CN111434305A - Fetal electrocardiogram extraction system and method based on convolutional coding and decoding neural network - Google Patents

Abstract

The invention discloses a fetal electrocardiogram extraction system based on a convolutional coding and decoding neural network and a method thereof. The system comprises a data acquisition device: used to collect real abdominal electrical signals of pregnant women; a maternal electrocardiogram component estimation device: used for The convolutional codec neural network is used to estimate the maternal ECG component in the abdominal electrical signal of pregnant women. During training, the simulated abdominal electrical signal is used as the input of the neural network, and the maternal ECG component in the simulated abdominal electrical signal is used as the network label for training. , during the test, the real abdominal electrical signals of pregnant women are used as the input of the neural network, and the output of the neural network is the estimated maternal ECG components in the abdominal electrical signals; The maternal ECG component obtained above is subtracted from the electrical signal, thereby extracting the fetal ECG component in the abdominal electrical signal of the pregnant woman. The method includes data preprocessing, estimation of maternal ECG components and extraction of fetal ECG components. Through the technical solution, the efficiency and accuracy of fetal ECG extraction can be effectively improved.

Description

Fetal ECG extraction system and method based on convolutional codec neural network

technical field

The invention relates to the technical field of fetal ECG extraction, in particular to a fetal ECG extraction system based on a convolutional codec neural network, and further provides a method for using the system.

Background technique

Fetal heart rate monitoring is a means of fetal intrauterine monitoring, which can reflect the biophysical activity of the fetus in real time, and is widely used in clinical practice. plays an important role in reducing perinatal mortality. Perinatal mortality to some extent reflects the comprehensive economic development and health care conditions of a country and region.

The fetal ECG signal can not only be used to calculate the fetal heart rate, but also provide more morphological information, which records the movement of the fetal heart and objectively reflects the various states of the fetal intrauterine physiological activity. It can judge the developmental degree, position, acidosis or arrhythmia of the fetus, so as to obtain the current health status of the fetus. One of the difficulties in extracting the fetal ECG signal is that the abdominal electrical signal contains the mother's ECG component. The mother's ECG component usually has a larger amplitude than that of the fetus, and the mother's ECG component is at the same time as the fetal ECG component. Both the domain and the frequency domain overlap, therefore, it brings great interference to the extraction of fetal ECG. Therefore, it is a challenging task to extract fetal ECG signals from the mixed signals in the maternal abdomen.

SUMMARY OF THE INVENTION

In order to overcome the deficiencies of the prior art, the technical problem solved by the present invention is to provide a fetal ECG extraction system based on a convolutional codec neural network, which can improve extraction efficiency and accuracy, and a method for using the system.

In order to solve the above-mentioned first technical problem, the content of the technical solution adopted in the present invention is as follows:

A fetal ECG extraction system based on convolutional codec neural network, including the following devices:

Data acquisition device: used to collect real abdominal electrical signals of pregnant women;

Maternal ECG component estimation device: used to construct the convolutional codec neural network to be trained. During training, the simulated abdominal electrical signal is used as the input of the neural network, and the maternal ECG component in the simulated abdominal electrical signal is used as the network label for volume Integrate the encoder-decoder neural network for training and change the parameters in the neural network (ie, the weight coefficient and the amount of deviation), until the error of the neural network is completely converged and the training is terminated (the error is completely converged. Compare the maternal ECG components in the known simulated abdominal electrical signals until there is no error or the error is within a reasonable range), and save all parameters to obtain a trained convolutional codec neural network unit. When measuring the real abdominal electrical signals of pregnant women, the real single-channel pregnant abdominal electrical signals collected by the data acquisition device are substituted into the trained convolutional codec neural network unit as the input of the neural network, and the obtained output of the neural network is Estimated maternal ECG components in abdominal electrical signals.

Wherein, the convolutional coding and decoding neural network in the execution program of the maternal ECG component estimation device is composed of multiple convolution-deconvolution modules and a fully connected module in series; the convolutional layer in the convolution-deconvolution module The convolution kernel size is 1×3, 1×4 or 1×5.

Preferably, in the technical solution, the convolutional encoder-decoder neural network is trained by using the gradient descent method and the back-propagation algorithm. The simulated abdominal electrical signal is obtained according to the data in the existing simulation database, and can be divided into a training set and a verification set.

