A heart sound segmentation method and system based on synchronous squeezed short-time Fourier transform and BiGRU

Abstract

The present invention discloses a heart sound segmentation method and system based on synchronized squeezed short-time Fourier transform and BiGRU, the method comprising: constructing a heart sound segmentation model based on synchronized squeezed short-time Fourier transform (SSTFT) and BiGRU, the model consisting of SSTFT unit, BiGRU network, Softmax classifier; training the heart sound segmentation model using samples of heart sound data set; inputting the heart sound signal to be segmented into the trained heart sound segmentation model for state classification, and dividing it into four state categories of the first heart sound (S1), the second heart sound (S2), S1-S2 interval, and S2-S1 interval; correcting the wrong state category predicted by the heart sound segmentation model; the system comprises: a model construction module, a model training module, a heart sound state classification module, and a state correction module. The present invention can effectively improve the positioning precision and accuracy of heart sound segmentation.

Description

Technical Field

[0001] The invention relates to a heart sound segmentation method and system based on synchronous extrusion short-time Fourier transform and BiGRU, belonging to the technical field of heart sound analysis and artificial intelligence. Background Art

[0002] At present, doctors can diagnose the heart condition of the subject with the help of X-rays, magnetic resonance images, electrocardiograms, heart sounds, etc. Among them, heart sound auscultation is widely used as an auxiliary diagnostic method for cardiovascular diseases in clinical practice due to its convenience, simplicity, low cost and harmlessness. It plays an important role in early screening of diseases and daily physical examinations. However, the results of heart sound auscultation usually depend on the auscultation experience and diagnostic level of medical staff, and are greatly affected by subjective factors. In recent years, with the emergence of electronic stethoscopes, combined with machine learning and pattern recognition technology, the digital collection, storage, analysis of heart sound signals and the intelligent screening of cardiovascular diseases have become possible.

[0003] Heart sound signals are quasi-periodic, and the period of each heart sound signal is called a cardiac cycle. Each cardiac cycle contains four states: the first heart sound (S1), the S1-S2 interval, the second heart sound (S2), and the S2-S1 interval. Among them, S1 and the S1-S2 interval constitute the cardiac systole (Systole), S1 occurs at the beginning of the systole, its frequency range is mainly distributed in 50-150Hz, and its duration is about 100-150ms; S2 and the S2-S1 interval constitute the cardiac diastole (Diastole), S2 occurs at the beginning of the diastole, its frequency range is mainly distributed in 50-200Hz, and its duration is about 80-100ms. Heart sound segmentation refers to determining the position intervals of the four states of S1, S1-S2 interval, S2, and S2-S1 interval in the heart sound signal, providing a positioning benchmark for the extraction of heart sound feature parameters and heart sound classification. The specific location of heart murmurs in the cardiac cycle is related to specific cardiovascular diseases. After determining the locations of S1 and S2, the abnormal heart sound category can be directly detected based on the location of the heart murmur. Therefore, the first and most important step in heart sound analysis is to segment the heart sound signal. Accurate heart sound segmentation is the premise and basis for automatic analysis of heart sounds.

[0004] At present, methods for heart sound segmentation can be divided into three categories: (1) The heart sound segmentation method that combines electrocardiogram as reference information has a good segmentation effect, but in the actual electrocardiogram acquisition environment, it has certain limitations due to high requirements on hardware equipment and the inability of some special groups to acquire electrocardiograms; (2) The heart sound segmentation method based on the envelope of the heart sound signal needs to set a suitable detection threshold to detect the peak values ​​of S1 and S2. However, since the heart sound signal is inevitably interfered by external noise, especially for those heart sound signals containing heart murmurs, it is likely that the peak values ​​of S1 and S2 cannot be detected. Therefore, this type of method has great limitations; (3) The heart sound segmentation method based on the hidden Markov model (HMM) or the hidden semi-Markov model requires unsupervised learning through a large amount of training data, which easily results in incomplete extraction of detailed features and low segmentation accuracy. Summary of the invention

[0005] The purpose of the present invention is to address the defects and shortcomings of the above-mentioned prior art and propose a heart sound segmentation method and system based on synchronous squeezed short-time Fourier transform and BiGRU. The method uses synchronous squeezed short-time Fourier transform to extract the instantaneous frequency characteristics of heart sounds, and uses the BiGRU network to extract the contextual time-frequency characteristics of heart sounds. After the heart sound features are classified into states, the error categories are corrected according to the constraint relationship between the heart sound states, which can effectively improve the positioning accuracy and precision of heart sound segmentation.

