CN106096649A - Sense of taste induced signal otherness feature extracting method based on core linear discriminant analysis - Google Patents

Abstract

The present invention provides a kind of sense of taste induced signal otherness feature extracting method based on core linear discriminant analysis, and method includes: utilize electronic tongues to detect Tea Samples, obtains sensor response clock signal；Principal component residual sum mahalanobis distance method is used to be analyzed exceptional sample and reject according to described response clock signal；The parameter of core linear discriminant analysis method is optimized, with Longjing tea quality grade correct recognition rata for according to the parameter selecting core linear discriminant analysis method；Use core linear discriminant analysis method to carry out Nonlinear feature extraction to sensor response signal, obtain the flavor characteristics of Tea Samples；By the flavor characteristics input grader of Tea Samples, carry out tea leaf quality grade judgement.Abnormality value removing is carried out to Tea Samples, utilizes the core linear discriminant analysis method after Optimal Parameters can preferably characterize the nonlinear characteristic of different brackets Tea Samples, promote the signal difference opposite sex in high-dimensional feature space for the sample after Nonlinear Mapping.

Description

Kernel linearity discriminant analysis-based taste sense signal difference feature extraction method

Technical Field

The invention relates to the technical field of tea detection, in particular to a taste sense induction signal difference characteristic extraction method based on nuclear linear discriminant analysis.

Background

In recent years, tea quality testing has been a difficult task because tea contains many ingredients and their effects on tea quality are very different. West lake Longjing tea is a typical representative of Chinese green tea. Some vendors fry other green tea into flat shape to serve as the dragon well tea or use the dragon well of other production places of Zhejiang to serve as the west lake dragon well, which disturbs the market of the dragon well tea and damages the benefits of consumers, therefore, the method has important significance for scientific detection and evaluation of the quality of the west lake dragon well tea.

Sensory evaluation is an important method for evaluating the quality of tea leaves for a long time, but the method needs rich tea science knowledge and evaluation experience, and the sensory organ sensitivity of professional tea tasters is easily changed by the interference of external factors. Many analytical tools are therefore used to analyse tea leaf chemicals, such as high performance liquid chromatography, gas chromatography and the like. But the traditional linear feature extraction method cannot effectively explore the intrinsic regularity existing in the nonlinear data.

Disclosure of Invention

The invention aims to provide a taste sense induction signal difference characteristic extraction method based on nuclear linear discriminant analysis, which can effectively extract difference characteristics of tea.

In order to solve the above technical problems, an embodiment of the present invention provides a taste sensation signal difference feature extraction method based on kernel linear discriminant analysis, including:

detecting a tea sample by using an electronic tongue to obtain a sensor response time sequence signal;

analyzing and eliminating abnormal samples by adopting a principal component residual error and Mahalanobis distance method according to the response time sequence signal;

optimizing parameters of a kernel linear discriminant analysis method, and selecting the parameters of the kernel linear discriminant analysis method according to the quality grade correct recognition rate of the Longjing tea;

performing nonlinear feature extraction on the sensor response signals by adopting a kernel linear discriminant analysis method to obtain the flavor features of the tea samples;

inputting the taste characteristics of the tea sample into a classifier, and judging the quality grade of the tea.

Preferably, the sensor response timing signal comprises: at least one of a ZA sensor response time sequence signal, a BB sensor response time sequence signal, a JE sensor response time sequence signal, a GA sensor response time sequence signal, an HA sensor response time sequence signal, a JB sensor response time sequence signal, a CA sensor response time sequence signal and an Ag/AgCl reference electrode sensor response time sequence signal.

Preferably, the detecting the tea sample by using the electronic tongue comprises the following steps:

placing the sample and the cleaning solution on an autosampler of the electronic tongue in sequence;

the collection of each sample was repeated, and each collection was performed according to the procedure of "tea soup sample → washing solution 1 → washing solution 2".

Preferably, the analyzing and rejecting abnormal samples by using principal component residual error and mahalanobis distance method according to the response time sequence signal includes:

for data set X ═ X₁,x₂,…,x_N]∈R^m×NThe centralization is carried out, and the device is,；

calculating a covariance matrix of the centralized data:

calculating eigenvalues and eigenvectors of the covariance matrix: cv ═ λ v;

the eigenvalue lambda of the covariance matrix_iSorting according to the sequence from large to small, and sorting the eigenvectors corresponding to the eigenvalues according to the sequence from large to small;

by usingProjecting the data sample onto a feature vector obtained in Cv ═ λ v;

by usingCalculating the estimated value of the sample, wherein the principal component residual is the difference between the true value and the estimated value of the sample, i.e.

