A method for ECG pattern fusion recognition taking into account the degree of neural network excitation

Abstract

The present invention discloses an ECG pattern fusion recognition method taking into account the degree of neural network excitation, comprising the following steps: obtaining 12-lead ECG signals on the body surface, constructing a lattice neural network to extract deep nonlinear electrocardiodynamic features under the ECG signals; calculating the distance metric based on the nonlinear dynamics difference between different nonlinear electrocardiodynamic features, and determining the set of ECG patterns to be fused according to the distance metric value; assigning different weights to the ECG patterns to be fused according to the different degrees of neural network excitation between the ECG patterns to be fused, and storing the fusion results in the form of a constant matrix. The ECG pattern fusion recognition method proposed in the present invention can better represent different ECG patterns and improve the ability to distinguish ECG signals. At the same time, the lattice neural network recognition technology is used for approximation and identification, and the extracted nonlinear electrocardiodynamic features can well reflect the intrinsic characteristics of the original signal.

Description

Technical Field

[0001] The invention belongs to the technical field of pattern recognition, and in particular relates to an electrocardiogram pattern fusion recognition method taking into account the degree of neural network excitation. Background Art

[0002] ECG pattern fusion technology mainly aims at effectively fusing the intrinsic features of different ECG signals, and the fused ECG patterns need to have good discrimination capabilities. At present, ECG pattern fusion has made a lot of progress, but most of them are based on static images and the fusion of static features. Since the features of static images are very limited, they cannot well describe the intrinsic time characteristics of ECG signals. Time characteristics are an important characteristic of signals. Each time point contains rich relationship information with the previous and next time points. In response to the above problems, one solution is to extract the nonlinear dynamic features of ECG signals and establish an ECG pattern fusion framework with good discrimination capabilities. However, how to extract nonlinear dynamic features that can well characterize ECG signals is a difficult problem, and how to fuse nonlinear dynamic features into ECG patterns is also a key issue. Summary of the invention

[0003] The purpose of the present invention is to overcome the above-mentioned problems and to provide a more concise and accurate ECG pattern fusion recognition method based on the degree of neural network excitation by extracting the nonlinear dynamic characteristics of ECG signals.

[0004] To solve the above problems, the specific solution of the present invention is achieved by the following steps:

[0005] Step 1: Obtain 12-lead ECG signals from the body surface and construct a lattice neural network to extract the deep nonlinear electrocardiographic dynamics characteristics of the ECG signals;

[0006] Step 2: Calculate the distance metric between different nonlinear electrocardiographic dynamic features based on the nonlinear dynamics difference, and determine the electrocardiographic pattern set to be fused according to the distance metric value;

[0007] Step 3: According to the different degrees of neural network excitation between the ECG patterns to be fused, different weights are assigned to perform weighted fusion on the ECG patterns to be fused, and the fusion results are stored in the form of a constant matrix.

[0008] Preferably, the step 1 specifically includes the following process:

[0009] 1-1. Perform linear transformation on the 12-lead ECG signal to complete data dimension reduction, and filter out noise and artifacts on the signal after dimension reduction;

[0010] 1-2. Perform deep decomposition of the signal, extract key parameters based on signal components, and use the extracted key parameters as nonlinear ECG dynamic parameters;

[0011] 1-3. Carry out nonlinear electrocardiodynamic modeling, construct a lattice neural network based on the nonlinear electrocardiodynamic model, input nonlinear electrocardiodynamic parameters into the lattice neural network, and store the nonlinear electrocardiodynamic characteristics in the form of a constant neural network weight matrix through the learning of the lattice neural network.

[0012] Preferably, the data dimension reduction is specifically calculated as follows:

[0013]

[0014] Among them, v1, v2, v3 represent the signals of each dimension after data dimensionality reduction, x1, x2, x3, x4, x5, x6, x7, x8, x9, x 10 ,x 11 ,

[0015] x 12 They represent the I lead, II lead, V1 lead, V2 lead, V3 lead, V4 lead, V5 lead, V6 lead, III lead, avF lead, avL lead, and avR lead signals in the ECG signal respectively.

