An electrocardio time-frequency feature map extraction method based on multi-scale wavelet transform

Abstract

The application discloses a kind of electrocardiogram time-frequency feature map extraction methods based on multiscale wavelet transform, comprising the following steps: removing the noise of electrocardiogram signal;Combining different kernel functions, electrocardiogram data is mapped to high-dimensional data space through nonlinear change, and the dominant feature of electrocardiogram information is extracted;The dominant feature of electrocardiogram is carried out multiscale wavelet transform, and multidimensional wavelet transform coefficient is obtained, according to the time of signal, wavelet transform coefficient, different wavelet transform scale, wavelet transform scale chart is drawn, and further form electrocardiogram time-frequency feature map that can be used for subsequent electrocardiogram signal classification identification.The present application directly extracts information dominant component of 12-lead electrocardiogram signal, and it is simple, convenient and easy to operate.The extracted time-frequency feature map can effectively reduce the irrelevant information of original multi-lead signal, and can reflect the multi-resolution frequency domain feature under the multiscale wavelet transform, realize the electrocardiogram signal time-frequency feature extraction under the multiscale wavelet transform.

Description

Technical Field

[0001] This invention belongs to the field of electrocardiogram (ECG) signal feature extraction technology, specifically relating to a method for extracting time-frequency feature maps of ECG signals under multi-scale wavelet transform. Background Technology

[0002] Cardiovascular diseases are common human ailments that seriously threaten people's lives and health. Therefore, the diagnosis and prevention of these diseases have become crucial issues facing modern medicine. Electrocardiogram (ECG) signals can not only be used to analyze and differentiate various diseases such as arrhythmias and myocardial infarction, but also reflect ventricular function and structure, making them an effective diagnostic tool for cardiovascular diseases.

[0003] Current solutions to the problem of time-frequency feature extraction from ECG signals mostly involve directly transforming the 12-lead ECG signal and fusing the results of all lead transformations as features, or only transforming lead II as features. While many valuable works have emerged, these methods still have the following problems: First, fusing features from all leads may incorporate information from irrelevant leads, thus affecting the discriminative nature of the extracted features. Second, using only lead II may ignore information from other leads, resulting in extracted features that do not fully represent the characteristics of the 12 leads. In fact, since the 12-lead ECG signal is essentially a detection of different parts of the body, directly extracting the dominant information components from the 12-lead ECG signal and performing multi-scale wavelet transforms to demonstrate its frequency domain characteristics at multiple resolutions would obviously greatly help in the realization of time-frequency extraction from multi-lead ECG signals. Summary of the Invention

[0004] The purpose of this invention is to overcome the problems existing in the prior art and provide a simpler and more convenient method for extracting time-frequency features of electrocardiogram signals under multi-scale wavelet transform.

[0005] This invention performs denoising on the original electrocardiogram (ECG) signal data; maps the denoised ECG signal to a high-dimensional space through nonlinear transformation using different kernel functions to extract the advantageous features of ECG information; performs multi-scale wavelet transform on the advantageous features of ECG information to construct a multi-dimensional wavelet transform scale map, and plots the time-frequency feature map of the ECG signal, thereby realizing the extraction of time-frequency features of ECG signal under multi-scale wavelet transform.

[0006] The specific technical solution of the present invention is achieved through the following steps:

[0007] Step 1: Remove noise from ECG signal data;

[0008] The original electrocardiogram (ECG) signal was filtered using mean filtering, median filtering, wavelet decomposition, and reconstruction to remove electromyographic interference, baseline drift, and power frequency interference, thus obtaining the filtered ECG signal.

[0009] The noise reduction process for the raw electrocardiogram signal is as follows:

[0010] 1-1. Eliminate electromyographic interference in electrocardiogram signal data using a mean filter;

[0011] 1-2. Baseline drift in ECG signal data is eliminated using a median filter;

[0012] 1-3. Using discrete wavelet decomposition with Daubechies8 as the wavelet basis, replace the wavelet decomposition coefficients with the following parameters by filtering out those with a threshold d less than a specified value:

[0013]

[0014] Where x represents the electrocardiogram signal data for each lead, x i For each lead, there are ECG signal data sample points, with d = 0.04; then, wavelet decomposition coefficients are reconstructed using wavelet to obtain the data after eliminating power frequency interference.

