CN107122643B - Identification method based on feature fusion of PPG signal and respiratory signal - Google Patents

Abstract

The invention discloses an identity recognition method based on the feature fusion of PPG signal and breathing signal, which mainly solves the problem of low recognition rate of the existing method. Its implementation steps: 1) preprocess the training PPG signal and the breathing signal; 2) perform peak detection on the processed signals respectively, and obtain the waveform samples of the two signals; Extracting features from the two signal samples to obtain a feature matrix; 4) Fusion of the feature matrices of the two signals to obtain a training template library; 5) Projecting the test PPG samples and breathing samples respectively to obtain test features; 6) Performing feature fusion to obtain fusion Test samples; 7) Predict the test sample categories to obtain recognition results. The present invention can generate stable and effective fusion feature samples. The simulation results show that the recognition rate reaches 100%, and it can be used in medical, security and defense applications that require high identification accuracy.

Description

Identification method based on feature fusion of PPG signal and respiratory signal

technical field

The invention belongs to the technical field of information processing, and in particular relates to an identity recognition method based on feature fusion, which can be used in application fields such as medical treatment, security defense and the like.

Background technique

With the development of wireless networks, wireless network applications such as telemedicine and e-commerce are playing an increasingly important role in people's lives. Since these applications involve important information such as personal and property, security issues are particularly important. The traditional identity authentication uses ID cards or passwords to identify and authorize, and such security protection is not secure enough. The ID card or password information is easy to be stolen or forgotten, and the biometric system is widely used because of its uniqueness, reliability and privacy. However, some biometric identification methods still have certain security risks. For example, fingerprints can be extracted with latex, face recognition can be deceived with fake photos, and voices can also be imitated. Electrode collection is not widely available.

The photoplethysmography PPG signal is acquired by a non-invasive method, which is convenient and simple to collect, and reflects the abundant microcirculation physiological information of the human body. In addition, the respiration signal contains various information such as the movement of human internal organs, which complements the human body information contained in the PPG signal. At present, the recognition rate of the identification technology based on PPG signal is low, and it is difficult to meet the application occasions with high accuracy requirements.

The proposed identification methods based on PPG signals are:

In the paper "An Enhanced Intrinsic Biometric in Identifying People by Photopleythsmography Signal" presented by N.S.Girish Rao Salanke, N.Maheswari and Andrews Samraj et al at the 2012 "theFourth International Conference on Signal and Image Processing" Component analysis KPCA method is a method of extracting PPG signal features for identification. This method first divides the PPG signal into multiple single cycles, and then denoises and normalizes the multiple single cycles, and then uses the kernel principal component analysis method. The method extracts the features of the single-cycle waveform, and finally uses the Mahalanobis distance for identification. This paper analyzes the recognition performance of PPG signals under tension and relaxation, but does not give specific identification rates of individuals, and the number of experimental subjects is small, which cannot fully demonstrate the recognition of this method.

The method proposed in the article "Photoplethysmogram Based Biometric Recognition for Twins" by NI Mohammed Nadzr, M Sulaimi, LF Umadi, KA Sidek, etc., published in the "IndianJournal of Science and Technology" in 2016, uses the single-cycle waveform of PPG signal for identity recognition . In this method, the original PPG signal is denoised by a low-pass filter, and then the waveform of the PPG signal is segmented to extract the single-cycle waveform, and then the single-cycle waveform is identified by the radial basis function network and the naive Bayes classifier. Classification, the final identification rate of correct identification reaches more than 97%. This method verifies the effectiveness of the single-cycle waveform feature of the PPG signal for individual identification, but does not fully exploit and utilize the single-cycle waveform feature of the PPG signal, and the identification rate still has room for improvement.

SUMMARY OF THE INVENTION

Aiming at the deficiencies of the above-mentioned prior art, the present invention proposes an identity recognition method based on feature fusion of PPG signal and breathing signal, so as to improve the correct recognition rate of identity.