When using the gradient descent method and the back-propagation algorithm to train the convolutional encoder-decoder neural network, during normal operation, the simulated pregnant abdominal electrical signals of the training set are preprocessed and input into the convolutional encoder-decoder neural network, Train the convolutional encoder-decoder neural network, evaluate the maternal ECG components through the loss function, and then calculate the gradient of all parameters in the convolutional neural network through the back-propagation algorithm (gradient descent method), and change according to the gradient (each Derivative of parameters) to update all parameters, re-substitute the training set into the convolutional encoder-decoder neural network for training, until the iteration is completed to find the model with the smallest loss function, save all parameters and keep the model. When it is necessary to test the real abdominal electrical signals of pregnant women, the convolutional neural network model is called, and the real abdominal electrical signals of pregnant women collected by the data acquisition device are preprocessed and input into the convolutional codec neural network to obtain the estimated maternal heart rate. Then, the estimated maternal ECG component obtained above is subtracted from the actual collected abdominal electrical signal of the pregnant woman, so as to extract the fetal ECG component in the abdominal electrical signal of the pregnant woman.

The device for extracting fetal ECG components is used to subtract the estimated maternal ECG components obtained above from the collected abdominal electrical signals of pregnant women, so as to extract the fetal ECG components in the abdominal electrical signals of pregnant women.

It should be noted that, in order to improve the efficiency and accuracy of fetal ECG extraction during the detection process, the inventor provides an extraction system in this technical solution, which is convenient for operators to realize signal processing through the system and facilitate subsequent detection. and diagnosis and treatment. In this technical solution, the inventor innovatively uses a convolutional coding and decoding neural network to estimate the maternal ECG components in the abdominal electrical signals of pregnant women, and subtracts the above-obtained maternal cardiac electrical signals from the collected abdominal electrical signals of pregnant women. The scheme of extracting the fetal ECG components in the abdominal electrical signals of pregnant women, the system separately sets up a data acquisition device, a maternal ECG component estimation device and a fetal ECG component extraction device to realize the above process.

It should be noted that, in the data acquisition device, the electrical signal of the pregnant woman's abdomen will also be converted into a digital signal, which is convenient for processing on a digital device such as a computer.

Preferably, the data acquisition device also executes the following procedures:

After the pregnant woman's abdominal electrical signal is processed by processing procedures such as amplification and/or filtering, it is then converted into a digital signal by an A/D conversion processing procedure.

More preferably, the data acquisition device also executes the following procedures:

Crop the digital signal to a size that matches the input of the convolutional codec neural network. Preferably, the cropped size of the digital signal is 1×1000, and the value in each 1×1 signal is the electrocardiogram value (amplitude) of a certain point, and 1000 has 1000 points, and each point represents a different time point. The value (amplitude) of the ECG.

By cropping to a matching size, the following training of neural networks using convolutional encoders and decoders can be performed more efficiently.

Preferably, the device for estimating the maternal ECG components executes the following procedures during the training process:

The network model is trained using an open full-layer network parameter update and using the mean squared error as the loss function. That is, the present application adopts the back-propagation algorithm (gradient descent method), and the gradient of all parameters in the convolutional neural network needs to be calculated during calculation, and all parameters are updated according to the gradient change (derivative of each parameter).

Preferably, the loss function is:

为网路输出的对母体心电成分的预估，X为网络标签，即仿真腹部电信号中的母体心电成分。采用此损失函数，可以完整而快速地学习到母体心电的特征。Among them, M is the number of training samples,

is the estimation of the maternal ECG component output by the network, and X is the network label, that is, the maternal ECG component in the simulated abdominal electrical signal. Using this loss function, the characteristics of the maternal ECG can be learned completely and quickly.

Preferably, during training, the maximum number of iterations is 10,000, and the learning rate is 0.0001.

