WO2020228420A1

WO2020228420A1 - Denoising autoencoder training method, electrocardiography signal denoising method, and apparatuses

Info

Publication number: WO2020228420A1
Application number: PCT/CN2020/080880
Authority: WO
Inventors: 王贵锦; 黄勇锋; 丁子建; 张宇
Original assignee: 清华大学; 华为技术有限公司
Priority date: 2019-05-14
Filing date: 2020-03-24
Publication date: 2020-11-19
Also published as: CN110141215B; CN110141215A

Abstract

A denoising autoencoder training method, an electrocardiography signal denoising method, and related apparatuses and devices, which use artificial intelligence to perform electrocardiography signal denoising, and may be applied to fields such as smart electrocardiogram detection. A reference electrocardiography signal to be denoised comprising significant features such as R peak positions and R-R distances of an electrocardiography signal to be denoised is extracted from said electrocardiography signal, and denoising is performed on a remaining electrocardiography signal to be denoised by means of a target denoising autoencoder after said reference electrocardiography signal has been removed from said electrocardiography signal, to prevent the target denoising autoencoder from performing denoising processing on the significant features in said electrocardiography signal, so that said reference electrocardiography signal and the denoised remaining electrocardiography signal are superimposed to obtain a denoised electrocardiography signal, wherein the R peak positions in said electrocardiography signal are better maintained, and distortion of the denoised electrocardiography signal is reduced.

Description

Training method of denoising autoencoder, denoising method and device of electrocardiogram signal

This application requires that it be submitted to the State Intellectual Property Office of China on May 14, 2019, the application number is 201910399830.9, and the application name is "Training method of noise reduction autoencoder, method of noise reduction of ECG signal and related devices and equipment" China The priority of the patent application, the entire content of which is incorporated in this application by reference.

Technical field

The present invention relates to the technical field of artificial intelligence, and in particular to a method for training a noise reduction autoencoder, a method for noise reduction of an electrocardiogram signal and a device thereof.

Background technique

With the development of artificial intelligence technology, it has gradually become possible to help doctors perform ECG diagnosis through artificial intelligence technology. The quality of ECG signals directly affects the accuracy of ECG signal diagnosis. The collection of ECG signals is usually obtained through electrodes attached to the surface of the skin. Since the ECG signal on the skin is relatively weak and easily interfered by noise, the collected ECG signal contains a lot of noise, which reduces the accuracy and reliability of the ECG diagnosis. In particular, the ECG collected by the wearable ECG device when the user is in a non-stationary state contains a large amount of myoelectric noise. At this time, the noise reduction processing of the ECG signal is particularly important.

In the prior art, the denoising of the ECG signal can be achieved by training a convolutional autoencoder (CAE). The specific training method is: a convolutional autoencoder that inputs noisy ECG signals, The convolution autoencoder processes the input noisy information signal and then outputs the predicted ECG signal, and adjusts the deconvolution auto-encoder according to the error between the predicted ECG signal and the noise-reduced ECG signal corresponding to the noisy ECG signal The parameters of the encoder cause the error to converge, and a target autoencoder capable of reducing the noise of the ECG signal is obtained, and further, the target autoencoder is used to reduce the noise of the ECG signal to be noise-reduced. However, since the ECG signal is directly used as input during the training process of the convolutional autoencoder, it is difficult for the autoencoder to obtain the encoding representation of the complete ECG signal, which leads to the denoised ECG signal through the target autoencoder. The detailed information in the ECG signal is lost, and the waveform is distorted. How to eliminate noise without losing the detailed information of the ECG signal is a technical problem that needs to be solved urgently.

Summary of the invention

The technical problem to be solved by the embodiments of the present invention is to provide a method for training a noise-reducing autoencoder, a method for reducing noise of an electrocardiogram signal, and a device thereof, so as to avoid the technical problems of difficult learning of the autoencoder and distortion of the predicted waveform.

In the first aspect, an embodiment of the present application provides a method for training a noise-reducing autoencoder, including: a training device superimposes a noise-free ECG signal and an EMG noise signal to obtain a synthetic ECG signal, the noise-free ECG signal including H beat ECG signal, each beat ECG signal contains a QRS complex, H is a positive integer greater than 1, and the signal-to-noise ratio of the noise-free ECG signal is not less than the first threshold; using average beat subtraction Decompose the synthesized ECG signal into a reference ECG signal and a noise-containing residual ECG signal; remove the reference ECG signal from the noise-free ECG signal to obtain a noise-free residual ECG signal; further, according to the noise The residual ECG signal and the noise-free residual ECG signal corresponding to the noisy residual ECG signal training noise-reducing autoencoder, where the noise-containing residual ECG signal is the training input, and the noise-free residual ECG signal corresponds to the noise-free The remaining ECG signal is the training label.

It should be understood that the R peak position of the reference ECG signal obtained by decomposing the synthetic ECG signal using the average beat subtraction method is the same as the R peak position of the synthetic ECG signal.

It should be understood that the obtained noisy residual ECG signal and the noisy residual ECG signal corresponding to the noisy residual ECG signal form a training sample, and the specific realization of generating the training sample can also be completed by other equipment or devices before the training device. The training device may receive training samples sent by the device or device that generates the training samples, which is not limited here.

In the above method, the training equipment uses the noisy residual ECG signal as the input of the noise reduction autoencoder, and uses the noise-free residual ECG signal as the label to train the noise reduction autoencoder. At this time, the training input noisy residual ECG signal is removed In addition to the obvious features in the synthesized ECG signal (such as R peak position, RR interval, etc.), the noise reduction autoencoder only needs to extract the encoding representation of the detailed information of the synthesized ECG signal, and does not need to obtain the entire synthesized ECG signal. Encoding representation, in turn, reduces the difficulty of training, so that the trained target denoising autoencoder can better extract the detailed features in the noisy residual ECG signal, thereby improving the noise reduction performance of the obtained target denoising autoencoder .

In the first implementation in which the training device uses average beat subtraction to decompose the synthesized ECG signal into the reference ECG signal and the noisy residual ECG signal:

The training device averages the W beat ECG signal in the target ECG signal to obtain the second average ECG signal. The target ECG signal can be a synthetic ECG signal, a noiseless ECG signal, or from the first user (also That is, the user who has been collected to obtain noise-free ECG signals) historically collected ECG signals, etc., W≤H, W is a positive integer; replace the second average ECG signal with the H beat ECG signal in the synthetic ECG signal A _j , the reference ECG signal is obtained, A _j represents the ECG signal in the interval of Δt taken around R _j as the center in the H-beat ECG signal, and R _j is the apex of the QRS complex in the W-beat ECG signal , J=1,2...H; further, remove the reference ECG signal to be noise-reduced from the synthesized ECG signal to obtain the noise-containing residual ECG signal.

The signal averaging method can remove noise. Therefore, the reference ECG signal formed above retains the obvious characteristics of the R peak position and R-R distance in the synthesized ECG signal, and does not contain noise.

The specific implementation of the above-mentioned second average ECG signal may be:

Second average ECG signal

It is obtained by averaging the W beat ECG signal in the target ECG signal. which is:

among them,

Represents the second average ECG signal, _Ak represents the ECG signal in the interval of Δt taken from the left and right sides of the W-beat ECG signal with R _k as the center, R _k is the apex of the QRS complex in the W-beat ECG signal, k= 1,2……W.

Further, the second average ECG signal and the ECG signal in the Δt interval may include V sampling points, where V is a positive integer greater than 1, and the signal average can be determined by the formula

To calculate,

Is the second average ECG signal

The value of the v-th sampling point in the W beat ECG signal segment, A _k (v) is the value of the v-th sampling point in the ECG signal in the Δt interval in the W beat ECG signal segment with R _k as the center, 1≤v ≤V, 1≤k≤W, v and k are positive integers.

