• <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 ?

    波信號(hào)的解調(diào)和人工神經(jīng)網(wǎng)絡(luò)的損傷識(shí)別算法

    2010-12-04 08:56:54SaravananJuGuo
    無(wú)損檢測(cè) 2010年8期
    關(guān)鍵詞:人工神經(jīng)網(wǎng)絡(luò)信號(hào)算法

    S.Saravanan,F(xiàn).Ju,N.Q.Guo

    (1.School of Mechanical &Aerospace Engineering,Nanyang Technological University,Singapore;2.School of Engineering,Monash University,Bandar Sunway 46150,Selangor,Malaysia)

    1 Introduction

    Composite materials are widely and increasingly used due to their low weight,high specific stiffness and strength,good fatigue performance and excellent corrosion resistance. The major disadvantage of composite structures is the high probability of severe degrading of mechanical properties in the presence of damage especially in the through-thickness direction.The unique anisotropic,low-conductivity and low-permeability characteristics of composite materials have limited the applications of traditional nondestructive evaluation(NDE)techniques in damage detection.In addition,traditional NDE testing is usually timeconsuming point testing thus not suitable for large structure inspection.

    The long-range damage detection potential of Lamb waves has been studied extensively1-9.The interaction between Lamb wave and damage will modify the response wave signal from which information related to damage can be extracted for automated damage detection. However, the interpretation of the response wave signal is not easy due to the complex nature of the wave-damage interaction.

    Artificial neural network(ANN)is a powerful computational tool for pattern recognition and function approximation and has been tried in Lamb wave-based damage detection10,11.However,the accuracy of the network has been limited by the redundant highdimensional features and the very complicated network architecture.In this paper,a new damage detection scheme is proposed which uses signal demodulation for feature extraction,an unsupervised neural network for clustering and feature dimensionality reduction,and a supervised neural network for damage characterization.

    2 Numerical simulation of lamb waves

    2.1 Numerical model for simulation

    A 2D unidirectional composite laminate with a notch defect is modeled for simulation.The laminate is 300mm long and 1mm thick as shown in Figure 1.The notch is located at x mm from the left end of the laminate with width 0.4 mm and depth y mm.The element size is 0.2mm×0.125mm,and the time step is 0.011 099 9μs.

    Figure 1.A 2Dnumerical model of a unidirectional composite laminate

    A five-cycle sinusoidal signal modulated by a hanning window with the center frequency of 500kHz is applied at the left end in the x-direction in the form of pressure to excite S0Lamb wave mode,as shown in Figure 2.The phase velocity dispersion curve for S0in Figure 2 is quite flat at the chosen operation frequency.

    2.2 Damage cases

    Thex-andy-displacement response signals at sensing point 1(75mm)and 2(225mm)in Figure 1 are recorded during simulation.A simulation without damage is performed first and the results are plotted in Figure 3.Then simulations are performed for 21notch locations(x:100-200mm,Δx:5mm)and 7notch depths(y:0.125~0.875mm,Δy:0.125mm),the combination of them results in 147damage cases in total.For example,the response signals when the notch is at the middle(x=150mm andy=0.5mm)are plotted in Figure 4.By comparing with Figure 3,there are additional wave packets and change of amplitude due to wave-damage interaction such as reflection,diffraction and mode conversion.So the response signals contain the information associated with damage state from which features can be extracted for damage characterization.

    3 Feature extraction

    3.1 Baseline subtraction

    The signals in Figure 3are used as the baselines and subtracted from the signals in Figure 4thus the resulting signals indicate the effect of damage on the response signals,as shown in Figure 5.

    3.2 Wave signal demodulation

    Each wave packet in Figure 5can be considered as a low-frequency envelope signalx(n)modulated by a high-frequency sine carrier signalc(n)as shown in Figure 6.The modulation operation is carried out by multiplication in the time domainy(n)=x(n)·c(n),which results in the convolution in the frequency domainY(ω)=X(ω)*C(ω),as shown in Figure 7.The modulated signal is then transmitted through the plate and received by the sensor.In order to retrieve the envelope signal from the received signal,the received signal is convoluted by itself in the frequency domainZ(ω)=Y(jié)(ω)*Y(ω)as shown in Figure 7.It is obvious that the resulting signalz(n)has a lowfrequency component and a high-frequency component in 0≤ω≤π/2 which can be separated by a lowpass filter with the cutoff frequency in between.The magnitude response of the filter is shown in Figure 8 and the system function is:

    The signalz(n)is then filtered with this lowpass filter and the result is shown in Figure 8,which is a slightly delayed envelope signal.This demodulation algorithm for envelope extraction is computationally more efficient than the conventional Hilbert transform and do not have discrepancies at two ends.

    3.3 Peak extraction

    The signals in Figure 5are demodulated with the above algorithm and the results are shown in Figure 9,which are related to the energy change due to the wave-damage interaction. The peak values and locations are extracted by finding the local maxima and combined into the following 8-dimensional feature vector:

    wherepkixandpkiyare the peak values at sensor i in thex-andy-direction,respectively.locixandlociyare the peak locations at sensor i in thex-andydirection,respectively.The feature vectors will be used as input vectors to the unsupervised and supervised artificial neural networks for pattern recognition.

