Heart sound data augmentation method and system based on generative adversarial network

Abstract

This invention discloses a method and system for enhancing heart sound data based on generative adversarial networks (GANs). The method includes: acquiring heart sound signals; filtering and denoising the acquired heart sound signals; normalizing the denoised heart sound signals; dividing the normalized signals into several heart sound segments of equal length according to the starting position of the first heart sound; inputting the heart sound segments into a generative adversarial network (GAN) model; training the model to obtain a trained generative network and a trained discriminator network; acquiring Gaussian white noise randomly generated through a normal distribution; and inputting the Gaussian white noise into the trained generative network to obtain enhanced heart sound data.

Description

Technical Field

[0001] This invention relates to the field of heart sound signal data processing technology, and in particular to a method and system for enhancing heart sound data based on generative adversarial networks. Background Technology

[0002] The statements in this section merely refer to the background art related to this invention and do not necessarily constitute prior art.

[0003] Cardiovascular disease is one of the most common diseases worldwide. Currently, 17.9 million people die from cardiovascular disease each year, accounting for 32% of all deaths globally. This number is projected to rise to 23 million annually by 2030. Early detection of cardiovascular disease is crucial for more effective treatment; therefore, rapid and accurate diagnosis is of paramount importance.

[0004] Heart sounds are the sounds produced by the mechanical waves caused by myocardial contraction, valvular closure, and blood impacting the walls of the ventricles and aortas. They can be heard with a stethoscope at specific points on the chest wall, or the mechanical waves of heart sounds can be recorded using instruments; the graph showing the changes in these mechanical waves over time is called a phonocardiogram. Heart sound signals contain a wealth of cardiac physiological information, and heart sound detection results can reveal signs of organic heart disease earlier than other detection results. Therefore, analyzing heart sound signals can aid in the diagnosis of cardiovascular diseases.

[0005] In recent years, automated heart sound analysis based on biosignal processing and artificial intelligence technologies has become a hot research topic, aiming to reduce expensive and time-consuming manual examinations. Constructing a heart sound dataset is a key issue in computer-aided heart sound analysis based on artificial intelligence. Currently, the number of existing heart sound datasets is relatively small, and the quality and quantity of heart sounds in most datasets are insufficient to meet the requirements for training classification models. The PhysioNet / CinC 2016 Heart Sound Recognition Challenge released a large-scale heart sound dataset with 3240 heart sound signals, including 665 normal heart sounds and 2575 abnormal heart sounds, with each signal ranging in length from 5 to 120 seconds. However, this dataset is also insufficient for deep learning models, due to both insufficient data volume and data imbalance.

[0006] When training data is insufficient, the model obtains less information from the original data, leading to overfitting. This means the model can only recognize data in the training set, but performs poorly on the test set, indicating weak generalization ability. When encountering imbalanced data, classification algorithms aiming for overall accuracy may focus excessively on the majority class, causing a decline in the performance of minority class samples and consequently reducing overall recognition performance. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a method and system for enhancing heart sound data based on generative adversarial networks. This method utilizes generative adversarial networks to generate a sufficient amount of heart sound data to meet the needs of training deep learning models.

[0008] In a first aspect, the present invention provides a method for enhancing heart sound data based on adversarial generative networks;

[0009] Heart sound data augmentation methods based on generative adversarial networks include:

[0010] Acquire heart sound signals;

[0011] The acquired heart sound signal is filtered and denoised. Then, the denoised heart sound signal is normalized. Then, the normalized signal is divided into several heart sound segments of equal length according to the starting position of the first heart sound in the heart sound signal.

[0012] Heart sound segments are input into a generative adversarial network model, and the model is trained to obtain a trained generative network and a trained discriminative network.

[0013] Gaussian white noise, randomly generated through a normal distribution, is obtained and input into the trained generative network to obtain enhanced heart sound data.