It should be understood that when the second average ECG signal is replaced by A _j corresponding to the H beat ECG signal in the synthesized ECG signal to obtain the reference ECG signal, it must be ensured that the position of the R peak of the second average ECG signal is consistent with the ECG signal. The R peak positions of the signal A _j are aligned.

It can be seen that in the second average ECG signal

It is obtained by averaging the W beat ECG signal in the synthetic ECG signal, for different synthetic ECG signals, the obtained second average ECG signal is different, and then obtained by the second average ECG signal The reference ECG signal is also different. In the embodiment of this application, the reference ECG signal is adaptively selected for different synthetic ECG signals, and the obtained reference ECG signal can extract the obvious characteristics of the synthetic ECG signal, so that the target noise reduction autoencoder obtained by training It can adapt to different ECG signals, thereby improving the noise reduction performance of the target noise reduction autoencoder.

It should also be understood that signal averaging can reduce the noise of the ECG signal. By averaging the synthesized ECG signal or the noise-free ECG signal, the second average ECG signal is a one-beat ECG signal after noise reduction processing. , The reference ECG signal obtained from the second average ECG signal can be considered to contain no EMG noise, and the average amplitude of the R peak of the target ECG signal is retained, and the reference ECG signal obtained above is for a specific The synthetic ECG signal is generated, which can more accurately represent the obvious characteristics of the synthetic ECG signal.

Average ECG signal in the second

It is obtained by averaging the W beat ECG signal in the noiseless ECG signal. Since the synthetic ECG signal is synthesized by the noise-free ECG signal and the noise signal, the noise-free ECG signal has the same characteristics as the R peak position and RR distance of the synthesized ECG signal, so it is combined with the synthesized ECG signal. Compared with the second average ECG signal obtained by averaging, the second average ECG signal obtained by averaging the noiseless ECG signal has less noise and EMG noise, and then more EMG noise is retained to contain noise In the remaining ECG signal, the trained target noise reduction autoencoder can learn the noise reduction function for this part of the EMG noise.

Average ECG signal in the second

It is obtained by averaging the W-beat ECG signals from the historically collected ECG signals of the first user. For the ECG signal of the same user, the same second average ECG signal is used. The reference ECG signal generated by the second average ECG signal obtained in this way takes into account individual differences, so that the reference ECG signal can more accurately represent the obvious characteristics of the synthesized ECG signal. For the reference ECG signal obtained by the same user, only A calculation is required to improve calculation efficiency.

In the second implementation that uses average beat subtraction to decompose the synthesized ECG signal into the reference ECG signal and the noisy residual ECG signal:

The training equipment can detect the R peak of each beat of the ECG signal in the target ECG signal (that is, the apex of the R wave) in the W beat ECG signal, and the R peak is the sampling point with the largest energy value in the one beat ECG signal ; Perform average processing on the W-beat ECG signal in the target ECG signal to obtain the second average ECG signal, W≤H, W is a positive integer; Replace the H-beat ECG signal in the synthesized ECG signal with the second average ECG signal A _j corresponding to the signal, the reference ECG signal is obtained, A _j represents the ECG signal in the interval obtained by taking R _j as the reference in the H-beat ECG signal, taking Δt1 on the left and taking Δt2 on the right, and R _j is the W-beat ECG The apex of the QRS complex in the signal, j = 1, 2...H, and then remove the reference ECG signal to be noise-reduced from the synthesized ECG signal to obtain the noisy residual ECG signal.

At this time, the R peak of the second average ECG signal obtained includes the ECG signal of Δt1 on the left and the ECG signal of Δt2 on the right. The target ECG signal can be a synthetic ECG signal, a noiseless ECG signal, Or the historically collected ECG signals from the first user (that is, the user whose noise-free ECG signals are collected).

A specific implementation of the second average ECG signal may be to select the ECG signal in the interval obtained by taking the left Δt1 and the right taking Δt2 based on the R peak for each beat of the W beat ECG signal, and obtain W ECG signal segments. Furthermore, the W ECG signal segments are averaged to obtain a second average ECG signal. Among them, each ECG signal segment includes a QRS complex, and the R peak position of all ECG signal segments has the same distance from the starting position of the ECG signal segment where it is located, that is, W ECG signal segments have the same R peak The number of sampling points on both sides are the same to ensure that the W ECG signal segments are aligned.

For example, the selected W ECG signal segments all include V sampling points, where the R peaks are all located at the Z-th sampling point, V and Z are integers greater than 1, and Z is less than V.

Second average ECG signal

It is obtained by averaging W ECG signal segments C _k (k=1, 2...W). At this time, C _k represents the ECG signal in the interval obtained by taking R _k as the reference Δt1 and Δt2 in the W-beat ECG signal, R _k is the apex of the QRS complex in the W-beat ECG signal, k=1, 2 ...W.

among them,

Further, the second average ECG signal and the ECG signal in the Δt interval may include V sampling points, where V is a positive integer greater than 1, and the signal average may be obtained by the formula:

among them,

Is the second average ECG signal

The value of the v-th sampling point in the W beat ECG signal, C _k (v) is the value of the v-th sampling point in the ECG signal in the interval obtained by taking R _k as the reference Δt1 in the W beat ECG signal, 1≤ v≤V, 1≤k≤W, v and k are positive integers.

In another implementation of the second average ECG signal, the training device can divide the W-beat ECG signal from the target ECG signal, and then detect the R peak of each beat of the ECG signal in the W-beat ECG signal, The R peak is the sampling point with the largest amplitude in the one-beat ECG signal; furthermore, the W-beat ECG signal is aligned based on the R peak. At this time, after the alignment, the number of sampling points at each position is not Is greater than W, and then the aligned W beat ECG signals are averaged to obtain a second average ECG signal. Specifically, the average amplitude of one or more sampling points at each position can be calculated, the average amplitude of each position can be calculated, and the average ECG signal can be obtained, and then the average ECG signal is selected based on R _k Δt1 takes the ECG signal in the interval obtained by Δt2 to the right to obtain the second average ECG signal.

At this time, the position of the R peak in the divided N-beat ECG signal may be different, and the length of the N-beat ECG signal may be the same or different. Number the positions obtained after the W-beat ECG signal is aligned, and k represents the index of the position obtained after the W-beat ECG signal is aligned.

at this time,

Average ECG signal

The value of the v-th sampling point in can be expressed as:

among them,

Average ECG signal

The value of the v-th sampling point in the W-beat ECG signal, D _k (v) is the value of the sampling point at position v in the W-beat ECG signal with R _k as the location, 1≤v≤V, 1≤k ≤W, v and k are positive integers. If there is no sampling point distribution at position v in the ECG signal D _k , then D _k (v) is zero.

Further, select the average ECG signal

Take R _k as the reference Δt1 and take the ECG signal in the interval obtained by Δt2 to get the second average ECG signal

In the second aspect, the embodiment of the present application also provides an ECG signal noise reduction method, including: the training device superimposes the noiseless ECG signal and the EMG noise signal to obtain a synthetic ECG signal, wherein the noiseless ECG signal The signal-to-noise ratio (SNR) of the medium myoelectric noise is not less than the first threshold. The noise-free ECG signal may include multi-beat ECG signals, H-beat ECG signals, and each ECG signal includes a QRS Wave group, H is an integer greater than 1; use average beat subtraction to decompose the noiseless ECG signal into the reference ECG signal and the noiseless residual ECG signal; remove the reference ECG signal from the synthesized ECG signal to obtain the noise The residual ECG signal; further, the noise-reducing autoencoder is trained according to the noisy residual ECG signal and the noise-free residual ECG signal corresponding to the noisy residual ECG signal, wherein the noisy residual ECG signal is the training input, the The noise-free residual ECG signal corresponding to the noisy residual ECG signal is the training label.

