ECG Identity Recognition Method Based on Local Segment Sparse Representation

Abstract

The present invention discloses an ECG identity recognition method based on local segment sparse representation, which includes collecting an electrocardiogram signal, using wavelet transform based on a weight threshold to perform signal denoising to obtain the denoised ECG signal, implementing signal division on the denoised ECG signal by using a sliding window method, extracting local segments, obtaining local features of each signal by using principal component analysis to achieve data dimensionality reduction, using the orthogonal matching pursuit algorithm to find the optimal matching atoms during the sparse representation process for the dimensionality-reduced data, using the K-singular value decomposition algorithm to construct a dictionary for sparse representation to obtain a processed sparse coefficient matrix, and performing probability neural network recognition on the final sparse coefficient matrix to obtain the recognition accuracy of the ECG signal. The present invention can well capture global and local information, and sparse representation and dictionary construction can improve the reliability of identity recognition based on electrocardiogram signals.

Description

Technical Field

[0001] The present invention belongs to the technical field of biometric identification, and relates to an ECG identity recognition method based on local fragment sparse representation. Background Art

[0002] In recent years, science and technology have been the primary productive forces in our country, and people's lives have gradually become more information-based. In the past, the means of identity recognition were knowledge-based or token-based, and some drawbacks have gradually emerged in their related applications. For example, passwords and codes are easily forgotten, or, like identity verification documents, they may be stolen, tampered with, traded, or shared. Therefore, biometric data, which serves as the basis for identity recognition, will be safer and more reliable.

[0003] Biometric features have characteristics such as uniqueness, stability, portability, and good anti-counterfeiting properties, and have become a new medium in the field of personal identity recognition. Currently, physiological methods such as face recognition, fingerprint recognition, and voice recognition have been widely studied in society. Face recognition uses people's facial features for differentiation, fingerprint recognition utilizes the uniqueness of each person's fingerprint to ensure distinguishability between people, and voice recognition is based on the different timbre characteristics of each person for identification. However, with the development of related technologies, counterfeiting of biometric features such as faces, fingerprints, and voices has emerged, which has led to a reduction in the security and reliability of biometric identification. Therefore, it is very necessary to find a non-forgeable and distinguishable biometric feature for identity recognition.

[0004] The progress of technology has improved people's living standards, but at the same time, it has also led to many bad living habits, which can cause some heart diseases. Electrocardiogram is an effective means for the diagnosis of heart diseases. Initially, electrocardiogram signals were used for disease diagnosis, but in the past decade, they have been widely used in biometric identification because of their multiple functions for biometric recognition. Different individuals have different physiological and cardiac structures, so individuals can be identified from electrocardiogram signals. For other traditional biometric features, such as gait, face, voice, etc., it is easy to leave traces in daily life. Although they have been widely used in the market, due to their easy collection by cameras and recording devices, their security is not very high, and they can also be imitated. In contrast, electrocardiogram signals belong to internal physiological features, which means they have higher security. Moreover, the devices for collecting electrocardiogram signals mostly rely on medical devices and are not easily stolen. Their weak signals, low frequencies, and variability make them not easily imitated. Therefore, identity recognition based on electrocardiogram signals is more conducive to the information management of modern society.

[0005] Although certain achievements have been made in the identification of electrocardiogram (ECG) signals for identity recognition, when using existing ECG signal identity recognition methods for identity recognition, it will be affected by factors such as changes in the human heart rate and the acquisition environment equipment, etc., which will cause certain deviations in identity recognition and reduce the reliability of ECG signal identity recognition. Summary of the Invention

[0006] The purpose of the present invention is to provide an ECG identity recognition method based on local segment sparse representation, which solves the problem of large deviations in existing ECG signal identity recognition methods.

[0007] The technical solution adopted by the present invention is an ECG identity recognition method based on local segment sparse representation, including the following steps:

[0008] Step 1, collect ECG signals, and use wavelet transform based on weight threshold to perform signal denoising to obtain the denoised ECG signal;

[0009] Step 2, use the sliding window method to divide the denoised ECG signal, extract local segments, and then use the principal component analysis method to obtain the local features of each signal to achieve data dimensionality reduction;

[0010] Step 3, use the orthogonal matching pursuit algorithm to find the optimal matching atoms during the sparse representation process for the data after dimensionality reduction, and use the K-singular value decomposition algorithm to construct the dictionary for sparse representation to obtain the processed sparse coefficient matrix;

[0011] Step 4, perform probability neural network recognition on the final sparse coefficient matrix obtained in Step 3 to obtain the recognition accuracy of the ECG signal.

