CN116250844B - Electrocardiosignal noise reduction optimization method and system based on condition generation countermeasure network - Google Patents

Abstract

The invention discloses an electrocardiosignal noise reduction optimization method and system based on a condition generation countermeasure network, comprising the following steps: acquiring an electrocardiosignal, dividing the electrocardiosignal according to the sample length T, and carrying out maximum and minimum normalization processing on the divided data fragments; inputting the processed data segment into a trained electrocardiosignal noise reduction model to obtain a noise-reduced electrocardiosignal; the electrocardiosignal noise reduction model is constructed based on a deep neural network and a condition generation countermeasure network; and determining the optimal values of modeling parameters such as the sample length, the number of layers of the neural network of the encoder and the decoder and the like according to the fitting function curve of the corresponding modeling parameters and the index on the premise of ensuring the noise reduction performance of the model by calculating the noise reduction performance and the calculation complexity ratio index of the model, so that the noise reduction performance and the calculation cost of the noise reduction model are optimal. The invention provides a noise reduction system with better cost performance by providing a noise reduction performance and calculation complexity ratio index.

Description

ECG signal denoising optimization method and system based on conditional generative adversarial network

Technical Field

The present invention relates to the technical field of electrocardiogram signal noise reduction, and in particular to an electrocardiogram signal noise reduction optimization method and system based on a conditional generative adversarial network.

Background technique

The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

Electrocardiogram (ECG) is a technology that reflects heart activity by collecting electrophysiological signals at fixed locations on the body surface. It has been widely used in the diagnosis of heart disease and is one of the gold standards for medical diagnosis of heart disease. However, the acquisition process of ECG signals (hereinafter referred to as ECG signals) is easily interfered by various noises such as body activity and circuit noise, which in turn affects the use of ECG signals for the diagnosis of heart disease.

Traditional ECG signal denoising methods, such as filters, Fourier transforms, wavelet decomposition, etc., are generally targeted at a single type of noise, and the time-frequency characteristics of the noise should be clearly different from the ECG signal itself. With the development of deep learning, there are more and more deep learning denoising algorithms for various noises, such as denoising autoencoders, convolutional autoencoders, etc. Among them, adversarial denoising based on fully connected networks and generative adversarial denoising based on convolutional autoencoders have, to a certain extent, solved the problems in the above algorithms, such as the need to segment ECG signals according to heart beats before denoising and the inability to handle mixed noise.

However, the above algorithms still have limitations: the model learning data distribution potential is not fully explored, the model lacks reasonable and systematic optimization, the model calculation complexity is high, and it is difficult to deploy on telemedicine equipment.

for example:

The prior art discloses methods for denoising ECG signals based on generative adversarial networks, wherein the generator networks mostly adopt fully connected, convolutional and other network structures. However, regardless of whether a convolutional neural network (CNN) or a bidirectional long short-term memory network (BiLSTM) is used for modeling, due to the lack of an optimization method for modeling an efficient denoising network model, the models established by these prior arts have a complex structure, a large scale, and a high computational complexity, and have stringent requirements on the deployment environment.

Summary of the invention

In order to solve the above problems, the present invention proposes an ECG signal denoising optimization method and system based on conditional generative adversarial network, constructs an ECG signal denoising model architecture based on deep neural network and conditional generative adversarial network, and proposes an indicator that takes into account both denoising performance and computational overhead. The internal structure, learning framework, data segmentation, model complexity and other aspects of the denoising model are systematically optimized and designed, so that better denoising effect can be achieved with lower computational overhead.

In some embodiments, the following technical solutions are adopted:

An electrocardiogram signal denoising optimization method based on a conditional generative adversarial network comprises:

Acquire an electrocardiogram signal, segment the electrocardiogram signal according to a sample length T, and perform maximum and minimum value normalization processing on the segmented data segments;

Input the processed data fragments into the trained ECG signal denoising model to obtain the denoised ECG signal;

Among them, the ECG signal denoising model is constructed based on deep neural network and conditional generative adversarial network, including generator and discriminator; for each modeling parameter in the ECG signal denoising model, denoising models with different values of modeling parameters (usually ≥3 cases) are established at a certain step size to obtain the corresponding denoising performance and computational complexity ratio index (The ratio of Signal-to-NoiseRatio to Computational Complexity, SNR-CC), and the fitting function is used to predict the corresponding optimal value of each modeling parameter, and through verification, it is determined that the denoising performance and computational overhead of the model are optimal;

The modeling parameters include: at least one or more of the sample length T, the number of deep neural network layers N of the encoder and decoder in the generator and the number of neurons _Lx in each layer, and the number of deep neural network layers M of the discriminator and the number of neurons _Rx in each layer.

As a further solution, the generator is composed of a denoising autoencoder (DAE) composed of a deep neural network, the DAE includes an encoder composed of N layers of deep neural networks and a decoder composed of N layers of deep neural networks, and the discriminator in the training process is composed of a binary classification deep neural network composed of M layers of deep neural networks;

The input of the generator is an ECG signal data segment of length T, and the output is a denoised signal; the discriminator performs anti-game learning together with the generator during the training of the ECG signal denoising model.