The technical solution for realizing the purpose of the present invention is to extract the stable features of the breathing signal and the PPG signal for fusion respectively, so that the effective information of the breathing signal and the PPG signal complement each other, and generate a fusion feature containing sufficient and stable identity information of the individual, and then carry out the identification. Identification, the implementation steps are as follows:

(1) Perform denoising and normalization on the PPG training data set X={X ₁ , X ₂ ,...,X _n } and the breathing training data set Y={Y ₁ , Y ₂ ,..., Y _n } respectively Processing, the normalized PPG signal set Z={Z ₁ , Z ₂ ,...,Z _n } and the normalized respiratory signal set R={R ₁ , R ₂ ,..., R _n } are obtained, where n represents the total number of people;

(2) Perform peak detection on the PPG signal set Z to obtain the PPG signal peak position set Loc={Loc ₁ ,Loc ₂ , . . . Loc _n } of all people; perform peak detection on the respiratory signal set R to obtain The set of peak positions of respiratory signals of all people L={L ₁ , L ₂ ,...,L _n };

(3) with all elements in the PPG signal peak position set Loc as the reference point, extract the waveform sample of the PPG signal; With all the elements in the respiratory signal peak position set L as the reference point, extract the waveform sample of the respiratory signal;

(4) remove the samples with large differences in the PPG waveform samples of all people, and obtain the PPG training set M _p ; remove the samples with large differences in the respiratory signal waveform samples of all people, and obtain the respiratory training set _MR ;

(5) Using the smooth non-negative matrix decomposition method with exponential sparsity constraint to extract the features of the PPG training set _{M p} and the breathing training set MR respectively, and obtain the PPG feature matrix C _P , the PPG projection matrix _WP , and the breathing feature matrix C _R and breathing projection matrix _WR ;

(6) each column in the PPG feature matrix CP is connected in series with the corresponding column of the respiratory feature matrix _CR , and a fusion training template is generated, and the training template library _F is formed by all the fusion training templates;

(7) utilize the PPG projection matrix _WP and the breathing projection matrix WR to obtain the PPG test feature matrix _TP and the _{breathing feature test matrix TR ;}

(8) connecting each column of the PPG test feature matrix _TP and the corresponding column of the breathing test feature matrix _TR in series to obtain a fusion feature column, taking the fusion feature column as a fusion test sample, and forming a test sample library Lb from all the fusion test samples;

(9) Perform category prediction on all the fusion test samples in the sample library Lb, and obtain the identification result of the identified person.

Compared with the prior art, the present invention has the following advantages:

First, the present invention utilizes the exponential sparse constrained smooth non-negative matrix decomposition method to perform feature extraction on the PPG signal waveform samples and the respiratory signal waveform samples, and fully obtains the effective information of the PPG signal and the effective information of the respiratory signal, thereby improving the identity recognition. accuracy rate.

Second, the present invention obtains fusion features by extracting the main features of the individual PPG signal and respiratory signal waveform samples, and fusing the PPG signal features with the respiratory signal features, so that the effective information of the PPG signal and the effective information of the respiratory signal complement each other. , using fusion features for identification, which improves the accuracy of identification.

Description of drawings

Fig. 1 is the realization flow chart of the present invention;

Fig. 2 is the result figure of the identification rate of MIMIC database;

Figure 3 shows the result of the identification rate of MIMIC2 database.

Detailed ways

The implementation and effects of the present invention will be described in further detail below with reference to the accompanying drawings.

1, the implementation steps of the present invention are as follows:

Step 1. Preprocess the training data.

The commonly used public PPG signal databases include MIMIC database, MIMIC2 database, etc. In this example, the PPG signal and respiratory signal data of 50 people in the MIMIC database are selected, and the PPG signal with a length of 400 seconds and the respiratory signal with a length of 400 seconds per person are used as training data. , the sampling frequency f of the PPG signal and the breathing signal is 125Hz, then the length of the PPG signal data of each person is 400*125=50000, the length of the breathing signal data is 400*125=50000, and the PPG training data set is composed of the PPG signals of 50 people X={X ₁ ,X ₂ ,...,X _i ,...,X _n }, the respiratory training data set is composed of respiratory signals of 50 people Y={Y ₁ ,Y ₂ ,...,Y _i ,...,Y _n } , where X _i represents the PPG signal data of the ith person, Y _i represents the respiratory signal data of the ith person, i=1,2,...,n,n represents the total number of people, for the PPG training data set X and breathing training data Set Y for preprocessing:

Perform denoising and normalization on the PPG training data set X and breathing training set Y respectively, so that the normalized values of all sampling points are within the interval [0, 1], and the normalized PPG is obtained. signal set Z={Z ₁ , Z ₂ ,...,Z _n } and normalized respiratory signal set R={R ₁ ,R ₂ ,...,R _n };

Commonly used denoising methods for PPG signals and breathing signals include low-pass filtering, wavelet transform denoising, and adaptive morphological filtering, etc. This example adopts but is not limited to low-pass filtering.