    4 Unsupervised learning(SOM)

    4.1 SOM neural network

    The self-organizing map (SOM)12is an unsupervised artificial neural network model for nonlinearly mapping the high-dimensional input vectors onto a low-dimensional,topologically ordered array of neurons,inspired by the topographical mapping ability of the human brain cortex.It has become a powerful tool for clustering,feature selection and highdimensional data visualization due to its properties of input space approximation,topological ordering and density matching13.In this paper,a two-dimensional 4×4Kohonen SOM neural network is used and the architecture is shown in Figure 10.

    Figure 10.The architecture of the SOM neural network

    The training of the SOM is based on unsupervised competitive learning consisting of five essential processes:

    (1)Initialization.The initial synaptic weight vectorswj(0)of neurons are first assigned small random values which should be different from each other.

    (2)Competition.A discriminant function related to the distance between the input vectorxand the weight vectorwjis selected and its value is calculated for each neuron in the network.The neuroni(x)with the minimum discriminant function value is called the winning neuron:

    where‖·‖is the Euclidean norm,andi(x)is the index of the winning neuron.

    (3)Cooperation.A topological neighborhoodhj,iof cooperating neurons centered on the winning neuron is determined with the following Gaussian topological neighborhood function:

    wheredj,iis the lateral distance between winning neuroniand cooperating neuronj,σ(n)is the width of the Gaussian function with initial valueσ0and time constantτ1,nis the number of iterations.

    (4)Weights adaptation.The synaptic weight vectors of all neurons are updated using the following equations:

    whereη(n)is the learning rate with initial valueη0and time constantτ2.

    (5)Iteration.Repeat processes(2)~(4)by randomly presenting the training samples to the network until the stopping criterion is met,which can be the predefined number of iterations or the small rate of changes in the map's weights.

    4.2 Damage clustering using SOM

    An important ability of the SOM neural network is clustering the data into different categories in an unsupervised manner.For example,the peak values of all samples are used as 4-dimensional input patterns to the SOM network in Figure 10.After training without the damage information,the network converges and 7 clusters are formed with 11samples in each,as shown in Figure 11.By checking the samples in each cluster,it is found that samples with the same severity of damage(depth of the notch)are sorted into the same cluster.Now the clusters can be labeled with damage level 1-7.If a new input pattern with unknown damage severity is presented to the SOM,it will be sorted into the cluster which is most activated and the damage severity can be estimated from the cluster label.Similarly,if the peak locations of all samples are used as input patterns to the SOM network,11 clusters are formed with 7samples in each,as shown in Figure 11.It is observed that samples with the same damage location are clustered together.

    Figure 11.Clusters formed by the SOM neural network

    4.3 Feature dimensionality reduction using SOM

    Another important application of the SOM neural network in pattern recognition is the dimensionality reduction of features. This can save a lot of computational cost in the damage detection algorithm if the features are high-dimensional and correlated.Although dimensionality reduction can also be achieved using traditional principal component analysis(PCA)14,the SOM is more advantageous in visualization and will not lose the real meaning of features.In order to reduce the dimensionality of the features acquired in the feature extraction session,the feature vectors of all damage cases are used as input vectors to the SOM neural network.After training and convergence of the network,the weights are plotted in the weight planes in Figure 12.The 8weight planes correspond to the 8elements of the input vector.Each element in the weight plane represents the connection(weight)between one input element and one neuron,with the darkness of color indicating the magnitude of the weight.If the weight planes of two input elements are very similar,the two input elements are highly correlated.It is observed from Figure 12that the weight plane 1,3,5,7are almost the same,this means the input element 1,3,5,7are correlated.Also the correlation between input element 2and 4can be found from weight plane 2and 4.Therefore,the dimensionality of the 8-dimensional feature vector in Equation (2)can be reduced by eliminating the correlated elements,resulting in the following 4-dimensional feature vector which will be used as input to the supervised neural network:

    Figure 12.Weight planes of the SOM neural network

    5 Supervised learning(MLP)

    5.1 MLP neural network and BP learning algorithm

    The multi-layer perceptron(MLP)neural network is the most widely used model of supervised neural network due to its excellent performance in function approximation, associative memory and pattern classification.The typical MLP neural network is a feed-forward network containing one input layer,one or more hidden layers and one output layer,as shown in Figure 13.The input layer neurons do not perform any computation and just distribute the input vectors to the hidden layer.The hidden layer and output layer neurons are computational neurons with a continuously differentiable nonlinear activation functionφ(·),which can be sigmoidal functions such as the following logistic or hyperbolic tangent function:

    Figure 13.The architecture of the MLP neural network

    wherevj(n)=(n)is the activation signal of neuronjat iterationn.