As a further solution, the loss function of the generator adds the difference l _dist between the denoised ECG signal and the original ECG signal, and the maximum local error l _max between the denoised ECG signal and the original ECG signal, based on the loss function of the conditional generative adversarial network.

As a further solution, the denoising performance and computational complexity ratio indicator of the ECG signal denoising model is specifically: the ratio of the average signal-to-noise ratio of the denoised signal during model testing to the time required for the model to process a single data sample.

As a further solution, the specific process of using the fitting function to predict the optimal value of the sample length T is as follows:

The ECG signal is segmented according to the sample length T, and the data is normalized to obtain a data set corresponding to the sample length T;

Set multiple different sample lengths T to obtain data sets corresponding to each sample length T;

Other modeling parameters are fixed unchanged, and the denoising model is modeled and trained according to the set T value and the corresponding data set;

For the trained ECG signal denoising model, the denoising performance and computational complexity ratio of the model are calculated respectively; and the fitting function f(T) is obtained by curve fitting based on different sample lengths T and corresponding SNR-CC data, and then the fitting function is derived to obtain f′(T), and the value of T when f′(T)= ₀ is calculated. The database is re-established using T ₀ , and the denoising model is trained, tested and verified. If the average signal-to-noise ratio of the signal after denoising by the model is not less than the set expected value, the current sample length T ₀ is the optimal value of T.

At the same time, when optimizing modeling parameters, other indicators that may affect model deployment can also be combined to optimize the model, such as the memory size occupied by model parameters, etc. Combined with the above modeling parameter optimization process, the memory size occupied by model parameters is restricted, and the corresponding piecewise fitting function can be obtained, which is then analyzed and predicted to determine the optimal modeling parameter value within the segment.

For the number of deep neural network layers N of the encoder and decoder in the generator, and the number of neurons L _x in each layer of the deep neural network in the generator, the same optimization method as the sample length T value is used, the non-optimized modeling parameters are fixed, and the optimal values of the corresponding modeling parameters to be optimized are selected.

For the number of layers M of the deep neural network in the discriminator and the number of neurons R _x in each layer of the deep neural network in the discriminator, the same optimization method as the sample length T is used to select the corresponding optimal value;

After the optimization parameters of the deep neural network in the discriminator are optimized, the denoising model is rebuilt together with the optimized generator parameters to further verify the denoising effect of the optimized parameters until the model denoising performance and computational complexity ratio index stabilizes near the maximum value position.

Different types of deep neural networks and conditional generative adversarial networks are selected to construct corresponding ECG signal denoising models. The above-mentioned optimization methods are used to optimize the modeling parameters of the corresponding denoising models, and the denoising performance of the optimized corresponding models and the ratio of denoising performance to computational complexity are compared. The denoising model with the best performance is selected as the final ECG signal denoising model.

In other embodiments, the following technical solutions are adopted:

An electrocardiogram signal denoising optimization system based on conditional generative adversarial network, comprising:

A data acquisition module is used to acquire an ECG signal, segment the ECG signal according to a sample length T, and perform maximum and minimum value normalization processing on the segmented data segments;

A denoising module, used to input the processed data segments into a trained ECG signal denoising model to obtain a denoised ECG signal;

Among them, the ECG signal denoising model is constructed based on deep neural networks and conditional generative adversarial networks, including generators and discriminators. For each modeling parameter in the ECG signal denoising model, different values are selected according to the set step size, and denoising models with different modeling parameters are established respectively, and the corresponding denoising performance and computational complexity ratio indicators are obtained. The fitting function is used to predict the optimal value of each modeling parameter, so that the denoising performance and computational overhead of the model are optimized.

The modeling parameters include: at least one or more of the sample length T, the number of deep neural network layers N of the encoder and decoder in the generator and the number of neurons _Lx in each layer, and the number of deep neural network layers M of the discriminator and the number of neurons _Rx in each layer.

In other embodiments, the following technical solutions are adopted:

A terminal device comprises a processor and a memory, wherein the processor is used to implement instructions; the memory is used to store multiple instructions, wherein the instructions are suitable for being loaded by the processor and executing the above-mentioned electrocardiogram signal denoising optimization method based on conditional generative adversarial network.

Compared with the prior art, the present invention has the following beneficial effects:

In summary, an indicator that is currently lacking and can take into account both computational complexity and denoising performance is established and proposed, as well as a method for fully, reasonably and systematically optimizing the model based on the indicator. This can solve the problems of long modeling time, high computational complexity of the denoising model, and high computational overhead that arise when implementing the above-mentioned existing denoising methods.