Wherein,v is a characteristic vector corresponding to the characteristic value;

the mahalanobis distance between sample points is: d_ij＝[(x_i-x_j)^T[Cov(X)]^-1(x_i-x_j)]^1/2；

And judging the sample points which are away from the same type sample points and are distributed in the whole manner as abnormal sample elimination according to the principal component residual value and the Mahalanobis distance between the sample points and the same type sample mean value.

Preferably, the optimizing the parameters of the kernel linear discriminant analysis method, and selecting the parameters of the kernel linear discriminant analysis method based on the correct recognition rate of the quality grade of the tea leaves, includes:

taking a Gaussian kernel function as a nonlinear conversion function of a kernel linear discriminant analysis method, and calculating the linear discriminant analysis of the Gaussian kernel function k (x, y) as exp (- | | x-y | |)²/2σ²) Parameter σ in²Carrying out optimization selection;

and selecting parameter values according to the correct recognition rate determined by the quality grade of the tea during parameter selection.

Preferably, the gaussian kernel function is:

$k(x,y)=\exp (-\frac{\left|\right|x-y|{|}^2}{2{\mathrm{\σ}}^2})$

preferably, the performing nonlinear feature extraction on the sensor response signal by using a kernel linear discriminant analysis method to obtain the flavor features of the tea sample includes:

by a non-linear transformationMapping the input data to high-dimensional feature space, and obtaining a data point phi (x) after nonlinear transformation₁),Φ(x₂),…,Φ(x_N)；

In a high-dimensional feature space, converting the problem of maximization of the Fisher criterion function into a problem of solving a feature value and a feature vector of a feature equation;

and carrying out nonlinear characteristic extraction on the sensor response signals to obtain the taste characteristics of the tea sample.

Preferably, said transformation is by a non-linear transformationMapping the input data to high-dimensional feature space, and obtaining a data point phi (x) after nonlinear transformation₁),Φ(x₂),…,Φ(x_N) The method comprises the following steps:

inter-class dispersion matrix of training samples in high-dimensional feature spaceAnd intra-class dispersion matrixComprises the following steps:

$S_b^{\mathrm{\Φ}}=\underset{i=1}{\overset{L}{\Σ}}{P}_i(m_{\mathrm{\Φ},i}-m_{\mathrm{\Φ}}){(m_{\mathrm{\Φ},i}-m_{\mathrm{\Φ}})}^T$

$S_w^{\mathrm{\Φ}}=\underset{i=1}{\overset{L}{\Σ}}\underset{x_k\∈c_i}{\Σ}\left(\mathrm{\Φ}\right(x_k)-m_{\mathrm{\Φ},i}){\left(\mathrm{\Φ}\right(x_k)-m_{\mathrm{\Φ},i})}^T$

wherein m is_ΦAnd m_Φ,iRespectively representing the mean values of all training samples in the high-dimensional feature space and the mean value of the ith class of training samples;

the Fisher criterion function in the high-dimensional feature space is:

$J_f\left(W\right)=\left|\frac{W^T{S}_b^{\mathrm{\Φ}}W}{W^T{S}_w^{\mathrm{\Φ}}W}\right|$

in the high-dimensional feature space, the problem of maximizing the Fisher criterion function is converted into a problem of solving the feature value and the feature vector of the feature equation, and the method comprises the following steps:

define the kernel matrix K ═ K of N × N_ij]Then the above formula becomes

KBKα＝λKWKα

Wherein, K_ij＝k(x_i,x_j)＝Φ(x_i)^TΦ(x_j)，B＝GCG^T，

$C=diag(n_1,n_2,...,n_L)\∈R^{L\×L},G=diag(\frac{1}{n_1}{1}_{n_1\×1},...,\frac{1}{n_L}{1}_{n_L\×1}),$

$W=diag(I_{n_1}-\frac{1}{n_1}{1}_{n_1\×n_1},...,I_{n_L}-\frac{1}{n_L}{1}_{n_L\×n_L})\∈R^{N\×N}$

Preferably, the performing nonlinear feature extraction on the sensor response signal to obtain the taste features of the tea sample comprises:

calculating a kernel matrix K-K of the training sample set according to the determined kernel function and the optimized kernel function parameters_ij]In which K is_ij＝k(x_i,x_j)＝Φ(x_i)^TΦ(x_j)；

The Fisher criterion function maximization is converted into a problem of solving generalized eigenvalues, and the eigenvalues of KBK α -lambda KWK α and corresponding eigenvectors α - α are solved₁,α₂,…,α_N]^TSorting according to the sequence of the characteristic values from large to small;

will train sample phi (x)_i) The most sampled nonlinear feature projected onto the kth feature vector:

$\mathrm{\Φ}{\left(x_i\right)}^T{v}^k=\mathrm{\Φ}{\left(x_i\right)}^T\underset{j=1}{\overset{N}{\Σ}}{\mathrm{\α}}_j^k\mathrm{\Φ}\left(x_j\right)=\underset{j=1}{\overset{N}{\Σ}}{\mathrm{\α}}_j^k\mathrm{\Φ}{\left(x_i\right)}^T\mathrm{\Φ}\left(x_j\right)=\underset{j=1}{\overset{N}{\Σ}}{\mathrm{\α}}_j^k{K}_{ij}$

calculating a kernel matrix K-K of the training sample set according to the determined kernel function and the optimized kernel function parameters_ij]In which K is_ij＝k(x_i,x_j)＝Φ(x_i)^TΦ(x_j)；

The eigenvalues of KBK α λ KWK α and the corresponding eigenvectors α [ α ]₁,α₂,…,α_N]^TSorting according to the sequence of the characteristic values from large to small;

will train sample phi (x)_i) The most sampled nonlinear feature projected onto the kth feature vector:

$\mathrm{\Φ}{\left(x_i\right)}^T{v}^k=\mathrm{\Φ}{\left(x_i\right)}^T\underset{j=1}{\overset{N}{\Σ}}{\mathrm{\α}}_j^k\mathrm{\Φ}\left(x_j\right)=\underset{j=1}{\overset{N}{\Σ}}{\mathrm{\α}}_j^k\mathrm{\Φ}{\left(x_i\right)}^T\mathrm{\Φ}\left(x_j\right)=\underset{j=1}{\overset{N}{\Σ}}{\mathrm{\α}}_j^k{K}_{ij}$

calculating test samplesAnd a kernel matrix K' between the training set samples, projecting the test samples onto the feature vectors

Preferably, the inputting of the taste characteristics of the tea sample into the classifier to perform the quality grade determination of tea comprises:

for the sample to be testedAnd training image sample x_iCalculating the similarity between the image sample to be tested and the training image sample

$d({\tilde{x}}_i,x_i)=\sqrt{\underset{k=1}{\overset{d}{\Σ}}{(x_{ik}^{\′}-x_{ik})}^2}$

If it isSample x_iIf it belongs to class k, the test sample is testedIs decided as class k.

The technical scheme of the invention has the following beneficial effects:

according to the scheme, abnormal values of tea samples can be eliminated, nonlinear characteristics of the tea samples in different grades can be better represented by using the kernel linear discriminant analysis method after parameters are optimized, and the signal difference of the samples subjected to nonlinear mapping in a high-dimensional characteristic space is improved.

Drawings

FIG. 1 is a flowchart of a taste sensation signal difference feature extraction method based on kernel linear discriminant analysis according to an embodiment of the present invention;

FIG. 2 is an electronic tongue response map of a tea sample according to an embodiment of the invention;

FIGS. 3a-3d are graphs of the principal component residue-Mahalanobis distance distribution for an embodiment of the present invention;

FIG. 4 shows the correct recognition rate and parameter σ of tea samples according to the embodiment of the present invention²Selecting a relation graph;

FIGS. 5a-5d are comparative results of the KLDA and linear dimensionality reduction method of the present invention with respect to tea sample differentiation;

fig. 6a and 6b are the results of comparing the correct recognition rate curves of KLDA and linear dimensionality reduction methods of embodiments of the present invention as a function of dimensionality reduction.

Detailed Description

In order to make the technical problems, technical solutions and advantages of the present invention more apparent, the following detailed description is given with reference to the accompanying drawings and specific embodiments.

As shown in fig. 1, the method for extracting difference features of taste sensation signal based on kernel linear discriminant analysis according to the embodiment of the present invention includes:

step 101: detecting a tea sample by using an electronic tongue to obtain a sensor response time sequence signal;

step 102: analyzing and eliminating abnormal samples by adopting a principal component residual error and Mahalanobis distance method according to the response time sequence signal;

step 103: optimizing parameters of a kernel linear discriminant analysis method, and selecting the parameters of the kernel linear discriminant analysis method according to the quality grade correct recognition rate of the Longjing tea;

step 104: performing nonlinear feature extraction on the sensor response signals by adopting a kernel linear discriminant analysis method to obtain the flavor features of the tea samples;

step 105: inputting the taste characteristics of the tea sample into a classifier, and judging the quality grade of the tea.

The gustation sensing signal difference characteristic extraction method based on the kernel linear discriminant analysis can remove abnormal values of tea samples, the kernel linear discriminant analysis method after parameters are optimized can better represent nonlinear characteristics of the tea samples in different grades, and the signal difference of samples subjected to nonlinear mapping in a high-dimensional characteristic space is improved.