    The training of the MLP neural network is in a supervised manner.During the training process,the input vectorxis presented to the network and the outputois generated by the network.By comparing the output with the a priori desired outputd,an error signale=d-ocan be obtained.Then the adjustments to the synaptic weights of the network are calculated based on the error signal so that the network output can approximate the desired output.The weight adjustments of the output layer neurons can be easily determined using optimization methods such as the gradient descent.However,the calculation of the weight adjustments of the hidden layer neurons becomes a problem which has not been solved until the development of the back-propagation (BP)algorithm15.This algorithm has become the most popular learning algorithm for the training of MLPs due to its high computational efficiency.In the BP algorithm,two passes of computation are identified13:the forward pass and the backward pass.In the forward pass when the synaptic weights remain unchanged,the function signals come in at the input layer,propagate forward on a neuron-by-neuron,layer-by-layer basis and emerge at the output layer as output signals in Equation(8):

    In the backward pass,the error signals are computed at the output layer,propagate backward,layer-by-layer, accompanied by the recursive calculation of the local gradient for each neuron which enables the adjustment of synaptic weight in the following delta rule:

    whereΔwji(n)is the correction applied to the synaptic weight connecting neuronito neuronj,ηis the learning rate,δj(n)is the local gradient,andyi(n)is the input signal of neuronj.The calculation of the local gradient is shown in Equation (10),depending on whether neuronjis an output or a hidden neuron:

    The training process is iteratively performed by presenting epochs of training samples to the network until the stopping criterion is met,which can be the predefined number of iterations or the small rate of change in the mean square error.

    5.1 Damage characterization using MLP

    According to the universal approximation theorem16,any continuous nonlinear function with a finite number of discontinuities can be approximated arbitrarily well by a MLP neural network having one hidden layer of sufficient neurons.In this paper,a MLP neural network is used to approximate the unknown inverse model of the structure in order to estimate the notch parameters providing the features extracted from the response signals.The network has the architecture shown in Figure 13 with one input layer ofm0neurons,one hidden layer ofm1neurons and one output layer ofm2neurons.Herem0=4(number of dimensions of the input vector)andm2=2(number of notch parameters to be estimated).The number of hidden neuronsm1 which governs the express power of the network depends on the complexity of the function to be approximated.According to the Ockham's razor principle,the simplest network which can adequately fit the training set is more preferred.Unless special conditions of the problem are given,a three-layer network(one hidden layer)is sufficient to approximate any arbitrary function.Complex network with more hidden layers and neurons are more susceptible to overfitting which leads to poor generalization.Therefore only 6hidden neurons are used in this paper and this simple network performs well.

    The number of training samples should be more than the number of weights in the network.A rule of thumb for determining the number of training samples is14:

    Since there are 147damage cases,the training samples are enough.

    The elements in the feature vector extracted from the response signal have different orders of magnitude.If they are directly presented to the network for training,the elements with higher orders of magnitude will have dominant effect on the weight adjustments.In order to avoid this,standardization is performed to transform the input and target data into standardized data with zero mean and unit variance using the following equations:

    wherex(j)iis the ith element of the jth sample,is the mean,σ2iis the variance,andz(j)iis the standardized data.The boxplots of input data before and after standardization are shown in Figure 14.

    The standardized data set is then randomly divided into a training set and a testing set,with the ratio of 0.8to 0.2.The testing set is not used during training,but provides an independent test of the network's generalization ability.

    The stopping criterion is set asMSE=1×10-5,which is the mean square error in the standardized space.After 55 epochs of training,the stopping criterion is met and the performance history of the network is plotted in Figure 15and 16.The result shows a good generalization of the network since the performance on the testing set are quite close to that on the training set.And no evidence of overfitting is observed.

    In order to further assess the performance of the network,the entire data set is presented to the trained network and a linear regression analysis is carried out between the network outputs and the corresponding targets.The result is shown in Figure 15.Since the output vector is 2-dimensional,there are two plots.Both of them show the strong linear relationship between outputs and targets,with the correlation coefficients of 1and 0.999 99respectively.This means that the network fits the entire data set well.

    Finally,the errors between the network outputs and corresponding targets are calculated in the original space in terms of the root mean square error(RMSE),the normalized root mean square error(NRMSE)and the maximum absolute error (MAE),which are defined as follows:

    And the results in Table 1show that the errors are quite small,therefore the accuracy of the trained network is reasonably high.

    Table 1.Errors between the network outputs and corresponding targets

    Now the trained network can be used to detect unknown notches by extracting feature vectors form the response signals and presenting them to the network.

    6 Conclusion

    A damage detection algorithm based on Lamb wave signal demodulation and ANNs has been proposed in this paper, consisting of feature extraction,clustering,feature dimensionality reduction and damage characterization.The validity of this damage detection algorithm is verified using a FE model of a composite laminate with notch defects.The wave signal demodulation algorithm is able to demodulate the response Lamb wave signal into the envelope signal and extract peaks which are related to the energy change due to damage.Then the peak values and locations are combined into an 8-dimensional feature vector which is used as the input vector to ANNs.A 4×4SOM neural network is first employed in an unsupervised manner and it is shown that this network is capable of clustering the damage cases into categories according to the damage severity or location.Feature dimensionality reduction is also performed by this network to reduce the original highly correlated 8-dimensional feature vector into a 4-dimensional one.The 4-dimensional feature vector is then used as the input to a MLP neural network with simple architecture for damage characterization.The training of this network is in a supervised manner and based on BP algorithm.The performance of this network is then assessed using an independent testing set,the regression analysis and the evaluation of errors.It is shown that this network has high accuracy and good generalization ability.The developed system and methodology will be used and tested for future experimental signals and 3Dsimulation signals.