[0016] Preferably, the key parameter extraction specifically includes the following steps:

[0017] (1) Given an original input signal V t =[v1,v2,v3] T , assuming V t The local extreme point of k ,k=1,2,3,…}, define τ 0＝0 , define a baseline extraction factor L, L in [0,τ k ] The definition of the moment can be expressed as follows:

[0018]

[0019] Among them, v it Represents the signal v of the i-th dimension i Take the value at point t, L k Indicates that in [0,τ k ] The definition of the baseline extraction factor at the moment L. The next moment L k+1 Satisfy the following formula:

[0020]

[0021] Among them, a represents the gain parameter, the value range is between (0,1), and it represents the amplitude of the component obtained by this decomposition;

[0022] (2) Input signal Vt Perform deep decomposition and obtain the baseline extraction factor L according to step (1). t decomposed into a baseline L and a component H t The sum form, each decomposition can be expressed as follows:

[0023] V t =Lv it +(1-L)v it =L t +H t

[0024] The input signal of this decomposition comes from the baseline signal L obtained from the previous decomposition. t , and the component H t As the obtained signal component, it contains information such as the instantaneous frequency of the signal;

[0025] (3) Continue with step (2). When the signal component H t When a monotonic trend is shown, the decomposition is terminated, and the overall decomposition process satisfies the following formula:

[0026]

[0027] Where p represents the number of decompositions.

[0028] (4) Perform Pearson correlation analysis on all decomposed signal components and the original input signal in turn:

[0029]

[0030] Among them, r t It means that the component H obtained after the signal is decomposed t times is different from the input signal V before decomposition. t The coefficient obtained by Pearson correlation analysis, H represents the number of signal components, n represents the total length of the signal components, and the Pearson correlation coefficient r is taken t The largest signal component is used as the key parameter to obtain the nonlinear ECG dynamic parameters.

[0031] Preferably, the nonlinear electrocardiodynamic modeling specifically includes the following steps:

[0032] (1) Modeling nonlinear electrocardiographic dynamics:

[0033]

[0034] Where H = [H1, H2, H3] Tis the nonlinear electrocardiodynamic parameter, z is the model parameter; F(H; z) is the main part of the nonlinear dynamics, l(H; z) represents the uncertainty of the model; F(H; z) and l(H; z) together constitute a complete nonlinear electrocardiodynamic model. Since l(H; z) is inseparable in some cases, φ(H; z): = F(H; z) + l(H; z) is used to represent the complete nonlinear electrocardiodynamic model.

[0035] (2) The model is learned using a lattice neural network, which is expressed as follows:

[0036]

[0037] in, Represents the state variable of the lattice neural network; A is a diagonal matrix whose diagonal elements are gain parameters, with values ​​between [-1,1]; It is the weight learned by the neural network according to the parameter trajectory at different distances of neurons. The specific calculation formula is: in, is the state error, Γ i , σ i They are all network adjustment parameters; S(H) is the Gaussian radial basis function, which mainly includes the number and distribution information of neurons.

[0038] (3) Extract the deep nonlinear electrocardiographic dynamics characteristics under the ECG signal. The main calculation formula is as follows:

[0039]

[0040] Among them, φ(H;z) is the complete nonlinear electrocardiographic information, yes After learning, it will gradually tend to the convergent mean of the convergence trend, It is the nonlinear electrocardiographic dynamics feature learned by each neuron from the local nonlinear electrocardiographic parameter trajectory in the lattice neural network, and ∈ represents the learning error.

[0041] Preferably, the step 2 specifically includes the following steps:

[0042] 2-1. Calculate the distance metric between different nonlinear electrocardiographic dynamic features based on the difference in nonlinear dynamics;

[0043] 2-2. Determine the set of ECG patterns to be fused based on the distance metric value.

[0044] Preferably, the distance metric is calculated as follows:

[0045]

[0046] Among them, ε is the comprehensive identification error, ε i represents the error in the i-th dimension between two nonlinear electrocardiodynamic features, where the error in the i-th dimension ε i The system satisfies the following equation:

[0047]

[0048] Among them, b i is the gain parameter, f i '(H',z') represents the feature of the i-th dimension of nonlinear electrocardiodynamics; given a feature of the i-th dimension of nonlinear electrocardiodynamics Based on the above error system, we can obtain With another nonlinear dynamics i-th dimension feature f i The error of '(H',z')ε i .