Preferably, the sensor response timing signal comprises: at least one of a ZA sensor response time sequence signal, a BB sensor response time sequence signal, a JE sensor response time sequence signal, a GA sensor response time sequence signal, an HA sensor response time sequence signal, a JB sensor response time sequence signal, a CA sensor response time sequence signal and an Ag/AgCl reference electrode sensor response time sequence signal.

Preferably, the detecting the tea sample by using the electronic tongue comprises the following steps:

placing the sample and the cleaning solution on an autosampler of the electronic tongue in sequence;

the collection of each sample was repeated, and each collection was performed according to the procedure of "tea soup sample → washing solution 1 → washing solution 2".

The invention can adopt ASTREE electronic tongue system of French Alpha MOS company to detect the Longjing tea sample, and is specially designed for gustatory analysis technology.

Before data acquisition, the electronic tongue system can be subjected to steps of self-checking, activation, training, calibration, diagnosis and the like, so that the acquired data is ensured to have reliability and stability. After the collection, seven electronic tongue response fingerprints are obtained from each tea sample, as shown in figure 2. The horizontal axis represents measurement time, and the vertical axis represents the acquired induced voltage value. The point on the curve represents the change of the potential difference with time when the tea soup flavor substance passes through the sensor channel. In the measurement process, each detection time is 120s, a set of data is acquired every 0.5s by the electronic tongue, and each sample is detected to finally obtain 7 time sequence signals which change along with time, such as response signals of the electronic tongue sensor shown in the attached figure 2. Thus, for each sample test, the data obtained is a matrix of 7 × 240 dimensions. The stable value of the sensor response at 120s can be selected as a characteristic point for the subsequent establishment of the tea quality model.

Preferably, the analyzing and rejecting abnormal samples by using principal component residual error and mahalanobis distance method according to the response time sequence signal includes:

for data set X ═ X₁,x₂,…,x_N]∈R^m×NThe centralization is carried out, and the device is,

calculating a covariance matrix of the centralized data:

calculating eigenvalues and eigenvectors of the covariance matrix: cv ═ λ v;

the eigenvalue lambda of the covariance matrix_iSorting according to the sequence from big to small, the characteristic value corresponds toThe eigenvectors are sorted in the order from big to small;

by usingProjecting the data sample onto a feature vector obtained in Cv ═ λ v;

by usingCalculating the estimated value of the sample, wherein the principal component residual is the difference between the true value and the estimated value of the sample, i.e.

Wherein,v is a characteristic vector corresponding to the characteristic value;

the mahalanobis distance between sample points is: d_ij＝[(x_i-x_j)^T[Cov(X)]^-1(x_i-x_j)]^1/2；

And judging the sample points which are far away from the same type sample point and are distributed in the whole manner as abnormal sample elimination according to the principal component residual value and the Mahalanobis distance between the sample points and the same type sample mean value.

Wherein, the principal component analysis and mahalanobis distance value calculation are performed on the fine sample set, the special sample set, the first sample set and the second sample set, respectively, as shown in fig. 3a-3 d. The points marked in the graph are abnormal sample points. And performing subsequent data processing on the tea sample data from which the abnormal sample points are removed.

Preferably, the optimizing the parameters of the kernel linear discriminant analysis method, and selecting the parameters of the kernel linear discriminant analysis method based on the positive recognition rate of the quality grade of the tea leaves, includes:

linear discrimination using gaussian kernel function as kernelThe nonlinear conversion function of the analysis method is used for solving the problem that the Gaussian kernel function k (x, y) is exp (- | | x-y | | | survival rate²/2σ²) Parameter σ in²Carrying out optimization selection;

and selecting parameter values according to the correct recognition rate determined by the quality grade of the tea during parameter selection.

Preferably, the gaussian kernel function is:

$k(x,y)=\exp (-\frac{\left|\right|x-y|{|}^2}{2{\mathrm{\σ}}^2})$

the parameter selection is an important factor influencing the algorithm discrimination effect, the selection of proper parameters can enhance the effectiveness of the algorithm, while the selection of improper parameters can greatly weaken the function of the algorithm and even make the algorithm effective. For the kernel linear discriminant analysis algorithm (KLDA), the construction of the kernel function is the core of the algorithm. The high-dimensional map has no explicit form and needs to be computed by means of a kernel function. All (Φ (x) · Φ (y)) are replaced by a kernel function k (x, y). The choice of kernel functions determines the transformation function Φ and the feature space F.

The invention can adopt the most widely applied Gaussian radial basis kernel function, and the kernel function needs to be corresponding to the parameter sigma²And (4) carrying out optimization selection, and determining the optimal value of the parameter through a group of experiments. Respectively take sigma²0.5,5,50,500,5000,50000, the parameters take six specific values, and the characteristic dimension is increased continuously in a certain step lengthIs large. So as to obtain the change of the correct recognition rate in the process of continuously increasing the feature dimension. The results of the correct recognition rate for different parameters as the features increase tertiary are shown in fig. 3.

