Recognition of crude drugs based on SVM

Recognition of Crude Drugs Based on SVM *Zhiyuan Ming , Jin He and Chao Huang Yu Lei College of Electrical and Information Engineering College of Chemical and BiologicalYunnan University of Nationality Yunnan University of Nationality Kunming, Yunnan Province ,China Kunming, Yunnan Province ,Chinahjmwh@ 531907844@*Fund: National Natural Science Foundation (6116361/f 020809)About the author : Ming-Zhiyuan(1985.9-) graduate in Yunnan University of Nationalities . He Jin(1965.9-) Professor in Yunnan University of Nationalities (Corresponding Author).Abstract - Suppor t Vector Machine (SVM) is a machine lear ning theor y based on statistical lear ning algor ithms, SVM based on kernel function has lots of unique advantages on solving the small sample, nonlinea and high dimensional patte n recognition. This article al so uses BP neural networks ǃSupport Vector Machine based on PSO algor ithm and so on to be compa ed to identify p opolis in Yunnan. Compa ed with tr aditional algor ithms, it can solve the small sample, nonlinear and other issues. The experiments show the performance is good when using SVM ke r nel function on solving the he r bs recognition.Index Terms - Support vector machine; kernel function; forecast; medicine identification.I. I NTRODUCTIONYunnan is one of the provinces which are rich in traditional medicine resources in China. Like other places, Chinese faces "status", "standard", "intellectual property" and other problems when exports drugs. If we establish quality standards of drugs, we can evaluate the origin, price, and improve the quality of medicines. In this paper, we use the drug propolis which grows in Yunnan as the main target, the data is got by High Performance Liquid Chromatography, and then use the method of Support Vector Machine to train the sample and assort the data. In this paper, another method called BP neural network is used to be compared in accuracy and time.These samples of propolis were got in two place of Yunnan province, and cold soak ultrasound is used to get the drugs data, purify 0.1g propolis samples, place in 10mL volumetric flask, add 10mL of methanol in it, shake, use ȝP WR inject the filter, shake and then use the High-performance liquid chromatography to treat. The liquid chromatography condition is: Column: Agilent XDB C18 FROXPQ î PP ȝP PRELOH SKDVH D FHWRQLWULOH methanol and 0.001% phosphoric acid solution, flow rate of 0.8mL/min, by a certain percentage of mobile phase gradient HOXWLRQ FROXPQ WHPSHUDW ure : 30 ɗ VFDQ ZDYHOHQJWK 320nm. The liquid processing reached its peak time, peak height, peak area, and so on. T KLV LV WKH ILUVW GDWD LGHQWLI\ LW because the data we got only in 36 minutes, the sample is not enough to train the neural network. So based on the original standards data, I add 1% random noise as the second training sample, and then add 5% random noise as the test sample. Every times the test results are different. After several simulations, the highest rate of extraction of accurate is as the result, the following data extracted in the same way.II. H ELPFUL H INTSA. Neural Network BP neural network (Back Propagation NN) is a multilayer feed forward neural network, whose main feature is the signal transmit from back to pass, the error transmit to forward. During the processing of transmit processing, the input signal passes to the input layer, and then processes in the hidden layer, and finally goes into the output layer. Every state of the neurons is only influenced by the last layer of neurons. If the output layer can not get the desired output, the signal turn back. According to the predicted error we can get the weight and threshold value, so the BP network’s prediction output get closer and closer to the desired output[1]. Figure 2-1 shows the topology of BP neural network.Fig. 1 Magnetization The topology of BP neural network.Neural network is a mapping implement function frominput to output essentially. If the structure of BPNN is excellent, generally, 20 input functions in the following problems can converge to the minimum learning error vicinity in 50,000 times, and in theory, a three-layer neural network can approximate to arbitrary accuracy given function. But the learning speed of BP network is very slow, which is determined by its gradient descent method essentially, and because of its complicated optimization objective function, it will appear some flat areas when the output of neuron get close to 1 or 0. In these areas, the weight to the change