[0014] Secondly, this invention provides a heart sound data enhancement system based on generative adversarial networks;

[0015] A heart sound data augmentation system based on generative adversarial networks includes:

[0016] The acquisition module is configured to acquire heart sound signals.

[0017] The preprocessing module is configured to: filter and reduce noise in the acquired heart sound signal, then normalize the noise-reduced heart sound signal, and then divide the normalized signal into several heart sound segments of equal length according to the starting position of the first heart sound in the heart sound signal.

[0018] The training module is configured to: input heart sound segments into the generative adversarial network model, train the model, and obtain the trained generative network and the trained discriminative network;

[0019] The output module is configured to: acquire Gaussian white noise randomly generated by a normal distribution, input the Gaussian white noise into the trained generative network, and obtain enhanced heart sound data.

[0020] Thirdly, the present invention also provides an electronic device, comprising:

[0021] Memory, used for non-transitory storage of computer-readable instructions; and

[0022] Processor, for executing the computer-readable instructions,

[0023] When the computer-readable instructions are executed by the processor, they perform the method described in the first aspect above.

[0024] Fourthly, the present invention also provides a storage medium for non-transitory storage of computer-readable instructions, wherein, when the non-transitory computer-readable instructions are executed by a computer, the instructions for the method described in the first aspect are executed.

[0025] Fifthly, the present invention also provides a computer program product, including a computer program that, when run on one or more processors, is used to implement the method described in the first aspect above.

[0026] Compared with the prior art, the beneficial effects of the present invention are:

[0027] (1) The present invention can effectively solve the problem of insufficient training data for heart sound signals by using generative adversarial networks to generate a large amount of data.

[0028] (2) This invention can obtain enhanced data of various heart sound types by training different types of heart sound data, which can solve the problem of imbalance in heart sound datasets, that is, the problem of inconsistent data volume of different types of heart sounds. Attached Figure Description

[0029] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0030] Figure 1 The flowchart shown is a process for enhancing heart sound data based on generative adversarial networks according to the present invention.

[0031] Figure 2 The diagram shown is a basic structural diagram of a generative adversarial network.

[0032] Figure 3 The diagram shown is a basic structural diagram of a generative network;

[0033] Figure 4 The diagram shown is a basic structural diagram of the discrimination network;

[0034] Figure 5 The image shows the confusion matrix of the experimental results. Detailed Implementation

[0035] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0036] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments of the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. Furthermore, it should be understood that the terms “comprising” and “having”, and any variations thereof, are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0037] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0038] All data acquisition in this embodiment is carried out in accordance with laws and regulations and with user consent, and the data is used legally.

[0039] Terminology Explanation:

[0040] First heart sound: In cardiac physiology, the first heart sound (S1) is a heart sound, and it is the first heart sound to occur during the cardiac cycle. It is usually emitted from the mitral (M) and tricuspid (T) valves, slightly synchronized with the carotid pulse, and typically occurs simultaneously with the apical impulse, having the largest amplitude at the apex. The first heart sound is a complex wave, composed of the mechanical waves of atrioventricular valve closure and semilunar valve opening. It begins with atrioventricular valve closure and ends with semilunar valve opening, and is usually the heart sound with the highest amplitude, longest duration, and longest wavelength. The first heart sound occurs during systole and marks its beginning.

[0041] Example 1

[0042] This embodiment provides a method for enhancing heart sound data based on generative adversarial networks;

[0043] like Figure 1 As shown, the method for enhancing heart sound data based on generative adversarial networks includes:

[0044] S101: Acquire heart sound signals;

[0045] S102: The acquired heart sound signal is filtered and denoised, then the denoised heart sound signal is normalized, and then the normalized signal is divided into several heart sound segments of equal length according to the starting position of the first heart sound in the heart sound signal.