As can be seen from FIG. 4, when σ is²When the number of samples is 0.5,5, and 500, the correct recognition rate of the sample is low. Sigma²When the number is 50, the correct recognition rate of the sample is the highest. Sigma²The correct recognition rate for the sample is similar when 5000,50000. Therefore, the invention selects 50 as the parameter σ in the KLDA algorithm²The value of (a).

Preferably, the performing nonlinear feature extraction on the sensor response signal by using a kernel linear discriminant analysis method to obtain the flavor features of the tea sample includes:

by a non-linear transformationMapping the input data to high-dimensional feature space, and obtaining a data point phi (x) after nonlinear transformation₁),Φ(x₂),…,Φ(x_N)；

In a high-dimensional feature space, converting the problem of maximization of the Fisher criterion function into a problem of solving a feature value and a feature vector of a feature equation;

and carrying out nonlinear characteristic extraction on the sensor response signals to obtain the taste characteristics of the tea sample.

Preferably, said transformation is by a non-linear transformationMapping the input data to high-dimensional feature space, and obtaining a data point phi (x) after nonlinear transformation₁),Φ(x₂),…,Φ(x_N) The method comprises the following steps:

inter-class dispersion matrix of training samples in high-dimensional feature spaceAnd intra-class dispersion matrixComprises the following steps:

$S_b^{\mathrm{\Φ}}=\underset{i=1}{\overset{L}{\Σ}}{P}_i(m_{\mathrm{\Φ},i}-m_{\mathrm{\Φ}}){(m_{\mathrm{\Φ},i}-m_{\mathrm{\Φ}})}^T$

$S_w^{\mathrm{\Φ}}=\underset{i=1}{\overset{L}{\Σ}}\underset{x_k\∈c_i}{\Σ}\left(\mathrm{\Φ}\right(x_k)-m_{\mathrm{\Φ},i}){\left(\mathrm{\Φ}\right(x_k)-m_{\mathrm{\Φ},i})}^T$

wherein m is_ΦAnd m_Φ,iRespectively representing the mean values of all training samples in the high-dimensional feature space and the mean value of the ith class of training samples;

the Fisher criterion function in the high-dimensional feature space is:

$J_f\left(W\right)=\left|\frac{W^T{S}_b^{\mathrm{\Φ}}W}{W^T{S}_w^{\mathrm{\Φ}}W}\right|$

in the high-dimensional feature space, the problem of maximizing the Fisher criterion function is converted into a problem of solving the feature value and the feature vector of the feature equation, and the method comprises the following steps:

define the kernel matrix K ═ K of N × N_ij]Then the above formula becomes

KBKα＝λKWKα

Wherein, K_ij＝k(x_i,x_j)＝Φ(x_i)^TΦ(x_j)，B＝GCG^T，

$C=diag(n_1,n_2,...,n_L)\∈R^{L\×L},G=diag(\frac{1}{n_1}{1}_{n_1\×1},...,\frac{1}{n_L}{1}_{n_L\×1}),$

$W=diag(I_{n_1}-\frac{1}{n_1}{1}_{n_1\×n_1},...,I_{n_L}-\frac{1}{n_L}{1}_{n_L\×n_L})\∈R^{N\×N}$

Preferably, the performing nonlinear feature extraction on the sensor response signal to obtain the taste features of the tea sample comprises:

calculating a kernel matrix K-K of the training sample set according to the determined kernel function and the optimized kernel function parameters_ij]In which K is_ij＝k(x_i,x_j)＝Φ(x_i)^TΦ(x_j)；

Optimizing the Fisher criterion functionThe maximization is converted into the problem of solving the generalized eigenvalue, and the eigenvalue of KBK α ═ λ KWK α and the corresponding eigenvector α ═ α₁,α₂,…,α_N]^TSorting according to the sequence of the characteristic values from large to small;

will train sample phi (x)_i) The most sampled nonlinear feature projected onto the kth feature vector:

$\mathrm{\Φ}{\left(x_i\right)}^T{v}^k=\mathrm{\Φ}{\left(x_i\right)}^T\underset{j=1}{\overset{N}{\Σ}}{\mathrm{\α}}_j^k\mathrm{\Φ}\left(x_j\right)=\underset{j=1}{\overset{N}{\Σ}}{\mathrm{\α}}_j^k\mathrm{\Φ}{\left(x_i\right)}^T\mathrm{\Φ}\left(x_j\right)=\underset{j=1}{\overset{N}{\Σ}}{\mathrm{\α}}_j^k{K}_{ij}$

calculating a kernel matrix K-K of the training sample set according to the determined kernel function and the optimized kernel function parameters_ij]In which K is_ij＝k(x_i,x_j)＝Φ(x_i)^TΦ(x_j)；