    [1] Alleyne DN,Cawley P.The interaction of Lamb waves with defects[J].IEEE Trans Ultrason Ferroelectr Freq Control,1992,39(3):381-397.

    [2] Guo N,Cawley P.The interaction of Lamb waves with delaminations in composite laminates[J].J Acoust Soc Am,1993,94(4):2240-2246.

    [3] Guo N,Cawley P.Lamb wave-propagation in composite laminates and its relationship with acousto-ultrasonics[J].Ndt &E International,1993,26(2):75-84.

    [4] Guo NQ,Cawley P.Lamb wave reflection for the quick nondestructive evaluation of large composite laminates[J].Mater Eval,1994,52(3):404-411.

    [5] Alleyne D,Cawley P.The long range detection of corrosion in pipes using Lamb waves[G].Annual Review of Progress in Quantitative Nondestructive Evaluation.Snowmass Village:Plenum Press Div Plenum Publishing,1994:2073-2080.

    [6] Scudder LP,Hutchins DA,Guo NQ.Laser-generated ultrasonic guided waves in fiber-reinforced plates -Theory and experiment[J].Ieee Transactions on Ultrasonics Ferroelectrics and Frequency Control,1996,43(5):870-880.

    [7] Lemistre M,Balageas D.Structural health monitoring system based on diffracted Lamb wave analysis by multiresolution processing[J].Smart Mater Struct,2001,10(3):504-511.

    [8] Jian XM,Guo N,Li MX,et al.Characterization of bonding quality in a multilayer structure using segment adaptive filtering [J]. Journal of Nondestructive Evaluation,2002,21(2):55-65.

    [9] Bartoli I,F(xiàn)L di Scalea,F(xiàn)ateh M,et al.Modeling guided wave propagation with application to the long-range defect detection in railroad tracks[J].NDT E Int,2005,38(5):325-334.

    [10] Su ZQ, Ye L. Lamb wave-based quantitative identification of delamination in CF/EP composite structures using artificial neural algorithm[J].Compos Struct,2004,66(1-4):627-637.

    [11] Lu Y,Ye L,Su ZQ,et al.Artificial Neural Network(ANN)-based Crack Identification in Aluminum Plates with Lamb Wave Signals[J].J Intell Mater Syst Struct,2009,20(1):39-49.

    [12] Kohonen T.The self-organizing map[J].Proc IEEE,1990,78(9):1464-1480.

    [13] Haykin SS. Neural networks: a comprehensive foundation,Prentice Hall,Upper Saddle River,NJ(1999).

    [14] Zang C,Imregun M.Structural damage detection using artificial neural networks and measured FRF data reduced via principal component protection[J].J Sound Vibr,2001,242(5):813-827.

    [15] Rumelhart DE,Hinton GE,Williams RJ.Learning representations by back-propagating errors[J].Nature,1986,323(6088):533-536.

    [16] Cybenko G.Approximation by superpositions of a sigmoidal function[J].Mathematics of Control,Signals,and Systems,1989,2(4):303-314.