(1) The present invention combines a deep neural network with a conditional generative adversarial network to achieve noise reduction optimization of ECG signals; wherein the deep neural network can learn the distribution of real data in the sample, thereby removing noise; the conditional generative adversarial network is an improved generative adversarial network, which combines the noisy signal as a condition with the original input signal (the denoised signal or the original signal) as the input of the discriminator, which enables the generator and the discriminator to better engage in adversarial game, thereby controlling the learning direction of the generator to be closer to the real data distribution in the sample.

(2) The present invention proposes a noise reduction performance and computational complexity ratio index to evaluate the model noise reduction performance and model computational overhead, and uses the proposed index to perform curve fitting with the experimental data of the corresponding modeling parameters to obtain the fitting function and derive it, thereby predicting the optimal value of the modeling parameter, and then verifying the value of each modeling parameter to be optimized through experiments. This optimization method can use a small amount of experiments to achieve combined optimization of model parameters, so that the model can achieve the smallest possible model computational overhead while achieving the same noise reduction performance.

(3) The loss function of the noise reduction model generator of the present invention is based on the loss function of the conventional CGAN network, and also adds the distance between the noise reduction signal and the pure ECG signal and the maximum local error. The distance between the noise reduction signal and the pure ECG signal reflects the overall difference between the noise reduction signal and the pure ECG signal, and can control the general direction of the data generated by the generator; while the maximum local error between the noise reduction signal and the pure ECG signal reflects the local difference between the noise reduction signal and the pure signal that needs to be improved, that is, the signal details that need to be improved when the generator generates data. When the above two parts are added to the loss function of the generator, the data generation process of the generator can be controlled in terms of the overall direction and local details, thereby obtaining a better noise reduction effect.

Other features and advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through the practice of the present invention.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a schematic diagram of an optimization process of an ECG signal denoising model in an embodiment of the present invention;

FIG2 is a schematic diagram of the overall structure of an ECG signal denoising model according to an embodiment of the present invention;

FIG3 is a schematic diagram of a fitting function curve of a modeling parameter T and SNR-CC in an embodiment of the present invention;

FIG4 is a schematic diagram of an electrocardiogram signal denoising process based on a conditional generative adversarial network in an embodiment of the present invention.

Detailed ways

It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used in the present invention have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and/or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and/or combinations thereof.

Embodiment 1

In one or more embodiments, a method for optimizing electrocardiogram signal noise reduction based on a conditional generative adversarial network is disclosed, which specifically includes the following:

(1) obtaining an ECG signal, segmenting the ECG signal according to a sample length T, and performing maximum and minimum value normalization processing on the segmented data segments;

(2) inputting the processed data segments into the trained ECG signal denoising model to obtain the denoised ECG signal;

The ECG signal denoising model is constructed based on a deep neural network and a conditional generative adversarial network. The denoising performance and computational complexity ratio of the ECG signal denoising model is calculated, and the optimal modeling parameter value is predicted by curve fitting. The modeling parameter is verified to determine whether the denoising performance and computational overhead of the model are optimal. The modeling parameter may be one or more of the sample length T, the number of deep neural network layers N of the encoder and decoder and the number of neurons _Lx in each layer, the number of deep neural network layers M of the discriminator and the number of neurons _Rx in each layer.

In this embodiment, in combination with Figure 2, the ECG signal denoising model is constructed based on a combination of a deep neural network (such as a bidirectional long short-term memory network, a convolutional neural network, a recurrent neural network, etc.) and a conditional generative adversarial network; the ECG signal denoising model consists of two parts: a generator and a discriminator. The generator can be composed of a denoising autoencoder (DAE) composed of a deep neural network, and the DAE includes an encoder composed of N layers of deep neural networks and a decoder composed of N layers of deep neural networks. The discriminator is composed of a binary classification deep neural network composed of M layers of deep neural networks.

The input of the generator is a noisy signal of length T The output is the signal after noise reduction/> The input of the discriminator is a combination of the noisy signal and the denoised signal or a combination of the noisy signal and the original signal, and the output is true or false.

The training and optimization process of the ECG signal denoising model is described below in conjunction with FIG1 .

a) Signal noise addition.

Select a relatively clean ECG signal from a public or specific ECG database as the original signal (x _n ), and select a variety of common ECG noises from the ECG noise database as the noise signal (e _n ). The sampling frequency of the original signal and the noise signal is f. According to different input signal-to-noise ratios (SNR), multiple single noises and multiple mixed noises are superimposed on the original signal to obtain a noisy signal.

b) Setting of modeling parameter set to be optimized.

The set of modeling parameters C to be optimized for the ECG signal denoising model includes: generator modeling parameters (sample length T, number of encoder and decoder deep neural network layers N, number of neurons in each layer L _x ), and discriminator modeling parameters (number of discriminator deep neural network layers M, number of neurons in each layer R _x ).

First, the modeling parameters to be optimized of the above-mentioned denoising model should be preset according to the following rules.