[0006] The technical solution adopted by the present invention to solve the technical problem is: a heart sound segmentation method based on synchronous squeezed short-time Fourier transform and BiGRU, the method comprising the following steps:

[0007] Step 1: Build a heart sound segmentation model based on synchronous squeezed short-time Fourier transform and BiGRU;

[0008] Step 2: Use samples of the heart sound dataset to train the heart sound segmentation model, and adjust the parameters of the heart sound segmentation model to the optimal value through the error back propagation algorithm;

[0009] Step 3: Input the heart sound signal to be segmented into the trained heart sound segmentation model for state classification, and divide it into four state categories: the first heart sound (S1), the second heart sound (S2), the S1-S2 interval, and the S2-S1 interval;

[0010] Step 4: According to the constraint that the four state categories of the heart sound signal always appear cyclically in the order of S1, S1-S2 interval, S2, and S2-S1 interval, the erroneous state categories predicted by the heart sound segmentation model are corrected and the final segmentation result is output.

[0011] Furthermore, the step 1 includes a synchronous squeezing short-time Fourier transform unit performing a short-time Fourier transform (STFT) and a synchronous squeezing transform (SST) on the input heart sound signal to extract the instantaneous frequency characteristics of the heart sound signal. The specific steps are as follows:

[0012] Step 1-1: Assume that the input discrete heart sound signal is x(n), the window function is h(n), the window width of the window function is L, and the moving step length of two adjacent windows is H sampling time intervals. Then the short-time Fourier transform expression of x(n) is:

[0013]

[0014] Where m represents the time step number when the window is moved, m = 0, 1, 2, ..., M-1, corresponding to the time parameter of STFT; k represents the frequency sampling point number in the frequency domain, k = 0, 1, 2, ..., L-1; assuming that the sampling frequency of the heart sound signal x(n) is f s , then STFT(m,k) is x(n) at time mH / f s The frequency spectrum of each k is kf s / L;

[0015] Step 1-2: Calculate the local instantaneous frequency at (m,k) Its expression is:

[0016]

[0017] Among them, Re(·) represents the operation of taking the real part of a complex number, and Round(·) represents the operation of rounding off;

[0018] Step 1-3: In the time-frequency domain, The mapping relationship of STFT (m, k) time spectrum is synchronously squeezed and transformed, and its expression is:

[0019]

[0020] in, represents the frequency sample point after squeezing, δ(·) is the impulse response function;

[0021] Step 1-4: Perform complex modulo operation to obtain the instantaneous frequency characteristics of the heart sound signal Its expression is:

[0022]

[0023] Here, |·| represents the complex modulus operation.

[0024] Furthermore, the window function in step 1-1 is a Kaiser window, and its expression is:

[0025]

[0026] Where I0(·) is the first-kind modified Besso function of zero order, and α represents the shape parameter of the Kaiser window.

[0027] Furthermore, m in step 1-1 corresponds to the time step of BiGRU, and the instantaneous frequency features of the heart sounds extracted from the short-term window of each time step are A time-frequency feature sequence is constructed along the time dimension as the input of the Bi-GRU network.

[0028] Furthermore, the specific method of correcting the erroneous state category predicted by the heart sound segmentation model in step 4 is as follows:

[0029] If the four state categories of the heart sound signal S1, S1-S2 interval, S2, S2-S1 interval are numbered with "0", "1", "2", and "3" respectively, for an input heart sound segment to be segmented, the category number sequence predicted by the model is y0, y1, ..., y m ,…,y M-1 ,y m ∈{0,1,2,3}, m=0,1,2,…,M-1, let:

[0030]

[0031] Check and judge y0,y1,…,y M-1 , if y m ≠y m-1 , and y m+1 =y m-1 , it indicates that the predicted category of the mth time step may be wrong, let y m =y m-1 If y m ≠y m-1 ,and Then If y m ≠y m-1 ,and Then let y m =y m-1 .