    猜你喜歡
    人工神經(jīng)網(wǎng)絡(luò)信號(hào)算法
    信號(hào)
    鴨綠江(2021年35期)2021-04-19 12:24:18
    完形填空二則
    利用人工神經(jīng)網(wǎng)絡(luò)快速計(jì)算木星系磁坐標(biāo)
    基于MapReduce的改進(jìn)Eclat算法
    Travellng thg World Full—time for Rree
    人工神經(jīng)網(wǎng)絡(luò)實(shí)現(xiàn)簡(jiǎn)單字母的識(shí)別
    電子制作(2019年10期)2019-06-17 11:45:10
    進(jìn)位加法的兩種算法
    滑動(dòng)電接觸摩擦力的BP與RBF人工神經(jīng)網(wǎng)絡(luò)建模
    基于FPGA的多功能信號(hào)發(fā)生器的設(shè)計(jì)
    電子制作(2018年11期)2018-08-04 03:25:42
    一種改進(jìn)的整周模糊度去相關(guān)算法
    av在线播放免费不卡| 国产精品久久久人人做人人爽| 香蕉久久夜色| 欧美黄色淫秽网站| 精品免费久久久久久久清纯 | 水蜜桃什么品种好| 国产成人免费无遮挡视频| 国产成人啪精品午夜网站| tocl精华| 亚洲一区二区三区欧美精品| 岛国毛片在线播放| 宅男免费午夜| 午夜福利,免费看| 伊人久久大香线蕉亚洲五| 超碰97精品在线观看| 亚洲一区中文字幕在线| 久久午夜综合久久蜜桃| 1024视频免费在线观看| 国产亚洲欧美在线一区二区| 日韩欧美三级三区| 无遮挡黄片免费观看| 纯流量卡能插随身wifi吗| 欧美乱码精品一区二区三区| 十分钟在线观看高清视频www| 国产av精品麻豆| 精品亚洲乱码少妇综合久久| 一区二区三区激情视频| 精品少妇久久久久久888优播| 视频在线观看一区二区三区| 精品久久久久久久毛片微露脸| 亚洲美女黄片视频| 亚洲国产成人一精品久久久| 亚洲成人手机| 2018国产大陆天天弄谢| 欧美成狂野欧美在线观看| 欧美激情高清一区二区三区| 一区福利在线观看| 超碰97精品在线观看| 一本大道久久a久久精品| 亚洲精品国产色婷婷电影| 香蕉国产在线看| 免费观看a级毛片全部| 日韩人妻精品一区2区三区| 波多野结衣av一区二区av| 韩国精品一区二区三区| 国产精品熟女久久久久浪| 久久午夜综合久久蜜桃| 欧美中文综合在线视频| 亚洲国产中文字幕在线视频| 大码成人一级视频| 精品久久久久久电影网| 久久精品亚洲av国产电影网| 国产有黄有色有爽视频| 国产精品.久久久| 国产成人系列免费观看| 变态另类成人亚洲欧美熟女 | 亚洲av日韩在线播放| 久久久久久久大尺度免费视频| 变态另类成人亚洲欧美熟女 | 最新美女视频免费是黄的| 国产欧美日韩一区二区精品| 国产男女超爽视频在线观看| 欧美激情极品国产一区二区三区| 十八禁网站免费在线| 女人高潮潮喷娇喘18禁视频| 天堂俺去俺来也www色官网| 国产成人av激情在线播放| 婷婷丁香在线五月| 国产欧美日韩综合在线一区二区| 国产欧美亚洲国产| 久久国产精品男人的天堂亚洲| 国产精品免费视频内射| 热99国产精品久久久久久7| 在线观看免费午夜福利视频| 国产黄频视频在线观看| av网站免费在线观看视频| 国产亚洲精品久久久久5区| 日韩熟女老妇一区二区性免费视频| 91精品三级在线观看| 国产91精品成人一区二区三区 | 国产成人av教育| 精品国产一区二区久久| 18禁黄网站禁片午夜丰满| 久久香蕉激情| 久久九九热精品免费| 亚洲精品自拍成人| 欧美亚洲 丝袜 人妻 在线| 国产成人精品久久二区二区91| 一级片免费观看大全| 亚洲av国产av综合av卡| 别揉我奶头~嗯~啊~动态视频| 一区二区三区精品91| 国产黄频视频在线观看| 亚洲色图 男人天堂 中文字幕| 99riav亚洲国产免费| 国产黄频视频在线观看| 成人黄色视频免费在线看| 欧美 日韩 精品 国产| av一本久久久久| 色婷婷av一区二区三区视频| 99久久精品国产亚洲精品| 操美女的视频在线观看| 老司机午夜十八禁免费视频| 久久久久久亚洲精品国产蜜桃av| 性少妇av在线| 极品人妻少妇av视频| 麻豆国产av国片精品| 99热国产这里只有精品6| 老司机福利观看| 亚洲国产欧美在线一区| 老司机福利观看| 精品一区二区三区av网在线观看 | 国产国语露脸激情在线看| 国产成人影院久久av| 亚洲五月婷婷丁香| 老司机午夜十八禁免费视频| 欧美日韩成人在线一区二区| 欧美日韩一级在线毛片| 日本wwww免费看| 国产黄频视频在线观看| 免费观看a级毛片全部| 免费不卡黄色视频| 2018国产大陆天天弄谢| 我的亚洲天堂| 