Preferably, the convolutional coding and decoding neural network in the execution program of the maternal ECG component estimation device is composed of a plurality of convolution-deconvolution modules and a fully connected module in series;

Among them, the fully-connected module directly connected to the output layer consists of one or more fully-connected layers, and the convolution-deconvolution module connected to the fully-connected module consists of a convolutional layer, a deconvolutional layer and a nonlinear layer In addition to the composition, all other convolution-deconvolution modules include one convolution layer, one deconvolution layer and two nonlinear layers; the two adjacent convolution-deconvolution modules are directly connected. The way. Preferably, the activation function of the nonlinear layer is a tanh function. Preferably, the size of the convolution kernel of the convolution layer is 1×3, 1×4 or 1×5. Ensure that the convolution kernel is relatively suitable, and will not cause too much local feature extraction, nor will some local features be lost due to the size of the convolution kernel.

It should be noted that the convolutional encoding and decoding neural network adopts the above structure, which can effectively capture the characteristics of the maternal ECG, speed up the convergence speed of network training, and also has a good effect on the gradient dispersion problem. It should be noted that the direct connection method is adopted between two adjacent convolution-deconvolution modules, so that the features of the shallow layer can be transferred to the deep layer, thereby retaining more detailed features of the maternal ECG, and further Improve efficiency and accuracy.

Preferably, the system further includes a display device for splicing the extracted fetal ECG components into a complete signal and displaying it.

It should be noted that the method of cutting the signal into a shorter length for learning and splicing the network output into a complete signal after the learning is completed can enable the network to learn more detailed features and improve the effect of fetal ECG extraction.

Preferably, all parameters of the network are initialized with small random numbers before starting training. First of all, if the parameters are small, the network will not be saturated due to too large weights at the beginning, resulting in training failure. Second, the randomization parameters are implemented to ensure that the initial values of the parameters are different each time, so as to avoid getting the same parameters for each training. The value range of the small random number is between 0-1.

In order to solve the above-mentioned second technical problem, the content of the technical solution adopted in the present invention is as follows:

A method for extracting fetal ECG based on a convolutional codec neural network, comprising the following steps:

Data preprocessing: collecting abdominal electrical signals of pregnant women;

Maternal ECG component estimation: The convolutional codec neural network is used to estimate the maternal ECG component in the abdominal electrical signal of pregnant women, that is, the simulated abdominal electrical signal is used as the input of the neural network during training to simulate the maternal electrical component in the abdominal electrical signal. The ECG component is used as a network label for training, and the real single-channel pregnant abdominal electrical signal is used as the input of the neural network during testing, and the output of the neural network is the estimated maternal ECG component in the abdominal electrical signal;

Elimination of maternal ECG components: The maternal ECG components obtained above are subtracted from the collected abdominal electrical signals of pregnant women, so as to extract fetal ECG components in the abdominal electrical signals of pregnant women.

In order to improve the efficiency and accuracy of fetal ECG extraction, the inventor innovatively uses a convolutional codec neural network in this technical solution to estimate the maternal ECG components in the abdominal electrical signals of pregnant women, and use the collected data to estimate the maternal ECG components. The scheme of subtracting the maternal ECG components obtained above from the abdominal electrical signals of pregnant women, thereby extracting the fetal ECG components in the abdominal electrical signals of pregnant women. In this technical solution, a convolutional codec neural network is used, which has strong robustness and can still extract fetal ECG signals under interference conditions such as contact noise and breathing noise. Furthermore, the method also performs well when the signal-to-noise ratio between the fetal signal and the maternal signal is low and when the fetal QRS complex overlaps with the maternal QRS complex. Through this technical means, the relevant noise can be excluded to a greater extent, so that the extraction is more accurate and the efficiency is higher. The clear fetal ECG signal extracted by the invention can provide information such as fetal heart rate, and has important clinical application value and considerable social benefit.

More preferably, the digital signal is obtained through A/D conversion after processing including amplification and/or filtering.

Further, the data preprocessing step further includes trimming the digital signal to a size matching the input of the convolutional codec neural network.

Preferably, the training method in the estimation of the maternal ECG components is to use an open full-layer network parameter update method and use the mean square error as a loss function to train a convolutional encoder-decoder neural network model;

Preferably, the loss function is:

为网路输出的预估的母体心电成分，X为网络标签，即仿真腹部电信号中的母体心电成分。Among them, M is the number of training samples,

is the estimated maternal ECG component output by the network, and X is the network label, that is, the maternal ECG component in the simulated abdominal electrical signal.