[0012] Among them, the specific process of Step 1 is as follows:

[0013] Step 1.1, obtain the original ECG data through the device reading or database acquisition method, and then use the drawing algorithm to perform drawing processing on the obtained original ECG data to obtain a matrix storing ECG data, and this matrix is the ECG signal to be processed;

[0014] Step 1.2, use the denoising method based on wavelet weight threshold shrinkage to decompose and reconstruct the ECG signal obtained in Step 1.1 to obtain the denoised ECG signal.

[0015] The specific process of Step 1.2 is as follows:

[0016] Step 1.2.1, use the mallat wavelet algorithm to decompose the ECG signal obtained in Step 1.1 to obtain the wavelet coefficient values of each layer;

[0017] Step 1.2.2: Screen the wavelet coefficient values obtained in Step 1.2.1 using the dynamic soft-threshold formula to obtain the processed wave coefficients;

[0018] Step 1.2.3: Merge the wave coefficients processed in Step 1.2.2 into the wave structure to receive the signal, and then the denoised ECG signal can be obtained.

[0019] The dynamic soft-threshold formula in Step 1.2.2 is as follows:

[0020]

[0021] In the formula, j represents the decomposition scale, T j represents the critical threshold, W represents the wave coefficient, and sign(ω) is the sign function.

[0022] The specific process of Step 2 is as follows:

[0023] Step 2.1: Use the sliding window method to divide the denoised ECG signal and extract local segments to obtain the corresponding segmented matrix of the local segments;

[0024] Step 2.2: Process the segmented matrix obtained in Step 2.1 using the principal component analysis method to obtain the local features of each signal and achieve data dimensionality reduction.

[0025] The specific process of Step 2.2 is as follows:

[0026] Step 2.2.1: Calculate the mean of each column of the segmented matrix obtained in Step 2.1, and then subtract the corresponding mean from each column of data to calculate the characteristic covariance matrix;

[0027] Step 2.2.2: Calculate the eigenvalues and eigenvectors according to the characteristic covariance matrix, and then perform normalization processing on the eigenvectors;

[0028] Step 2.2.3: Select the eigenvector corresponding to the largest eigenvalue obtained in Step 2.2.2, project the sample points onto the selected eigenvector, and obtain the matrix Y after dimensionality reduction.

[0029] The specific process of Step 3 is as follows:

[0030] Step 3.1: Use the orthogonal matching pursuit algorithm for the matrix Y after dimensionality reduction to find the optimal matching atoms in the matrix Y during the sparse representation process;

[0031] Step 3.2: Construct a dictionary for sparse representation using the K-singular value decomposition algorithm according to the optimal matching atoms to obtain the processed sparse coefficient matrix.

[0032] The specific process of Step 3.2 is as follows:

[0033] Step 3.2.1: Describe the optimal matching atoms obtained in Step 3.1 with linear expressions of basis elements to form a basis atom dictionary. Initialize the dictionary and use this dictionary to sparsely represent the given samples to obtain the corresponding matrix.

[0034] Step 3.2.2: When the matrix obtained in Step 3.2.1 is sparse, iteratively update the dictionary column by column to continuously reduce the error between the sample data and the data represented by the coefficient matrix, approaching the optimal basis atom dictionary infinitely.

[0035] Step 3.2.3: Update the basis atoms obtained in Step 3.2.2 column by column.

[0036] Step 3.2.4: After the dictionary update in Step 3.2.3, perform sparse coding with the new dictionary and stop updating the dictionary when the number of iterations reaches the specified value or the error rate reaches the specified range. The processed sparse coefficient matrix can be obtained, that is, the reconstructed electrocardiogram signal Yˊ = Dx after sparse representation. Yˊ is the data signal matrix after dimensionality reduction to be processed, D ∈ R m×n is the dictionary matrix, and x is the sparse coefficient obtained after processing, x ∈ R m .

[0037] The specific process of Step 4 is as follows:

[0038] Step 4.1: Input the final sparse coefficient matrix obtained in Step 3 into the input layer of the probabilistic neural network and then enter the hidden layer.

[0039] Step 4.2: Input the data output from the hidden layer into the summation layer for calculation to obtain the maximum category calculated in the summation layer.