精品卡一卡二卡四卡免费| 欧美国产精品一级二级三级| 久久人妻av系列| 亚洲欧美日韩高清在线视频 | 一本大道久久a久久精品| 欧美性长视频在线观看| 成人影院久久| 精品国产乱码久久久久久小说| 精品国产一区二区久久| 亚洲国产av影院在线观看| 在线av久久热| 亚洲综合色网址| 首页视频小说图片口味搜索| 亚洲成a人片在线一区二区| 十分钟在线观看高清视频www| 国产精品秋霞免费鲁丝片| 国产成人一区二区三区免费视频网站| 国产欧美日韩综合在线一区二区| 精品熟女少妇八av免费久了| 国产伦人伦偷精品视频| 国产一区二区三区在线臀色熟女 | e午夜精品久久久久久久| 最新在线观看一区二区三区| 伦理电影免费视频| 欧美日韩亚洲国产一区二区在线观看 | 亚洲天堂av无毛| bbb黄色大片| 高清av免费在线| 两个人看的免费小视频| 国产av国产精品国产| 大片免费播放器 马上看| 大片免费播放器 马上看| 亚洲国产中文字幕在线视频| 亚洲中文日韩欧美视频| 亚洲国产欧美在线一区| 汤姆久久久久久久影院中文字幕| 一区在线观看完整版| 三级毛片av免费| 欧美一级毛片孕妇| 最近最新中文字幕大全电影3 | 啦啦啦在线免费观看视频4| 正在播放国产对白刺激| 男女高潮啪啪啪动态图| 国产人伦9x9x在线观看| 黑人欧美特级aaaaaa片| 亚洲欧美日韩高清在线视频 | 国产精品久久久人人做人人爽| 宅男免费午夜| 亚洲性夜色夜夜综合| 久久久久久久久久久久大奶| 两性夫妻黄色片| 91麻豆av在线| 欧美精品av麻豆av| 免费一级毛片在线播放高清视频 | 黄色a级毛片大全视频| 精品国产国语对白av| 激情在线观看视频在线高清 | 超碰成人久久| 超碰97精品在线观看| 国产xxxxx性猛交| 日韩熟女老妇一区二区性免费视频| 日本vs欧美在线观看视频| 一进一出抽搐动态| 欧美一级毛片孕妇| e午夜精品久久久久久久| 国产精品欧美亚洲77777| 免费av中文字幕在线| 午夜福利视频精品| 日韩免费高清中文字幕av| 麻豆乱淫一区二区| 亚洲精品中文字幕在线视频| 妹子高潮喷水视频| 91成人精品电影| 日韩欧美一区视频在线观看| 91麻豆精品激情在线观看国产 | 高清av免费在线| 黄色丝袜av网址大全| 国产在线精品亚洲第一网站| av福利片在线| 成人免费观看视频高清| 丝袜美腿诱惑在线| a在线观看视频网站| 一区二区av电影网| 亚洲欧洲精品一区二区精品久久久| 最近最新中文字幕大全免费视频| 免费看十八禁软件| 免费久久久久久久精品成人欧美视频| 亚洲成a人片在线一区二区| 欧美中文综合在线视频| 亚洲色图综合在线观看| 丝袜美足系列| 国产老妇伦熟女老妇高清| 麻豆乱淫一区二区| 宅男免费午夜| 国产亚洲精品久久久久5区| 无人区码免费观看不卡 | 久久热在线av| 丝袜美足系列| 黄色视频在线播放观看不卡| 飞空精品影院首页| 午夜福利乱码中文字幕| 久久久久精品人妻al黑| 91精品三级在线观看| 国产精品自产拍在线观看55亚洲 | 国产精品一区二区在线观看99| 纵有疾风起免费观看全集完整版| 午夜老司机福利片| 国产精品欧美亚洲77777| 国产99久久九九免费精品| 不卡av一区二区三区| 亚洲精品中文字幕在线视频| 久久狼人影院| 欧美日本中文国产一区发布| 免费人妻精品一区二区三区视频| 老司机午夜十八禁免费视频| 两人在一起打扑克的视频| 国产激情久久老熟女| 久久久久精品国产欧美久久久| 成人影院久久| 成人精品一区二区免费| 黄色成人免费大全| 日韩欧美国产一区二区入口| 国产欧美日韩一区二区三| 男女之事视频高清在线观看| 亚洲人成电影免费在线| 十分钟在线观看高清视频www| 国产高清videossex| 91精品国产国语对白视频| 99国产精品一区二区三区| 黄网站色视频无遮挡免费观看| 午夜视频精品福利| 日韩精品免费视频一区二区三区| 亚洲av日韩精品久久久久久密| cao死你这个sao货| 国产精品成人在线| 亚洲成a人片在线一区二区| 夫妻午夜视频| 中文字幕精品免费在线观看视频| 热re99久久精品国产66热6| 久久久国产欧美日韩av| 亚洲精品国产精品久久久不卡| 一级,二级,三级黄色视频| 一本色道久久久久久精品综合| 日本av手机在线免费观看| 亚洲专区字幕在线| 一个人免费看片子| 日韩熟女老妇一区二区性免费视频| 大型av网站在线播放| 色婷婷av一区二区三区视频| 成人18禁在线播放| 欧美成人午夜精品| 欧美日韩黄片免| 亚洲成人国产一区在线观看| 69av精品久久久久久 | 欧美+亚洲+日韩+国产| 国产精品久久久人人做人人爽| 国产精品免费一区二区三区在线 | 精品国产亚洲在线| 国产精品香港三级国产av潘金莲| 日韩熟女老妇一区二区性免费视频| 欧美日韩成人在线一区二区| 国产国语露脸激情在线看| 