Preferably, the convolutional coding and decoding neural network in the device for estimating maternal ECG components is composed of multiple convolution-deconvolution modules and a fully connected module in series;

Among them, the fully-connected module directly connected to the output layer consists of one or more fully-connected layers, and the convolution-deconvolution module connected to the fully-connected module consists of a convolutional layer, a deconvolutional layer and a nonlinear layer In addition to the composition, all other convolution-deconvolution modules include one convolution layer, one deconvolution layer and two nonlinear layers; the two adjacent convolution-deconvolution modules are directly connected. The way.

Preferably, in the technical solution, the convolutional encoder-decoder neural network is trained by using the gradient descent method and the back-propagation algorithm.

Preferably, the method further comprises splicing the extracted fetal ECG components into a complete signal and displaying it.

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

1. The fetal ECG extraction system based on the convolutional coding and decoding neural network of the present invention adopts the convolutional coding and decoding neural network. The fetal ECG signal can eliminate relevant noise to a greater extent, making the extraction more accurate and more efficient;

2. The present invention is based on a fetal ECG extraction system based on a convolutional coding and decoding neural network. The signal is cut into a shorter length for learning, and the network output is spliced into a complete signal after the learning is completed, so that the network can learn more details. Features, improve the effect of fetal ECG extraction;

3. The fetal ECG extraction method based on the convolutional coding and decoding neural network of the present invention, and the above-mentioned system is used to execute the corresponding process, and the extraction purpose can be achieved.

The above description is only an overview of the technical solutions of the present invention, in order to be able to understand the technical means of the present invention more clearly, it can be implemented according to the content of the description, and in order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand , the following specific preferred embodiments, and in conjunction with the accompanying drawings, are described in detail as follows.

Description of drawings

Fig. 1 is a schematic diagram of the process frame of a preferred embodiment of a fetal ECG extraction method based on a convolutional codec neural network of the present invention;

2 is a schematic diagram of a simulated abdominal signal and a corresponding maternal ECG component during training in one of the embodiments of the method of the present invention;

3 is a schematic diagram of the fetal ECG signal obtained according to the present method, the corresponding real abdominal signal and the maternal ECG component output by the network;

Figure 4 is a schematic diagram of the structure of a convolutional encoder-decoder neural network.

Detailed ways

In order to further illustrate the technical means and effects adopted by the present invention to achieve the predetermined purpose of the invention, below in conjunction with the accompanying drawings and preferred embodiments, the specific embodiments, structures, features and effects according to the present invention are described in detail as follows:

Embodiment 1 (a fetal ECG extraction system based on convolutional codec neural network)

A fetal ECG extraction system based on a convolutional codec neural network, comprising the following devices:

Data acquisition device: used to collect a route of abdominal electrical signals of pregnant women;

Maternal ECG component estimation device: It is used to estimate the maternal ECG component in the abdominal electrical signal of pregnant women by using the convolutional codec neural network, that is, the simulated abdominal electrical signal is used as the input of the neural network during training to simulate the abdominal electrical signal. The maternal ECG component is used as a network label for training, and a trained convolutional codec neural network is obtained; when measuring the real fetal ECG component, the real single-channel pregnant abdominal electrical signal is input into the trained convolutional codec. In the neural network, the estimated maternal ECG component in the real abdominal electrical signal is obtained through the trained convolutional encoder-decoder neural network. ;

The device for extracting fetal ECG components is used to subtract the maternal ECG components obtained above from the collected abdominal electrical signals of pregnant women, so as to extract the fetal ECG components in the abdominal electrical signals of pregnant women.

The above is the basic implementation of the technical solution. In order to improve the efficiency and accuracy of fetal ECG extraction in the detection process, the inventor provides an extraction system in this technical solution, which is convenient for operators to realize signal processing through the system and facilitate subsequent detection and diagnosis and treatment. In this technical solution, the inventor innovatively uses a convolutional coding and decoding neural network to estimate the maternal ECG components in the abdominal electrical signals of pregnant women, and subtracts the above-obtained maternal cardiac electrical signals from the collected abdominal electrical signals of pregnant women. The scheme of extracting the fetal ECG components in the abdominal electrical signals of pregnant women, the system separately sets up a data acquisition device, a maternal ECG component estimation device and a fetal ECG component extraction device to realize the above process. In the data acquisition device, the electrical signal of the pregnant woman's abdomen is also converted into a digital signal, which is convenient for processing on a digital device such as a computer.