[0049] Preferably, the set of ECG patterns to be fused specifically includes the following steps:

[0050] (1) Take any two nonlinear electrocardiodynamic features, denoted as P and Q, and calculate the comprehensive identification error ε between P and Q;

[0051] (2) Based on the comprehensive recognition error ε, it is determined whether the ECG patterns between PQ can be fused. The specific determination method is:

[0052] When the comprehensive identification error ε takes a value of [0,0.5], it is determined that the two sets of nonlinear electrocardiodynamic features to be compared are relatively close, and are set to the waiting fusion mode, which is recorded as the waiting fusion set C1(P,Q);

[0053] When the comprehensive identification error ε is between (0.5,∞), it is considered that the two groups of nonlinear electrocardiographic dynamic characteristics to be compared are different and do not participate in the fusion of electrocardiographic patterns;

[0054] (3) Take the nonlinear dynamic feature P and any nonlinear electrocardiodynamic feature M that has not been selected in this round, and repeat steps (1) and (2). When the comprehensive error ε of P and M is between [0, 0.5], update the set C1(P, …, M), otherwise do not update. This process continues until P and all nonlinear dynamic features have completed the fusion judgment of the electrocardiodynamic pattern. At this time, the pattern to be fused is C1(P, …);

[0055] (4) Select any two nonlinear dynamic features R, S that do not appear in the pattern to be fused C1 (P, ...), repeat steps (1), (2), and (3), and record the pattern to be fused as C2 (R, ...), until all nonlinear electrocardiographic dynamic features are in the pattern to be fused C1 (P, ...). i End the judgment.

[0056] Preferably, the step three specifically includes the following steps:

[0057] 3-1. Calculate the degree of neural network excitation between ECG patterns to be fused;

[0058] 3-2. Perform numerical normalization based on the degree of excitation, use the normalized excitation value as the weight to perform weighted fusion on the ECG patterns to be fused, and store the fusion results in the form of a constant matrix.

[0059] Preferably, the degree of neural network excitation between the ECG patterns to be fused is specifically calculated as follows:

[0060]

[0061] Among them, S is the excitation counter, which is used to indicate the excitation degree of the neural network weight matrix; W ijk It represents the weight of the i-th layer, j-th row, and k-th column in the m×n×l constant-valued neural network weight space matrix; T is the excitation threshold.

[0062] Preferably, the weighted fusion of the ECG patterns to be fused specifically comprises the following steps:

[0063] (1) Select the nonlinear electrocardiodynamic feature set C in the mode to be fused i (c1,c2,…,c n ), respectively calculate the excitation degree S1, S2, ..., S of the constant value neural network weight space matrix n , and normalize them numerically, denoted as Q1, Q2, …, Q n ;

[0064] (2) Treatment of fused ECG pattern C i Perform weighted fusion:

[0065] y i =Q1c i1 +Q2c i2 +…+Q n c in

[0066] Among them, y i Represents the set to be fused C i The nonlinear electrocardiographic dynamic characteristics of the i-th dimension after fusion; Q n Represents the set to be fused C i The excitation degree of the nth nonlinear electrocardiodynamic characteristic value after normalization; c in Represents the set to be fused C i The feature of the i-th dimension of the n-th nonlinear electrocardiodynamic feature; the fusion result y is stored in the form of a constant matrix.

[0067] 1. The present invention adopts 12-lead ECG signals, reduces the dimensionality of the data, and completes a deeper signal decomposition of the reduced dimensionality signal. The extracted signal components can efficiently reflect the intrinsic time characteristics of the ECG signals, fully consider the intrinsic information of the signals, and improve the discrimination ability of subsequent mode fusion.

[0068] 2. The present invention uses lattice neural network recognition technology for approximation and identification. The extracted nonlinear electrocardiographic dynamics characteristics can well reflect the intrinsic characteristics of the original signal. The fusion of ECG patterns is completed by determining the recognition error and calculating the degree of neural network excitation. The fused ECG patterns can make the recognition more versatile and robust. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 The present invention proposes a flow chart of an electrocardiogram pattern fusion recognition method taking into account the degree of neural network excitation.