[0040] Step 4.3: Normalize the maximum category data calculated in the summation layer to obtain the probability estimate of each category, that is, the recognition accuracy of the ECG signal.

[0041] The beneficial effects of the present invention are as follows: Extract local features from the training signal to construct a compact and discriminative dictionary. This method can well capture global and local information, and sparse representation and dictionary construction can improve the reliability of electrocardiogram signal-based identity recognition; Using the soft threshold wavelet denoising method to denoise the signal can effectively complete the filtering process of the electrocardiogram signal. The algorithm is relatively stable and the recognition accuracy is relatively high; The present invention is based on the original ECG identity recognition steps including preprocessing, feature extraction, and classification recognition of electrocardiogram signals. Adding the sparse representation link can well filter out the invalid information of electrocardiogram signals, reduce the signal complexity, and also reduce costs and time. Description of the Drawings

[0042] Figure 1It is the overall flowchart of the ECG identity recognition method based on local segment sparse representation of the present invention;

[0043] Figure 2 It is the specific flowchart of wavelet threshold denoising in the ECG identity recognition method based on local segment sparse representation of the present invention;

[0044] Figure 3 It is the specific flowchart of feature extraction of the electrocardiogram signal in the ECG identity recognition method based on local segment sparse representation of the present invention;

[0045] Figure 4 It is the sparse representation flowchart in the ECG identity recognition method based on local segment sparse representation of the present invention;

[0046] Figure 5 It is the basic structure diagram of the probabilistic neural network (PNN) used in the ECG identity recognition method based on local segment sparse representation of the present invention. Specific embodiments

[0047] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0048] The ECG identity recognition method based on local segment sparse representation of the present invention, referring to Figure 1 , includes the following steps:

[0049] Step 1, referring to Figure 2 , collect the electrocardiogram signal, and use wavelet transform based on weight threshold to denoise the signal to obtain the denoised ECG signal. The specific process is as follows:

[0050] Step 1.1, obtain the original electrocardiogram data through the arrhythmia electrocardiogram database (MIT-BIH), and then use the plotting algorithm to plot the obtained original electrocardiogram data to obtain a matrix storing ECG data. This matrix is the ECG signal to be processed. In this embodiment, the ECG algorithm reading program written by Robert Tratnig of the Vorarlberg University of Applied Sciences is selected. By borrowing this program, any set of sample data in the electrocardiogram sample database can be read, plotted, and a matrix storing ECG data can be obtained. This matrix is the ECG signal to be processed;

[0051] Step 1.2, use the denoising method based on wavelet weight threshold shrinkage to decompose and reconstruct the ECG signal obtained in Step 1.1 to obtain the denoised ECG signal. The specific process is as follows:

[0052] Step 1.2.1: First, select the wavelet basis scaling function db5 that can best highlight the characteristics of the ECG signal. At the same time, select the decomposition principle of 8 scales, and apply the fast dyadic orthogonal wavelet transform algorithm based on multi-resolution analysis, that is, use the mallat wavelet algorithm to perform wavelet decomposition and reconstruction of the ECG signal obtained in Step 1.1 at 8 scales to obtain the wavelet coefficient values of each layer;

[0053] Step 1.2.2: Use the dynamic soft threshold formula to screen the wavelet coefficient values obtained in Step 1.2.1 to obtain the processed wave coefficients;

[0054] The dynamic soft threshold formula in Step 1.2.2 is as follows:

[0055]

[0056] In the formula, j represents the decomposition scale, T j represents the critical threshold, W represents the wave coefficient, and sign(ω) is the sign function;

[0057] Step 1.2.3: Incorporate the wave coefficients processed in Step 1.2.2 into the wave structure to receive the signal, and then the denoised ECG signal can be obtained.