The eigenvalues of KBK α λ KWK α and the corresponding eigenvectors α [ α ]₁,α₂,…,α_N]^TSorting according to the sequence of the characteristic values from large to small;

will train sample phi (x)_i) The most sampled nonlinear feature projected onto the kth feature vector:

$\mathrm{\Φ}{\left(x_i\right)}^T{v}^k=\mathrm{\Φ}{\left(x_i\right)}^T\underset{j=1}{\overset{N}{\Σ}}{\mathrm{\α}}_j^k\mathrm{\Φ}\left(x_j\right)=\underset{j=1}{\overset{N}{\Σ}}{\mathrm{\α}}_j^k\mathrm{\Φ}{\left(x_i\right)}^T\mathrm{\Φ}\left(x_j\right)=\underset{j=1}{\overset{N}{\Σ}}{\mathrm{\α}}_j^k{K}_{ij}$

calculating test samplesAnd a kernel matrix K' between the training set samples, projecting the test samples onto the feature vectors

Preferably, the inputting of the taste characteristics of the tea sample into the classifier to perform the quality grade determination of tea comprises:

for the sample to be testedAnd training image sample x_iCalculating the similarity between the image sample to be tested and the training image sample

$d({\tilde{x}}_i,x_i)=\sqrt{\underset{k=1}{\overset{d}{\Σ}}{(x_{ik}^{\′}-x_{ik})}^2}$

If it isSample x_iIf it belongs to class k, the test sample is testedIs decided as class k.

According to the taste sensation induction signal difference characteristic extraction method based on the kernel linear discriminant analysis, the discrimination capability of the kernel linear discriminant analysis method and the discrimination capability of the traditional linear dimensionality reduction method on tea samples are compared, KLDA, PCA, LDA and LPP are respectively adopted to reduce dimensionality of data of an electronic tongue intelligent sensory instrument, and the characteristic dimensionality is selected to be 2. The distribution diagram of the tea sample points after the dimensionality reduction is shown in figure 5.

As can be seen from fig. 5a-5d, for the linear dimensionality reduction method, whether unsupervised (PCA) or supervised (LDA, LPP) algorithms, the different grades of tea samples are severely aliased in the two-dimensional dimensionality reduction space. Although the LDA and LPP algorithms are supervised methods, the reduced-dimension samples are still severely aliased in the two-dimensional space. Experimental results show that the KLDA algorithm achieves the best sample separation, sample points of the same class are grouped together, while sample points of different classes are correctly separated. The algorithm utilizes the category information of tea samples to optimize a discriminant function, and the kernel-based feature extraction algorithm can mine the nonlinear features of the tea sample data, so that samples which cannot be correctly classified in an original data space are correctly classified in a high-dimensional space after being subjected to high-dimensional mapping.

The invention also analyzes and compares the quality grade classification results of different grades of Longjing tea by using the KLDA algorithm. After the abnormal value is eliminated, the number of the samples for tea grade discrimination is 212. In order to verify the adaptability and generalization of the kernel principal component analysis method, the judgment range of the algorithm is widened. In the experiment, 20 (case 1) samples and 30 (case 2) samples are randomly selected from tea samples of each grade for training, and the rest samples are tested. FIGS. 6a-6b show the comparison of the KLDA algorithm and the PCA, LDA, LPP algorithms on the correct classification recognition rates (as a function of the dimensionality reduction) for different grades of tea samples under two experimental conditions. It can be seen that the correct recognition rate of the PCA algorithm is overall lower than LDA and LPP. Both LDA and LPP algorithms adopt a supervision type calculation method, and as can be seen from the figure, when the feature dimension is lower, the correct recognition rate of LPP is better than that of LDA, and the difference between the two gradually decreases with the increase of the feature dimension. In general, the accurate recognition rate of the KLDA algorithm is higher than that of other linear dimension reduction methods. As can be seen from the variation trend of the graph, the highest correct recognition rate of the algorithm usually does not occur in the situation that the feature dimension is the largest

In this case, the recognition rate curve tends to increase in the early stage and generally tends to decrease or level in the latter half. Therefore, it can be understood that a small number of features can effectively represent original sample information, and the mapping of high-dimensional samples in a low-dimensional space can extract sample effective information and eliminate noise.

While the foregoing is directed to the preferred embodiment of the present invention, it will be understood by those skilled in the art that various changes and modifications may be made without departing from the spirit and scope of the invention as defined in the appended claims.