The specific implementation method of decomposing the noise-free ECG signal into the reference ECG signal and the noise-free residual ECG signal using the average beat subtraction method can be referred to in the first aspect above, using the average beat subtraction method to decompose the noise-free ECG signal into the reference ECG signal. The related description in the signal and the noise-free residual ECG signal will not be repeated here.

In a third aspect, an embodiment of the present application also provides an ECG signal denoising method, including: executing a device to obtain an ECG signal to be denoised. The ECG signal to be denoised includes M beats of the ECG signal. The signal contains a QRS complex, and M is a positive integer greater than 1. Use average beat subtraction to decompose the ECG signal to be denoised into the reference ECG signal to be denoised and the remaining ECG signal to be denoised; The remaining ECG signal to be denoised is input into the target noise reduction autoencoder to obtain the denoised residual ECG signal; the denoised reference ECG signal and the denoised residual ECG signal are superimposed to obtain the denoised heart electric signal.

It should be understood that the calculation method of the reference ECG signal to be denoised is the same as the calculation method of the reference ECG signal involved in the target denoising autoencoder obtained by training.

It should also be understood that the above-mentioned execution device may specifically be a wearable device such as a smart bracelet or a smart watch, a terminal such as a mobile phone, a tablet computer, and a personal computer, or a server, a cloud, and the like.

Perform the above method, remove from the ECG signal to be denoised, including the R peak position of the ECG signal to be denoised, the RR interval and other obvious characteristics of the reference ECG signal to be denoised, and use the target noise reduction autoencoder to be denoised. The residual ECG signal to be denoised after the denoising reference ECG signal is removed from the denoising ECG signal, to avoid the target denoising autoencoder from denoising the obvious features in the denoising ECG signal, making the denoising The reference ECG signal and the residual ECG signal after noise reduction are superimposed to obtain the noise-reduced ECG signal, which can better retain the obvious characteristics of the R peak position and RR interval in the ECG signal to be noise-reduced, and reduce the noise after noise reduction. Distortion of the ECG signal.

Optionally, the above-mentioned target noise reduction autoencoder may be obtained through the training method of the noise reduction autoencoder in the first aspect or the second aspect. For the specific training method, refer to the related description in the above-mentioned first aspect, which is not repeated in the embodiment of the present application.

Furthermore, the target noise reduction autoencoder used for noise reduction uses the noisy residual ECG signal obtained after removing the reference ECG signal from the synthetic ECG signal as the input of the noise reduction autoencoder, and the noiseless ECG signal The noise-free residual ECG signal obtained after the reference ECG signal is removed is used as the label, and the noise reduction autoencoder is trained. At this time, the noisy residual ECG signal input for training removes the obvious features in the synthetic ECG signal ( For example, R peak position, RR interval), the noise reduction autoencoder only needs to extract the coding representation of the detailed information of the synthesized ECG signal, and does not need to obtain the coding representation of the entire synthesized ECG signal, thereby reducing the difficulty of training, making The trained target denoising autoencoder can better extract the detailed features in the noisy residual ECG signal, thereby improving the noise reduction performance of the obtained target denoising autoencoder.

In a possible implementation, the execution device may be a wearable device or terminal equipped with an ECG sensor, such as a smart bracelet, smart watch, etc. At this time, the execution device obtains a specific realization of the ECG signal to be noise-reduced It may be: the execution device collects the analog ECG signal on the surface of the user's skin through the ECG sensor; further, the analog ECG signal is processed through the digital-to-analog conversion module to obtain the digital ECG signal to be denoised.

In another possible implementation, the execution device may be a server or a terminal, etc. At this time, a specific implementation for the execution device to obtain the ECG signal to be noise-reduced may be: the execution device receives the noise-reduction sent by the ECG acquisition device ECG signal. Among them, the ECG acquisition device may be a wearable device or a terminal equipped with an ECG sensor.

In another possible implementation, the execution device uses average beat subtraction to decompose the ECG signal to be denoised into the reference ECG signal to be denoised and the first realization of the remaining ECG signal to be denoised. Yes:

The execution device averages the N beats of the ECG signal to be noise-reduced to obtain the first average ECG signal, where N is less than M, and N is a positive integer; replace the first average ECG signal in the ECG signal to be noise-reduced B _j corresponding to the M beat ECG signal, obtain the reference ECG signal to be noise-reduced, B _j represents the ECG signal in the interval of Δt taken from the left and right sides of the M beat ECG signal with R _j as the center, and R _j is M Snap the apex of the QRS complex in the ECG signal, j=1, 2...M; then, remove the reference ECG signal to be denoised from the ECG signal to be denoised to obtain the remaining ECG signal to be denoised.

The signal averaging method can remove noise. Therefore, the above-mentioned reference ECG signal to be denoised retains the R peak position and R-R distance in the ECG signal to be denoised and does not contain noise.

In the first average ECG signal realization method 1:

First average ECG signal

It is obtained by averaging the N-beat ECG signal in the ECG signal to be noise-reduced, namely:

among them,

Represents a first average ECG, B _i represents N ECG beat to about the center R _i in the ECG signal from each interval Δt, R _i shot vertex ECG QRS complex to the N, i=1, 2……N.

Further, the first average ECG signal and the ECG signal in the Δt interval may include V sampling points, where V is a positive integer greater than 1, and the signal average can be determined by the formula

To calculate,

Is the first average ECG signal

In the v-th sampling point values, B _i (v) about the center from each of W Sign v-th sampling point in the ECG interval Δt electrical signals to heart fragment is the value R _i, v = 1 ,2……V,i=1,2……N.

It can be seen that for different ECG signals to be denoised, the obtained first average ECG signals are different, and the reference ECG signals obtained from the first average ECG signals are also different. In the embodiment of the present application, the reference ECG signal is adaptively selected for different ECG signals to be denoised, and the obtained reference ECG signal can more accurately extract the obvious features of the ECG signal to be denoised, thereby reducing The noisy ECG signal can better retain the obvious features of the ECG signal to be denoised, reduce the distortion of the ECG signal after noise reduction, and improve the quality of the ECG signal after noise reduction.

In the implementation of the first average ECG signal:

First average ECG signal

It is obtained by averaging the N-beat ECG signals from the historically collected ECG signals of the second user (the user whose ECG signals to be denoised is collected). The specific calculation method is the same as that of the above-mentioned first average electrocardiogram signal implementation manner 1, please refer to the related description in the above-mentioned implementation manner 1, which will not be repeated here.

At this time, for the ECG signals of the same user, the same first average ECG signal is used. The implementation manner 5 obtains the reference ECG signal to be denoised generated by the first average ECG signal, taking into account individual differences, so that the reference ECG signal to be denoised can more accurately represent the obvious characteristics of the ECG signal to be denoised, relatively For the implementation of the first average ECG signal 4, for the reference ECG signal to be noise-reduced obtained by the same user, only one calculation is required, which improves the calculation efficiency.