Step 2. Carry out peak detection to the normalized PPG signal Z _i of the ith person, obtain the PPG signal peak position sequence Loc _i of the ith person, and then obtain the PPG signal peak position set consisting of the peak position sequence of the PPG signal of all people. Loc={Loc ₁ ,Loc ₂ ,...,Loc _i ,...,Loc _n }; perform peak detection on the normalized respiratory signal R _i of the ith person to obtain the respiratory signal peak position sequence Li of the _ith person, Then, a set of respiratory signal peak positions L= _{ L ₁ , L ₂ , . . . , L _i , .

The peak detection method of PPG signal and respiration signal here adopts the dynamic differential threshold peak detection method. "A text.

Step 3. Obtain PPG signal waveform samples and respiratory signal waveform samples.

(3a) Taking each element in the i-th person's PPG signal peak position sequence Loc _i as a reference point, extracting a sample point before a and a back a ₂ sample points of each reference point in the i _- th person's PPG signal Z _i Sampling point and reference point, composed of (a ₁ +a ₂ +1) sampling points to form a PPG signal waveform sample, obtain the PPG signal waveform sample with the same number of elements in the sequence Loc _i , and then obtain the PPG signal waveform sample of everyone , where a ₁ and a ₂ are two positive integers;

(3b) Taking each element in the i-th person's respiratory signal position sequence Li as a reference point, extract b ₁ sampling point before and b ₂ sampling points after each reference point in the _i -th person's breathing signal R _i point and reference point, a respiratory signal waveform sample is composed of (b ₁ +b ₂ +1) sampling points, and the respiratory signal waveform samples with the same number of elements as in the sequence _Li are obtained, and then the respiratory signal waveform samples of all people are obtained, Among them, b ₁ and b ₂ are two positive integers.

Step 4. Remove waveform samples with large differences to obtain a training set.

表示(a₁+a₂+1)维的实数向量空间，gn表示PPG训练集合M_p中样本个数；(4a) Perform waveform averaging on all PPG signal waveform samples of the ith person to obtain an average waveform sample, calculate the Euclidean distance between each waveform sample of the ith person and the average waveform sample, arrange all the Euclidean distances from small to large, and select For the waveform samples corresponding to the first m _i Euclidean distances, the selected waveform samples are taken as column vectors, and the PPG training set M _p = {SP ₁ , SP ₂ ,..., SP _gn } is formed by the selected PPG signal waveform samples of all people. where m _i is a positive integer,

represents the (a ₁ +a ₂ +1)-dimensional real vector space, and gn represents the number of samples in the PPG training set M _p ;

表示(b₁+b₂+1)维的实数向量空间。(4b) Perform waveform averaging on all respiratory signal waveform samples of the ith person to obtain an average waveform sample, calculate the Euclidean distance between each waveform sample of the ith person and the average waveform sample, arrange all the Euclidean distances from small to large, and select For the waveform samples corresponding to the first m _i Euclidean distances, the selected waveform samples are taken as column vectors, and the respiratory training set is composed of the selected respiratory signal waveform samples of all people M _R ={SR ₁ ,SR ₂ ,...,SR _gn } ,in,

Represents a (b ₁ +b ₂ +1)-dimensional real vector space.

Step 5. Perform feature extraction on the PPG training set M _p and the breathing training set MR, respectively, to obtain a PPG feature matrix _CP , a PPG projection matrix _WP , a breathing feature matrix _CR and a breathing projection matrix WR _.

