| 国产成人a∨麻豆精品| 亚洲精品乱码久久久久久按摩| 天堂影院成人在线观看| 成人漫画全彩无遮挡| 亚洲一区二区三区色噜噜| 国产一级毛片在线| 国产精品三级大全| 欧美最新免费一区二区三区| 日本撒尿小便嘘嘘汇集6| 日韩三级伦理在线观看| 久久鲁丝午夜福利片| 国产精品一区www在线观看| 国产伦精品一区二区三区视频9| 亚洲欧美成人精品一区二区| 午夜福利视频1000在线观看| 3wmmmm亚洲av在线观看| 亚洲av中文字字幕乱码综合| 插阴视频在线观看视频| 亚洲精品日韩av片在线观看| 免费一级毛片在线播放高清视频| 午夜福利在线观看吧|