In another possible implementation, the execution device uses average beat subtraction to decompose the ECG signal to be noise-reduced into the reference ECG signal to be noise-reduced and the second implementation of the remaining ECG signal to be noise-reduced Can be:

The above-mentioned method uses the average beat subtraction method to decompose the ECG signal to be denoised into the reference ECG signal to be denoised and the remaining ECG signal to be denoised. In the first implementation, B _i represents the M beat ECG signal with R _j as the reference Δt1 Take the ECG signal in the interval obtained by Δt2 on the right, _Ak represents the ECG signal in the interval obtained by taking R _k as the reference in the W beat ECG signal, Δt1 and take the ECG signal in the interval obtained by taking Δt2 on the right, R _k is the QRS wave in the W beat ECG signal For the vertices of the group, k = 1, 2...W.

The specific calculation method of the second average electrocardiogram signal is the same as the calculation method of the second average electrocardiogram signal described in the first implementation. You can refer to the related description in the first implementation above, and will not be repeated here.

In a fourth aspect, an embodiment of the present application also provides a training device for noise reduction and self-encoding, and the device includes a module for executing the method in the first aspect.

In a fifth aspect, a training device for noise reduction and self-encoding is provided. The device includes: a memory for storing a program; a processor for executing a program stored in the memory, and when the program stored in the memory is executed, the processor is configured to execute The method in the first aspect.

In a sixth aspect, an embodiment of the present application also provides a training device for noise reduction and self-encoding, which includes a module for executing the method in the second aspect.

In a seventh aspect, a training device for noise reduction and self-encoding is provided. The device includes: a memory for storing programs; a processor for executing programs stored in the memory, and when the programs stored in the memory are executed, the processor is configured to execute The method in the second aspect.

In an eighth aspect, an embodiment of the present application also provides an electrocardiographic signal noise reduction device, which includes a module for executing the method in the third aspect.

In a ninth aspect, an electrocardiographic signal noise reduction device is provided. The device includes: a memory for storing a program; a processor for executing a program stored in the memory. When the program stored in the memory is executed, the processor is configured to execute the first The method in three aspects.

In a tenth aspect, a computer-readable medium is provided, and the computer-readable medium stores program code for device execution, and the program code includes the method for executing the method in the first aspect.

In an eleventh aspect, a computer program product containing instructions is provided, when the computer program product runs on a computer, the computer executes the method in the first aspect.

In a twelfth aspect, a computer-readable medium is provided, and the computer-readable medium stores program code for device execution, and the program code includes a method for executing the method in the second aspect.

In a thirteenth aspect, a computer program product containing instructions is provided, which when the computer program product runs on a computer, causes the computer to execute the method in the above second aspect.

In a fourteenth aspect, a computer-readable medium is provided, and the computer-readable medium stores program code for device execution, and the program code includes the method for executing the method in the third aspect.

In a fifteenth aspect, a computer program product containing instructions is provided, which when the computer program product runs on a computer, causes the computer to execute the method in the third aspect.

In a sixteenth aspect, a chip is provided. The chip includes a processor and a data interface. The processor reads instructions stored in a memory through the data interface, and executes the method in the first aspect.

Optionally, as an implementation manner, the chip may further include a memory in which instructions are stored, and the processor is configured to execute instructions stored on the memory. When the instructions are executed, the The processor is used to execute the method in the first aspect.

In a seventeenth aspect, a chip is provided. The chip includes a processor and a data interface. The processor reads instructions stored in a memory through the data interface and executes the method in the second aspect.

Optionally, as an implementation manner, the chip may further include a memory in which instructions are stored, and the processor is configured to execute instructions stored on the memory. When the instructions are executed, the The processor is used to execute the method in the second aspect.

In an eighteenth aspect, a chip is provided. The chip includes a processor and a data interface, and the processor reads instructions stored in a memory through the data interface, and executes the method in the third aspect.

Optionally, as an implementation manner, the chip may further include a memory in which instructions are stored, and the processor is configured to execute instructions stored on the memory. When the instructions are executed, the The processor is used to execute the method in the third aspect.

In a nineteenth aspect, an electronic device is provided, which includes the noise reduction self-encoding training device in any one of the foregoing fourth to fifth aspects.

According to a twentieth aspect, an electronic device is provided, the electronic device including the noise reduction self-encoding training device in any one of the above-mentioned sixth to seventh aspects.

In a twenty-first aspect, there is provided an electronic device, which includes the electrocardiographic signal noise reduction device in any one of the eighth to ninth aspects.

Description of the drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the background art, the following will describe the drawings that need to be used in the embodiments of the present invention or the background art.

FIG. 1A is a schematic flowchart of an automatic analysis process of an ECG signal provided by an embodiment of the present invention;

FIG. 1B is a schematic structural diagram of a smart watch provided by an embodiment of the present invention;

2 is a schematic explanatory diagram of an ECG waveform provided by an embodiment of the present invention;

Fig. 3 is a schematic diagram of the principle of a self-encoder provided by an embodiment of the present invention;

4 is a schematic diagram of a system framework provided by an embodiment of the present invention;

Figure 5 is a network structure diagram of a noise reduction autoencoder provided by an embodiment of the present invention;

Figure 6 is a schematic diagram of a chip hardware structure provided by an embodiment of the present invention;

FIG. 7A is a schematic flowchart of a training method for a noise reduction autoencoder according to an embodiment of the present invention;

FIG. 7B is a schematic explanatory diagram of a method for training a noise reduction autoencoder according to an embodiment of the present invention;

FIG. 7C is a schematic diagram of the principle of calculating a reference ECG signal according to an embodiment of the present invention;

7D is a schematic flowchart of another method for training a noise reduction autoencoder according to an embodiment of the present invention;

FIG. 8A is a schematic flowchart of a method for reducing noise of an ECG signal according to an embodiment of the present invention;

8B is a schematic explanatory diagram of an ECG signal noise reduction method provided by an embodiment of the present invention;

FIG. 8C is a schematic diagram of a principle of calculating a reference ECG signal to be denoised according to an embodiment of the present invention;

8D is a schematic explanatory diagram of a noise reduction result of an ECG signal of a target noise reduction autoencoder according to an embodiment of the present invention;

FIG. 9A is a schematic block diagram of a training device for a noise reduction autoencoder according to an embodiment of the present invention;

FIG. 9B is a schematic block diagram of another apparatus for training a noise reduction autoencoder according to an embodiment of the present invention;

10 is a schematic block diagram of an electrocardiographic signal noise reduction device provided by an embodiment of the present invention;

FIG. 11 is a schematic diagram of the hardware structure of a training device for a noise reduction autoencoder provided by an embodiment of the present application;

FIG. 12 is a schematic diagram of the hardware structure of an electrocardiographic signal noise reduction device provided by an embodiment of the present application.

Detailed ways

The technical solution in this application will be described below in conjunction with the drawings.

The ECG signal noise reduction method provided in the embodiments of the present application can be applied to scenarios such as ECG signal analysis, identification, and diagnosis. Specifically, the ECG signal noise reduction method of the embodiment of the present application can be applied in the following scenarios:

Application scenario A

Machine learning technology can help doctors perform ECG diagnosis, and the quality of ECG signals directly affects the accuracy of ECG signal analysis. As shown in Fig. 1A, a schematic flow diagram of the automatic analysis process of the ECG signal. The automatic analysis of the ECG signal mainly includes two processing processes, namely the noise reduction of the ECG signal and the analysis of the ECG signal after the noise reduction. The ECG signal noise reduction method provided in the embodiments of the present application can be applied to the noise reduction process of the ECG signal.