(5a) In order to effectively improve the sparseness of the decomposition results of the non-smooth non-negative matrix factorization method and the interpretability of local features, and reduce the decomposition error, an exponential sparse constraint is added to the objective function of the non-smooth non-negative matrix factorization method to obtain the exponential The objective function of the smooth non-negative matrix factorization method for sparsity constraints:

Constraints need to be met:

W ^T W=I, S ^T S=I,

Among them, V represents the data matrix to be decomposed, W represents the basis matrix, S represents the smooth matrix, H represents the coefficient matrix, V _qs represents the element of the qth row and the sth column of the matrix V to be decomposed, and W· _κ represents the matrix W. κ column, S· _v represents the vth column of matrix S, ||·|| ₂ represents the L2 norm of the vector, β, λ represent two constraint parameters, (·) ^T represents the transpose of the matrix, and I represents the unit matrix;

(5b) Decompose the PPG training set M _p by using the smooth non-negative matrix decomposition method with exponential sparsity constraint to obtain the PPG feature matrix C _P and the PPG projection matrix W _P , and the specific operation steps are as follows:

和

分别表示(a₁+a₂+1)×r₁维、r₁×r₁维和r₁×gn维的实数矩阵空间，r₁表示PPG信号的分解维数；(5b1) Randomly initialize the basis matrix W ⁽⁰⁾ , the smoothing matrix S ⁽⁰⁾ and the coefficient matrix H ⁽⁰⁾ so that all elements are in the interval (0,1), where,

and

respectively represent (a ₁ +a ₂ +1)×r ₁ -dimensional, r ₁ ×r ₁ -dimensional and r ₁ ×gn-dimensional real matrix spaces, and r ₁ represents the decomposition dimension of the PPG signal;

进行更新：(5b2) According to the following iterative formula, for the elements of the basis matrix W ^(t)

To update:

First, update according to the following formula to get the intermediate variable value

Then, for intermediate variable values Perform column normalization to get

表示迭代t次后的基矩阵W^(t)的第

行第θ列元素，t表示迭代次数，t∈[1,iter]，iter为预先定义的最大迭代次数，

θ＝1,2,…,r₁，

表示迭代t-1次后的基矩阵W^(t-1)的第

行第θ列元素，S^(t-1)表示迭代t-1次后的平滑矩阵，H^(t-1)表示迭代t-1次后的系数矩阵，e表示自然常数；in,

^Represents the th

The element of row θth column, t represents the number of iterations, t∈[1,iter], iter is the predefined maximum number of iterations,

θ=1,2,...,r ₁ ,

Represents the base matrix W ^(t-1) after iteration t-1 times

The element of row θth column, S ^(t-1) represents the smooth matrix after t-1 iterations, H ^(t-1) represents the coefficient matrix after t-1 iterations, and e represents a natural constant;

进行更新：(5b3) According to the following iterative formula, for the elements of the coefficient matrix H ^(t)

To update:

表示迭代t次后系数矩阵H^(t)的第θ行第

列元素，

表示迭代t-1次后系数矩阵H^(t-1)的第θ行第

列元素；in,

Represents the θth row of the coefficient matrix H ^(t) after t iterations

column element,

Represents the θth row of the coefficient matrix H ^(t-1) after t-1 iterations

column element;

进行更新：(5b4) According to the following formula, for the elements of the smoothing matrix S ^(t)

To update:

First, update according to the following formula to get the intermediate variable value

Then, to the intermediate variable value Column normalization, we get

表示迭代t次后平滑矩阵S^(t)的第u行第v列元素，u,v＝1,2,…,r₁，

迭代t-1次后平滑矩阵S^(t-1)的第u行第v列元素；in,

represents the element of the uth row and the vth column of the smooth matrix S ^(t) after t iterations, u,v=1,2,...,r ₁ ,

After iterating t-1 times, the element of the uth row and the vth column of the smooth matrix S ^(t-1) ;

(5b5) Repeat steps (5b2)-(5b4), when the maximum iteration number iter is reached, the iteration is stopped, and the basis matrix W ^(iter) and the coefficient matrix H ^(iter) are output; The basis matrix W ^{(iter) is} used as the PPG projection matrix W _P , take the coefficient matrix H ^(iter) as the PPG feature matrix C _P , where,

和

分别表示r₂×gn维和(b₁+b₂+1)×r₂维的实数矩阵空间，r₂表示呼吸信号分解维数。(5b) According to step (5a), the breathing training set _MR is decomposed to obtain the breathing characteristic matrix _CR and the breathing projection matrix WR, _wherein ,

and

respectively represent r ₂ ×gn-dimensional and (b ₁ +b ₂ +1)×r ₂ -dimensional real matrix spaces, and r ₂ represents the respiratory signal decomposition dimension.