[0058] Step 2: Refer to Figure 3 , use the sliding window method to divide the denoised ECG signal, extract local segments, and then use the principal component analysis method to obtain the local features of each signal to achieve data dimensionality reduction. Specifically:

[0059] Step 2.1: For the denoised ECG signal, use the non-reference point method, that is, the sliding window method. First, preset a window with a fixed length, and the width of the sliding window should be greater than the width of one cycle. Then move from the first segment to the last segment of the denoised ECG signal obtained in Step 1 to extract local segments, and obtain the segmented matrix X m×n ={x1, x2,..., x n}, where x n is each column of data in the data matrix;

[0060] Step 2.2: Use the principal component analysis method to process the segmented matrix obtained in Step 2.1 to obtain the local features of each signal and achieve data dimensionality reduction. The specific process is as follows:

[0061] Step 2.2.1: Calculate the mean value of each column of the segmented matrix obtained in Step 2.1 Then subtract the corresponding mean value from each column of data to obtain Finally, calculate the characteristic covariance matrix

[0062] Step 2.2.2, obtain eigenvalues and eigenvectors based on the characteristic covariance matrix: A cov λ = λv, A cov = E∑E -1 , where ∑ is a diagonal matrix, λ are the eigenvalues, v is the eigenvector, and the eigenvectors are normalized to obtain the matrix E m×n ';

[0063] Step 2.2.3, select the eigenvectors corresponding to the largest several eigenvalues obtained in Step 2.2.2 to form the matrix E n×k ′, project the sample points onto the selected eigenvectors, and the reduced - dimensional matrix Y = X m×k = X m×n ×E n×k ′.

[0064] Step 3, refer to Figure 4 , use the orthogonal matching pursuit algorithm for the reduced - dimensional data to find the optimal matching atoms during the sparse representation process, and use the K - singular value decomposition algorithm to construct the dictionary for sparse representation to obtain the processed sparse coefficient matrix. The specific process is as follows:

[0065] Step 3.1, use the orthogonal matching pursuit algorithm for the reduced - dimensional matrix Y to find the optimal matching atoms in the matrix Y during the sparse representation process. Specifically:

[0066] Step 3.1.1, Y is the original signal to be decomposed obtained in Step 2, D = {x i} ∈ R n×K is an over - complete dictionary matrix, and the norms of all atoms in D are equal to 1. R×f is the data set after the k - th step of iteration. When initializing, f0 = 0, R′f = f, x = 0, a 0 = 0, k = 0. It can be assumed that after the k - th step of decomposition of the signal data set, the signal becomes:

[0067]

[0068] and x n , R k f = 0 n = 1, 2,, k... a n k represents the coefficient obtained after the k - th step of decomposition;

[0069] Step 3.1.2, according to Equation (3 - 1) obtained in Step 3.1.1, it can be deduced that at the (k + 1) - th step, the above formula can be decomposed to obtain:

[0070] and <x n , R k+1 f> = 0 n = 1, 2, k + 1... (3 - 2)

[0071] and <γ k , x n > = 0, n = 1, 2, ..., k (3 - 3)

[0072] wherein, represents the projection of x k+1 onto {x1, x2, ..., x k}}. represents the component obtained by vertically mapping x k+1 relative to the matrix {x1, x2, ..., x k}}, where

[0073] a n k+1 = a n k - a k b n k , n = 1, 2, ..., k, and

[0074]

[0075] Step 3.1.3, according to the remaining unprocessed signal R k+1 f in Equation (3 - 2) in Step 3.1.2 has certain conditions satisfied, and we can obtain:

[0076] R k+1 f = R k f - α k γ k , and

[0077] when the iteration reaches the requirement of the control condition in the k - th loop, the iteration ends: k = K, where K is the sparsity, and the optimal matching atom can be obtained;

[0078] Step 3.2, use the K - singular value decomposition algorithm to construct a dictionary for sparse representation based on the optimal matching atom, and obtain the processed sparse coefficient matrix, specifically:

[0079] Step 3.2.1, describe the optimal matching atom obtained in Step 3.1 with a linear expression of basis elements to form a basis atom dictionary, initialize the dictionary, use this dictionary to sparsely represent the given sample to obtain the corresponding matrix, and multiply the basis atom dictionary by the sparse matrix, that is:

[0080]

[0081] d i represents the column of D, xi The row representing X, and then update the columns of the dictionary.