The sample length T determines both the amount of information input to the model and the number of neurons in the starting layer of the generator and discriminator (i.e., the starting size of the model). The heart rate usually ranges from 50 to 150 beats per minute, i.e., the cycle of a single heartbeat is 0.4 to 1.2 seconds. When the sampling rate is f, in order to ensure that each sample segment contains at least one heartbeat cycle, T should usually be selected from multiple values within the range of ≥1.2f for optimization.

In order to make the generator and discriminator of the denoising model better able to compete against each other during training, so as to achieve better denoising effect. Generally speaking, the neural network structures used by the generator and the discriminator should be similar, and M should be greater than or equal to N. That is, a suitable N should be selected first according to the characteristics of the deep neural network that constitutes the generator, and then M should be selected as a value greater than or equal to N, and the number of neurons (L _x , R _x ) of the corresponding network layer should be set according to the selected deep neural network type.

c) Data segmentation and normalization: The noisy signal is segmented according to the sample length (T). The original signal (x _n ) is segmented to obtain ECG signal data segments of corresponding lengths, and the maximum and minimum values are normalized to form the corresponding data set (D _T ). In the data set (D _T ), C% of the total data is used for training, and 1-C% of the data is used for testing.

Among them, the calculation formula for maximum and minimum normalization of data is as follows:

Wherein, _xn is the value of the nth sampling point of the signal to be normalized, _xmin is the minimum value among the sampling points of the signal to be normalized, and _xmax is the maximum value among the sampling points of the signal to be normalized.

d) Model training.

Using the prepared data set (D _T ), the constructed ECG signal denoising model is trained through adversarial game learning. When the discriminator cannot distinguish whether the input is the original signal or the denoised signal, it is considered that the adversarial learning has reached a Nash equilibrium and the model has converged.

e) Noise reduction model test: Use the trained generator to input the noisy signal in the data set to obtain the denoised signal, and calculate its signal-to-noise ratio (SNR), root mean square error (RMSE) and other noise reduction performance indicators for noise reduction effect evaluation. Subsequently, the number of parameters of the noise reduction model at this time, the memory occupied by the parameters, the amount of calculation and other parameters are statistically used to evaluate the computational consumption (overhead) of the model. By comprehensively comparing the model noise reduction performance indicators and model calculation overhead under different values of the modeling parameters to be optimized, the values of the modeling parameters to be optimized are preferred.

In order to facilitate the comprehensive comparison of the above-mentioned noise reduction performance and computational complexity indicators, the ratio of noise reduction performance to computational complexity (SNR-CC) is proposed in this embodiment, and its calculation method is as follows:

Among them, O is the value of SNR-CC, which represents the ratio between the model denoising performance and the computational complexity. The larger the index is, the smaller the computational time (computational overhead) consumed when the denoised signal reaches the same signal-to-noise ratio. SNR _Average is the average signal-to-noise ratio of the denoised signal during model testing. t _s is the time required for the trained denoising model (using the generator part) to calculate a single sample, which also reflects the comprehensive influence of the parameters of the denoising model, the memory occupied by the parameters, the amount of computation, and other indicators to a certain extent. When SNR _Average is greater than or equal to the expected value S ₀ , the SNR-CC of the denoising model is as large as possible. Among them, S ₀ is the expected value of the denoising performance of the optimized model, that is, the average SNR, which can be obtained by referring to the performance indicators of other advanced denoising algorithms.

When optimizing the algorithm model architecture, firstly, curve fitting is performed based on the modeling parameter p and the corresponding SNR-CC data to obtain the fitting function f(p). Figure 3 shows a schematic diagram of the fitting function for different values of sample length T and its corresponding SNR-CC value. The fitting function is derived to obtain f′(p). Then, when f′(p)＝0, the value of p is obtained by calculation p _0. The common third-order polynomial fitting function and its derivative are shown below:

O＝f(p)＝ap ³ +bp ² +cp+d(3)

O′＝f′(p)＝3ap ² +2bp+c(4)

Then, the model is re-modeled and trained and tested according to the modeling parameter p ₀ , and the signal-to-noise ratio SNR value at this time and the SNR-CC value O ₀ corresponding to this parameter are calculated to verify the noise reduction performance and calculation loss of the model at this time. That is, when f′(p)＝0 and SNR _Average ≥S ₀ , it is considered to be the optimal value of the modeling parameter under the current model situation.

At this point, if there are requirements on the memory size occupied by the denoising model parameters, the obtained fitting function f(p) should be transformed into a piecewise function according to its specific restrictions.

O = f(p) = ap ³ + bp ² + cp + d (SNR _Average ≥ S ₀ and MC ≤ MC ₀ ) (5)

Where MC is the memory size occupied by the model parameters, and MC ₀ is the maximum value of the memory occupied by the model parameters. When SNR _Average ≥ S ₀ and MC ≤ MC ₀ , p _min, which is the minimum value that f′(p) can take, is considered to be the optimal value of the modeling parameter under the current model situation.