Step 6. Each column of the PPG feature matrix CP and the corresponding column in the respiratory feature matrix _CR are connected in series one-to-one to generate a fusion training template; the training template library _F is formed by all the fusion training templates;

Step 7. Obtain the PPG test matrix _TP and the breath test matrix _TR .

和

分别表示(a₁+a₂+1)×G维和(b₁+b₂+1)×G维的实数矩阵空间，G表示PPG信号测试样本个数；(7a) The PPG signal test data and the respiratory signal test data of the appraiser are respectively processed in steps 1-4 to obtain a PPG test set P and a respiratory test set Q, wherein,

and

Represent (a ₁ +a ₂ +1)×G-dimensional and (b ₁ +b ₂ +1)×G-dimensional real matrix spaces, respectively, and G represents the number of PPG signal test samples;

表示r₁×G维的实数矩阵空间，inv(·)表示矩阵求逆运算；(7b) Use the PPG projection matrix WP to perform feature extraction on the PPG test set _P , and obtain the PPG test feature matrix T _P =inv((W _P ) ^T ×W _P )×(W _P ) ^T ×P, where,

Represents the real matrix space of r ₁ ×G dimension, and inv( ) represents the matrix inversion operation;

(7c) Use the breathing projection matrix WR to perform feature extraction on the breathing test set Q, and obtain the breathing test feature matrix T _R =inv _{( (} WR ) ^T × _WR )×( _WR ) ^T ×Q, where Represents a real matrix space of r ₂ ×G dimension.

Step 8. Connect each column of the PPG test feature matrix _TP and the corresponding column of the breathing test feature matrix _TR in series to obtain a fusion feature column, and use the fusion feature column as a fusion test sample, and all the fusion test samples form a test sample library Lb. .

Step 9. Use the K-nearest neighbor classifier for identification.

(9a) Calculate the Euclidean distance between each fusion test sample in the test sample library Lb and all the training templates in the training template library, arrange all Euclidean distances from small to large, select the training templates corresponding to the first K Euclidean distances, and then count The number of occurrences of each category in the K training templates, and the category with the largest occurrence frequency is selected as the predicted category of the fusion test sample, where K is a positive integer;

(9b) According to step (9a), perform category prediction on all the fusion test samples in the test sample library Lb, count the number of occurrences of each predicted category in all the fusion test samples, and use the category with the highest frequency of occurrence as the identity of the person to be authenticated .

Classifiers for class prediction for fusion test samples are not limited to K-nearest neighbor classifiers, and support vector machine SVM classifiers, Bayesian classifiers, etc. can also be used.

The effect of the present invention can be further illustrated by the following simulation.

1. Simulation conditions

The simulation experiment of the present invention uses the PPG signal and the breathing signal in the two public databases MIMIC and MIMIC2 respectively as the experimental data, and simulates the PPG signal and the breathing signal collected from the human body. conduct.

2. Simulation content

First, randomly select the PPG signals and respiratory signals of 50 people from the databases MIMIC and MIMIC2, respectively, use the method of the present invention to identify each person in the two databases, calculate the identification rate of each person, and obtain the identity of each database. Recognition rate curve, as shown in Figure 2 and Figure 3, where the calculation method of each person's identification rate is as follows:

Identification rate = the number of fusion test samples with correct category prediction / the total number of test samples of the identified person;

Then, take the average of the identification rates of everyone in each database as the identification rate for that library.

It can be seen from Fig. 2 and Fig. 3 that the average identification rate of each library reaches 100%, which fully demonstrates the effectiveness and high correct identification rate of the present invention.

The above description is only a specific example of the present invention, and does not constitute any limitation to the present invention. Obviously, for those skilled in the art, after understanding the content and principle of the present invention, it is possible to do so without departing from the principle and structure of the present invention. Below, various corrections and changes in form and details are made, but these corrections and changes based on the idea of the present invention are still within the protection scope of the claims of the present invention.