[0032] The present invention also provides a heart sound segmentation system based on synchronous squeezed short-time Fourier transform and BiGRU, the system comprising:

[0033] A model building module, used to build a heart sound segmentation model, wherein the heart sound segmentation model includes a synchronous squeezed short-time Fourier transform unit, a BiGRU network and a Softmax classifier;

[0034] The synchronous squeezing short-time Fourier transform unit is used to perform short-time Fourier transform and synchronous squeezing transform on the input heart sound signal to extract the instantaneous frequency characteristics of the heart sound signal;

[0035] The BiGRU network is composed of a forward gated recurrent unit (GRU) layer and a backward GRU layer, and is used to extract contextual time-frequency features of the heart sound signal;

[0036] The Softmax classifier is used to classify the heart sound features extracted by the BiGRU network at each time step, and output four state categories: the first heart sound (S1), the second heart sound (S2), the S1-S2 interval, and the S2-S1 interval;

[0037] A model training module is used to train the constructed heart sound segmentation model using samples of the heart sound data set, and adjust the parameters of the heart sound segmentation model to the optimal value through an error back propagation algorithm;

[0038] The heart sound state classification module is used to classify the newly input heart sound signal to be segmented using the trained heart sound segmentation model, and divide it into four state categories: the first heart sound (S1), the second heart sound (S2), the S1-S2 interval, and the S2-S1 interval;

[0039] The state correction module is used to correct the erroneous state category predicted by the heart sound state classification module and output the final segmentation result.

[0040] Furthermore, the system includes at least one computing device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, and when the computer program is loaded into the processor, the method for segmenting heart sounds based on synchronized squeezed short-time Fourier transform and BiGRU is implemented.

[0041] Beneficial effects:

[0042] 1. The present invention discloses a heart sound segmentation method based on synchronized squeezed short-time Fourier transform and BiGRU. The method uses synchronized squeezed short-time Fourier transform (SSTFT) technology to extract instantaneous frequency characteristics of heart sound signals. SSTFT performs synchronized squeeze transform (SST) processing on the STFT time-frequency spectrum based on the short-time Fourier transform (STFT), so that the energy of the signal is concentrated near the instantaneous frequency, which can effectively improve the concentration of time-frequency spectrum energy. Therefore, compared with the STFT technology, the instantaneous frequency characteristics of the heart sound signal extracted using the SSTFT technology can more effectively characterize the heart sound signal.

[0043] 2. The present invention discloses a heart sound segmentation method based on synchronous squeezed short-time Fourier transform and BiGRU. This method constructs a time-frequency feature sequence along the time dimension from the instantaneous frequency features of the heart sound signal extracted in each time step window as the input of a bidirectional gated recurrent unit (Bidirectional GRU, BiGRU) network, and uses the time series modeling capability of the BiGRU network to extract the contextual time-frequency features of the heart sound signal, which is beneficial to improving the positioning accuracy of the heart sound segmentation. GRU is a variant of the Long Short-Term Memory (LSTM) neural network. Compared with LSTM, GRU has no cell state and directly uses hidden state for information transmission. Therefore, the structure of GRU is simpler, fewer parameters are required for training, faster training speed, and easier to converge.

[0044] 3. The present invention discloses a heart sound segmentation method based on synchronous squeezed short-time Fourier transform and BiGRU. According to the constraint relationship that the four state categories of heart sound signals always appear cyclically in the order of S1, S1-S2 interval, S2, and S2-S1 interval, the method corrects the erroneous state categories predicted by the heart sound segmentation model and outputs the final segmentation result, which can further improve the positioning precision and accuracy of heart sound segmentation. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 The figure is a flow chart of the method of the present invention.

[0046] Figure 2 This is a schematic diagram of the structure of the heart sound segmentation model constructed in the present invention.

[0047] Figure 3 Schematic diagram of the structure of the BiGRU network of the present invention.

[0048] Figure 4 It is a schematic diagram of the four heart sound states included in each cardiac cycle of the present invention.