In some embodiments, the data collection device further performs the following procedures:

After the electrical signal of the pregnant woman's abdomen is processed through a processing program of amplification and/or filtering, the electrical signal is converted into a digital signal through an A/D conversion program.

In some embodiments, the data acquisition device further performs the procedure of cropping the digital signal to a size matching the input of the convolutional codec neural network. By cropping to a matching size, the following training of neural networks using convolutional encoders and decoders can be performed more efficiently.

In some embodiments, the apparatus for estimating maternal ECG components performs the following procedure for the training process: using an open full-layer network parameter update manner and using mean square error as a loss function to train a network model.

In some embodiments, the loss function is:

为网路输出的对母体心电成分的预估，X为网络标签，即仿真腹部电信号中的母体心电成分。采用此损失函数，可以完整而快速地学习到母体心电的特征。Among them, M is the number of training samples,

is the estimation of the maternal ECG component output by the network, and X is the network label, that is, the maternal ECG component in the simulated abdominal electrical signal. Using this loss function, the characteristics of the maternal ECG can be learned completely and quickly.

In some embodiments, the convolutional codec neural network in the execution program of the device for estimating the maternal ECG component is composed of multiple convolution-deconvolution modules and a fully connected module in series; The fully-connected module consists of one or more fully-connected layers, and the convolution-deconvolution module connected to the fully-connected module consists of a convolutional layer, a deconvolutional layer and a nonlinear layer. All other convolutional layers -Deconvolution modules all contain one convolution layer, one deconvolution layer and two non-linear layers; two adjacent convolution-deconvolution modules are directly connected. The convolutional encoding and decoding neural network adopts the above structure, which can effectively capture the characteristics of the maternal electrocardiogram, accelerate the convergence speed of network training, and also has a good effect on the gradient dispersion problem. It should be noted that the direct connection method is adopted between two adjacent convolution-deconvolution modules, so that the features of the shallow layer can be transferred to the deep layer, thereby retaining more detailed features of the maternal ECG, and further Improve efficiency and accuracy.

Preferably, the cropped size of the digital signal is 1×1000, and the value in each 1×1 signal is the electrocardiogram value (amplitude) of a certain point, and 1000 has 1000 points, and each point represents a different time point. The value (amplitude) of the ECG. The convolution kernel size of the convolution layer is 1×3, 1×4 or 1×5, and the number of convolution kernels in each convolution layer is 32, 64 or 128, preferably 64.

In one of the embodiments, as shown in FIG. 4 , three convolution-deconvolution modules are included, namely the first convolution-deconvolution module, the second convolution-deconvolution module and the third convolution-deconvolution module. The deconvolution module, in the first convolution-deconvolution module, convolves the cropped digital signal with the convolution kernel, and the matrix obtained by the convolution then completes the calculation of the activation function through the nonlinear layer, and performs deconvolution After that, the calculation of the activation function is completed through the nonlinear layer, and the output of the first convolution-deconvolution module can be used as the input of the second convolution-deconvolution module or the third convolution-deconvolution module, while The output of the second convolution-deconvolution module can then be used as the input of the third convolution-deconvolution module. Finally, the output of the third convolution-deconvolution module is used as the input of the fully connected layer (dense), and the final output image is obtained through the processing of the fully connected layer. In this embodiment, the number of neurons in the fully connected layer is 1000.

In some embodiments, the system further includes a display device for splicing the extracted fetal ECG components into a complete signal and displaying it. The method of cutting the signal into a shorter length for learning, and splicing the network output into a complete signal after the learning is completed, can enable the network to learn more detailed features and improve the effect of fetal ECG extraction.