[0070] Figure 2 It is a schematic diagram of signal components obtained by decomposing one dimension of the signal after dimensionality reduction in the embodiment.

[0071] Figure 3 Schematic diagram of the convergence of the neural network weights in the embodiment.

[0072] Figure 4 Schematic diagram of nonlinear electrocardiographic dynamics characteristics obtained in a lattice neural network in an embodiment.

[0073] Figure 5 Schematic diagram of recognition error obtained by calculating two nonlinear electrocardiographic dynamic characteristics in the embodiment.

[0074] Figure 6 Schematic diagram of the degree of excitation of the nonlinear electrocardiodynamic characteristic trajectory in the weight matrix in the embodiment. DETAILED DESCRIPTION

[0075] The present invention is further described in detail below in conjunction with embodiments and drawings, but the embodiments of the present invention are not limited thereto.

[0076] Example

[0077] like Figure 1 As shown, a method for fusion recognition of electrocardiogram patterns taking into account the degree of neural network excitation comprises the following steps:

[0078] Step 1: Obtain 12-lead ECG signals from the body surface and construct a lattice neural network to extract the deep nonlinear electrocardiodynamic characteristics of the ECG signals. The specific steps include:

[0079] 1-1. Perform linear transformation on the 12-lead ECG signal to complete data dimension reduction, and filter out noise and artifacts from the reduced-dimensional signal. The specific calculation method for data dimension reduction is as follows:

[0080]

[0081] Among them, v1, v2, v3 represent the signals of each dimension after data dimensionality reduction, x1, x2, x3, x4, x5, x6, x7, x8, x9, x 10 ,

[0082] x 11 ,x 12 They represent the I lead, II lead, V1 lead, V2 lead, V3 lead, V4 lead, V5 lead, V6 lead, III lead, avF lead, avL lead, and avR lead signals in the ECG signal respectively.

[0083] 1-2. Perform deep decomposition of the signal and extract key parameters based on the signal components. The extracted key parameters are used as nonlinear ECG dynamic parameters. The specific process is as follows:

[0084] (1) Given an original input signal V t =[v1,v2,v3] T , assuming V t The local extreme point of k ,k=1,2,3,…}, define τ 0＝0 , define a baseline extraction factor L, L in [0,τ k ] The definition of the moment can be expressed as follows:

[0085]

[0086] Among them, v it Represents the signal v of the i-th dimension i Take the value at point t, L k Indicates that in [0,τ k ] The definition of the baseline extraction factor at the moment L. The next moment L k+1 Satisfy the following formula:

[0087]

[0088] Among them, α represents the gain parameter, the value range is between (0,1), and it represents the amplitude of the component obtained by this decomposition;

[0089] (2) Input signal V t Perform deep decomposition and obtain the baseline extraction factor L according to step (1). t The baseline L tand H t The sum form, each decomposition can be expressed as follows:

[0090] V t =Lv it +(1-L)v it =L t +H t

[0091] The input signal of this decomposition comes from the baseline signal L obtained from the previous decomposition. t , and the component H t As the obtained signal component, it contains information such as the instantaneous frequency of the signal;

[0092] (3) Continue with step (2). When the signal component H t When a monotonic trend is shown, the decomposition is terminated, and the overall decomposition process satisfies the following formula:

[0093]

[0094] Where p represents the number of decompositions.

[0095] (4) Perform Pearson correlation analysis on all decomposed signal components and the original input signal in turn:

[0096]

[0097] Among them, r t It means that the component H obtained after the signal is decomposed t times is different from the input signal V before decomposition. t The coefficient obtained by Pearson correlation analysis, N represents the number of signal components, n represents the total length of the signal components, and the Pearson correlation coefficient r is taken t The largest signal component is used as the key parameter to obtain the nonlinear ECG dynamic parameters.