[0046] S103: Input the heart sound segment into the generative adversarial network model, train the model, and obtain the trained generative network and the trained discriminative network;

[0047] S104: Obtain Gaussian white noise randomly generated through a normal distribution, and input the Gaussian white noise into the trained generative network to obtain enhanced heart sound data.

[0048] This invention utilizes a trained generative network to generate more data to meet the needs of training a classification model.

[0049] Furthermore, the acquisition of heart sound signals involves collecting the heart sound signals of the test subject in a resting state as training data for the generative adversarial network model.

[0050] For example, the subject lies supine on the experimental bed. The signal acquisition personnel attach a heart sound transducer to the subject, placing it in the mitral valve auscultation area (medial to the fifth intercostal space on the left midclavicular line) and securing it with an elastic band. Then, the staff sets the operating parameters of the RM-6280C multifunctional physiological experimental system to a sampling frequency of 1000Hz. Signal acquisition is then checked to ensure normal signal reception. The subject is instructed to remain calm during the acquisition process and refrain from any limb movements. After all preparations are complete, the record button is clicked to acquire heart sound signals for five minutes, observing and recording any abnormalities during the acquisition process. Then, the end button is clicked, and the signal is exported and stored.

[0051] Furthermore, the filtering and noise reduction process employs a Chebyshev bandpass filter with a frequency range of 10-400Hz.

[0052] For example, since the frequency range of the heart sound signal is 20-200Hz, while the sampling frequency of the acquisition device is 1000Hz, according to the Nyquist sampling theorem, a Chebyshev bandpass filter with a frequency range of 10-400Hz is used to remove low-frequency and high-frequency noise signals from the heart sound signal.

[0053] Furthermore, the normalization process employs a maximum-minimum value normalization algorithm, the specific formula of which is as follows:

[0054]

[0055] Where x(n) is the original signal and y(n) is the signal after normalization.

[0056] Furthermore, the normalized signal is divided into several equal-length heart sound segments based on the starting position of the first heart sound in the heart sound signal, specifically including:

[0057] The normalized original signal is then subjected to Hilbert transform to obtain the analytic signal of the original signal; the amplitude of the analytic signal is the Hilbert envelope of the original signal.

[0058] The normalized original signal is processed by the average Shannon energy to obtain the average Shannon energy envelope.

[0059] Both the Hilbert envelope and the average Shannon energy envelope are normalized by the maximum-minimum method to obtain the first and second intermediate signals after normalization.

[0060] Multiply the first intermediate signal and the second intermediate signal to obtain the envelope signal of the heart sound signal;

[0061] The maximum amplitude of the envelope signal of the heart sound signal is multiplied by the set threshold parameter a to obtain the set threshold Th;

[0062] For points where the signal amplitude is greater than the set threshold Th, the local maximum algorithm is used to select the segmentation point, and the selected segmentation point is regarded as the starting position of the first heart sound in the heart sound signal.

[0063] It should be understood that the local maximum algorithm means that if a sample point x(n) is known, and the value of x(n-1) is less than x(n), and the value of x(n+1) is less than x(n), then x(n) is a local maximum point.

[0064] It should be understood that the Hilbert envelope is obtained from the Hilbert transform, and the formula for the Hilbert transform is:

[0065]

[0066] Where x(t) is the original signal and h(t) is the signal after Hilbert transform.

[0067] The analytic signal of the original signal is obtained by using the Hilbert transform:

[0068] s(t) = x(t) + jh(t)

[0069] The amplitude of s(t)

[0070]

[0071] It is the Hilbert envelope of the signal.

[0072] The average Shannon energy envelope is obtained from the Shannon energy envelope of the signal, and the specific formula is as follows:

[0073]

[0074] Where x(t) is the original signal, y(t) is the average Shannon energy envelope, and N is the number of data points for mean filtering, which is set to 10 in this invention.