The ECG noise reduction method provided by the embodiments of the present application first uses the average beat subtraction method to decompose the ECG signal to be denoised into the reference ECG signal to be denoised and the noise-containing residual ECG signal, that is, first according to the ECG to be denoised The signal obtains a reference ECG signal that retains obvious features such as the R peak position and RR interval of the ECG signal to be denoised (that is, the reference ECG signal to be denoised in the embodiment of this application), and then obtains the reference ECG signal to be denoised The reference ECG signal is removed from the system to obtain the remaining ECG signal to be denoised. Further, the remaining ECG signal to be denoised is input to the target noise reduction autoencoder, and the target noise reduction autoencoder The ECG signal undergoes noise reduction processing to obtain the noise-reduced residual ECG signal, and then the noise-reduced residual ECG signal is added to the reference ECG signal to be noise-reduced (also referred to as "superposition" in this article) The obtained ECG signal is the ECG signal after the noise reduction of the ECG signal to be reduced.

Among them, the target denoising autoencoder is a trained neural network. The target denoising autoencoder is obtained by training an initialized denoising autoencoder through multiple training samples. The training samples include noisy residual ECG signals and The noise-free residual ECG signal corresponding to the noisy residual ECG signal, where the noise-containing residual ECG signal is the signal obtained after removing the reference ECG signal from the synthesized ECG signal, and the synthesized ECG signal is the noise-free residual ECG signal The signal superimposed with the EMG noise signal, the noise-free residual ECG signal corresponding to the noise-containing residual ECG signal is the signal obtained by removing the reference ECG signal from the noise-free ECG signal. Among them, the reference ECG signal is extracted from the synthetic ECG signal or the noise-free ECG signal, and the obvious features such as the R peak position and R-R interval of the synthetic ECG signal are retained.

After the denoised ECG signal is obtained, the denoised ECG signal can be analyzed. The specific analysis process can be: Identify the characteristic points of the denoised ECG signal, and then identify the Input the feature points of the into the feature diagnosis model, and predict the diagnosis result of the denoised ECG signal through the feature diagnosis model based on the identified feature points. The feature diagnosis model is a trained machine learning model, and the feature diagnosis The model takes the characteristic points of the ECG signal as input, and the true diagnosis result of the ECG signal is the machine learning model obtained by label training. It should be understood that the identification of the characteristic points of the ECG signal is not a necessary step in the analysis process of the ECG signal. In another implementation of this application, the ECG signal can also be input to the signal diagnosis model, and the signal diagnosis model Directly predict the diagnosis result for the ECG signal. Among them, the signal diagnosis model is a trained machine learning model. The signal diagnosis model takes the ECG signal as input, and the true diagnosis result of the ECG signal is the label for training. Machine learning model.

Application scenario B:

Wearable devices with ECG sensors, such as smart bracelets and smart watches, can be worn on the user's wrist. The smart bracelet and smart watch can be equipped with an ECG sensor to collect the user's ECG data. The smart watch in the first embodiment of the present application is described as an example. Usually an ECG sensor includes two electrodes for collecting ECG signals. Please refer to the schematic diagram of the smart watch shown in Figure 1B. The smart watch 11 may include one electrode 111 of the electrocardiogram sensor disposed on the back of the smart watch 11, and the other electrode 112 is disposed on the side of the smart watch 11. The smart watch 11 may include a digital-to-analog conversion module 113. The digital-to-analog conversion module 113 can perform analog-to-digital conversion on the analog ECG signals collected through the

electrodes

111 and 112 to obtain discrete digital ECG signals. The processing module inside the smart watch 11 can use the digitized ECG signal as the ECG signal to be noise-reduced by applying the ECG noise reduction method in the embodiment of the present application for noise reduction processing to obtain the noise-reduced ECG signal.

It should be understood that when the user uses the smart watch 11 to collect the ECG signal, he can press his finger on the electrode 112, and the electrode 111 contacts the user's arm.

The smart watch 11 or the smart bracelet 12 can also analyze the noise-reduced ECG signal to obtain the analysis result. Further, the smart watch 11 or the smart bracelet 12 can also output the analysis result through an output device, such as a display, a loudspeaker, and the like.

The smart watch 11 or smart bracelet 12 can also send the ECG signal to be noise-reduced to the terminal or server bound to it, and the terminal or server applies the ECG noise-reduction method in the embodiments of this application to perform the noise-reduction ECG signal Noise reduction processing to obtain the ECG signal after noise reduction. The terminal or the server can send the noise-reduced ECG signal to the smart watch 11 or the smart bracelet 12, or send an analysis result obtained by analyzing the noise-reduced ECG signal.

The smart watch 11 or smart bracelet 12 equipped with an ECG sensor can monitor the wearer's ECG data in real time to monitor the wearer's physical condition.

The training method of the noise-reducing autoencoder provided by the embodiment of the application involves data processing, and can be specifically applied to data processing methods such as data training, machine rest, and deep learning. For training samples (such as the noisy residual ECG signal in this application) The noise-free residual ECG signal corresponding to the noisy residual ECG signal) performs symbolic and formalized intelligent information modeling, extraction, preprocessing, training, etc., and finally obtains the target noise reduction autoencoder; and, this application The ECG signal noise reduction method provided by the embodiment can use the above-mentioned target noise reduction autoencoder to convert the input data (such as the residual ECG signal to be denoised obtained after removing the reference ECG signal from the ECG signal to be denoised in this application) ) Is input to the target noise reduction autoencoder to obtain output data (such as the residual ECG signal after noise reduction in this application). It should be noted that the training method for the denoising autoencoder and the method for denoising the ECG signal provided by the embodiments of this application are inventions based on the same concept, and can also be understood as two parts in a system, or a whole Two stages of the process: such as model training stage and model application stage.

Since the embodiments of this application involve a large number of applications of neural networks, in order to facilitate understanding, the following first introduces related terms and neural networks and other related concepts involved in the embodiments of this application.

(1) Beat

Electrocardiography (ECG or EKG), also known as electrocardiogram, records the bioelectric signals generated during the contraction and relaxation of the heart. Each time the heart completes a complete electrical activity, it corresponds to an ECG waveform as shown in Figure 2, including P wave, QRS wave group (including Q wave, R wave and S wave) and T wave. Among them, the first wave on the ECG that deviates from the baseline in the positive direction is the P wave, and the second waveband is the QRS complex. The QRS complex consists of a series of 3 deviations, reflecting the currents related to the depolarization of the left and right ventricles. The first negative deviation in the QRS complex is called the Q wave, the first positive deviation in the QRS complex is called the R wave, and the negative deviation after the R wave is called the S wave. The round and blunt waveform at the top that appears after the QRS complex is the T wave, which represents the state of ventricular repolarization. A complete waveform including the above-mentioned waves is called a beat.

(2) Basic ECG (basic ECG)

The reference ECG signal includes obvious features such as the R peak position and RR interval in the ECG signal. The reference ECG signal may include at least QRS complex. It is understandable that the reference ECG signal includes the characteristics of the part of the ECG signal, usually Obvious features that are easier to extract, such as R peak position, average R peak amplitude, etc. In the embodiment of the present application, the R peak position in the ECG signal is the same as the R peak position in the reference ECG signal corresponding to the ECG signal.

(3) Residual ECG (residual ECG)

After removing the reference ECG signal corresponding to the ECG signal from the ECG signal, the remaining ECG signal is obtained. Since the reference ECG signal includes the obvious features of the ECG signal, the remaining ECG signal includes the hidden features that are not easy to extract from the ECG signal, also known as the detailed features.

(4) Electromyography (EMG)

EMG noise, also known as EMG noise signal, is the superposition of muscle fibers in the body's motor unit action potential (MUAP) in time and space. EMG noise is caused by human activity and muscle tension. The frequency of EMG noise is mainly concentrated in 0.01-100Hz.