Claims (7)

1. An identification method based on the fusion of PPG signal and respiratory signal features, comprising the following steps: (1) Perform denoising and normalization on the PPG training data set X={X ₁ , X ₂ ,...,X _n } and the breathing training data set Y={Y ₁ , Y ₂ ,..., Y _n } respectively Processing, the normalized PPG signal set Z={Z ₁ , Z ₂ ,...,Z _n } and the normalized respiratory signal set R={R ₁ , R ₂ ,..., R _n } are obtained, where n represents the total number of people; (2) Perform peak detection on the PPG signal set Z to obtain the PPG signal peak position set Loc={Loc ₁ ,Loc ₂ , . . . Loc _n } of all people; perform peak detection on the respiratory signal set R to obtain The set of peak positions of respiratory signals of all people L={L ₁ , L ₂ ,...,L _n }; (3) with all elements in the PPG signal peak position set Loc as the reference point, extract the waveform sample of the PPG signal; With all the elements in the respiratory signal peak position set L as the reference point, extract the waveform sample of the respiratory signal; (4) remove the samples with large differences in the PPG waveform samples of all people, and obtain the PPG training set M _p ; remove the samples with large differences in the respiratory signal waveform samples of all people, and obtain the respiratory training set _MR ; (5) Using the smooth non-negative matrix decomposition method with exponential sparsity constraint to extract the features of the PPG training set _{M p} and the breathing training set MR respectively, and obtain the PPG feature matrix C _P , the PPG projection matrix W _P , and the breathing feature matrix C _R and the breathing projection matrix _WR ; (6) each column in the PPG feature matrix CP is connected in series with the corresponding column of the respiratory feature matrix _CR , and a fusion training template is generated, and the training template library _F is formed by all the fusion training templates; (7) Utilize the PPG projection matrix _WP and the breathing projection matrix WR to obtain the PPG test feature matrix _TP and the _{breathing feature test matrix TR ,} and the specific operations are as follows: (7a) Using the PPG projection matrix W _P , obtain the PPG test feature matrix T _P =inv((W _P ) ^T ×W _P )×(W _P ) ^T ×P, where inv( ) represents the matrix inversion operation, P represents the PPG test set; (7b) Using the respiration projection matrix W _R , obtain a respiration test feature matrix T _R =inv((W _R ) ^T ×W _R )×(W _R ) ^T ×Q, where Q represents a breathing test set; (8) connect each column of the PPG test feature matrix _TP and the corresponding column of the breath test feature matrix _TR in series to obtain a fusion characteristic column, and use the fusion characteristic column as a fusion test sample, and form the test sample library Lb by all the fusion test samples; (9) Perform category prediction on all the fusion test samples in the sample library Lb, and obtain the identification result of the identified person. 2. method according to claim 1, wherein extracting PPG signal waveform sample in step (3), take each element in the i-th person's PPG signal peak position sequence Loc _i as a reference point, in the i-th person's Extract a ₁ sampling point before and a ₂ sampling points after each reference point and the reference point from the PPG signal Z _i , and form a PPG signal waveform sample from these (a ₁ +a ₂ +1) sampling points, and obtain the PPG signal waveform samples with the same number of elements in the sequence Loc _i , and then obtain the PPG signal waveform samples of all people; where a ₁ , a ₂ are two positive integers, i=1, 2,...,n, n represents the total number of people . 3. method according to claim 1, wherein step (3) extracts respiration signal waveform sample, is the reference point with each element in the respiration signal peak position sequence Li of the _ith person, in the respiration of the ith person. Extract b ₁ sampling point before and b ₂ sampling points after each reference point and the reference point from the signal R _i , and form a respiratory signal waveform sample from these (b ₁ +b ₂ +1) sampling points, and obtain and sequence _Respiratory signal waveform samples with the same number of elements in Li, and then obtain respiratory signal waveform samples of all people; wherein, b ₁ and b ₂ are two positive integers. 4. The method according to claim 1, wherein step (4) obtains the PPG training set M _p , and the specific operations are as follows: Perform waveform averaging on all PPG signal waveform samples of the ith person to obtain an average waveform sample, calculate the Euclidean distance between each waveform sample of the ith person and the average waveform sample, arrange all the Euclidean distances from small to large, and select the first m _i The waveform samples corresponding to the Euclidean distances, the selected waveform samples are taken as column vectors, and the PPG training set M _p = {SP ₁ , SP ₂ ,..., SP _gn } is composed of the selected PPG signal waveform samples of all people, where, m _i is a positive integer,