[0082] Step 3.2.2, the matrix obtained in Step 3.2.1 is often not optimal, and there is a certain error between the sample data and the data represented by the coefficient matrix. When the matrix obtained in Step 3.2.1 is sparse, iteratively update the dictionary column by column to continuously reduce the error between the sample data and the data represented by the coefficient matrix, approaching the optimal basis atom dictionary infinitely. The formula for calculating the error is as follows:

[0083]

[0084] Step 3.2.3, since the dictionary update is column-based, update the basis atoms obtained in Step 3.2.2 column by column; that is, assume that X and D are known before derivation, and update the k-th column of the dictionary to obtain d k , and the k-th row in the corresponding coefficient matrix X is x k T , so that the multiplication term in Equation (3-8) obtained in Step 3.2.2 can be rewritten as the equation:

[0085]

[0086] Solve Equation (3-9). If directly decomposed by SVD, the obtained result will "diverge". After transforming Equation (3-9), the following equation is obtained:

[0087]

[0088] In Equation (3-10), Ω k is an N×|ω k | matrix. The characteristic of this matrix is that the value at (ω k (i),i) is equal to 1, and other points are equal to 0. According to the characteristics of its matrix, define where are respectively Y, E k the contraction results obtained after removing the zero input vectors in k . Perform SVD decomposition on E R k = UΔV T , and update the dictionary column by column;

[0089] Step 3.2.4, after the dictionary update is completed, use the new dictionary Perform sparse coding and stop updating the dictionary when the number of iterations reaches a specified value or the error rate reaches a specified range. Then, the processed sparse coefficient matrix can be obtained, that is, the reconstructed electrocardiogram (ECG) signal after sparse representation: Yˊ = Dx, where Yˊ is the data signal matrix after dimensionality reduction to be processed, D ∈ R m×n is the dictionary matrix, and x is the sparse coefficient obtained after processing, x ∈ R m .

[0090] Step 4: Refer to Figure 5 , and perform probability neural network (PNN) recognition on the final sparse coefficient matrix obtained in Step 3 to obtain the recognition accuracy of the ECG signal. The specific process is as follows:

[0091] Step 4.1: Input the final sparse coefficient matrix obtained in Step 3 into the input layer of the probability neural network (PNN), and then enter the hidden layer;

[0092] The PNN consists of an input layer x1, x2, x3,..., x f , a hidden layer (for calculating distances), a summation layer Ψ1, Ψ2, Ψ3,..., Ψ f , and an output layer y, arranged in sequence from left to right. The sample signal first enters the hidden layer through the input layer. Since each neuron node in the hidden layer has a center point, the distances between these vectors and the neuron centers can be calculated after detecting the input data. First, multiply the input layer vector by a weighting value and calculate using the corresponding function:

[0093] Z i = xq i (4-1)

[0094] where i = 1, 2,..., N, and N is the total number of types of training data. If the input layer vector and the weighting value are normalized, the data relationship determined by the jth pattern neuron of the ith type in the hidden layer is determined by Equation (4-2) as follows:

[0095]

[0096] Step 4.2: Input the data output from the hidden layer into the summation layer for calculation,

[0097]

[0098] where v i is the output of the calculation result of the ith class, H is the number of neurons of the ith class, and the number of types N is the same as the number of neurons in the summation layer;

[0099] Obtain the maximum type calculated in the summation layer,

[0100] y = argmax(v i) (4-4)

[0101] In the above formula (4-2), σ is a smoothing factor, which affects the final network and the calculation result. Each type in the summation layer corresponds to a neuron, and each type in the hidden layer also corresponds to a neuron. PNN uses the mechanism of Online Supervised Learning, and the data obtained by the output layer is positively correlated with the probability estimates of different types.

[0102] Step 4.3: Normalize the maximum type data calculated in the summation layer to obtain the probability estimate of each type, that is, the recognition accuracy of the ECG signal.

[0103] The number of types is the same as the number of neurons in the summation layer. The classification and recognition mainly obtain the output of the summation layer and distinguish samples based on it. Among all the neurons in the hidden layer, some have the highest probability density, and their output is 1, while the other neurons are 0. Finally, the final recognition category of the sample and a label vector are output. According to the label vector, the recognition accuracy after sparse representation can be obtained as 95.33% under the optimal parameters.