国产片内射在线| 我要看黄色一级片免费的| 自拍欧美九色日韩亚洲蝌蚪91| 日本wwww免费看| 精品一区二区三区av网在线观看 | 亚洲精品在线观看二区| 亚洲男人天堂网一区| 亚洲精品中文字幕一二三四区 | 午夜两性在线视频| 美女午夜性视频免费| 狠狠婷婷综合久久久久久88av| 国产黄色免费在线视频| 亚洲国产欧美在线一区| 国产精品熟女久久久久浪| 搡老岳熟女国产| 国产伦人伦偷精品视频| 日韩大片免费观看网站| 两性午夜刺激爽爽歪歪视频在线观看 | 9色porny在线观看| 老司机亚洲免费影院| 人人妻,人人澡人人爽秒播| 中文字幕人妻丝袜一区二区| 最黄视频免费看| 亚洲精品久久午夜乱码| 亚洲精品一卡2卡三卡4卡5卡| 一个人免费看片子| 一区二区日韩欧美中文字幕| 日韩视频一区二区在线观看| 久久久精品国产亚洲av高清涩受| 人人澡人人妻人| 欧美精品啪啪一区二区三区| 一个人免费看片子| 国产免费福利视频在线观看| 国产麻豆69| 精品第一国产精品| 亚洲成人免费电影在线观看| 考比视频在线观看| 欧美在线一区亚洲| 欧美成人免费av一区二区三区 | 99riav亚洲国产免费| 精品欧美一区二区三区在线| 91九色精品人成在线观看| 纯流量卡能插随身wifi吗| 亚洲男人天堂网一区| 日韩大码丰满熟妇| 777米奇影视久久| 成人永久免费在线观看视频 | 别揉我奶头~嗯~啊~动态视频| 大片免费播放器 马上看| 极品少妇高潮喷水抽搐| 99国产综合亚洲精品| 国产单亲对白刺激| 国产精品国产高清国产av | 欧美av亚洲av综合av国产av| 日韩视频一区二区在线观看| 精品一区二区三区四区五区乱码| 欧美乱码精品一区二区三区| 丝袜喷水一区| 一夜夜www| 伦理电影免费视频| 成人永久免费在线观看视频 | 99久久精品国产亚洲精品| 丝袜美足系列| 自线自在国产av| 久久精品人人爽人人爽视色| 久久这里只有精品19| 亚洲中文日韩欧美视频| 久久久国产欧美日韩av| 亚洲欧美一区二区三区黑人| 久久久久视频综合| 亚洲欧美激情在线| 国产av精品麻豆| 亚洲成a人片在线一区二区| 天天躁日日躁夜夜躁夜夜| 成人国产一区最新在线观看| 精品国产一区二区三区久久久樱花| 最新在线观看一区二区三区| 岛国在线观看网站| 99国产精品免费福利视频| 两性夫妻黄色片| 久久 成人 亚洲| 丁香六月天网| 夜夜爽天天搞| 黄色a级毛片大全视频| 午夜精品国产一区二区电影| 国产免费现黄频在线看| 最新在线观看一区二区三区| 在线 av 中文字幕| 18禁国产床啪视频网站| 午夜精品久久久久久毛片777| 欧美日韩国产mv在线观看视频| 午夜福利视频在线观看免费| 欧美精品一区二区大全| 精品第一国产精品| 国产欧美日韩一区二区三区在线| 国产免费现黄频在线看| 亚洲国产欧美在线一区| 一二三四在线观看免费中文在| 亚洲成av片中文字幕在线观看| 精品国产乱码久久久久久小说| 香蕉久久夜色| 国产不卡一卡二| 一区二区三区精品91| 亚洲av片天天在线观看| 在线观看一区二区三区激情| 亚洲一码二码三码区别大吗| 欧美大码av| 99久久人妻综合| 欧美精品人与动牲交sv欧美| 91精品国产国语对白视频| 波多野结衣av一区二区av| 夜夜爽天天搞| 久久久久久久久免费视频了| 女人高潮潮喷娇喘18禁视频| 久久久久久久精品吃奶| 黄片播放在线免费| 男女午夜视频在线观看| 交换朋友夫妻互换小说| 看免费av毛片| 新久久久久国产一级毛片| 日韩人妻精品一区2区三区| 精品一区二区三区av网在线观看 | 18禁国产床啪视频网站| 久久久精品国产亚洲av高清涩受| 色婷婷久久久亚洲欧美| 他把我摸到了高潮在线观看 | 狂野欧美激情性xxxx| 无限看片的www在线观看| a级片在线免费高清观看视频| 这个男人来自地球电影免费观看| 国产99久久九九免费精品| 天天躁夜夜躁狠狠躁躁| 亚洲欧美一区二区三区黑人| 色视频在线一区二区三区| 999精品在线视频| 视频区欧美日本亚洲| 亚洲第一欧美日韩一区二区三区 | 精品亚洲乱码少妇综合久久| 满18在线观看网站| 亚洲精品国产色婷婷电影| 狠狠精品人妻久久久久久综合| xxxhd国产人妻xxx| 欧美日韩黄片免| 夜夜爽天天搞| 欧美乱码精品一区二区三区| 老汉色av国产亚洲站长工具| 久久精品国产99精品国产亚洲性色 | 黄色视频,在线免费观看| av天堂久久9| 极品少妇高潮喷水抽搐| 久久九九热精品免费| 精品福利观看| 国产精品av久久久久免费| 脱女人内裤的视频| 一区二区日韩欧美中文字幕| 国产极品粉嫩免费观看在线| 国产精品99久久99久久久不卡| 亚洲精品一二三| 免费不卡黄色视频| 亚洲国产欧美日韩在线播放| 桃红色精品国产亚洲av| 国产精品98久久久久久宅男小说| 中文字幕另类日韩欧美亚洲嫩草| 老司机午夜福利在线观看视频 | 国产日韩欧美在线精品| 日本av免费视频播放| av不卡在线播放| 国产一区二区三区综合在线观看| www.