[0049] Figure 5 The figure is a schematic diagram for labeling the state categories of a heart sound sample according to the present invention. DETAILED DESCRIPTION

[0050] The technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings and preferred embodiments of the specification.

[0051] like Figure 1 As shown, the present invention discloses a heart sound segmentation method based on synchronous squeezed short-time Fourier transform and BiGRU, and the method mainly comprises the following steps:

[0052] Step 1: Build a heart sound segmentation model based on synchronous squeezing short-time Fourier transform and BiGRU. The model consists of synchronous squeezing short-time Fourier transform unit, BiGRU network and Softmax classifier. Figure 2 As shown, where:

[0053] The synchronous squeezing short-time Fourier transform unit is used to perform short-time Fourier transform and synchronous squeezing transform on the input heart sound signal to extract the instantaneous frequency characteristics of the heart sound signal.

[0054] The BiGRU network consists of a forward gated recurrent unit (GRU) layer and a backward GRU layer, such as Figure 3 As shown, it is used to extract the contextual time-frequency features of the heart sound signal.

[0055] The Softmax classifier is used to classify the heart sound features extracted by the BiGRU network at each time step and output four state categories: the first heart sound (S1), the second heart sound (S2), the S1-S2 interval, and the S2-S1 interval. Each cardiac cycle contains four heart sound states such as Figure 4 shown.

[0056] As a further optimization scheme of the heart sound segmentation method based on synchronous squeezing short-time Fourier transform and BiGRU of the present invention, in step 1, the synchronous squeezing short-time Fourier transform unit performs short-time Fourier transform and synchronous squeezing transform on the input heart sound signal to extract the instantaneous frequency characteristics of the heart sound signal. The specific steps are as follows:

[0057] Step 1-1: Assume that the input discrete heart sound signal is x(n), the window function is h(n), the window width of the window function is L, and the moving step length of two adjacent windows is H sampling time intervals. Then the short-time Fourier transform expression of x(n) is:

[0058]

[0059] Wherein, m represents the time step number when the window is moved, m = 0, 1, 2, ..., M-1, corresponding to the time parameter of STFT, and in the present invention, corresponds to the time step of BiGRU; k represents the frequency sampling point number in the frequency domain, k = 0, 1, 2, ..., L-1; assuming that the sampling frequency of the heart sound signal x(n) is f s , then STFT(m,k) is x(n) at time mH / f s The frequency spectrum of each k is kf s / L; As a preferred solution, the window function selected in the present invention is the Kaiser window, and its expression is:

[0060]

[0061] Where I0(·) is the first-order modified Beso function of zero order, α represents the shape parameter of the Kaiser window; in this embodiment, L = 64, H = 20, M = 198, f s =2000Hz,α=1.

[0062] The m in the above step 1-1 of the present invention corresponds to the time step of BiGRU, and the instantaneous frequency features of the heart sounds extracted from the short-term window of each time step are A time-frequency feature sequence is constructed along the time dimension as the input of the Bi-GRU network.

[0063] Step 1-2: Calculate the local instantaneous frequency at (m,k) Its expression is:

[0064]

[0065] Among them, Re(·) represents the operation of taking the real part of a complex number, and Round(·) represents the operation of rounding off.

[0066] Step 1-3: In the time-frequency domain, The mapping relationship of STFT (m, k) time spectrum is synchronously squeezed and transformed, and its expression is:

[0067]

[0068] in, represents the frequency sample point after squeezing, δ(·) is the impulse response function;

[0069] Step 1-4: Perform complex modulo operation to obtain the instantaneous frequency characteristics of the heart sound signal Its expression is:

[0070]

[0071] Here, |·| represents the complex modulus operation.

[0072] Step 2: Use samples of the heart sound dataset to train the heart sound segmentation model, and adjust the parameters of the heart sound segmentation model to the optimal value through the error back propagation algorithm.

[0073] In this embodiment, experiments were conducted using the PhysioNet / CinC Challenge 2016 dataset. This dataset provides heart sound (Phonocardiogram, PCG) and electrocardiogram (Electrocardiogram, ECG) data collected in a real clinical environment. There are 3153 heart sound records of 5 to 120 seconds from 764 subjects in the dataset. This embodiment selects 792 heart sound signals with relevant ECG records, including 386 normal heart sounds and 406 abnormal heart sounds. The heart sound signals in the dataset are divided into heart sound segments with a fixed duration of 2 seconds. After segmentation, a total of 5870 heart sound segments are obtained as training samples. In order to reduce the subsequent processing pressure of the data and fully retain the main frequency components of the heart sound, this embodiment resamples the heart sound signal and adjusts the sampling frequency to 2000Hz.