[0075] Both envelopes are simultaneously subjected to maximum-minimum normalization, and then the two signals are multiplied to obtain the envelope signal. A threshold Th is then set, and the formula for Th is as follows:

[0076] Th=a*M

[0077] Where M is the maximum value of the envelope signal, and a is the threshold parameter, which is set to 0.5 in this invention. Finally, the local maximum algorithm is used to select the segmentation point for sample points that are greater than the threshold.

[0078] Furthermore, such as Figure 2 As shown, the generative adversarial network model has the following network structure:

[0079] Interconnected generative and discriminative networks;

[0080] The generative network includes: six linear function layers connected in sequence, wherein the first five linear function layers use the ReLU function as the activation function, and the last linear function layer uses the tanh function as the activation function;

[0081] The discrimination network comprises five linear function layers connected in sequence, wherein the first four linear function layers use the LeakyReLU function as the activation function, and the last linear function layer does not use an activation function.

[0082] Generative networks generate samples that closely resemble real samples from noisy data, while discriminative networks compare real and generated samples to determine the quality of the generative network by judging whether the generated samples are genuine or not. The basic structure of a generative network is as follows: Figure 3 As shown, the generator network consists of six linear function layers, with the first five layers using the ReLU function as the activation function and the last layer using the tanh function. The basic structure of the discriminator network is as follows. Figure 4 As shown, the discriminant network consists of 5 linear function layers, with the first four layers using the LeakyRelu function with a parameter of 0.2 as the activation function, and the last layer not using an activation function.

[0083] Further, in step S103: the heart sound fragment is input into the generative adversarial network model, the model is trained, and the trained generative network and the trained discriminative network are obtained. The specific training process includes:

[0084] S103-1: Several Gaussian white noise samples are randomly generated through a normal distribution. The Gaussian white noise samples are then processed by the generation network to obtain several generated samples.

[0085] S103-2: Using heart sound segments as real samples, input both real and generated samples into the discrimination network. Calculate the loss value based on the loss function of the discrimination network, optimize the parameters of the discrimination network, and repeat the training N times, where N is a positive integer.

[0086] S103-3: After the discriminator network is trained N times, the loss value is calculated based on the loss function of the generator network, and the parameters of the generator network are optimized.

[0087] S103-4: Repeat S103-1 to S103-3 until the accuracy of identifying real samples and generated samples using other classification algorithms is equal to 50%, and the training is complete;

[0088] Other classification algorithms include: K-nearest neighbor algorithm, support vector machine (SVM) algorithm, or random forest algorithm.

[0089] Furthermore, the loss function of the discrimination network is:

[0090]

[0091] Among them, P r P represents the distribution of the real sample; g The distribution of generated samples is represented by D(x); the output of the discrimination network is represented by D(x). The gradient of the discriminant network is represented by ||·||2, which represents the L2 norm. express Distribution;

[0092] This represents the expected output value of the real sample after passing through the discrimination network, and N represents the number of real samples;

[0093] This represents the expected output value of the generated sample after passing through the discrimination network, and N represents the number of generated samples;

[0094] Let N represent the gradient penalty term. Number of samples;

[0095] The gradient of the discrimination network is represented by x1…x n Represents the components of each dimension of sample x;

[0096] in, ε represents a random number that follows a uniform distribution in the range [0,1]. rLet x represent a randomly selected real sample. g This represents a randomly selected generated sample. The optimization function chosen is the Adam optimization method.

[0097] The first two terms of the loss function formula for the discriminant network represent the Wasserstein distance between the real sample distribution and the generated sample distribution. The smaller the values ​​of the first two terms, the closer it is to fitting the Wasserstein distance between the real distribution and the generated distribution.

[0098] Wasserstein distance is a measure of the distance between two distributions. Its advantage is that even if the two distributions do not overlap, Wasserstein distance can still reflect their proximity.

[0099] The last term of the loss function of the discriminator network is called the gradient penalty term, which restricts the gradient of the discriminator network to a certain range. This makes the gradient highly controllable and easy to adjust to a suitable scale.