In addition, the denoising model constructed when the modeling parameters to be optimized take the optimal values should be used to denoise the test data, and then the ECG classification algorithm based on the characteristics of the ECG signal should be used to evaluate the classification performance of the original ECG signals before and after denoising, in order to evaluate the ability of the denoising model to retain information of medical value.

f) Generator and discriminator model optimization.

During denoising, only the generator part of the trained denoising model is used. Therefore, T and N are two important parameters in the model optimization process. However, since M can affect the performance of the discriminator and thus respond to the game situation during the entire training process, the M value must also be optimized.

After analysis, the values of T, N, M, the number of neurons in each deep neural network layer of the generator (L _x ) and the number of neurons in each deep neural network layer of the generator (R _x ) will greatly affect the denoising effect of the model. Therefore, it is necessary to optimize the modeling parameters of the above models.

The general optimization process of modeling parameters is as follows: fix the current non-optimized modeling parameters unchanged, establish denoising models with different values of the modeling parameters to be optimized (usually ≥3 cases) according to a certain step size, obtain the corresponding denoising performance and computational complexity ratio index according to the above steps b)-e), use the fitting function to predict the optimal value of the modeling parameter to be optimized, and then build a training denoising model with the optimal value, and test the comprehensive performance of the denoising model for verification. Then, optimize the remaining modeling parameter values in order.

Since the model ultimately uses the generator to achieve noise reduction, the generator model should be optimized first. Specifically, the following scheme can be adopted: first, the value of the modeling parameter T is optimized, and after the optimization is completed, the corresponding data set (D _T ) is used to cycle the above optimization steps and optimize the generator model (i.e., the value of N and L _{x )} in sequence within a certain range.

Then, the values of the modeling parameters (M, R _x ) to be optimized in the discriminator are optimized in the same way. Since the values of N and M can affect the performance of the generator and the discriminator in the adversarial game respectively, there is a certain mutual influence between the two. Therefore, first determine the preferred value of N in the generator optimization, and then after the optimization parameters of the deep neural network in the discriminator are optimized, the denoising model is rebuilt together with the optimized generator parameters to further verify the denoising effect of the optimized parameters until the model denoising performance and computational complexity ratio index tends to be stable at a high level (near the maximum value position).

Finally, the optimal values of T, N, M, L _x , and R _x and the trained denoising algorithm optimization model are obtained.

After the ECG signal denoising model is trained and optimized, when implementing the denoising process, the denoising process is performed through the generator part of the optimized model; the specific process is as follows:

The actual collected ECG signal data is segmented into fixed lengths according to the preferred T value, and then the segmented data segments are normalized to the maximum and minimum values to obtain the noisy data to be processed. Subsequently, it is input into the generator part of the trained denoising algorithm optimization model for calculation, and then the denoised signal is obtained.

Embodiment 2

The ECG denoising model of this embodiment is constructed by a bidirectional long short-term memory network (BiLSTM) and a conditional generative adversarial network (CGAN). The model adopts the learning architecture of CGAN and consists of a generator and a discriminator composed of a bidirectional long short-term memory network.

The initial structure of the model is shown in Figure 2, where the generator is composed of an encoder consisting of 2 layers (N) BiLSTM, a decoder consisting of 2 layers (N) BiLSTM, and a denoising autoencoder consisting of a fully connected layer. The discriminator is composed of a binary classification deep learning network consisting of 3 layers (M) BiLSTM, a fully connected layer, and an activation function Sigmoid. The input sample length of the generator and discriminator is T. The number of neurons in the encoder and decoder in the generator is L _x = {400, 200}, and the number of neurons in the discriminator is R _x = {400, 200, 100}. The calculation formula of the activation function Sigmoid is as follows:

The input of the generator is a noisy signal, and the output is a denoised signal; the input of the discriminator is a combination of a noisy signal and a denoised signal or a combination of a noisy signal and an original signal, and the output is true or false.

In this embodiment, the loss function of the ECG denoising model generator is based on the loss function of the conventional CGAN network, and further adds the difference l _dist between the denoised signal and the pure ECG signal and the maximum local error l _max .

The loss function of the generator is specifically:

in, is a noisy signal, obeying the noisy signal distribution/> The generator input is a noisy signal/> Output when, /> The discriminator input is /> and/> is the output when , λ ₁ and λ ₂ are the weight coefficients of l _dist and l _max , which are 0.7 and 0.2 respectively.

The calculation formula of the difference l _dist between the denoised signal and the pure ECG signal is as follows:

in, is the signal after noise reduction, _xn is the original signal, and T is the sample length.