[0074] In order to train the heart sound segmentation model, the state categories of the heart sound samples must first be labeled. In this embodiment, the heart sound segmentation algorithm based on the logistic regression hidden semi-Markov (LR-HSMM) model is first used to obtain the four state positions of the heart sound segment, and then the cardiothoracic experts correct the erroneous state positions based on the synchronously collected ECG signals, and the four state categories of the heart sound signal S1, S1-S2 interval, S2, and S2-S1 interval are labeled with "0", "1", "2", and "3" respectively. For a heart sound segment with a duration of 2s, when the sampling frequency is 2000Hz, each heart sound segment contains 4000 sampling points, and the window width L of the Kaiser window function is set to 64 sampling points, and the moving step H between two adjacent windows is 20 sampling points. In this way, the number of time steps M corresponding to a heart sound segment is 198, and its state category is labeled as follows. Figure 5 shown.

[0075] Suppose a training sample set of heart sound segments is {(g0,c0),(g1,c1),…,(g M-1 ,c M-1 )}, where g m is the heart sound feature vector extracted by the Bi-GRU network at the mth time step, m = 0, 1, 2, ..., M-1, c m is the number of the category to which the heart sound state belongs at the mth time step. If the four state categories of the heart sound signal S1, S1-S2 interval, S2, and S2-S1 interval are numbered with "0", "1", "2", and "3" respectively, then cm ∈{0,1,2,3}, the heart sound state at any time step is classified into one of the four categories with a certain probability.

[0076] The Softmax classifier is defined as:

[0077]

[0078] Among them, h(g m ; θ) is a column vector with 4 rows and 1 column, and p(c m =j|g m θ j ) represents the eigenvector g m The probability of being classified as the jth class, j = 0, 1, 2, 3; θ j are the parameters of the Softmax classifier that can be trained and learned, and constitute the parameter matrix θ, is θ j The transpose of .

[0079] For g m , select p(c m =j|g m θ j ) is the j corresponding to the maximum value as the output category number y of the Softmax classifier at the mth time step m ,Right now

[0080]

[0081] When training the model, y m The category number c annotated with the training sample m Make a comparison. If they are consistent, the classification is correct, otherwise the classification is wrong. The loss function of the Softmax classifier is defined as:

[0082]

[0083] Among them, I(c m =j) is the indicator function, when c m =j, I(c m =j) = 1; when c m ≠j, I(c m =j)=0.

[0084] In this embodiment, the Adam optimization algorithm is used to train the heart sound segmentation model.

[0085] Step 3: Input the heart sound signal to be segmented into the trained heart sound segmentation model for state classification, and divide it into four state categories: the first heart sound (S1), the second heart sound (S2), the S1-S2 interval, and the S2-S1 interval.

[0086] Step 4: According to the constraint that the four state categories of the heart sound signal always appear cyclically in the order of S1, S1-S2 interval, S2, and S2-S1 interval, the wrong state category predicted by the heart sound segmentation model is corrected and the final segmentation result is output. The specific method is as follows:

[0087] If the four state categories of the heart sound signal S1, S1-S2 interval, S2, S2-S1 interval are numbered with "0", "1", "2", and "3" respectively, for an input heart sound segment to be segmented, the category number sequence predicted by the model is y0, y1, ..., y m ,…,y M-1 ,y m ∈{0,1,2,3}, m=0,1,2,…,M-1, let

[0088]

[0089] Check and judge y0,y1,…,y M-1 , if y m ≠y m-1 , and y m+1 =y m-1 , it indicates that the predicted category of the mth time step may be wrong, let y m =y m-1 If y m ≠y m-1 ,and Then If y m ≠y m-1 ,and Then let y m =y m-1 .