Claims (10)

1. A taste sense signal difference feature extraction method based on nuclear linear discriminant analysis is characterized by comprising the following steps of:

detecting a tea sample by using an electronic tongue to obtain a sensor response time sequence signal;

analyzing and eliminating abnormal samples by adopting a principal component residual error and Mahalanobis distance method according to the response time sequence signal;

optimizing parameters of a kernel linear discriminant analysis method, and selecting the parameters of the kernel linear discriminant analysis method according to the quality grade correct recognition rate of the Longjing tea;

performing nonlinear feature extraction on the sensor response signals by adopting a kernel linear discriminant analysis method to obtain the flavor features of the tea samples;

inputting the taste characteristics of the tea sample into a classifier, and judging the quality grade of the tea.

2. The method of claim 1, wherein the sensor response timing signal comprises: at least one of a ZA sensor response time sequence signal, a BB sensor response time sequence signal, a JE sensor response time sequence signal, a GA sensor response time sequence signal, an HA sensor response time sequence signal, a JB sensor response time sequence signal, a CA sensor response time sequence signal and an Ag/AgCl reference electrode sensor response time sequence signal.

3. The method for extracting difference characteristics of taste sensation signals based on kernel linear discriminant analysis according to any one of claims 1 or 2, wherein the detecting of the tea sample by using the electronic tongue comprises:

placing the sample and the cleaning solution on an autosampler of the electronic tongue in sequence;

the collection of each sample was repeated, and each collection was performed according to the procedure of "tea soup sample → washing solution 1 → washing solution 2".

4. The method for extracting difference features of taste sensation signals based on kernel linear discriminant analysis according to claim 1, wherein the analyzing and removing abnormal samples according to the response time series signals by using principal component residuals and mahalanobis distance method comprises:

for data set X ═ X₁,x₂,…,x_N]∈R^m×NThe centralization is carried out, and the device is,

calculating a covariance matrix of the centralized data:

calculating eigenvalues and eigenvectors of the covariance matrix: cv ═ λ v;

the eigenvalue lambda of the covariance matrix_iSorting according to the sequence from large to small, and sorting the eigenvectors corresponding to the eigenvalues according to the sequence from large to small;

by usingProjecting the data sample onto a feature vector obtained in Cv ═ λ v;

by usingCalculating the estimated value of the sample, wherein the principal component residual is the difference between the true value and the estimated value of the sample, i.e.

Wherein,v is a characteristic vector corresponding to the characteristic value;

the mahalanobis distance between sample points is: d_ij＝[(x_i-x_j)^T[Cov(X)]^-1(x_i-x_j)]^1/2；

And judging the sample points which are far away from the same type sample point and are distributed in the whole manner as abnormal sample elimination according to the principal component residual value and the Mahalanobis distance between the sample points and the same type sample mean value.

5. The method of any one of claims 1 or 4, wherein the optimizing the parameters of the kernel linear discriminant analysis method to select the parameters of the kernel linear discriminant analysis method based on the correct recognition rate of the quality grade of the tea comprises:

taking a Gaussian kernel function as a nonlinear conversion function of a kernel linear discriminant analysis method, and calculating the linear discriminant analysis of the Gaussian kernel function k (x, y) as exp (- | | x-y | |)²/2σ²) Parameter σ in²Carrying out optimization selection;

and selecting parameter values according to the correct recognition rate determined by the quality grade of the tea during parameter selection.

6. The method of claim 5, wherein the Gaussian kernel function is:

$k(x,y)=\exp (-\frac{\left|\right|x-y|{|}^2}{2{\mathrm{\σ}}^2})$

7. the method for extracting difference characteristics of taste sensation signals based on kernel linear discriminant analysis according to claim 5, wherein the non-linear feature extraction of sensor response signals by using kernel linear discriminant analysis method to obtain flavor characteristics of tea samples comprises:

by a non-linear transformation Φ:mapping the input data to high-dimensional feature space, and obtaining a data point phi (x) after nonlinear transformation₁),Φ(x₂),…,Φ(x_N)；

In a high-dimensional feature space, converting the problem of maximization of the Fisher criterion function into a problem of solving a feature value and a feature vector of a feature equation;

and carrying out nonlinear characteristic extraction on the sensor response signals to obtain the taste characteristics of the tea sample.