The calculation formula of the maximum local error l _max between the denoised signal and the pure ECG signal is as follows:

At this time, the overall loss function of the model is:

Where x is the original signal, which obeys the original signal distribution p _data (x); is a noisy signal, which obeys the distribution of noisy signals

The model is implemented using Pytorch programming, and the RMSProp (Root Mean Square Prop) optimizer is used for parameter update, and the learning rate is set to 0.0001. Model training and testing are run on a Dell T640 server (NVIDIA GTX 3090 24GB). The model training and optimization process is the same as in Example 1, and the specific implementation process is as follows:

a) Select ECG signals numbered 103, 105, 111, 116, 122, 205, 213, 219, 223, and 230 from the MIT-BIH arrhythmia database as the original signals (x _n ). Select three common ECG noises, muscle artifacts, electrode motion artifacts, and baseline drift, from the ECG noise database as noise signals (e _n ). According to setting the input SNR to 0dB, 1dB, 2dB, 3dB, 4dB, and 5dB, three single noises, three mixed noises superimposed on each other, and one mixed noise superimposed on three noises are respectively superimposed on the original signal to obtain noisy signals with different input SNRs and different types of noises. The sampling rate of the above ECG signal is 360 Hz, so the sample length T should be ≥ 432 sampling points.

b) According to the optimization strategy, the modeling parameter set C to be optimized is first pre-set. According to the data feature extraction capability of the BiLSTM network, T=512, N=2, M=3, _Lx ={400, 200} of the generator, and _Rx ={400, 200, 100} of the discriminator are tentatively set.

c) The noisy signals and original signals with different input SNRs and noises are segmented according to the corresponding T value (512 in this case) to obtain ECG signal data segments of different lengths, and their maximum and minimum values are normalized respectively to form a corresponding data set (D ₅₁₂ ). 80% of the data in the data set is used for training and 20% of the data is used for testing.

d) The ECG denoising model of this embodiment is trained by using different data sets through the adversarial game learning process to continuously optimize the generator and the discriminator. When the discriminator cannot distinguish whether the input is the original signal or the denoised signal, or when the SNR of the generated denoised signal is high enough or the RMSE is low enough, the model converges and the training is completed.

e) At this time, after statistics and calculation, the performance of the noise reduction model constructed with the modeling parameters (T=512) is as follows: SNR=28.71dB, SNR-CC=97.32dB/ms. The model expected noise reduction performance index S ₀ =28dB is set, and the model noise reduction performance SNR>S ₀ .

f) Generator model optimization: First, taking the optimization of T value as an example, the process is as follows: set multiple sample length (T) values (such as 512, 1024, 2048, 4096, etc.), obtain data sets ( _DT ) of various sample lengths, fix other modeling parameters to build a denoising model, and use the data set ( _DT ) to train the denoising model. Then use the trained generator to input the corresponding data set ( _DT ) for testing the noisy signal to obtain the denoised signal, calculate its signal-to-noise ratio (SNR), SNR-CC denoising performance and computational overhead indicators, and perform curve fitting based on T value and SNR-CC value. The fitting results are shown in Figure 3. When f′(T)＝0, the fitting result T ₀ ＝2127. For the convenience of calculation, T ₀ ＝2048 can be taken here. After remodeling with T ₀ , comprehensive comparison of various indicators showed that the model denoising performance SNR＞S ₀ reached a good level (SNR＝36.21dB), SNR-CC＝109.74dB/ms, which verified that its computational overhead was small. Subsequently, an ECG classification algorithm based on ECG signal features was used to compare the classification performance of the original ECG signals before and after denoising to evaluate the ability of the denoising model to retain medically valuable information.

Repeat the above steps, and through comprehensive comparison of noise reduction performance indicators, the ability to retain medical information, and the computational overhead of the model, that is, calculating and analyzing SNR-CC, modeling parameters, and SNR-CC fitting functions and their derivatives, predict and optimize other modeling parameters (N and L _x ) in the generator. Determine the optimal values of N and L _x in turn.

Then, repeat the above steps, and then optimize M and R _x in turn by fixing the values of the modeling parameters to be optimized of the generator, that is, optimize the discriminator model. After determining the optimal value of M, repeat the above steps to optimize the value of N again to determine the effectiveness of N.

The results show that when the value range of N is {1, 2, 3, 4} and the value range of M is {3, 4, 5, 6}, it is concluded that when N=3, M=4, the generator's _Lx ={800, 400, 200}, and the discriminator's _Rx ={800, 400, 200, 100}, the average SNR of the denoised ECG signal is 45.12dB, and the model SNR-CC is 131.54dB/ms. At this time, the computational consumption of the model is lower than that of other algorithms with approximate denoising effects, achieving a good balance between computational consumption and denoising effect, that is, the SNR-CC value of the model can reach a high and stable level. In addition, an ECG classification algorithm (support vector machine) based on ECG signal features was used to evaluate the four-category performance of abnormal ECG before, after and after denoising. The classification accuracy of the denoised signal using the proposed model was close to that of the original signal (≥95%), which confirmed that the denoised signal retained a large amount of medically valuable information.

Combined with Figure 4, for the trained ECG denoising model, the ECG signal data is collected and segmented according to a fixed length of T = 2048, and then the segmented data segments are normalized to obtain the noisy ECG signal segments to be processed. Then, the segments are input into the generator in the trained and optimized ECG denoising model for denoising, and the denoised signal can be obtained.