(5) Signal-averaged

Signal averaging is a method to eliminate random interference by using the determinism (repeatability) of the signal and the randomness of noise. Signal averaging refers to the technology of superimposing signals and then averaging them. In order to avoid signal distortion after superposition, the signals must be strictly aligned during superposition. In the embodiment of the present application, the averaged multiple ECG signals or multiple ECG signal segments are aligned based on the R peak, that is, the average multiple ECG signals or multiple ECG signal segments are based on the R peak Aligned.

(6) Neural network (NN)

A neural network can be composed of neural units. A neural unit can refer to an arithmetic unit that takes xs and intercept 1 as inputs. The output of the arithmetic unit can be:

Among them, s=1, 2,...n, n is a natural number greater than 1, W _s is the weight of x _s , and b is the bias of the neural unit. f is the activation function of the neural unit, which is used to introduce nonlinear characteristics into the neural network to convert the input signal in the neural unit into an output signal. The output signal of the activation function can be used as the input of the next convolutional layer. The activation function can be a sigmoid function. A neural network is a network formed by connecting many of the above-mentioned single neural units together, that is, the output of one neural unit can be the input of another neural unit. The input of each neural unit can be connected with the local receptive field of the previous layer to extract the characteristics of the local receptive field. The local receptive field can be a region composed of several neural units.

(7) Deep neural network (deep neural network, DNN)

Deep neural network, also called multilayer neural network, can be understood as a neural network with many hidden layers. There is no special metric for "many" here. According to the location of different layers of DNN, the neural network inside DNN can be divided into three categories: input layer, hidden layer, and output layer. Generally speaking, the first layer is the input layer, the last layer is the output layer, and the number of layers in the middle are all hidden layers. The layers are fully connected, that is to say, any neuron in the i-th layer must be connected to any neuron in the i+1th layer. Although DNN looks complicated, it is not complicated in terms of the work of each layer. Simply put, it is the following linear relationship expression:

among them,

Is the input vector,

Is the output vector, b is the offset vector, W is the weight matrix (also called coefficient), and α() is the activation function. Each layer is just the input vector

After such a simple operation, the output vector is obtained

Due to the large number of DNN layers, the number of coefficients W and offset vectors b is also large. The definition of these parameters in the DNN is as follows: Take the coefficient W as an example: Suppose that in a three-layer DNN, the linear coefficients from the fourth neuron in the second layer to the second neuron in the third layer are defined as

The superscript 3 represents the number of layers where the coefficient W is located, and the subscript corresponds to the output third layer index 2 and the input second layer index 4. The summary is: the coefficient from the kth neuron of the L-1th layer to the jth neuron of the Lth layer is defined as

It should be noted that the input layer has no W parameter. In deep neural networks, more hidden layers make the network more capable of portraying complex situations in the real world. Theoretically speaking, a model with more parameters is more complex and has a greater "capacity", which means it can complete more complex learning tasks. Training a deep neural network is also a process of learning a weight matrix, and its ultimate goal is to obtain the weight matrix of all layers of the trained deep neural network (a weight matrix formed by vectors W of many layers).

(8) Autoencoder

Autoencoders are a type of neural network designed to copy their input to the output. It is used to do such a thing, that is to make the output as much as possible to simulate the input, so as to find the compressed representation of the input. The schematic diagram of the self-encoder shown in Figure 3, the self-encoder includes an encoder and a decoder, that is, the mapping between the input layer and the intermediate layer is called encoding, and the intermediate layer and the output layer (also called the reconstruction layer) The mapping between is called decoding. Among them, the encoder compresses the input into a latent space representation (also referred to as an encoding representation in this application), and the encoding function can be used

Where W is

The weight of, b is the bias of the neural unit, f(·) is the activation function of the encoder,

Is the output of the encoder (also called input

Latent space representation); the decoder aims to reconstruct the input from the latent space representation, you can use the decoding function

Get it for the decoder. Where W′ is

The weight of, b′ is the bias of the neural unit, g(·) is the activation function of the neural unit of the decoder, and the decoder reconstructs

The training of the autoencoder reduces the reconstruction error of the autoencoder by optimizing the parameters W, W′, b, b′, that is, reducing

with

The difference between. In the autoencoder, when the number of neurons in the hidden layer is smaller than the input layer and output layer (also called reconstruction layer), that is

Is less than

At this time, the parameters are optimized, if the output is reconstructed

With input

Very close, then it can be considered that the latent space represents

Caught

The effective characteristics of is

The effective compression representation can achieve the purpose of data dimensionality reduction and feature extraction. Data visualization and data noise reduction are two main application scenarios of autoencoders.

(9) Denoising autoencoder (denoising autoencoder)

For the autoencoder, optimization and training can only make the output of the autoencoder close to the input, and for lossy input (input containing noise), it cannot be reconstructed to obtain a lossless input. In order to increase the robustness and generalization ability of the implicit feature representation, the noise-reducing autoencoder is introduced. The network structure of the noise-reducing autoencoder is the same as that of the autoencoder, but the training method is improved, through the damaged input ( Also called noise-containing input) to train the autoencoder to reconstruct the input.

(10) Loss function

In the process of training a deep neural network, because it is hoped that the output of the deep neural network is as close as possible to the value that you really want to predict, you can compare the predicted value of the current network with the target value you really want, and then based on the difference between the two To update the weight vector of each layer of neural network (of course, there is usually an initialization process before the first update, that is, pre-configured parameters for each layer in the deep neural network), for example, if the predicted value of the network If it is high, adjust the weight vector to make its prediction lower, and keep adjusting until the deep neural network can predict the really wanted target value or a value very close to the really wanted target value. Therefore, it is necessary to predefine "how to compare the error between the predicted value and the target value". This is the loss function or objective function, which is used to measure the error between the predicted value and the target value. Important equation. Among them, take the loss function as an example. The higher the output value (loss) of the loss function, the greater the error. Then the training of the deep neural network becomes a process of reducing this loss as much as possible.

(11) Backpropagation algorithm

Convolutional neural networks can use backpropagation (BP) algorithms to modify the size of the parameters in the initial super-resolution model during the training process, so that the reconstruction error loss of the super-resolution model becomes smaller and smaller. Specifically, forwarding the input signal to the output will cause error loss, and the parameters in the initial super-resolution model are updated by backpropagating the error loss information, so that the error loss is converged. The backpropagation algorithm is a backpropagation motion dominated by error loss, and aims to obtain the optimal super-resolution model parameters, such as a weight matrix.

(12) Sampling point

In the embodiments of the present application, the ECG signal is usually a body surface bioelectric signal extracted by an ECG device through electrodes, and the acquired ECG signal is composed of multiple sampling points sorted in time. The value of the sampling point is the intensity or energy value of the bioelectric signal on the body surface at the time of collection. The multiple sampling points are arranged in the order of sampling time to form a data matrix (usually a row vector) to facilitate data processing of the ECG signal.

(13) ECG acquisition equipment

In the embodiments of the present application, the ECG acquisition device is a device that collects ECG signals and analyzes ECG signals equal to the acquisition and processing of ECG signals. It can be an ECG acquisition device, an ECG machine, etc., or it can be equipped with an ECG signal. Sensors such as wearable devices or terminals.

The following describes the system architecture provided by the embodiments of the present application.

Referring to FIG. 4, an embodiment of the present invention provides a system architecture 100.