[0098] 1-3. Perform nonlinear electrocardiodynamic modeling, construct a lattice neural network based on the nonlinear electrocardiodynamic model, input nonlinear electrocardiodynamic parameters into the lattice neural network, and store nonlinear electrocardiodynamic characteristics in the form of a constant neural network weight matrix through the learning of the lattice neural network. Specifically, the following steps are included:

[0099] (1) Modeling nonlinear electrocardiographic dynamics:

[0100]

[0101] Where H = [H1, H2, H3] Tis the nonlinear electrocardiodynamic parameter, z is the model parameter; F(H; z) is the main part of the nonlinear dynamics, l(H; z) represents the uncertainty of the model; F(H; z) and l(H; z) together constitute a complete nonlinear electrocardiodynamic model. Since l(H; z) is inseparable in some cases, φ(H; z): = F(H; z) + l(H; z) is used to represent the complete nonlinear electrocardiodynamic model.

[0102] (2) The model is learned using a lattice neural network, which is expressed as follows:

[0103]

[0104] in, represents the state variable of the lattice neural network; A is a diagonal matrix, whose diagonal elements are gain parameters, and the value is set to 0.9; It is the weight learned by the neural network according to the parameter trajectory at different distances of neurons. The specific calculation formula is: in, is the state error, Γ i , σ i are all network adjustment parameters, which are set to 1.5 and 100 respectively; S(H) is a Gaussian radial basis function, which mainly includes the number and distribution information of neurons. In this embodiment, the lattice neurons are set to be evenly distributed in the area [-1,1]×[-1,1]×[-1,1], and the width is set to 0.5.

[0105] (3) Extract the deep nonlinear electrocardiographic dynamics characteristics under the ECG signal. The main calculation formula is as follows:

[0106]

[0107] Among them, φ(H;z) is the complete nonlinear electrocardiographic information, yes After learning, it will gradually tend to the convergent mean, such as Figure 3 As shown; In the lattice neural network, each neuron learns the nonlinear electrocardiographic dynamics characteristics of the local nonlinear electrocardiographic parameter trajectory. The nonlinear electrocardiographic dynamics characteristics learned by the lattice neural network are as follows: Figure 4 As shown; ∈ represents the learning error.

[0108] Step 2: Calculate the distance metric between different nonlinear electrocardiographic dynamic features based on nonlinear dynamic differences, and determine the electrocardiographic pattern set to be fused according to the distance metric value, which specifically includes the following steps:

[0109] 2-1. Calculate the distance metric between different nonlinear electrocardiographic dynamic features based on the difference in nonlinear dynamics;

[0110] The distance metric is calculated as follows:

[0111]

[0112] Among them, ε is the comprehensive identification error, ε i represents the error in the i-th dimension between two nonlinear electrocardiodynamic features, where the error in the i-th dimension ε i The system satisfies the following equation:

[0113]

[0114] Among them, b i is the gain parameter. In this embodiment, b i =0.2, f i '(H',z') represents the feature of the i-th dimension of nonlinear electrocardiodynamics; given a feature of the i-th dimension of nonlinear electrocardiodynamics Based on the above error system, we can obtain With another nonlinear dynamics i-th dimension feature f i The error of '(H',z')ε i ,like Figure 5 shown.

[0115] 2-2. Determine the set of ECG patterns to be fused based on the distance measurement value. Specifically, the following steps are included:

[0116] (1) Take any two nonlinear electrocardiodynamic features, denoted as P and Q, and calculate the comprehensive identification error ε between P and Q;

[0117] (2) Based on the comprehensive recognition error ε, it is determined whether the ECG patterns between PQ can be fused. The specific determination method is:

[0118] When the comprehensive identification error ε takes a value of [0,0.5], it is determined that the two sets of nonlinear electrocardiodynamic features to be compared are relatively close, and are set to the waiting fusion mode, which is recorded as the waiting fusion set C1(P,Q);

[0119] When the comprehensive identification error ε is between (0.5,∞), it is considered that the two groups of nonlinear electrocardiographic dynamic characteristics to be compared are different and do not participate in the fusion of electrocardiographic patterns;