熟女人妻精品国产| 波多野结衣av一区二区av| 亚洲人成电影观看| 搡老熟女国产l中国老女人| kizo精华| 亚洲国产av新网站| 老司机在亚洲福利影院| 啦啦啦 在线观看视频| 色综合欧美亚洲国产小说| 精品国产亚洲在线| 亚洲av第一区精品v没综合| 亚洲欧美日韩高清在线视频 | 中文欧美无线码| av不卡在线播放| www.999成人在线观看| 欧美黄色淫秽网站| 建设人人有责人人尽责人人享有的| 亚洲av电影在线进入| 色尼玛亚洲综合影院| 大型黄色视频在线免费观看| 夜夜爽天天搞| 99国产精品99久久久久| 精品视频人人做人人爽| 国产极品粉嫩免费观看在线| 亚洲色图av天堂| 国产在线免费精品| √禁漫天堂资源中文www| 久久精品国产a三级三级三级| 欧美乱妇无乱码| 另类精品久久| 国产精品 国内视频| 成人精品一区二区免费| 中文字幕人妻丝袜一区二区| 亚洲成国产人片在线观看| 久久久久久人人人人人| 亚洲精品自拍成人| 亚洲va日本ⅴa欧美va伊人久久| 国产精品自产拍在线观看55亚洲 | 中文字幕av电影在线播放| 蜜桃国产av成人99| 香蕉国产在线看| 在线观看免费视频日本深夜| 亚洲av欧美aⅴ国产| 亚洲自偷自拍图片 自拍| 成人亚洲精品一区在线观看| 亚洲 欧美一区二区三区| 一本一本久久a久久精品综合妖精| 精品欧美一区二区三区在线| videos熟女内射| 99精品久久久久人妻精品| 亚洲精品成人av观看孕妇| 男女免费视频国产| 一区二区三区国产精品乱码| 精品少妇黑人巨大在线播放| 丰满迷人的少妇在线观看| av视频免费观看在线观看| 日本撒尿小便嘘嘘汇集6| 国产成人欧美| 久久九九热精品免费| 国产精品亚洲一级av第二区| 在线看a的网站| 一本大道久久a久久精品| 亚洲久久久国产精品| 欧美中文综合在线视频| 妹子高潮喷水视频| 国产精品 国内视频| a级毛片黄视频| 国产精品久久久久久人妻精品电影 | av电影中文网址| 女性生殖器流出的白浆| 亚洲情色 制服丝袜| 精品人妻在线不人妻| 亚洲综合色网址| 亚洲美女黄片视频| 精品国产一区二区久久| 另类精品久久| 中亚洲国语对白在线视频| 在线 av 中文字幕| 免费观看av网站的网址| 99九九在线精品视频| 亚洲性夜色夜夜综合| 99国产精品99久久久久| 捣出白浆h1v1| 欧美中文综合在线视频| 亚洲第一av免费看| 大片电影免费在线观看免费| 国产精品麻豆人妻色哟哟久久| 99精品久久久久人妻精品| 在线观看www视频免费| 高清毛片免费观看视频网站 | 国产日韩欧美亚洲二区| 精品国产乱码久久久久久小说| 老司机靠b影院| 在线观看免费午夜福利视频| 亚洲全国av大片| 最黄视频免费看| 999久久久精品免费观看国产| 亚洲精品国产精品久久久不卡| 国产片内射在线| 午夜福利,免费看| 日日爽夜夜爽网站| 久久精品人人爽人人爽视色| 看免费av毛片| 国产精品国产av在线观看| 日韩中文字幕视频在线看片| 女人高潮潮喷娇喘18禁视频| 人人妻人人澡人人看| 少妇粗大呻吟视频| 成人手机av| 一区二区三区激情视频| 国产成人影院久久av| av免费在线观看网站| 国产精品久久久久久人妻精品电影 | 色婷婷久久久亚洲欧美| 中文字幕人妻丝袜制服| 成人永久免费在线观看视频 | 久久99热这里只频精品6学生| 亚洲精品一二三| 色尼玛亚洲综合影院| 午夜福利乱码中文字幕| 高清av免费在线| 男女无遮挡免费网站观看| 老司机福利观看| 精品视频人人做人人爽| 色综合欧美亚洲国产小说| 国产在线观看jvid| 91麻豆av在线| 精品人妻在线不人妻| 中文字幕最新亚洲高清| 免费在线观看完整版高清| 两个人看的免费小视频| 精品人妻熟女毛片av久久网站| 国产99久久九九免费精品| 国产精品九九99| 午夜日韩欧美国产| 肉色欧美久久久久久久蜜桃| 欧美国产精品va在线观看不卡| 国产成人精品久久二区二区91| a级毛片黄视频| 久久精品91无色码中文字幕| 日韩一区二区三区影片| 嫁个100分男人电影在线观看| 国产精品久久久久久精品古装| 欧美人与性动交α欧美精品济南到| 老汉色∧v一级毛片| 日本一区二区免费在线视频| 天天躁夜夜躁狠狠躁躁| 亚洲免费av在线视频| 日韩成人在线观看一区二区三区| 久久久国产一区二区| 国产免费av片在线观看野外av| 国产一区二区 视频在线| 国产欧美日韩精品亚洲av| 中文字幕av电影在线播放| 国产精品自产拍在线观看55亚洲 | 丰满少妇做爰视频| 亚洲av欧美aⅴ国产| 精品国产超薄肉色丝袜足j| 日韩一卡2卡3卡4卡2021年| 丰满人妻熟妇乱又伦精品不卡| 久热这里只有精品99|