As shown in the system architecture 100, the data collection device 160 is used to collect data (for example, noise-free ECG signals, noise signals, etc.), and can also generate training data (also referred to as training samples in this application) based on the collected data In the embodiment of the application, the training data includes the noisy residual ECG signal and the noise-free residual ECG signal corresponding to the noisy residual ECG signal, where the noisy residual ECG signal is the result of the synthetic ECG signal removing the reference ECG signal The obtained signal, the synthetic ECG signal is obtained by superimposing the noiseless ECG signal and the EMG noise signal, the noiseless ECG signal is the ECG signal without EMG noise; the noiseless residual ECG signal is the noiseless heart The electrical signal is obtained by removing the reference ECG signal; the reference ECG signal is obtained by the average beat subtraction, which retains the obvious characteristics of the composite ECG signal such as the R peak position of the composite ECG signal. It should be understood that a noise-free ECG signal refers to an ECG signal that contains no or almost no EMG noise, which can be collected by an ECG acquisition device in a stationary state, but it is not excluded that the noise-free ECG signal includes such as power frequency noise. , Baseline drift or other noise, etc.

The data acquisition device 160 may store the training data in the database 130, and the training device 120 obtains the target noise reduction autoencoder 101 by training based on the training data maintained in the database 130. The following will describe in more detail how the training device 120 obtains the target denoising autoencoder 101 based on the training data with the first embodiment. The target denoising autoencoder 101 can be used to implement the ECG signal denoising method provided in the embodiments of the present application. , That is, the ECG signal to be denoised is obtained, the reference ECG signal to be denoised is obtained by the average beat subtraction method, and the reference ECG signal to be denoised is removed from the ECG signal to be denoised to obtain the remaining ECG signal to be denoised. The reference ECG signal to be denoised includes the R peak position of the ECG signal to be denoised, the RR interval and the obvious characteristics of the ECG signal to be denoised. The remaining ECG signal to be denoised is input to the target noise reduction autoencoder 101, The residual ECG signal after noise reduction can be obtained, and the residual ECG signal after noise reduction is superimposed with the reference ECG signal to be noise-reduced to obtain the noise-reduced ECG signal. The target noise reduction autoencoder 101 in the embodiment of this application may specifically be an autoencoder. In the embodiment provided in this application, the target noise reduction autoencoder 101 is obtained by training an initialized noise reduction autoencoder. It should be noted that in actual applications, the training data maintained in the database 130 may not all come from the collection or generation of the data collection device 160, and may also be received from other devices (for example, training devices). In addition, it should be noted that the training device 120 may not necessarily train the target noise reduction autoencoder 101 based on the training data maintained by the database 130. It may also obtain training data from the cloud or other places for model training. The above description should not be used as Limitations of the embodiments of this application. Optionally, the training device 120 may also generate training data based on the noise-free ECG signal and the noise signal, and the training data may be stored in the database 130 by the training device 120, which is not limited in the embodiment of the present application.

The training device 120 trains the target denoising autoencoder 101 obtained by the denoising autoencoder 121 according to the training samples. The target denoising autoencoder 101 can be applied to different systems or devices, as shown in FIG. 4 Device 110. The execution device 110 may be a terminal, such as a mobile phone terminal, a tablet computer, a notebook computer, AR/VR, a vehicle-mounted unit, a wearable device, such as a smart bracelet, a smart watch, etc., or a server or a cloud. In FIG. 4, the execution device 110 may be configured with an I/O interface 112 for data interaction with external devices. The user may input data to the I/O interface 112 through the user device 140, and the input data is implemented in this application. Examples can include: ECG signal to be noise-reduced.

The preprocessing module 113 is configured to perform preprocessing according to the input data (such as the ECG signal to be denoised) received by the I/O interface 112. In the embodiment of the present application, the preprocessing module 113 can be used to generate the to be denoised The reference ECG signal and the reference ECG signal to be denoised from the residual ECG signal to be denoised are removed to obtain the residual ECG signal to be denoised.

The signal superposition module 114 is used to add the noise-reduced reference ECG signal obtained by the preprocessing module 113 and the noise-reduced residual ECG signal output by the target noise-reduction autoencoder 101 to obtain the noise-reduced ECG signal.

When the execution device 110 preprocesses input data, or when the calculation module 111 of the execution device 110 performs calculations and other related processing, the execution device 110 may call data, codes, etc. in the data storage system 150 for corresponding processing , The data, instructions, etc. obtained by corresponding processing may also be stored in the data storage system 150.

Finally, the I/O interface 112 returns the processing result, such as the target ECG signal obtained above, to the user equipment 140, so as to provide it to the user.

It is worth noting that the training device 120 can generate the corresponding target noise reduction autoencoder 101 based on different training data for different targets or different tasks, and the corresponding target noise reduction autoencoder 101 can be used to achieve The above goals or completion of the above tasks, so as to provide users with the desired results.

In the case shown in FIG. 4, the user can manually set input data, and the manual setting can be operated through the interface provided by the I/O interface 112. In another case, the user equipment 140 can automatically send input data to the I/O interface 112. If the user equipment 140 is required to automatically send input data and the user's authorization is required, the user can set the corresponding authority in the user equipment 140. The user can view the result output by the execution device 110 on the user device 140, and the specific presentation form may be a specific manner such as display and sound.

In one implementation, the user equipment 140 may specifically be a terminal equipped with an ECG sensor, such as a mobile phone, a smart bracelet, a smart watch, etc., and the terminal may send an ECG signal to be noise-reduced to the executing device, and the executing device may respond to the signal to be reduced. Noise reduction is performed on the noisy ECG signal, and the noise-reduced ECG signal is obtained. The execution device can send the noise-reduction ECG signal to the terminal, and the terminal can receive the noise-reduction ECG signal sent by the execution device, or according to The ECG signal after noise reduction is diagnosed and analyzed.

In another implementation, the execution device 110 can be a terminal such as a mobile phone, a tablet computer, or a server, a cloud, etc. After the execution device obtains the noise-reduced ECG signal, it can be based on the noise-reduced ECG signal Perform diagnostic analysis. The execution device may send the diagnosis result to the user device 140.

It is worth noting that FIG. 4 is only a schematic diagram of a system architecture provided by an embodiment of the present invention, and the positional relationship among the devices, devices, modules, etc. shown in the figure does not constitute any limitation. For example, in FIG. 4 The data storage system 150 is an external memory relative to the execution device 110. In other cases, the data storage system 150 may also be placed in the execution device 110.

As shown in FIG. 4, the target noise reduction autoencoder 101 is trained according to the training device 120. Specifically, the target noise reduction autoencoder 101 provided in the embodiment of the present application may include an encoder and a decoder. In the target noise reduction autoencoder 101 provided in the embodiment of the present application, both the encoder and the decoder may be a neural network, a convolutional neural network, or a deep neural network.

As shown in FIG. 5, the denoising autoencoder 200 may include an input layer 21, an encoder 22, a decoder 23, and an output layer 24. The encoder 22 may include one or more sets of convolutional layers/pooling layers 220. (The pooling layer is optional), the decoder 23 may include one or more sets of convolutional layers/upsampling layers 230. Generally, the pooling layer in the encoder 22 is used for dimensionality reduction, and the decoder 23 is used for dimensionality reduction. The sampling layer is used for dimension upgrading.

Convolutional layer/pooling layer 220:

Convolutional layer:

As shown in FIG. 5, the convolutional layer/pooling layer 220 may include layers 221-226, for example: in an implementation, layer 221 is a convolutional layer, layer 222 is a pooling layer, and layer 223 is a convolutional layer. Layers, 224 is the pooling layer, 225 is the convolutional layer, and 226 is the pooling layer; in another implementation, 221 and 222 are the convolutional layers, 223 is the pooling layer, and 224 and 225 are the convolutional layers. Layer, 226 is the pooling layer. That is, the output of the convolutional layer can be used as the input of the subsequent pooling layer, or as the input of another convolutional layer to continue the convolution operation.