[0015] Step 2: Combine different kernel functions to map ECG data to a high-dimensional data space through nonlinear transformation, and extract the advantageous features of ECG information;

[0016] The process of extracting dominant features from electrocardiogram information is as follows:

[0017] 2-1. Based on the ECG signal data obtained after denoising in step 1, construct a data matrix x:

[0018]

[0019] Where, x i =(x i1 ,x i2 ,…,x in ), i = 1, 2, ..., n, represent the column vectors of the data matrix X, and n represents the length of the electrocardiogram signal data.

[0020] 2-2. Introducing a kernel function φ to project the data matrix X into a higher-dimensional space. From the original 12-dimensional feature space to the M-dimensional feature space:

[0021] φ:R 12 →R M M>>12

[0022] Calculate the kernel matrix K based on the kernel function φ:

[0023]

[0024] Where x represents a column vector of the ECG data matrix X, and the kernel function includes the polynomial kernel function and the Cosine kernel function;

[0025] Polynomial kernel function:

[0026] k(x,y)=(px T y+c) h

[0027] Where p, c, and h are all adjustable parameters, p>0, c≥0;

[0028] Cosine kernel function:

[0029]

[0030] 2-3. Center the kernel matrix K:

[0031]

[0032]

[0033] Where A is a 12×12 matrix after centering;

[0034] The eigenvalue λ is calculated using the following formula:

[0035]

[0036]

[0037] Where E is the same as Identity matrices of the same order;

[0038] Using the obtained eigenvalues ​​λ, calculate the eigenvector V according to the following formula;

[0039] Each obtained eigenvalue λ i Substitute into the following formula to solve for the eigenvector V corresponding to each eigenvalue. i :

[0040]

[0041] Where i represents the number of features of the electrocardiogram signal data, i = 1, 2, ..., 12.

[0042] 2-4. Calculate the cumulative contribution rate based on the eigenvalues.

[0043]

[0044] in The eigenvalues ​​representing the top k principal components are used as the contribution rate to the proportion of the original electrocardiogram signal data.

[0045] Under different kernel functions, the top three principal components with a cumulative contribution rate of over 90% are selected as dominant features, i.e., ECG dominance features.

[0046] Step 3: Perform multi-scale wavelet transform on the dominant ECG features to obtain multi-dimensional wavelet transform coefficients. Based on the signal time, wavelet transform coefficients, and different wavelet transform scales, draw a wavelet transform scale map to further form an ECG time-frequency feature map that can be used for subsequent ECG signal classification and recognition.

[0047] Based on the advantageous features of the ECG information obtained in step 2, wavelet transform coefficients are calculated for each dimension. Combining the scale of the wavelet transform with the time of the signal, a wavelet transform scale map is plotted. By image stitching, the wavelet transform scale maps of each dimension of data are stitched together to achieve extraction of time-frequency features of ECG signals under multi-scale wavelet transform.

[0048] The process of extracting time-frequency features using wavelet transform is as follows:

[0049] 3-1. Selecting the wavelet mother function: Based on the waveform of the electrocardiogram signal data, the Mexicanhat function is selected as the wavelet mother function ψ(x):

[0050]

[0051] Then, by scaling and translating the wavelet mother function, we obtain the wavelet sub-function ψ. a,b (t):

[0052]

[0053] Where a represents the scale parameter of the wavelet transform, and b represents the time shift parameter;

[0054] Employing multi-scale wavelet transform, where 'a' is a one-dimensional matrix:

[0055] a = [a0, a1, ..., a m-1 ] = [1,2,…,m]

[0056] Where m is a positive integer greater than 1;

[0057] subwave function ψ a,b (t) can be represented as:

[0058]

[0059] 3-2. Calculate the wavelet transform coefficients. Expand each dimension of the ECG dominant feature data x(t) under the wavelet basis. This expansion is the wavelet transform coefficient matrix C(b,a) of the ECG signal x(t):

[0060]

[0061] in <x(t),ψ a,b (t)> represents the electrocardiogram signal x(t) and the wavelet function ψ a,b The inner product of (t), where n represents the length of the electrocardiogram signal.

[0062] 3-3. Process the wavelet transform results. After wavelet transform, obtain the wavelet transform coefficient matrix C(b,a) and the corresponding frequency f. Convert the frequency f into the scale s:

[0063] s = 1 / f.

[0064] 3-4. Draw a contour map. Construct a three-dimensional contour map using the following XYZ coordinates to obtain a wavelet transform scaling map:

[0065] X = t

[0066] Y = log2s

[0067] Z = log2|C(b,a)| 2 .