of the error is too small, which makes the training process almost stopped[2].The prediction accuracy of neural network is related to the number of training samples. For a multi-inputs and multi-outputs network, if it lack of enough training samples, there may be much errors in the predictive value, so this article enumerates two situations when using the BP neural network3URFHHGLQJV RI WKH WK:RUOG &RQJUHVV RQ ,QWHOOLJHQW &RQWURO DQG $XWRPDWLRQ -XO\ %HLMLQJ &KLQDalgorithm: 1. Take the first sample as the training sample. It is standard on the 5% random noise as the test sample, the training times are 1000 (Attachment: As tested, the training times when use 1000 times or 5000 times in this condition is almost same, in order to reduce time loss, so set 1000 as the training times), test data are as follows: TABLE ICorrespondence between different prediction and transfer function Hiddenlayer Output layer Normalize Accuracy Mean- SquarederrorTime(s) tansig purelin Yes 57.38% 0.2238 37.672No 67.21% 0.2164 32.235tansig tansig Yes 42.62% 0.5744 31.452No 42.62% 0.5743 32.322logsig purelin Yes 59.02% 0.2283 31.840No 55.74% 0.2331 30.227logsig tansig Yes 42.62% 0.5739 31.323no 42.62% 0.573831.513Fig. II The actual output and forecast output of the BP net.2. As a training sample of the two samples, take 122samples with 5% noise data as test sample, hidden layer nodes is 10, hidden layer and output function is tansig and purelin respectively(chose the best result according to test), the data are as follows:TABLE IIDifferent training times results of forecastTest times Accuracy Mean squared errorTine(s)5000 72.13% 0.2787 42.957 10000 78.69% 0.2131 85.24420000 79.32% 0.2013 207.08 40000 80.33% 0.1967 301.8650000 90.16% 0.0984 453.74From the two tests, the accrual network is dependent on sample size and training times, but when the sample size and training times increase, the time LQFUHDVH this is a great defect in neural network algorithm.III. S UPPORT V ECTOR M ACHINE1. OverviewSVM (support vector machine) is based on statistical learning theory, which was proposed structural risk minimization principle and can maximize the generalization ability of learning machine. In additional, support vector machine is a problem to optimize convex quadratic, which can guarantee that the extreme solution---the global optimal solution as introduce kernel function, it solutes inner product in high-dimensional space well. So as to solve the problem of non-linear pattern recognition, kernel is the core in support vector machine, and it impacts the performance of support vector machines directly[3] . Schematic diagram of the kernel function as shown below:Fig. III Schematic diagram of the kernel function It assumed there are two types of samples which can be divided into a collection: (,),1,2,,,,{1,1}d i i i i x y i n x R y ", (1) the linear discriminate function of the general form is ()f x x b Z , so the correct classifications of all samples meet the following equation:[()]10,1,2,i i y x b i n Z t ", (2)So the optimal separating surface is:^`**1()()sgn ()n i i i i f x sng w x b a y x x b ­½®¾¯¿¦(3)2. Kernel functionIn general, the core function that only satisfies the Mercer condition could be used to do inner product in the high-dimensional space. These are some commonly used core functions:Linear kernel: (,)Ti K x x x x (4) Polynomial kernel: (,)[(,)1]qj i K x x x x (5)Radial basis kernel : 22(,)exp i i x x K x x V ­½ °° ®¾°°¯¿(6) Sigmoid kernel : >@(,)tanh ()j i K x x v x x a (7)2. K-CV cross validation functionAs the selection of SVM parameters, there is no generally uniform way that is accepted. The common method is tolet the c and g (2V ) change in a certain range, and then use K-CV method to get the classification accuracy with the c and g. This method sets the samples into K groups first, take K-1 groups as training samples and the next as the test sample. It would receive K models, and take the average of classification accuracy of the K models as the K-CV classifier performance: 1. Take the first sample as the training sample, then add 5% random noise as the test sample, and use the method of K-CV to test. The test results are as follows: (1). c, g range [-5,5]:TABLE IIIDifferent accuracy in different kernel functions(2). c, g range [-15, 15]:Table IVDifferent accuracy in different kernel functionsKernel Function Accuracy Linear kernel 86.8852% Polynomial kernel 86.8852% Radial basis kernel function 80.3279% Sigmoid kernel function83.6066%The method of using K-CV can found the optimal