[0100] Furthermore, the loss function of the generator network is:

[0101]

[0102]

[0103] The loss function of the generator network is the negative of the second term of the loss function of the discriminator network. The smaller the loss function of the generator network, the greater the difference between the distribution of real samples and the distribution of generated samples. The Adam optimization method is also chosen as the optimization function.

[0104] Using training samples with different labels, corresponding generative networks can be trained separately. For a class with a small number of samples, the generative network can generate a sufficient number of samples to make the number of each class consistent and solve the problem of data imbalance.

[0105] To verify the effectiveness of the heart sound data generated by this invention, the following experiment was conducted:

[0106] A convolutional neural network classification method was used to classify real samples using generated data as training data. The effectiveness of the heart sound data was verified by the classification accuracy of the real samples.

[0107] The training set used 14,000 generated samples, including 7,000 normal heart sounds and 7,000 abnormal heart sounds. The test set contained 4,819 real heart sound samples, including 2,460 normal samples and 2,359 abnormal samples. Each sample lasted 2 seconds, and the sampling frequency was 1000 Hz. The real heart sound samples were obtained through laboratory sampling.

[0108] The overall experimental steps include:

[0109] The samples are filtered and denoised; the Mel-frequency cepstral coefficients (MFCCs) of the samples are extracted; the MFCC features are used as input data to train the neural network model; and the recognition accuracy of the neural network model is tested.

[0110] The classification network uses a 5-layer convolutional neural network, consisting of a first convolutional layer, a second convolutional layer, a third convolutional layer, a first linear layer, and a second linear layer connected in sequence; the ReLU activation function is chosen for each layer. The Adam optimization method is used to optimize the neural network parameters, and the cross-entropy function is used as the loss function of the neural network.

[0111] Experimental results are as follows Figure 5 As shown in the confusion matrix, the classification model accurately identified 2177 normal samples and 1936 abnormal samples. The model's precision was 87.2%, recall was 82.1%, and overall accuracy was 85.3%. The experimental results demonstrate that the heart sound data generated by this invention is effective, and the heart sounds generated by the generative adversarial network have features similar to real heart sounds. This method utilizes generative adversarial networks to generate a sufficient amount of heart sound data to meet the needs of training deep learning models.

[0112] Example 2

[0113] This embodiment provides a heart sound data enhancement system based on generative adversarial networks;

[0114] A heart sound data augmentation system based on generative adversarial networks includes:

[0115] The acquisition module is configured to acquire heart sound signals.

[0116] The preprocessing module is configured to: filter and reduce noise in the acquired heart sound signal, then normalize the noise-reduced heart sound signal, and then divide the normalized signal into several heart sound segments of equal length according to the starting position of the first heart sound in the heart sound signal.

[0117] The training module is configured to: input heart sound segments into the generative adversarial network model, train the model, and obtain the trained generative network and the trained discriminative network;

[0118] The output module is configured to: acquire Gaussian white noise randomly generated by a normal distribution, input the Gaussian white noise into the trained generative network, and obtain enhanced heart sound data.

[0119] It should be noted that the acquisition module, preprocessing module, training module, and output module described above correspond to steps S101 to S104 in Embodiment 1. The examples and application scenarios implemented by these modules and their corresponding steps are the same, but they are not limited to the content disclosed in Embodiment 1. It should also be noted that these modules, as part of the system, can be executed in a computer system, such as a set of computer-executable instructions.

[0120] The descriptions of each embodiment in the above embodiments have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0121] The proposed system can be implemented in other ways. For example, the system embodiments described above are merely illustrative, and the division of modules described above is only a logical functional division. In actual implementation, there may be other division methods. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not executed.

[0122] Example 3

[0123] This embodiment also provides an electronic device, including: one or more processors, one or more memories, and one or more computer programs; wherein, the processor is connected to the memory, and the one or more computer programs are stored in the memory. When the electronic device is running, the processor executes the one or more computer programs stored in the memory to cause the electronic device to perform the method described in Embodiment 1.