[0068] Based on whether the extracted ECG time-frequency feature map can display correlation characteristics, the kernel function selection strategy is as follows:

[0069] If the contribution rate of the first principal component is less than 90% and less than 80%, choose the cosine kernel function;

[0070] If 90% < the contribution rate of the first principal component < 100%, choose the polynomial kernel function;

[0071] By stitching together the wavelet transform scale maps of each dimension, the time-frequency feature maps of ECG signals under multi-scale wavelet transform can be extracted.

[0072] If the contribution rate of the first principal component is greater than 80% and less than 90%, then the contribution rates of the second and third principal components to the data are relatively high, the data correlation is relatively poor, the absolute value of the correlation coefficient is small, and the correlation characteristics between the extracted ECG time-frequency feature maps are not obvious. If the contribution rate of the first principal component is greater than 90%, then the contribution rates of the second and third principal components to the data are relatively low, the data correlation is relatively strong, the absolute value of the correlation coefficient is large, and the correlation characteristics between the extracted ECG time-frequency feature maps are obvious.

[0073] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0074] 1. Based on the denoising process of electrocardiogram (ECG) signals, this invention combines different kernel functions to map ECG signals to a high-dimensional data space through nonlinear transformation, and extracts the advantageous features of ECG information by calculating the information ratio.

[0075] 2. Converting time-domain signals into time-frequency features can reflect the similarity and correlation of signal waveforms at different times, and also reflect the multi-resolution frequency domain characteristics at different scales. By stitching images together the wavelet transform scale maps of each dimension after dimensionality reduction, the time-frequency characteristics of the 12-lead ECG signal can be characterized. Attached Figure Description

[0076] Figure 1 This is a flowchart of an electrocardiogram time-frequency feature map based on multi-scale wavelet transform proposed in this invention.

[0077] Figure 2(a) is a schematic diagram of a lead in the original electrocardiogram signal in the embodiment.

[0078] Figure 2(b) is a schematic diagram of the ECG signal after filtering in the embodiment.

[0079] Figure 3 This is a three-dimensional schematic diagram of the advantageous features of electrocardiogram information in the embodiment.

[0080] Figure 4 This is a continuous wavelet transform scaling diagram of a single-dimensional signal in the embodiment.

[0081] Figure 5(a) is a correlation coefficient graph of the dominant features of electrocardiogram information using the cosine kernel function in the embodiment.

[0082] Figure 5(b) is a correlation coefficient diagram of the dominant features of electrocardiogram information using a polynomial kernel function in the embodiment.

[0083] Figure 6(a) is a time-frequency feature diagram of ECG information using the cosine kernel function in the embodiment.

[0084] Figure 6(b) is a time-frequency feature diagram of ECG information dominance features using a polynomial kernel function in the embodiment.

[0085] Figure 7(a) is a time-frequency feature diagram of ECG information at different scales using the cosine kernel function in the embodiment.

[0086] Figure 7(b) is a time-frequency feature diagram of ECG information at different scales using polynomial kernel functions in the embodiment. Detailed Implementation

[0087] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0088] Example

[0089] like Figure 1 As shown, a method for extracting ECG time-frequency feature maps based on multi-scale wavelet transform includes the following steps:

[0090] Step 1: Remove noise from ECG signal data;

[0091] The original electrocardiogram (ECG) signal was filtered using mean filtering, median filtering, wavelet decomposition, and reconstruction to remove electromyographic interference, baseline drift, and power frequency interference, thus obtaining the filtered ECG signal.

[0092] The noise reduction process for the raw electrocardiogram signal is as follows:

[0093] 1-1. Eliminate electromyographic interference in electrocardiogram signal data using a mean filter;

[0094] 1-2. Baseline drift in ECG signal data is eliminated using a median filter;

[0095] 1-3. Using discrete wavelet decomposition with Daubechies8 as the wavelet basis, replace the wavelet decomposition coefficients with the following parameters by filtering out those with a threshold d less than a specified value:

[0096]

[0097] Where x represents the electrocardiogram signal data for each lead, x i For each lead, there are ECG signal data sample points, with d = 0.04; then, wavelet decomposition coefficients are reconstructed using wavelet to obtain the data after eliminating power frequency interference.

[0098] like Figure 2a and Figure 2b The image shows a comparison of ECG signals before and after preprocessing.