parameters c and g in some sense, and here "in some sense" means the time that parameters c and g under the idea of cross-validation and gets the highest classification accuracy. But it cannot be guaranteed to make the test sample to achieve the highest classification accuracy. From the above two tables, the parameters c and g in K-CV algorithm are found in a grid, certainly and cannot be found some parameters, so its performance is influenced by this, and with the range of c and g increase, the consumed time increases too.3. Particle Swarm Optimization Particle Swarm Optimization (PSO) is an iterative computing algorithm derived from the behavior of the bird’s prey. Imagine such a scenario: a group of birds search the food randomly, and there is only one food in this region all the birds do not know where the food is, but they know how far from the current location to the food, then what is the bestway to find it? The most simple and effective way is to searchfor the area current around the bird where close to the food, which the inspiration get from in PSO to used to solve optimization problems. In PSO algorithm, each optimization problem is to search a bird in the space, which we call it"particle". All the particles are decided by an optimized fitnessfunction, and each particle has a speed which determines the direction and distance when it flies, then particles follow the current optimal particles in the solution space to search. Mathematical representation of PSO algorithm is asfollows:Set the searching space as D-dimensional, the totalnumber of particles is n, the position of the i-th particle express is 12(,,)i i i iD X x x x ", its optimal position where it flight in the history (correspond to the optimal solution) is 12(,,,)i i iD Pi p p p ", the optimal location which the g–th particle past (1,)i P i n " is the best optimal VROXWLRQ The i-th particle’s position changes rate (speed) for the vector 12(, ,, )i i i iD V v v v }, the position of each particle changed by the following formula:1(1)()()[()()]id id id id v t w v t c rand p t x t u u u2()[()()]gd id c rand p t x t u u (8)Among them, 1c , 2c are normal numbers, called accelerated IDFWRU ()randis random number LQ > @ w is inertia factor. Process the data with this algorithm, test results are as follows:Table VDifferent accuracy as different accuracy with PSOLinear kernel Accuracy Linear kernel 98.3607% Polynomial kernel 100% Radial basis kernel function 100% Sigmoid kernel function100%According to the analysis from the chart, the support vector machine in the PSO algorithm is more advantageous, and need less time to find c and g than the K-CV method.4. ConclusionSupport Vector Machine is a new theory based on statistical classification techniques, which is mainly for the small sample FDVH WKH RSWLPDO VROXWLRQ is based on a limited sample of information rather than in a large of sample infinity, which has strict theoretical and mathematical foundations. In this paper, only a small sample of cases by BP neural network, based on comparison among support vectormachine, K-CV algorithm and PSO algorithm, experimental results show that using the PSO algorithm to find parameters by the support vector machine, the accuracy rate on the search of parameters classification is greatly improved, and in the consumption of time, there are big advantages. R EFERENCES[1] Shi Neural network implementation [D]. Hunan College, 2010.31.5.[2] Han Liqun Artificial neural network theory, design and application[M].Beijing: Chemical Industry Press,2002.[3] Bai Peng, Zhang Xibin, Li Yan and so on. Support vector machine theoryand engineering application examples [M]. Xi'an: Xidian University Press,2008. [4] DONG Hui, FU He-lin, LENG Wu-ming. Support Vector Machines forTime Series Regression and Prediction. Journal of System Simulation[J], 2006, 18(7): 1785-1788.[5] John Shawe-Taylor, Nello Cristianini. Kernel Methods for PatternAnalysis[M]. Cambridge: Cambridge University Press, 2004. [6] Joachims T. Text Categorization with Support Vector Machines: Learning with Many Relevant Features[C]. Proc. of the 10thEuropean Conf Machine Learning, 1999.1372-142.[7] Kim, H.C. ,Pang, S., Je, H.M,Kim, D. and Bang, S. Y. , “Construction support vector machine ensemble”, Pattern Recognition, vol.36, pp. 27572767, 2003.[8] V. Vapnik. Statistical learning theory. John Wiley and Sons, New York, 1998.Kernel Function Accuracy Linear kernel 81.9672% Polynomial kernel 83.6066% Radial basis kernel function 81.9672% Sigmoid kernel function85.2459%。