It should be understood that during normal operation, the convolutional codec neural network is trained by first preprocessing the simulated pregnant abdominal electrical signals of the training set and/or the validation set and then inputting it into the convolutional codec neural network. The maternal ECG components are evaluated, and then the gradient of all parameters in the convolutional neural network is calculated through the back-propagation algorithm (gradient descent method), and all parameters are updated according to the gradient change (derivative of each parameter), and the training set is reset. Substitute it into the convolutional encoder-decoder neural network for training, until the model with the smallest loss function is found after iteration, save all parameters and keep the model. When it is necessary to test the real abdominal electrical signals of pregnant women, the convolutional neural network model is called, and the real abdominal electrical signals of pregnant women are preprocessed and input into the convolutional codec neural network to obtain the estimated maternal ECG components, and then use the real The estimated maternal ECG component obtained above is subtracted from the collected abdominal electrical signal of the pregnant woman, so as to extract the fetal ECG component in the abdominal electrical signal of the pregnant woman.

Embodiment 2 (a kind of fetal ECG extraction method based on convolutional codec neural network)

As shown in Figure 1, the present invention provides a method for extracting fetal ECG based on a convolutional codec neural network, comprising the following steps:

Data preprocessing: collecting abdominal electrical signals of pregnant women;

Maternal ECG component estimation: The convolutional codec neural network is used to estimate the maternal ECG component in the abdominal electrical signal of pregnant women, that is, the simulated abdominal electrical signal is used as the input of the neural network during training to simulate the maternal electrical component in the abdominal electrical signal. The ECG component is used as a network label for training, and the real single-channel pregnant abdominal electrical signal is used as the input of the neural network during testing, and the output of the neural network is the estimated maternal ECG component in the abdominal electrical signal;

Extraction of fetal ECG components: The estimated maternal ECG components obtained above are subtracted from the collected abdominal electrical signals of pregnant women, thereby extracting fetal ECG components in the abdominal electrical signals of pregnant women.

The above is the basic embodiment of the present invention. In this technical solution, in order to improve the efficiency and accuracy of fetal ECG extraction, the inventor innovatively uses a convolutional codec neural network in this technical solution to estimate the maternal ECG components in the abdominal electrical signals of pregnant women, The plan of extracting the fetal ECG component in the abdominal electrical signal of the pregnant woman is obtained by subtracting the estimated maternal electrocardiographic component obtained above from the collected abdominal electrical signal of the pregnant woman. Convolutional codec neural network is used, which has strong robustness and can still extract fetal ECG signals under interference conditions such as contact noise and breathing noise. Furthermore, the method also performs well when the signal-to-noise ratio between fetal and maternal signals is low and when fetal and maternal QRS complexes overlap. Through this technical means, the relevant noise can be excluded to a greater extent, so that the extraction is more accurate and the efficiency is higher. The clear fetal ECG signal extracted by the invention can provide information such as fetal heart rate, and has important clinical application value and considerable social benefit.

Embodiment 3 (a kind of fetal ECG extraction method based on convolutional codec neural network)

The implementation process of the above-mentioned basic implementation manner is described below with reference to a specific implementation manner. However, this embodiment is only used to illustrate the technical solution, and does not represent a limitation on the protection scope of the technical solution.

A method for extracting fetal ECG based on a convolutional codec neural network, comprising the following steps:

1) Data preprocessing: First, as shown in Figure 3, use electrodes to collect a signal in the abdomen of the mother with a sampling frequency of 250Hz, and again, use a bandpass filter with a passband of 0.5-100Hz and a notch filter with 50Hz The signal is filtered, and secondly, the signal is amplified by an amplifying circuit and converted into a digital signal by an analog-to-digital conversion circuit. Finally, the filtered data is cut into a fixed size (1×1000).

In a more specific preferred embodiment, the signal samples are uniformly cropped into a fixed size, and the size matches the size of the input of the convolutional codec neural network used in the method (or directly collects the signal of the corresponding size when collecting the signal, such as 2 seconds or 4 seconds in length).

2) Estimation of maternal ECG components: train a convolutional encoder-decoder neural network to estimate the maternal ECG components in the abdominal electrical signals of pregnant women. In some preferred embodiments, a single-channel input data processing framework is used to extract the ECG sample data, the order of the input samples is randomly disrupted, and the sample data is input with a batch_size of 64. In the training, the simulated abdominal electrical signal is used as the input of the neural network, the maternal ECG component in the simulated abdominal electrical signal is used as the network label for training, and the real single-channel pregnant abdominal electrical signal is used as the input of the neural network during testing. The output is the estimated maternal ECG component in the abdominal electrical signal, as shown in Figure 3.