[0120] (3) Take the nonlinear dynamic feature P and any nonlinear electrocardiodynamic feature that has not been selected in this round, record it as M, and repeat steps (1) and (2). When the comprehensive error ε of P and M is in [0, 0.5], update the set C1(P, …, M), otherwise do not update. This process continues until P and all nonlinear dynamic features have completed the fusion judgment of the electrocardiodynamic pattern. At this time, the pattern to be fused is C1(P, …);

[0121] (4) Select any two nonlinear dynamic features R, S that do not appear in the pattern to be fused C1 (P, ...), repeat steps (1), (2), and (3), and record the pattern to be fused as C2 (R, ...), until all nonlinear electrocardiographic dynamic features are in the pattern to be fused C1 (P, ...). i End the judgment.

[0122] Step 3: Normalize the values ​​based on the excitation level. The normalized excitation value is used as the weight to perform weighted fusion on the ECG pattern to be fused. The fusion result is stored in the form of a constant matrix. Specifically, the following steps are included:

[0123] 3-1. Calculate the degree of neural network excitation between the ECG patterns to be fused.

[0124] The specific calculation method for the degree of neural network excitation between the ECG patterns to be fused is as follows:

[0125]

[0126] Among them, S is the excitation counter, which is used to indicate the excitation degree of the neural network weight matrix; W ijk represents the weight of the i-th layer, j-th row, and k-th column in the m×n×l constant value neural network weight space matrix; T is the excitation threshold, which is set to 0.5 in this embodiment. The corresponding excitation matrix diagram is as follows Figure 6 As shown, the weight W close to the trajectory ijk will be greater than the excitation threshold T, and the weight W away from the trajectory ijk It will fall below the excitation threshold and be set to 0.

[0127] 3-2. Use the excitation degree as the weight to perform weighted fusion on the ECG patterns to be fused, and store the fusion results in the form of a constant matrix. Specifically, the following steps are included:

[0128] (1) Select the nonlinear electrocardiodynamic feature set C in the mode to be fused i (c1,c2,…,c n ), respectively calculate the excitation degree S1, S2, ..., S of the constant value neural network weight space matrix n , and normalize them numerically, denoted as Q1, Q2, …, Q n ;

[0129] (2) Treatment of fused ECG pattern C i Perform weighted fusion:

[0130] y i =Q1c i1 +Q2c i2 +…+Q n c in

[0131] Among them, y i Represents the set to be fused C i The nonlinear electrocardiographic dynamic characteristics of the i-th dimension after fusion; Q n Represents the set to be fused C i The excitation degree of the nth nonlinear electrocardiodynamic characteristic value after normalization; c in Represents the set to be fused C i The feature of the i-th dimension of the n-th nonlinear electrocardiodynamic feature; the fusion result y is stored in the form of a constant matrix.

[0132] In the embodiment, in the subsequent recognition stage, the comprehensive recognition error ε is calculated based on the nonlinear electrocardiodynamic characteristics of the signal to be recognized and the ECG fusion pattern, and the recognition of the ECG pattern fusion is completed by comparing ε. The ECG recognition accuracy is 92%. Without performing deep signal decomposition on the reduced-dimensional signal, the reduced-dimensional signal is directly used to extract the nonlinear electrocardiodynamic characteristics and distinguish them from the ECG fusion pattern. The ECG recognition accuracy is 83%.

[0133] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be equivalent replacement methods and are included in the protection scope of the present invention.