[0124] It should be understood that in this embodiment, the processor can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0125] Memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of memory may also include non-volatile random access memory. For example, memory may also store information about the device type.

[0126] In the implementation process, each step of the above method can be completed by the integrated logic circuits in the processor hardware or by instructions in software.

[0127] The method in Embodiment 1 can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor. The software modules can reside in readily available storage media in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, a detailed description is not provided here.

[0128] Those skilled in the art will recognize that the units and algorithm steps described in connection with the various examples of this embodiment can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this invention.

[0129] Example 4

[0130] This embodiment also provides a computer-readable storage medium for storing computer instructions, which, when executed by a processor, complete the method described in Embodiment 1.

[0131] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for enhancing heart sound data based on generative adversarial networks, characterized by: include: Acquire heart sound signals; The acquired heart sound signal is filtered and denoised. Then, the denoised heart sound signal is normalized. Then, the normalized signal is divided into several heart sound segments of equal length according to the starting position of the first heart sound in the heart sound signal. Heart sound segments are input into a generative adversarial network model, and the model is trained to obtain a trained generative network and a trained discriminative network. Gaussian white noise, randomly generated through a normal distribution, is obtained and input into the trained generative network to obtain enhanced heart sound data. Heart sound fragments are input into a generative adversarial network (GAN) model for training, resulting in a trained generative network and a trained discriminative network. The specific training process includes: (1): Several Gaussian white noise samples are randomly generated by normal distribution, and the Gaussian white noise samples are processed by the generating network to obtain several generated samples; (2): Using heart sound segments as real samples, input the real samples and generated samples into the discrimination network at the same time, calculate the loss value according to the loss function of the discrimination network, optimize the parameters of the discrimination network, and repeat the training N times, where N is a positive integer; (3): After the discriminant network is trained N times, the loss value is calculated based on the loss function of the generator network, and the parameters of the generator network are optimized. (4): Repeat (1) to (3) until the accuracy of identifying real samples and generated samples using other classification algorithms is equal to 50%, and the training is complete; Other classification algorithms include: K-nearest neighbor algorithm, support vector machine algorithm, or random forest algorithm; The loss function of the discrimination network is: ； in, P r Represents the distribution of the real samples; P g This represents the distribution of the generated samples; D This represents the output function of the discrimination network. This represents the gradient of the discrimination network; Represents the L2 norm; express Distribution; This represents the expected output value of a real sample after it has passed through the discrimination network. N Indicates the number of real samples; This represents the expected output value of the generated sample after passing through the discrimination network. N Indicates the number of samples generated; Represents the gradient penalty term. N express Number of samples; The gradient of the discrimination network is represented by... x 1… x n Indicates sample x The components of each dimension; in, ε represents a random number that follows a uniform distribution in the range [0,1]. x r This represents a randomly selected real sample. x g This represents a randomly selected generated sample.

2. The method for enhancing heart sound data based on generative adversarial networks as described in claim 1, characterized in that, The normalized signal is then divided into several equal-length heart sound segments based on the starting position of the first heart sound. Specifically, this includes: The normalized original signal is then subjected to Hilbert transform to obtain the analytic signal of the original signal; the amplitude of the analytic signal is the Hilbert envelope of the original signal. The normalized original signal is processed by the average Shannon energy to obtain the average Shannon energy envelope. Both the Hilbert envelope and the average Shannon energy envelope are normalized by the maximum-minimum method to obtain the first and second intermediate signals after normalization. Multiply the first intermediate signal and the second intermediate signal to obtain the envelope signal of the heart sound signal; The maximum amplitude of the envelope signal of the heart sound signal is compared with the set threshold parameter. a Multiply to obtain the set threshold. Th ; For signal amplitude greater than a set threshold Th The local maximum algorithm is used to select the segmentation point, and the selected segmentation point is regarded as the starting position of the first heart sound in the heart sound signal.