Embodiment 3

In this embodiment, two noise reduction models are constructed respectively:

The first type: a denoising model based on convolutional neural networks and conditional generative adversarial networks;

The second type: a denoising model based on a bidirectional long short-term memory network and a conditional generative adversarial network.

First, based on the same hardware computing platform or system, the two denoising models are optimized using the method proposed in Example 1 to obtain optimized denoising models composed of different deep neural networks. Then, the denoising performance and SNR-CC and other indicators of the two optimized denoising models are compared to determine the better denoising model.

The first noise reduction model is constructed under the optimal modeling parameters after optimization. After testing, its SNR=41.03dB and SNR-CC=2.28dB/ms.

The second noise reduction model is constructed under the optimal modeling parameters after optimization. After testing, its SNR=45.12dB and model SNR-CC=131.54dB/ms.

It can be seen that the denoising model constructed based on the bidirectional long short-term memory network and the conditional generative adversarial network has better denoising performance and denoising cost-effectiveness than the denoising model constructed based on the convolutional neural network and the conditional generative adversarial network.

Embodiment 4

In one or more embodiments, a system for optimizing electrocardiogram signal noise reduction based on a conditional generative adversarial network is disclosed, comprising:

A data acquisition module is used to acquire an ECG signal, segment the ECG signal according to a length T, and perform maximum and minimum value normalization processing on the segmented data segments;

A denoising module, used to input the processed data segments into a trained ECG signal denoising model to obtain a denoised ECG signal;

Among them, the ECG signal denoising model is constructed based on deep neural networks and conditional generative adversarial networks, including generators and discriminators. For each modeling parameter in the ECG signal denoising model, different values are selected according to the set step size, and denoising models with different modeling parameters are established respectively, and the corresponding denoising performance and computational complexity ratio indicators are obtained. The fitting function is used to predict the optimal value of each modeling parameter, so that the denoising performance and computational overhead of the model are optimized.

The modeling parameters include: at least one or more of the sample length T, the number of deep neural network layers N of the encoder and decoder in the generator and the number of neurons _Lx in each layer, and the number of deep neural network layers M of the discriminator and the number of neurons _Rx in each layer.

It should be noted that the specific implementation methods of the above modules have been described in detail in Example 1 and will not be described in detail here.

Embodiment 5

In one or more embodiments, a terminal device is disclosed, including a server, the server including a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the electrocardiogram signal denoising optimization method based on a conditional generative adversarial network in Embodiment 1 when executing the program. For the sake of brevity, it will not be described in detail here.

It should be understood that in this embodiment, the processor may be a central processing unit CPU, and the processor may also be other general-purpose processors, digital signal processors DSP, application-specific integrated circuits ASIC, off-the-shelf programmable gate arrays FPGA or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

The memory may include a read-only memory and a random access memory, and provide instructions and data to the processor. A portion of the memory may also include a non-volatile random access memory. For example, the memory may also store information about the device type.

In the implementation process, each step of the above method can be completed by an integrated logic circuit of hardware in a processor or an instruction in the form of software.

Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Technical personnel in the relevant field should understand that various modifications or variations that can be made by technical personnel in the field without creative work on the basis of the technical solution of the present invention are still within the scope of protection of the present invention.

Claims (9)

1. An electrocardiosignal noise reduction optimization method based on a condition generation countermeasure network is characterized by comprising the following steps:

acquiring an electrocardiosignal, dividing the electrocardiosignal according to a sample length T, and carrying out maximum and minimum normalization processing on the divided data fragments;

inputting the processed data segment into a trained electrocardiosignal noise reduction model to obtain a noise-reduced electrocardiosignal;

The electrocardiosignal noise reduction model is constructed based on a deep neural network and a condition generation countermeasure network and comprises a generator and a discriminator; for each modeling parameter in the electrocardiosignal noise reduction model, selecting different values according to a set step length, respectively establishing a noise reduction model of the modeling parameter under different values, and obtaining a corresponding noise reduction performance and calculation complexity ratio index of the noise reduction model, wherein the noise reduction performance and calculation complexity ratio index of the electrocardiosignal noise reduction model is specifically as follows: the average signal-to-noise ratio of the noise-reduced signal in the model test and the ratio of the time required by the model to process a single data sample are used for predicting the optimal value of each modeling parameter by adopting a fitting function so as to optimize the noise reduction performance and the calculation cost of the model;

The modeling parameters include: at least one or more of a sample length T, a number N of deep neural network layers of an encoder and a decoder in the generator, and a number L _x of neurons per layer, and a number M of deep neural network layers of the arbiter, and a number R _x of neurons per layer.