[0099] Step 2: Combine different kernel functions to map ECG data to a high-dimensional data space through nonlinear transformation, and extract the advantageous features of ECG information;

[0100] The process of extracting dominant features from electrocardiogram (ECG) information is as follows:

[0101] 2-1. Based on the denoised ECG signal data obtained in step 1, construct a data matrix X:

[0102]

[0103] Where, x i =(x i1 ,x i2 ,…,x i12 ) T ,i=1,2,…n, represent the column vectors of the data matrix X, and n represents the length of the electrocardiogram signal data.

[0104] 2-2. Introducing a kernel function φ to project the data matrix X into a higher-dimensional space. From the original 12-dimensional feature space to the M-dimensional feature space:

[0105] φ:T 12 →R M M>>12

[0106] Calculate the kernel matrix K based on the kernel function φ:

[0107]

[0108] Where x represents a column vector of the ECG data matrix X, and the kernel function includes the polynomial kernel function and the Cosine kernel function;

[0109] Polynomial kernel function:

[0110] k(x,y)=(px T y+c) h

[0111] Where p, c, and h are all adjustable parameters, p>0, c≥0;

[0112] Cosine kernel function:

[0113]

[0114] 2-3. Center the kernel matrix K:

[0115]

[0116]

[0117] Where A is a 12×12 matrix after centering;

[0118] The eigenvalue λ is calculated using the following formula:

[0119]

[0120]

[0121] Where E is the same as Identity matrices of the same order;

[0122] Using the obtained eigenvalues ​​λ, calculate the eigenvector V according to the following formula;

[0123] Each obtained eigenvalue λ i Substitute into the following formula to solve for the eigenvector V corresponding to each eigenvalue. i :

[0124]

[0125] Where i represents the number of features of the electrocardiogram signal data, i = 1, 2, ..., 12.

[0126] 2-4. Calculate the cumulative contribution rate based on the eigenvalues.

[0127]

[0128] in The eigenvalues ​​representing the top k principal components are used as the contribution rate to the proportion of the original electrocardiogram signal data.

[0129] Under different kernel functions, the top three principal components with a cumulative contribution rate of over 90% are selected as dominant features, i.e., ECG dominance features.

[0130] Table 1 shows the contribution rates and cumulative contribution rates of the first three principal components under different kernel functions.

[0131] Table 1 Feature contribution rate and contribution rate

[0132]

[0133] According to Table 1, this invention selects the cosine kernel function and the polynomial kernel function for time-frequency feature extraction of multi-scale wavelet transform.

[0134] The dominant features of the obtained electrocardiogram information, such as Figure 3 The image shown is a three-dimensional schematic diagram illustrating the advantageous features of electrocardiogram (ECG) information.

[0135] Step 3: Perform multi-scale wavelet transform on the dominant ECG features to obtain multi-dimensional wavelet transform coefficients. Based on the signal time, wavelet transform coefficients, and different wavelet transform scales, plot the wavelet transform scale diagram, as shown below. Figure 4 As shown, an ECG time-frequency feature map is further formed that can be used for subsequent ECG signal classification and recognition;

[0136] Based on the advantageous features of the ECG information obtained in step 2, wavelet transform coefficients are calculated for each dimension. Combining the scale of the wavelet transform with the time of the signal, a wavelet transform scale map is plotted. By image stitching, the wavelet transform scale maps of each dimension of data are stitched together to achieve extraction of time-frequency features of ECG signals under multi-scale wavelet transform.

[0137] The process of extracting time-frequency features using wavelet transform is as follows:

[0138] 3-1. Selecting the wavelet mother function: Based on the waveform of the electrocardiogram signal data, the Mexicanhat function is selected as the wavelet mother function ψ(x):

[0139]

[0140] Then, by scaling and translating the wavelet mother function, we obtain the wavelet sub-function ψ. a,b (t):

[0141]

[0142] Where a represents the scale parameter of the wavelet transform, and b represents the time shift parameter;

[0143] Employing multi-scale wavelet transform, where 'a' is a one-dimensional matrix:

[0144] a = [a0, a1, ..., a m-1 ] = [1,2,…,m]

[0145] Where m is a positive integer greater than 1;

[0146] subwave function ψ a,b (t) can be represented as:

[0147]

[0148] 3-2. Calculate the wavelet transform coefficients. Expand each dimension of the ECG dominant feature data x(t) under the wavelet basis. This expansion is the wavelet transform coefficient matrix C(b,a) of the ECG signal x(t):

[0149]

[0150] in <x(t),ψ a,b (t)> represents the electrocardiogram signal x(t) and the wavelet function ψ a,b The inner product of (t), where n represents the length of the electrocardiogram signal.