In some preferred embodiments, a convolutional codec neural network is used as a deep neural network model to estimate the maternal ECG components in the abdominal electrical signals of pregnant women. The adopted convolutional encoder-decoder neural network is composed of 5 convolution-deconvolution modules and a fully connected module in series; among them, the fully connected module directly connected to the output layer consists of a fully connected layer, which is connected to the fully connected module. Except that the convolution-deconvolution module consists of a convolutional layer, a deconvolutional layer and a nonlinear layer, all other convolutional-deconvolutional modules consist of a convolutional layer, a deconvolutional layer and a non-linear layer. Two nonlinear layers; at the same time, a direct connection is adopted between two adjacent convolution-deconvolution modules, so that the features of the shallow layers can be transferred to the deep layers.

In some preferred embodiments, the network model is trained using an open full-layer network parameter update approach and using mean squared error as a loss function.

In some more specific preferred embodiments, when training the convolutional codec neural network, the method of open full-layer network parameter update is adopted, and the mean square error is used as the loss function, and the function is expressed as:

为网路输出的预估的母体心电成分，X为网络标签，即仿真腹部电信号中的母体心电成分。Among them, M is the number of training samples,

is the estimated maternal ECG component output by the network, and X is the network label, that is, the maternal ECG component in the simulated abdominal electrical signal.

3) Extraction of fetal ECG components: subtracting the estimated maternal ECG components in the abdominal electrical signals from the collected abdominal electrical signals of pregnant women, so as to extract the fetal ECG components in the abdominal electrical signals of pregnant women.

More specifically, the estimated maternal ECG components output by the convolutional coding and decoding neural network are subtracted from the collected abdominal electrical signals, and the remaining part is the fetal ECG components, so as to realize the analysis of maternal abdominal ECG components. Extraction of fetal ECG components.

4) Display maternal ECG and fetal ECG: The extracted fetal ECG components are spliced into a complete signal, and the maternal ECG and fetal ECG are displayed, as shown in Figure 3.

In order to verify the effect of the present invention, the applicant has repeatedly substituted the test set into the trained convolutional codec neural network (the convolutional codec neural network is shown in Figure 4 at this time, and its convolution kernel size is 1 × 3). , the number of convolution kernels is 64), and the following data are obtained:

mean SE (%) mean PPV(%) mean F₁(%) trained neural network 93.08 91.21 92.13

TP是找对R峰(QRS峰的R峰)的数量，FN是漏掉R峰的数量，FP是找错R峰的数量，找到R峰的位置与真实的位置在50ms以内都算找对。in,

TP is the number of correct R peaks (the R peaks of QRS peaks), FN is the number of missing R peaks, FP is the number of wrong R peaks, and the position of the R peak found and the real position within 50ms are considered correct. .

The above-mentioned embodiments are only preferred embodiments of the present invention, and cannot be used to limit the scope of protection of the present invention. Any insubstantial changes and substitutions made by those skilled in the art on the basis of the present invention belong to the scope of the present invention. Scope of protection claimed.

Claims (10)