2. The method for optimizing the noise reduction of an electrocardiograph signal based on a condition generating countermeasure network according to claim 1, wherein the generator is composed of a noise reduction self-encoder composed of a deep neural network, and the noise reduction self-encoder comprises an encoder composed of an N-layer deep neural network and a decoder composed of an N-layer deep neural network; the discriminator consists of two classified deep neural networks consisting of M layers of deep neural networks;

the input of the generator is an electrocardiosignal data segment with the sample length T, and the electrocardiosignal data segment is output as a noise-reduced signal; the discriminator is combined with the generator to perform the countermeasure game learning when the electrocardiosignal noise reduction model is trained.

3. The method for optimizing the noise reduction of an electrocardiograph signal based on a condition generating countermeasure network according to claim 2, wherein the loss function of the generator increases a difference l _dist between the electrocardiograph signal after noise reduction and the original electrocardiograph signal and a maximum local error l _max between the electrocardiograph signal after noise reduction and the original electrocardiograph signal based on the condition generating countermeasure network.

4. The method for optimizing the noise reduction of an electrocardiographic signal based on a condition generating countermeasure network according to claim 1, wherein the specific process of predicting the optimal value of the sample length T by using a fitting function is as follows:

setting values of a plurality of different sample lengths T, and respectively obtaining a data set corresponding to each sample length T;

fixing other modeling parameters unchanged, and respectively modeling and training the noise reduction model according to the set T value and the corresponding data set;

For the trained electrocardiosignal noise reduction model, respectively calculating the noise reduction performance and the calculation complexity ratio index of the model; constructing a fitting function of the sample length T and the ratio index of the noise reduction performance and the computational complexity;

Deriving the fitting function, and calculating the value of the corresponding sample length T when the derivative of the fitting function is zero; and performing model noise reduction test based on the value of the sample length T, and if the average signal-to-noise ratio of the noise reduced signal is not less than the set expected value, obtaining the value of the sample length T as an optimal value.

5. The method for optimizing noise reduction of electrocardiosignal based on condition generation countermeasure network as claimed in claim 4, wherein the number N of deep neural network layers of encoder and decoder in the generator and the number L _x of neurons of each deep neural network layer in the generator are respectively optimized by adopting the same optimization method as the value of the sample length T, and the corresponding optimal value is selected.

6. The method for optimizing electrocardiosignal noise reduction based on a condition generation countermeasure network as claimed in claim 5, wherein the number M of layers of deep neural networks in the discriminator and the number R _x of neurons of each layer of deep neural networks in the discriminator are respectively optimized by adopting an optimization method which is the same as the value of the sample length T, and the corresponding optimal value is selected;

After optimization of the optimized parameters of the deep neural network in the discriminator is completed, the noise reduction model is reconstructed by combining the optimized generator parameters, and the noise reduction effect of the optimized parameters is further verified until the noise reduction performance and the calculation complexity ratio index of the model tend to be stable near the maximum value position.

7. The method for optimizing the noise reduction of an electrocardiograph signal based on a condition-generating countermeasure network of claim 1, further comprising:

Selecting different kinds of deep neural networks and condition generation countermeasure networks to construct corresponding electrocardiosignal noise reduction models; the optimization method of claim 1 is adopted to optimize modeling parameters of the corresponding noise reduction model, noise reduction performance and a ratio index of the noise reduction performance to the computational complexity of the optimized corresponding model are compared, and the noise reduction model with the optimal performance is selected to be used as a final electrocardiosignal noise reduction model.

8. An electrocardiosignal noise reduction optimization system based on a condition generation countermeasure network, which is characterized by comprising:

the data acquisition module is used for acquiring electrocardiosignals, dividing the electrocardiosignals according to the sample length T, and carrying out maximum and minimum normalization processing on the divided data fragments;

The noise reduction module is used for inputting the processed data fragments into a trained electrocardiosignal noise reduction model to obtain a noise-reduced electrocardiosignal;

The electrocardiosignal noise reduction model is constructed based on a deep neural network and a condition generation countermeasure network and comprises a generator and a discriminator; for each modeling parameter in the electrocardiosignal noise reduction model, selecting different values according to a set step length, respectively establishing a noise reduction model of the modeling parameter under different values, and obtaining a corresponding noise reduction performance and calculation complexity ratio index of the noise reduction model, wherein the noise reduction performance and calculation complexity ratio index of the electrocardiosignal noise reduction model is specifically as follows: the average signal-to-noise ratio of the noise-reduced signal in the model test and the ratio of the time required by the model to process a single data sample are used for predicting the optimal value of each modeling parameter by adopting a fitting function so as to optimize the noise reduction performance and the calculation cost of the model;

The modeling parameters include: at least one or more of a sample length T, a number N of deep neural network layers of an encoder and a decoder in the generator, and a number L _x of neurons per layer, and a number M of deep neural network layers of the arbiter, and a number R _x of neurons per layer.

9. A terminal device comprising a processor and a memory, the processor for implementing instructions; the memory for storing a plurality of instructions, wherein the instructions are adapted to be loaded by the processor and to perform the condition-based generation of an electrocardiographic signal noise reduction optimization method of any one of claims 1-7.