[0151] 3-3. Process the wavelet transform results. After wavelet transform, obtain the wavelet transform coefficient matrix C(b,a) and the corresponding frequency f. Convert the frequency f into the scale s:

[0152] s = 1 / f.

[0153] 3-4. Draw a contour map. Construct a three-dimensional contour map using the following XYZ coordinates to obtain a wavelet transform scaling map:

[0154] X = t

[0155] Y = log2s

[0156] Z = log2|C(b,a)| 2 .

[0157] Based on whether the extracted ECG time-frequency feature map can display correlation characteristics, the kernel function selection strategy is as follows:

[0158] If the contribution rate of the first principal component is less than 90% and less than 80%, choose the cosine kernel function;

[0159] If 90% < the contribution rate of the first principal component < 100%, choose the polynomial kernel function;

[0160] By stitching together the wavelet transform scale maps of each dimension, the time-frequency feature maps of ECG signals under multi-scale wavelet transform can be extracted.

[0161] If the contribution rate of the first principal component is greater than 80% and less than 90%, then the contribution rates of the second and third principal components to the data are relatively high, the data correlation is relatively poor, the absolute value of the correlation coefficient is small, and the correlation characteristics between the extracted ECG time-frequency feature maps are not obvious. If the contribution rate of the first principal component is greater than 90%, then the contribution rates of the second and third principal components to the data are relatively low, the data correlation is relatively strong, the absolute value of the correlation coefficient is large, and the correlation characteristics between the extracted ECG time-frequency feature maps are obvious.

[0162] The correlation analysis results of the dominant features of the obtained electrocardiogram information in various dimensions, such as Figure 5a and Figure 5b The figure shows the correlation coefficients of the dominant features of electrocardiogram information for different kernel functions.

[0163] The wavelet transform results of the dominant features of the obtained electrocardiogram information, such as Figure 6a and Figure 6b As shown, this is a time-frequency feature diagram of the ECG information dominance characteristics of different kernel functions.

[0164] By setting m=65 and m=200 respectively, time-frequency feature maps at different scales are obtained. The three maps on the left are for m=65, and the three maps on the right are for m=200. Figure 7a and Figure 7b As shown, this is a time-frequency feature map of ECG information at different scales and with different kernel functions.

[0165] This example uses the PTB-XL ECG dataset. This dataset contains 21,837 clinical 12-lead ECG records, each 10 seconds long, including 9,517 normal ECGs and 12,320 abnormal ECGs. Using the method of this invention, time-frequency feature maps of the ECG dataset are extracted, and then a general classification algorithm is applied to effectively classify the positive and abnormal ECGs, achieving an accuracy rate of over 90%.

[0166] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A method for extracting electrocardiogram time-frequency feature map based on multi-scale wavelet transform, characterized in that, The method comprises the following steps: Step 1, removing noise of electrocardio signal data; Step 2, mapping electrocardio signal data to high-dimensional data space through nonlinear change by combining different kernel functions, and extracting electrocardio dominant features; Step 3, performing multi-scale wavelet transform on the electrocardio dominant features to obtain multi-dimensional wavelet transform coefficients, and drawing a wavelet transform scale diagram according to the time of the signal, the wavelet transform coefficients, and different wavelet transform scales, to further form an electrocardio time-frequency feature diagram which can be used for subsequent electrocardio signal classification and recognition; The removing noise of electrocardio signal data comprises the following steps: 1-1. eliminating electromyographic interference in the electrocardio signal data through a mean filter; 1-2. eliminating baseline drift in the electrocardio signal data through a median filter; 1-3. replacing wavelet decomposition coefficients less than a specified value by screening a threshold d through discrete wavelet decomposition with a Daubechies 8 wavelet basis; Wherein, x is the electrocardiosignal data of each lead, x i is the electrocardiosignal data sample point of each lead, d is 0.04; then the wavelet reconstruction is performed on the wavelet decomposition coefficients to obtain the data after the power interference is eliminated; The extracting electrocardio dominant features of step 2 comprises the following steps: 2-1. constructing a data matrix X according to the electrocardio signal data obtained after denoising in step 1: wherein x i = (x i1 , x i2 ,..., x in ), i = 1, 2,... n, represents a column vector of the data matrix X, and n represents the length of the electrocardiosignal data; The extracting electrocardio dominant features further comprises the following steps: 2-2. Introducing a kernel function φ to project the data matrix X into a high-dimensional space From the original 12-dimensional feature space to the M-dimensional feature space: φ: R 12 → R M M > 12 calculating a kernel matrix K according to a kernel function φ: where x represents a column vector of the electrocardio data matrix X, and the kernel function includes a polynomial kernel function and a Cosine kernel function; the polynomial kernel function is: k(x,y) = (px T y+c) h where p, c, and h are adjustable parameters, and p>0 and c≥0; the Cosine kernel function is: 2.The method of claim 1, wherein The extracting electrocardio dominant features further comprises the following steps: 2-3. performing centralization processing on the kernel matrix K: wherein A is a 12x12 matrix after centering processing, is a kernel matrix after centering processing; calculating eigenvalues λ by using the following formula: where E is the kernel matrix after the centering process an identity matrix of the same order; the eigenvectors V are calculated using the eigenvalues λ found and according to the following formula: Each eigenvalue λ i is substituted into the following equation to solve for each eigenvalue λ i The corresponding eigenvector V i : where i represents the number of characteristics of the electrocardio signal data, and i=1, 2, …, 12. 3.The method of claim 2, wherein The extracting electrocardio dominant features further comprises the following steps: 2-4. Calculate the cumulative contribution rate according to the eigenvalue wherein, The eigenvalues representing the first k principal components are taken as the contribution rate of the original electrocardio signal data proportion; the first three main components with cumulative contribution rate reaching more than 90% under different kernel functions are selected as the dominant features, i.e. electrocardio dominant features.