8. The method of claim 7, wherein said difference features of taste sensation signals are extracted by a nonlinear transformation Φ:mapping the input data to high-dimensional feature space, and obtaining a data point phi (x) after nonlinear transformation₁),Φ(x₂),…,Φ(x_N) The method comprises the following steps:

inter-class dispersion matrix of training samples in high-dimensional feature spaceAnd intra-class dispersion matrixComprises the following steps:

$S_b^{\mathrm{\Φ}}=\underset{i=1}{\overset{L}{\Σ}}{P}_i(m_{\mathrm{\Φ},i}-m_{\mathrm{\Φ}}){(m_{\mathrm{\Φ},i}-m_{\mathrm{\Φ}})}^T$

$S_w^{\mathrm{\Φ}}=\underset{i=1}{\overset{L}{\Σ}}\underset{x_k\∈c_i}{\Σ}(\mathrm{\Φ}\left(x_k\right)-m_{\mathrm{\Φ},i}){\left(\mathrm{\Φ}\right(x_k)-m_{\mathrm{\Φ},i})}^T$

wherein m is_ΦAnd m_Φ,iRepresenting the mean and class i training samples of all training samples in the high-dimensional feature space, respectivelyMean value;

the Fisher criterion function in the high-dimensional feature space is:

$J_f\left(W\right)=\left|\frac{W^T{S}_b^{\mathrm{\Φ}}W}{W^T{S}_w^{\mathrm{\Φ}}W}\right|$

in the high-dimensional feature space, the problem of maximizing the Fisher criterion function is converted into a problem of solving the feature value and the feature vector of the feature equation, and the method comprises the following steps:

define the kernel matrix K ═ K of N × N_ij]Then the above formula becomes

KBKα＝λKWKα

Wherein, K_ij＝k(x_i,x_j)＝Φ(x_i)^TΦ(x_j)，B＝GCG^T，

C＝diag(n₁,n₂,…,n_L)∈R^L×L，

$W=diag(I_{n_1}-\frac{1}{n_1}{1}_{n_1\×n_1},\mathrm{...},I_{n_L}-\frac{1}{n_L}{1}_{n_L\×n_L})\∈R^{N\×N}$

9. The method of claim 8, wherein the performing nonlinear feature extraction on the sensor response signals to obtain the flavor features of the tea sample comprises:

calculating a kernel matrix K-K of the training sample set according to the determined kernel function and the optimized kernel function parameters_ij]In which K is_ij＝k(x_i,x_j)＝Φ(x_i)^TΦ(x_j)；

The Fisher criterion function maximization is converted into a problem of solving generalized eigenvalues, and the eigenvalues of KBK α -lambda KWK α and corresponding eigenvectors α - α are solved₁,α₂,…,α_N]^TAnd is from large according to the characteristic valueSorting to a small order;

will train sample phi (x)_i) The most sampled nonlinear feature projected onto the kth feature vector:

$\mathrm{\Φ}{\left(x_i\right)}^T{v}^k=\mathrm{\Φ}{\left(x_i\right)}^T\underset{j=1}{\overset{N}{\Σ}}{\mathrm{\α}}_j^k\mathrm{\Φ}\left(x_j\right)=\underset{j=1}{\overset{N}{\Σ}}{\mathrm{\α}}_j^k\mathrm{\Φ}{\left(x_i\right)}^T\mathrm{\Φ}\left(x_j\right)=\underset{j=1}{\overset{N}{\Σ}}{\mathrm{\α}}_j^k{K}_{ij}$

calculating a kernel matrix K-K of the training sample set according to the determined kernel function and the optimized kernel function parameters_ij]In which K is_ij＝k(x_i,x_j)＝Φ(x_i)^TΦ(x_j)；

The eigenvalues of KBK α λ KWK α and the corresponding eigenvectors α [ α ]₁,α₂,…,α_N]^TSorting according to the sequence of the characteristic values from large to small;

will train sample phi (x)_i) The most sampled nonlinear feature projected onto the kth feature vector:

$\mathrm{\Φ}{\left(x_i\right)}^T{v}^k=\mathrm{\Φ}{\left(x_i\right)}^T\underset{j=1}{\overset{N}{\Σ}}{\mathrm{\α}}_j^k\mathrm{\Φ}\left(x_j\right)=\underset{j=1}{\overset{N}{\Σ}}{\mathrm{\α}}_j^k\mathrm{\Φ}{\left(x_i\right)}^T\mathrm{\Φ}\left(x_j\right)=\underset{j=1}{\overset{N}{\Σ}}{\mathrm{\α}}_j^k{K}_{ij}$

calculating test samplesAnd a kernel matrix K' between the training set samples, projecting the test samples onto the feature vectors

10. The method for extracting difference features of taste sensation signals based on kernel linear discriminant analysis according to claim 1, wherein the step of inputting the taste features of the tea sample into a classifier to perform tea quality grade determination comprises:

for the sample to be testedAnd training image sample x_iCalculating the similarity between the image sample to be tested and the training image sample

$d({\tilde{x}}_i,x_i)=\sqrt{\underset{k=1}{\overset{d}{\Σ}}{(x_{ik}^{\′}-x_{ik})}^2}$

If it isSample x_iIf it belongs to class k, the test sample is testedIs decided as class k.