Claims

1. An ECG identity recognition method based on local segment sparse representation, characterized in that It includes the following steps: Step 1, collect electrocardiogram signals, and use wavelet transform based on weight threshold to perform signal denoising to obtain the denoised ECG signals; The specific process of Step 1 is as follows: Step 1.1, obtain the original electrocardiogram data by means of device reading or database acquisition, and then use a drawing algorithm to perform drawing processing on the obtained original electrocardiogram data to obtain a matrix storing ECG data, and this matrix is the ECG signal to be processed; Step 1.2, use a denoising method based on wavelet weight threshold shrinkage to decompose and reconstruct the ECG signal obtained in Step 1.1 to obtain the denoised ECG signal; The specific process of Step 1.2 is as follows: Step 1.2.1, use the mallat wavelet algorithm to decompose the ECG signal obtained in Step 1.1 to obtain the wavelet coefficient values of each layer; Step 1.2.2, use the dynamic soft threshold formula to screen the wavelet coefficient values obtained in Step 1.2.1 to obtain the processed wave coefficients; The dynamic soft threshold formula is: In the formula, represents the decomposition scale, T j represents the critical threshold, W represents the wave coefficient, is the sign function; Step 1.2.3, merge the wave coefficients processed in Step 1.2.2 into the wave structure to receive the signal, and then the denoised ECG signal can be obtained; Step 2, use the sliding window method to divide the denoised ECG signal, extract local segments, and then use the principal component analysis method to obtain the local features of each signal to achieve data dimensionality reduction; Step 3, use the orthogonal matching pursuit algorithm to find the optimal matching atoms in the sparse representation process for the data after dimensionality reduction, and use the K singular value decomposition algorithm to construct the dictionary for sparse representation to obtain the processed sparse coefficient matrix; Step 4, perform probability neural network recognition on the final sparse coefficient matrix obtained in Step 3 to obtain the recognition accuracy of the ECG signal.

2. The ECG identity recognition method based on local fragment sparse representation according to claim 1, wherein The specific process of Step 2 is as follows: Step 2.1, use the sliding window method to divide the denoised ECG signal, extract local segments, and obtain the segmented matrix corresponding to the local segments; Step 2.2, use the principal component analysis method to process the segmented matrix obtained in Step 2.1 to obtain the local features of each signal to achieve data dimensionality reduction.

3. The ECG identity recognition method based on local segment sparse representation according to claim 2, wherein The specific process of Step 2.2 is as follows: Step 2.2.1, calculate the mean value of each column of the segmented matrix obtained in Step 2.1, and then subtract the corresponding mean value from each column of data to calculate the characteristic covariance matrix; Step 2.2.2, calculate the eigenvalues and eigenvectors according to the characteristic covariance matrix, and then perform normalization processing on the eigenvectors; Step 2.2.3, select the eigenvector corresponding to the largest eigenvalue obtained in Step 2.2.2, project the sample points onto the selected eigenvector, and obtain the matrix Y after dimensionality reduction.

4. The ECG identity recognition method based on local fragment sparse representation according to claim 3, wherein The specific process of Step 3 is as follows: Step 3.1, use the orthogonal matching pursuit algorithm to find the optimal matching atoms in matrix Y in the sparse representation process for the matrix Y after dimensionality reduction; Step 3.2, construct the dictionary for sparse representation according to the optimal matching atoms by using the K singular value decomposition algorithm to obtain the processed sparse coefficient matrix.

5. The ECG identity recognition method based on local fragment sparse representation according to claim 4, characterized in that, The specific process of Step 3.2 is as follows: Step 3.2.1: Describe the optimal matching atoms obtained in Step 3.1 with linear expressions of basis elements to form a basis atom dictionary, initialize the dictionary, and use this dictionary to sparsely represent the given samples to obtain the corresponding matrix; Step 3.2.2: When the matrix obtained in Step 3.2.1 is sparse, iteratively update the dictionary column by column to continuously reduce the error between the sample data and the data represented by the coefficient matrix, approaching the optimal basis atom dictionary infinitely; Step 3.2.3: Update the basis atoms obtained in Step 3.2.2 column by column; Step 3.2.4, after the dictionary update in Step 3.2.3 is completed, perform sparse coding using the new dictionary, and stop updating the dictionary when the number of iterations reaches the specified value or the error rate reaches the specified range, then the processed sparse coefficient matrix can be obtained, that is, the reconstructed electrocardiogram signal Yˊ = D x , where Yˊ is the data signal matrix after dimensionality reduction to be processed, is the dictionary matrix, and x is the sparse coefficient obtained after processing .

6. The ECG identity recognition method based on local fragment sparse representation according to claim 5, characterized in that The specific process of Step 4 is as follows: Step 4.1: Input the final sparse coefficient matrix obtained in Step 3 into the input layer of the probabilistic neural network and then enter the hidden layer; Step 4.2: Input the data output from the hidden layer into the summation layer for calculation to obtain the maximum category calculated in the summation layer; Step 4.3: Normalize the maximum category data calculated in the summation layer to obtain the probability estimate for each category, that is, the recognition accuracy of the ECG signal.

Images