Claims

1. A method for fusion recognition of electrocardiogram patterns taking into account the degree of neural network excitation, characterized in that: The following steps are included: Step 1: Obtain 12-lead ECG signals from the body surface and construct a lattice neural network to extract the deep nonlinear electrocardiographic dynamics characteristics of the ECG signals; Step 2: Calculate the distance metric between different nonlinear electrocardiographic dynamic features based on the nonlinear dynamics difference, and determine the electrocardiographic pattern set to be fused according to the distance metric value; Step 3: According to the different degrees of neural network excitation between the ECG patterns to be fused, different weights are assigned to the ECG patterns to be fused for weighted fusion, and the fusion results are stored in the form of a constant matrix; The step 1 specifically includes the following steps: 1-1. Perform linear transformation on the 12-lead ECG signal to complete data dimension reduction, and filter out noise and artifacts on the signal after dimension reduction; 1-2. Perform deep decomposition of the signal, extract key parameters based on signal components, and use the extracted key parameters as nonlinear ECG dynamic parameters; 1-3. Perform nonlinear electrocardiodynamic modeling, construct a lattice neural network based on the nonlinear electrocardiodynamic model, input nonlinear electrocardiodynamic parameters into the lattice neural network, and store nonlinear electrocardiodynamic characteristics in the form of a constant neural network weight matrix through the learning of the lattice neural network; The key parameter extraction specifically includes the following steps: (1) Given an original input signal V t =[v1,v2,v3] T , v1, v2, v3 represent the signals of each dimension after data dimensionality reduction; assuming V t The local extreme point of k ,k=1,2,3,…}, define τ 0＝0 , define a baseline extraction factor L, L in [0,τ k ] The definition of the moment is expressed as follows: Among them, v it Represents the signal v of the i-th dimension i Take the value at point t, L k Indicates that in [0,τ k ] The definition of the baseline extraction factor at the moment L; the next moment L k+1 Satisfy the following formula: Among them, α represents the gain parameter, the value range is between (0,1), and it represents the amplitude of the component obtained by this decomposition; (2) Input signal V t Perform deep decomposition and obtain the baseline extraction factor L according to step (1). t The baseline L t and H t The sum form, each decomposition is expressed as follows: V t =Lv it +(1-L)v it =L t +H t The input signal of this decomposition comes from the baseline signal L obtained from the previous decomposition. t , and the component H t As the resulting signal component; (3) Continue with step (2). When the signal component H t When a monotonic trend is shown, the decomposition is terminated, and the overall decomposition process satisfies the following formula: Where p represents the number of decompositions; (4) Perform Pearson correlation analysis on all decomposed signal components and the original input signal in turn: Among them, r t It means the component H obtained after the signal is decomposed t times and the input signal V before decomposition t The coefficient obtained by Pearson correlation analysis, N represents the number of signal components, n represents the total length of the signal components, and the Pearson correlation coefficient r is taken t The largest signal component is used as the key parameter to obtain the nonlinear ECG dynamic parameters.

2. The ECG pattern fusion recognition method taking into account the degree of neural network excitation according to claim 1, characterized in that: The nonlinear electrocardiodynamic modeling specifically includes the following steps: (1) Modeling nonlinear electrocardiographic dynamics: Where H=[H1,H2,H3] T is a nonlinear electrocardiodynamic parameter, z is a model parameter; F(H; z) is the main part of nonlinear dynamics, l(H; z) represents the uncertainty of the model; F(H; z) and l(H; z) together constitute a complete nonlinear electrocardiodynamic model. Since l(H; z) is inseparable in some cases, φ(H; z): = F(H; z) + l(H; z) is used to represent the complete nonlinear electrocardiodynamic model; (2) The model is learned using a lattice neural network, which is expressed as follows: in, Represents the state variable of the lattice neural network; A is a diagonal matrix whose diagonal elements are gain parameters, with values ​​between [-1,1]; It is the weight learned by the neural network according to the parameter trajectory at different distances of neurons. The specific calculation formula is: in, is the state error, Γ i , σ i They are all network adjustment parameters; S(H) is the Gaussian radial basis function, which mainly includes the number and distribution information of neurons; (3) Extract the deep nonlinear electrocardiographic dynamics characteristics under the ECG signal. The main calculation formula is as follows: Among them, φ(H;z) is the complete nonlinear electrocardiographic information, yes After learning, it will gradually tend to the convergent mean of the convergence trend, It is the nonlinear electrocardiographic dynamics feature learned by each neuron from the local nonlinear electrocardiographic parameter trajectory in the lattice neural network, ∈ i represents the learning error.

3. The ECG pattern fusion recognition method taking into account the degree of neural network excitation according to claim 1, characterized in that: The step 2 specifically includes the following steps: 2-1. Calculate the distance metric between different nonlinear electrocardiographic dynamic features based on the difference in nonlinear dynamics; 2-2. Determine the set of ECG patterns to be fused based on the distance metric value.