3. The method for enhancing heart sound data based on generative adversarial networks as described in claim 1, characterized in that, The generative adversarial network model has the following network structure: Interconnected generative and discriminative networks; The generative network includes: six linear function layers connected in sequence, wherein the first five linear function layers use the ReLU function as the activation function, and the last linear function layer uses the tanh function as the activation function; The discrimination network comprises five linear function layers connected in sequence, wherein the first four linear function layers use the LeakyReLU function as the activation function, and the last linear function layer does not use an activation function.

4. The method for enhancing heart sound data based on generative adversarial networks as described in claim 1, characterized in that, The loss function of the generator network is: ； ； The loss function of the generator network is the negative of the second term of the loss function of the discriminator network. The smaller the loss function of the generator network, the greater the difference between the distribution of real samples and the distribution of generated samples.

5. The method for enhancing heart sound data based on generative adversarial networks as described in claim 1, characterized in that, The filtering and noise reduction process uses a Chebyshev bandpass filter with a frequency range of 10-400Hz.

6. A heart sound data enhancement system based on generative adversarial networks, characterized by: include: The acquisition module is configured to acquire heart sound signals. The preprocessing module is configured to: filter and reduce noise in the acquired heart sound signal, then normalize the noise-reduced heart sound signal, and then divide the normalized signal into several heart sound segments of equal length according to the starting position of the first heart sound in the heart sound signal. The training module is configured to: input heart sound segments into the generative adversarial network model, train the model, and obtain the trained generative network and the trained discriminative network; The output module is configured to: acquire Gaussian white noise randomly generated by a normal distribution, input the Gaussian white noise into the trained generative network, and obtain enhanced heart sound data; Heart sound fragments are input into a generative adversarial network (GAN) model for training, resulting in a trained generative network and a trained discriminative network. The specific training process includes: (1): Several Gaussian white noise samples are randomly generated by normal distribution, and the Gaussian white noise samples are processed by the generating network to obtain several generated samples; (2): Using heart sound segments as real samples, input the real samples and generated samples into the discrimination network at the same time, calculate the loss value according to the loss function of the discrimination network, optimize the parameters of the discrimination network, and repeat the training N times, where N is a positive integer; (3): After the discriminant network is trained N times, the loss value is calculated based on the loss function of the generator network, and the parameters of the generator network are optimized. (4): Repeat (1) to (3) until the accuracy of identifying real samples and generated samples using other classification algorithms is equal to 50%, and the training is complete; Other classification algorithms include: K-nearest neighbor algorithm, support vector machine algorithm, or random forest algorithm; The loss function of the discrimination network is: ； in, P r Represents the distribution of the real samples; P g This represents the distribution of the generated samples; D This represents the output function of the discrimination network. This represents the gradient of the discrimination network; Represents the L2 norm; express Distribution; This represents the expected output value of a real sample after it has passed through the discrimination network. N Indicates the number of real samples; This represents the expected output value of the generated sample after passing through the discrimination network. N Indicates the number of samples generated; Represents the gradient penalty term. N express Number of samples; The gradient of the discrimination network is represented by... x 1… x n Indicates sample x The components of each dimension; in, ε represents a random number that follows a uniform distribution in the range [0,1]. x r This represents a randomly selected real sample. x g This represents a randomly selected generated sample.

7. An electronic device, characterized in that it comprises: Memory is used to store computer-readable instructions in a non-transitory manner. as well as Processor, for executing the computer-readable instructions, When the computer-readable instructions are executed by the processor, they perform the method described in any one of claims 1-5.

8. A storage medium, characterized in that, The computer-readable instructions are stored non-transitory, wherein when the non-transitory computer-readable instructions are executed by a computer, the instructions of the method according to any one of claims 1-5 are executed.

Images