represents the (a ₁ +a ₂ +1)-dimensional real vector space, and gn represents the number of samples in the PPG training set M _p . 5. The method according to claim 1, wherein step (4) obtains a breathing training set _MR , and the specific operations are as follows: Perform waveform averaging on all respiratory signal waveform samples of the ith person to obtain an average waveform sample, calculate the Euclidean distance between each waveform sample of the ith person and the average waveform sample, arrange all Euclidean distances from small to large, and select the first m _i The waveform samples corresponding to the Euclidean distances, the selected waveform samples are used as column vectors, and the respiratory training set _MR = {SR ₁ ,SR ₂ ,...,SR _gn } is composed of the selected respiratory signal waveform samples of all people, where, m _i is a positive integer,

Represents a (b ₁ +b ₂ +1)-dimensional real vector space. 6. method according to claim 1, wherein step (5) utilizes the smooth non-negative matrix decomposition method of exponential sparse constraint to carry out feature extraction to PPG training set _{M p} and respiratory training set MR respectively, carry out according to the following steps: (5a) Randomly initialize the basis matrix W ⁽⁰⁾ , the smoothing matrix S ⁽⁰⁾ and the coefficient matrix H ⁽⁰⁾ ; (5b) According to the following iterative formula, for the elements of the basis matrix W ^(t) To update: First, update according to the following formula to get the intermediate variable value

Then, for intermediate variable values

Perform column normalization to get

in, Represents the first value of the basis matrix W ^(t) after t iteration

The element of row θth column, t represents the number of iterations, t∈[1, iter], iter is the predefined maximum number of iterations,

θ=1,2,...,r, r=r ₁ or r ₂ , r ₁ represents the PPG signal decomposition dimension, r ₂ represents the respiratory signal decomposition dimension, V represents the matrix to be decomposed, V ₌ MP or _MR , Represents the basis matrix W ^(t-1) after iteration t-1 times

The element of row θth column, S ^(t-1) represents the smooth matrix after t-1 iterations, H ^(t-1) represents the coefficient matrix after t-1 iterations, β represents the constraint parameter, ( ) ^T Represents the transpose of the matrix, and e represents a natural constant; (5c) According to the following formula, for the elements of the coefficient matrix H ^(t)

To update: in,

Represents the θth row of the coefficient matrix H ^(t) after t iterations

column element,

Represents the θth row of the coefficient matrix H ^(t-1) after iteration t-1 times

column element, λ represents the constraint parameter; (5d) According to the following formula, for the elements of the smoothing matrix S ^(t)

To update: First, update according to the following formula to get the intermediate variable value

Then, to the intermediate variable value Column normalization to get

in,

represents the element of the u-th row and the v-th column of the smooth matrix S ^(t) after t iterations, u,v=1,2,...,r ₁ ,

represents the element of the uth row and the vth column of the smooth matrix S ^(t-1) after iteration t-1; (5e) repeat steps (5b)-(5d), when reaching maximum iteration number iter, stop iteration, output base matrix W ^(iter) and coefficient matrix H ^(iter) ; When matrix V to be decomposed is PPG training set M During _P , the base matrix W ^{(iter) is} used as the PPG projection matrix W _P , and the coefficient matrix H ^{(iter) is} used as the PPG feature matrix C _P ; When the matrix V to be decomposed is the respiratory training set _MR , the base matrix W ^{( iter)} as the respiration feature matrix C _R , and the coefficient matrix H ^(iter) as the respiration projection matrix W _R . 7. The method according to claim 1, wherein step (9) obtains the identification result of the appraised person, and carries out as follows: (9a) Calculate the Euclidean distance between each fusion test sample in the test sample library Lb and all the training templates in the training template library, arrange all Euclidean distances from small to large, select the training templates corresponding to the first K Euclidean distances, and then count The number of occurrences of each category in the K training templates, and the category with the largest occurrence frequency is selected as the predicted category of the fusion test sample, where K is a positive integer; (9b) According to step (9a), perform category prediction on all fused test samples in the test sample library Lb, count the number of occurrences of each predicted category in all test samples, and use the category with the highest occurrence frequency as the identity of the identified person.

Images