4. The ECG pattern fusion recognition method taking into account the degree of neural network excitation according to claim 3, characterized in that: The distance metric is calculated as follows: Among them, ε is the comprehensive identification error, ε i represents the error in the i-th dimension between two nonlinear electrocardiodynamic features; where the error in the i-th dimension ε i Satisfy the following formula: Among them, b i is the gain parameter, f i '(H',z') represents the feature of the i-th dimension of nonlinear electrocardiodynamics; given a feature of the i-th dimension of nonlinear electrocardiodynamics Based on the above error formula, we can get With another nonlinear dynamics i-th dimension feature f i The error of '(H',z')ε i .

5. The ECG pattern fusion recognition method taking into account the degree of neural network excitation according to claim 3, characterized in that: The set of ECG patterns to be fused specifically includes the following steps: (1) Take any two nonlinear electrocardiodynamic features, denoted as P and Q, and calculate the comprehensive identification error ε between P and Q; (2) Based on the comprehensive recognition error ε, it is determined whether the ECG patterns between PQ can be fused. The specific determination method is: When the comprehensive identification error ε takes a value of [0,0.5], it is determined that the two sets of nonlinear electrocardiodynamic features to be compared are relatively close, and are set to the waiting fusion mode, which is recorded as the waiting fusion set C1(P,Q); When the comprehensive identification error ε is between (0.5,∞), it is considered that the two groups of nonlinear electrocardiographic dynamic characteristics to be compared are different and do not participate in the fusion of electrocardiographic patterns; (3) Take the nonlinear dynamic feature P and any nonlinear electrocardiodynamic feature M that has not been selected in this round, and repeat steps (1) and (2). When the value of the comprehensive error ε of P and M is in [0, 0.5], update the set C1(P, …, M), otherwise do not update; until P and all nonlinear dynamic features complete the fusion judgment of the electrocardiodynamic pattern, at this time the pattern to be fused is C1(P, …); (4) Select any two nonlinear dynamic features R, S that do not appear in the pattern to be fused C1 (P, ...), repeat steps (1), (2), and (3), and record the pattern to be fused as C2 (R, ...), until all nonlinear electrocardiographic dynamic features are in the pattern to be fused C1 (P, ...). i In, end the judgment.

6. The ECG pattern fusion recognition method taking into account the degree of neural network excitation according to claim 1, characterized in that: The step three specifically includes the following steps: 3-1. Calculate the degree of neural network excitation between ECG patterns to be fused; 3-2. Perform numerical normalization based on the degree of excitation, use the normalized excitation value as the weight to perform weighted fusion on the ECG patterns to be fused, and store the fusion results in the form of a constant matrix.

7. The ECG pattern fusion recognition method taking into account the degree of neural network excitation according to claim 6, characterized in that: The degree of neural network excitation between the ECG patterns to be fused is specifically calculated as follows: Among them, S is the excitation counter, which is used to indicate the excitation degree of the neural network weight matrix; W ijk It represents the neural network weight of the i-th layer, j-th row, and k-th column in the m×n×l constant neural network weight space matrix; T is the excitation threshold.

8. The ECG pattern fusion recognition method taking into account the degree of neural network excitation according to claim 6, characterized in that: The weighted fusion of the ECG patterns to be fused specifically includes the following steps: (1) Select the nonlinear electrocardiodynamic feature set C in the mode to be fused i (c1,c2,…,c n ), respectively calculate the excitation degree S1, S2, ..., S of the constant value neural network weight space matrix n , and normalize them numerically, denoted as Q1, Q2, …, Q n ; (2) Treatment of fused ECG pattern C i Perform weighted fusion: y i =Q1c i1 +Q2c i2 +…+Q n c in Among them, y i Represents the set to be fused C i The nonlinear electrocardiographic dynamic characteristics of the i-th dimension after fusion; Q n Represents the set to be fused C i The excitation degree of the nth nonlinear electrocardiodynamic characteristic value after normalization; c in Represents the set to be fused C i The feature of the i-th dimension of the n-th nonlinear electrocardiodynamic feature; the fusion result y is stored in the form of a constant matrix.

Images