1. the fetal electrocardiogram extraction system based on convolutional codec neural network, is characterized in that, comprises following device: Data acquisition device: used to collect abdominal electrical signals of pregnant women; Maternal ECG component estimation device: used to construct the convolutional codec neural network to be trained, and train the convolutional codec to be trained according to the simulated abdominal electrical signals of pregnant women; when measuring the real fetal ECG components, the data collected The single-channel abdominal electrical signal of pregnant women collected by the device is input into the trained convolutional codec neural network, and the estimated maternal ECG component of the real abdominal electrical signal is obtained through the trained convolutional codec neural network; Device for extracting fetal ECG components: used to subtract the estimated maternal ECG components from the collected abdominal electrical signals of pregnant women, so as to extract the fetal ECG components in the abdominal electrical signals of pregnant women; Wherein, the convolutional coding and decoding neural network in the execution program of the maternal ECG component estimation device is composed of multiple convolution-deconvolution modules and a fully connected module in series; the convolutional layer in the convolution-deconvolution module The convolution kernel size is 1×3, 1×4 or 1×5. 2. The fetal ECG extraction system as claimed in claim 1, wherein the data acquisition device also executes the following procedures: After the electrical signal of the pregnant woman's abdomen is processed through amplification and/or filtering, the electrical signal is converted into a digital signal through an AD conversion program. 3. The fetal ECG extraction system as claimed in claim 2, wherein the data acquisition device also executes the following procedures: Crop the digital signal to a size that matches the input of the convolutional codec neural network. 4 . The fetal ECG extraction system according to claim 1 , wherein the simulated abdominal electrical signal used by the maternal ECG component estimation device during training is composed of simulated maternal ECG components and simulated fetal ECG components. 5 . and some noise superposition. 5. The fetal ECG extraction system according to claim 1, wherein the maternal ECG component estimation device performs the following procedures for the training process: Using the open full-layer network parameter update method and using the mean square error as the loss function to train the convolutional decoding neural network; Preferably, the convolutional encoder-decoder neural network is trained using a gradient descent method and a back-propagation algorithm. 6. The fetal ECG extraction system according to claim 5, wherein the loss function is:

Among them, M is the number of training samples,

is the estimated maternal ECG component output by the network, and X is the network label, that is, the maternal ECG component in the simulated abdominal electrical signal. 7. The fetal ECG extraction system according to claim 1, wherein the fully connected module is also directly connected to the output layer, and the fully connected module is composed of one or more fully connected layers, which are connected to the fully connected layer. The convolution-deconvolution module connected to the module consists of a convolution layer, a deconvolution layer and a nonlinear layer. All other convolution-deconvolution modules include a convolution layer and a deconvolution layer. layer and two nonlinear layers; a direct connection is used between two adjacent convolution-deconvolution modules. 8 . The fetal ECG extraction system according to claim 1 , wherein the system further comprises a display device for splicing the extracted fetal ECG components into a complete signal and displaying it. 9 . 9. The fetal ECG extraction method based on convolutional codec neural network, is characterized in that, comprises the following steps: Data collection: collect all the abdominal electrical signals of pregnant women; Maternal ECG component estimation: construct the convolutional codec neural network to be trained and use the simulated abdominal electrical signal to train the convolutional codec neural network to be trained; when measuring the real fetal ECG component, the real single-channel pregnant abdomen The electrical signal is input into the trained convolutional encoder-decoder neural network, and the estimated maternal ECG component in the real abdominal electrical signal is obtained through the trained convolutional encoder-decoder neural network. Extraction of fetal ECG components: subtract the estimated maternal ECG components from the collected abdominal electrical signals of pregnant women, so as to extract the fetal ECG components in the abdominal electrical signals of pregnant women; Preferably, the process of converting the pregnant woman's abdominal electrical signal into a digital signal includes amplifying and/or filtering processing procedures and A/D conversion procedures; More preferably, the data preprocessing step also includes trimming the digital signal into a size matching the input of the convolutional codec neural network; Preferably, the training method in the estimation of the maternal ECG component is to use an open full-layer network parameter update method and use the mean square error as a loss function to train the network model; Preferably, the convolutional coding and decoding neural network in the estimation of the maternal ECG components is composed of a plurality of convolution-deconvolution modules and a fully connected module in series; Among them, the fully-connected module directly connected to the output layer consists of one or more fully-connected layers, and the convolution-deconvolution module connected to the fully-connected module consists of a convolutional layer, a deconvolutional layer and a nonlinear layer In addition to the composition, all other convolution-deconvolution modules include one convolution layer, one deconvolution layer and two nonlinear layers; the two adjacent convolution-deconvolution modules are directly connected. The way. Preferably, the activation function of the nonlinear layer is a tanh function. Preferably, the method further comprises splicing the extracted fetal ECG components into a complete signal and displaying it; Preferably, in the technical solution, the convolutional encoder-decoder neural network is trained by using the gradient descent method and the back-propagation algorithm. 10. The fetal ECG extraction method according to claim 9, wherein the loss function is:

Among them, M is the number of training samples,

is the estimated maternal ECG component output by the network, and X is the network label, that is, the maternal ECG component in the simulated abdominal electrical signal.

Images