4. The electrocardiogram time-frequency feature map extraction method based on multi-scale wavelet transform according to claim 3, characterized in that Step 3 of drawing the electrocardio signal time-frequency feature diagram comprises the following steps: 3-1. selecting a wavelet mother function, and selecting a Mexican hat function as the wavelet mother function ψ(x) according to the waveform of the electrocardio signal data: The wavelet function ψ is obtained by scaling and shifting the mother wavelet function φ a,b (t): where a represents a scale parameter of wavelet transform, and b represents a time translation parameter; a multi-scale wavelet transform is adopted, that is, a is a one-dimensional matrix: a = [a0, a1,..., am]T m-1 ] = [1, 2,..., m] where m is a positive integer greater than 1; wavelet function ψ a,b (t) is represented as:

5. The method of claim 4, wherein the method is based on a multi-scale wavelet transform. The drawing of the electrocardio signal time-frequency feature diagram further comprises the following steps: 3-2. calculating wavelet transform coefficients, and expanding each-dimensional data x(t) of the electrocardio dominant features under a wavelet basis, which is a wavelet transform coefficient matrix C(b, a) of the electrocardio signal x(t): where <x(t), ψ a,b (t) represents the inner product of the ECG signal x(t) and the wavelet function ψ a,b (t), and n represents the length of the ECG signal. 6.The method of extracting electrocardio time-frequency feature map based on multi-scale wavelet transform according to claim 5, characterized in that The drawing of the electrocardio signal time-frequency feature diagram further comprises the following steps: 3-3. processing the wavelet transform result, obtaining a wavelet transform coefficient matrix C(b, a) and a corresponding frequency f through the wavelet transform, and converting the frequency f into a scale s: s=1 / f; The drawing of the electrocardio signal time-frequency feature diagram further comprises the following steps: 3-4. drawing a contour map, and constructing a three-dimensional contour map with the following XYZ: X=t Y=log2s Z = log2|C(b,a)| 2 .

7. The electrocardiogram time-frequency feature map extraction method based on multi-scale wavelet transform according to claim 6, characterized in that The drawing of the electrocardio signal time-frequency feature diagram further comprises the following steps: According to the selection of whether the electrocardio signal time-frequency feature diagram to be extracted can display correlation characteristics, the kernel function selection strategy is as follows: if 80%<first principal component contribution rate<90%, select cosine kernel function; if 90%<first principal component contribution rate<100%, select polynomial kernel function; Through image splicing, the wavelet transform scale diagram of each dimension is spliced together to realize the extraction of the ECG time-frequency feature diagram under the multi-scale wavelet transform. If the contribution rate of the first principal component is greater than 80% and less than 90%, then the contribution rate of the second and third principal components to the data is relatively high, the data correlation is relatively poor, the absolute value of the correlation coefficient is small, and the correlation characteristics between the extracted ECG time-frequency feature diagrams are not obvious. If the contribution rate of the first principal component is greater than 90%, then the contribution rate of the second and third principal components to the data is relatively low, the data correlation is relatively strong, the absolute value of the correlation coefficient is large, and the correlation characteristics between the extracted ECG time-frequency feature diagrams are obvious.

Images