[0090] Based on the same inventive concept, the present invention also provides a heart sound segmentation system based on synchronous squeezed short-time Fourier transform and BiGRU, including: a model construction module, used to construct a heart sound segmentation model, the model is composed of a synchronous squeezed short-time Fourier transform unit, a BiGRU network and a Softmax classifier; the structure and function of each unit of the model refer to the above method embodiment; a model training module, used to train the heart sound segmentation model using samples of the heart sound data set, and adjust the parameters of the heart sound segmentation model to the optimal through the error back propagation algorithm; a heart sound state classification module, used to use the trained heart sound segmentation model to classify the state of the newly input heart sound signal to be segmented, and divide it into four state categories: the first heart sound (S1), the second heart sound (S2), the S1-S2 interval, and the S2-S1 interval; and a state correction module, used to correct the erroneous state category predicted by the heart sound state classification module, and output the final segmentation result.

[0091] Based on the same inventive concept, an embodiment of the present invention provides a heart sound segmentation system based on synchronized squeezed short-time Fourier transform and BiGRU, comprising at least one computing device, the computing device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when loaded into the processor, implements the above-mentioned heart sound segmentation method based on synchronized squeezed short-time Fourier transform and BiGRU.

[0092] The above scheme is only a description of the preferred implementation method of the present application, but the protection scope of the present application is not limited to this. Any person familiar with the technology can easily implement it within the scope of the description of the present application without changing the changes or substitutions involved in the basic principles of the claims, which should be covered by the protection scope of the present application, that is, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A heart sound segmentation method based on synchronous squeezed short-time Fourier transform and BiGRU, characterized in that: The steps include: Step 1: Construct a heart sound segmentation model based on synchronous squeezed short-time Fourier transform and BiGRU, including a synchronous squeezed short-time Fourier transform unit to perform short-time Fourier transform STFT and synchronous squeeze transform SST on the input heart sound signal to extract the instantaneous frequency characteristics of the heart sound signal. The specific steps are as follows: Step 1-1: Assume that the input discrete heart sound signal is x(n), the window function is h(n), the window width of the window function is L, and the moving step length of two adjacent windows is H sampling time intervals. Then the short-time Fourier transform expression of x(n) is: Where m represents the time step number when the window is moved, m = 0, 1, 2, ..., M-1, corresponding to the time parameter of STFT; k represents the frequency sampling point number in the frequency domain, k = 0, 1, 2, ..., L-1; assuming that the sampling frequency of the heart sound signal x(n) is f s , then STFT(m,k) is x(n) at time mH / f s The frequency spectrum of each k is kf s / L; Step 1-2: Calculate the local instantaneous frequency at (m,k) Its expression is: Among them, Re(·) represents the operation of taking the real part of a complex number, and Round(·) represents the operation of rounding off; Step 1-3: In the time-frequency domain, The mapping relationship of STFT (m, k) time spectrum is synchronously squeezed and transformed, and its expression is: in, represents the frequency sample point after squeezing, δ(·) is the impulse response function; Step 1-4: Perform complex modulo operation to obtain the instantaneous frequency characteristics of the heart sound signal Its expression is: Among them, |·| represents the complex modulus operation; Step 2: Use samples of the heart sound dataset to train the heart sound segmentation model, and adjust the parameters of the heart sound segmentation model to the optimal value through the error back propagation algorithm; Step 3: Input the heart sound signal to be segmented into the trained heart sound segmentation model for state classification, and divide it into four state categories: the first heart sound S1, the second heart sound S2, the S1-S2 interval, and the S2-S1 interval; Step 4: According to the constraint that the four state categories of the heart sound signal always appear cyclically in the order of S1, S1-S2 interval, S2, and S2-S1 interval, the erroneous state categories predicted by the heart sound segmentation model are corrected and the final segmentation result is output.

2. The heart sound segmentation method based on synchronous squeezed short-time Fourier transform and BiGRU according to claim 1, characterized in that: The window function in step 1-1 is a Kaiser window, and its expression is: Where I0(·) is the first-kind modified Besso function of zero order, and α represents the shape parameter of the Kaiser window.

3. The heart sound segmentation method based on synchronous squeezed short-time Fourier transform and BiGRU according to claim 1, characterized in that: The m in step 1-1 corresponds to the time step of BiGRU, and the instantaneous frequency features of the heart sounds extracted from the short-term window of each time step are A time-frequency feature sequence is constructed along the time dimension as the input of the Bi-GRU network.

4. The heart sound segmentation method based on synchronous squeezed short-time Fourier transform and BiGRU according to claim 1, characterized in that: The specific method of correcting the erroneous state category predicted by the heart sound segmentation model in step 4 is as follows: If the four state categories of the heart sound signal S1, S1-S2 interval, S2, S2-S1 interval are numbered with "0", "1", "2", and "3" respectively, for an input heart sound segment to be segmented, the category number sequence predicted by the model is y0, y1, ..., y m ,…,y M-1 ,y m ∈{0,1,2,3}, m=0,1,2,…,M-1, let: Check and judge y0,y1,…,y M-1 , if y m ≠y m-1 , and y m+1 =y m-1 , it means that the predicted category of the mth time step is wrong, let y m =y m-1 If y m ≠y m-1 ,and Then If y m ≠y m-1 ,and Then let y m =y m-1 .

5. A heart sound segmentation system based on synchronous squeezed short-time Fourier transform and BiGRU, characterized in that: include: A model building module, used to build a heart sound segmentation model, wherein the heart sound segmentation model includes a synchronous squeezed short-time Fourier transform unit, a BiGRU network and a Softmax classifier; The synchronous squeezing short-time Fourier transform unit is used to perform short-time Fourier transform and synchronous squeezing transform on the input heart sound signal to extract the instantaneous frequency characteristics of the heart sound signal. The specific steps include: The synchronous squeezing short-time Fourier transform unit performs short-time Fourier transform STFT and synchronous squeezing transform SST on the input heart sound signal to extract the instantaneous frequency characteristics of the heart sound signal. The specific steps are as follows: Step 1-1: Assume that the input discrete heart sound signal is x(n), the window function is h(n), the window width of the window function is L, and the moving step length of two adjacent windows is H sampling time intervals. Then the short-time Fourier transform expression of x(n) is: Where m represents the time step number when the window is moved, m = 0, 1, 2, ..., M-1, corresponding to the time parameter of STFT; k represents the frequency sampling point number in the frequency domain, k = 0, 1, 2, ..., L-1; assuming that the sampling frequency of the heart sound signal x(n) is f s , then STFT(m,k) is x(n) at time mH / f s The frequency spectrum of each k is kf s / L; Step 1-2: Calculate the local instantaneous frequency at (m,k) Its expression is: Among them, Re(·) represents the operation of taking the real part of a complex number, and Round(·) represents the operation of rounding off; Step 1-3: In the time-frequency domain, The mapping relationship of STFT (m, k) time spectrum is synchronously squeezed and transformed, and its expression is: in, represents the frequency sample point after squeezing, δ(·) is the impulse response function; Step 1-4: Perform complex modulo operation to obtain the instantaneous frequency characteristics of the heart sound signal Its expression is: Among them, |·| represents the complex modulus operation; The BiGRU network is composed of a forward gated recurrent unit GRU layer and a backward GRU layer, and is used to extract contextual time-frequency features of the heart sound signal; The Softmax classifier is used to classify the heart sound features extracted by the BiGRU network at each time step, and output four state categories: the first heart sound S1, the second heart sound S2, the S1-S2 interval, and the S2-S1 interval; A model training module is used to train the constructed heart sound segmentation model using samples of the heart sound data set, and adjust the parameters of the heart sound segmentation model to the optimal value through an error back propagation algorithm; The heart sound state classification module is used to classify the newly input heart sound signal to be segmented using the trained heart sound segmentation model into four state categories: the first heart sound S1, the second heart sound S2, the S1-S2 interval, and the S2-S1 interval; The state correction module is used to correct the erroneous state category predicted by the heart sound state classification module and output the final segmentation result.

6. A heart sound segmentation system based on synchronous squeezed short-time Fourier transform and BiGRU, characterized in that: The system includes at least one computing device, which includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the computer program is loaded into the processor, it implements a heart sound segmentation method based on synchronized squeezed short-time Fourier transform and BiGRU according to any one of claims 1-5.

Images