As shown in FIG. 5, the convolutional layer/upsampling layer 230 may include layers such as 231-236. For example, in an implementation, layer 231 is a convolutional layer, layer 232 is an upsampling layer, and layer 233 is a convolutional layer. Layers, 234 is an upsampling layer, 235 is a convolutional layer, and 236 is an upsampling layer; in another implementation, 231 and 232 are convolutional layers, 233 is an upsampling layer, and 234 and 235 are convolutional layers. Layer, 236 is the up-sampling layer. That is, the output of the convolution layer can be used as the input of the subsequent upsampling layer, or as the input of another convolution layer to continue the convolution operation.

The following will take the convolutional layer 221 as an example to introduce the internal working principle of a convolutional layer.

The convolution layer 221 can include many convolution operators. The convolution operator is also called a kernel. Its role in the processing of the ECG signal is equivalent to a filter that extracts specific information from the input data matrix. The essence of the convolution operator is The above can be a weight matrix. This weight matrix is usually predefined. In the process of convolution on the ECG signal, the weight matrix is usually one sample point after another sample point (or Two sampling points followed by two sampling points... it depends on the value of stride) to complete the work of extracting specific features from the ECG signal. The size of the weight matrix should be related to the number of sampling points in the ECG signal. It should be noted that the depth dimension of the weight matrix and the horizontal depth dimension of the input ECG signal are the same. Different weight matrices can be used to extract different features in the data matrix.

The weight values in these weight matrices need to be obtained through a lot of training in practical applications. Each weight matrix formed by the weight values obtained through training can be used to extract information from the input ECG signal, so that the noise reduction autoencoder 20 is Input the ECG signal for correct noise reduction.

When the noise reduction autoencoder 200 has multiple convolutional layers, the initial convolutional layer (such as 221) often extracts more general features, which can also be called low-level features; With the deepening of the depth of the self-encoder 200, the features extracted by the subsequent convolutional layers (for example, 226) become more and more complex, such as high-level semantic features. The features with higher semantics are more suitable for the problem to be solved. .

Pooling layer:

Since it is often necessary to reduce the number of training parameters, it is often necessary to periodically introduce a pooling layer after the convolutional layer. In the 221-226 layers as illustrated by 220 in Figure 2, it can be a convolutional layer followed by a layer The pooling layer can also be a multi-layer convolutional layer followed by one or more pooling layers.

In the encoder, the purpose of the pooling layer is to reduce the dimensionality of the input data. The pooling layer may include an average pooling operator and/or a maximum pooling operator for sampling the input data to obtain data with smaller dimensions. The average pooling operator can calculate the input data within a specific range to generate an average value as the result of average pooling. The maximum pooling operator can take the data with the largest value in a specific range as the result of maximum pooling. In addition, just as the size of the weight matrix used in the convolutional layer should be related to the dimension of the input data, the operators in the pooling layer should also be related to the dimension of the input data. The dimensionality of the data output after processing by the pooling layer can be smaller than the dimensionality of the data input to the pooling layer. Each sampling point in the data output by the pooling layer represents the average or maximum value of the corresponding subregion of the data input to the pooling layer. .

Up sampling layer (up sampling):

In the decoder, the purpose of the upsampling layer is to increase the dimensionality of the input data. Generally, the principle of upsampling is to insert new elements between the elements based on the original input data.

Output layer:

The output layer 240 can have a loss function similar to the classification cross entropy, which is specifically used to calculate the prediction error of the denoising autoencoder. Once the entire denoising autoencoder 200 has forward propagation (as shown in Figure 2 for the propagation from 210 to 240) For the forward propagation) is completed, the back propagation (as shown in Figure 2 from 240 to 210 direction propagation is back propagation) will start to update the weight values and deviations of the aforementioned layers to reduce the noise reduction autoencoder 200 The loss of the noise reduction autoencoder 200 through the output layer (the residual ECG signal can be predicted in the embodiment of the application) and the ideal result (the noise-free residual ECG signal in the embodiment of the application) Error.

It should be noted that the denoising autoencoder 200 shown in FIG. 2 is only used as an example of a denoising autoencoder. In specific applications, the denoising autoencoder may also exist in the form of other network models.

The following describes a chip hardware structure provided by an embodiment of the present application.

FIG. 6 is a hardware structure of a chip provided by an embodiment of the present invention. The chip includes a neural network processor 30. The chip can be set in the execution device 110 as shown in FIG. 4 to complete the calculation work of the calculation module 171. The chip may also be set in the training device 120 shown in FIG. 4 to complete the training work of the training device 120 and output the target noise reduction autoencoder 101. The algorithms of each layer in the noise reduction autoencoder as shown in FIG. 5 can be implemented in the chip as shown in FIG. 6.

The neural network processor 30 may be any processor suitable for large-scale XOR operation processing such as NPU, TPU, or GPU. Take the NPU as an example: the NPU can be mounted on the host CPU (Host CPU) as a coprocessor, and the host CPU assigns tasks to it. The core part of the NPU is the arithmetic circuit 303. The arithmetic circuit 303 is controlled by the controller 304 to extract matrix data in the memory (301 and 302) and perform multiplication and addition operations.

In some implementations, the arithmetic circuit 303 includes multiple processing units (Process Engine, PE). In some implementations, the arithmetic circuit 303 is a two-dimensional systolic array. The arithmetic circuit 303 may also be a one-dimensional systolic array or other electronic circuits capable of performing mathematical operations such as multiplication and addition. In some implementations, the arithmetic circuit 303 is a general-purpose matrix processor.

For example, suppose there is an input matrix A, a weight matrix B, and an output matrix C. The arithmetic circuit 303 fetches the weight data of the matrix B from the weight memory 302 and caches it on each PE in the arithmetic circuit 303. The arithmetic circuit 303 fetches the input data of matrix A from the input memory 301, and performs matrix operations based on the input data of matrix A and the weight data of matrix B, and the partial or final result of the obtained matrix is stored in the accumulator 308 .

The unified memory 306 is used to store input data and output data. The weight data is directly transferred to the weight memory 302 through the direct memory access controller (DMAC, Direct Memory Access Controller) 305 of the storage unit. The input data is also transferred to the unified memory 306 through the DMAC.

The bus interface unit (BIU, Bus Interface Unit) 310 is used for the interaction between the DMAC and the instruction fetch buffer (Instruction Fetch Buffer) 309; the bus interface unit 301 is also used for the instruction fetch memory 309 to obtain instructions from the external memory; the bus interface unit 301 also The storage unit access controller 305 obtains the original data of the input matrix A or the weight matrix B from the external memory.

The DMAC is mainly used to transfer the input data in the external memory DDR to the unified memory 306, or to transfer the weight data to the weight memory 302, or to transfer the input data to the input memory 301.

The vector calculation unit 307 has multiple arithmetic processing units, if necessary, further processing the output of the arithmetic circuit 303, such as vector multiplication, vector addition, exponential operation, logarithmic operation, size comparison and so on. The vector calculation unit 307 is mainly used for calculation of non-convolutional layers or fully connected layers (FC, fully connected layers) in the neural network. Specifically, it can process: Pooling (pooling), Normalization (normalization), etc. calculations. For example, the vector calculation unit 307 may apply a nonlinear function to the output of the arithmetic circuit 303, such as a vector of accumulated values, to generate the activation value. In some implementations, the vector calculation unit 307 generates a normalized value, a combined value, or both.

In some implementations, the vector calculation unit 307 stores the processed vector to the unified memory 306. In some implementations, the vector processed by the vector calculation unit 307 can be used as the activation input of the arithmetic circuit 303, for example, for use in subsequent layers in a neural network, as shown in FIG. 2, if the current processing layer is a hidden layer 1 (231), the vector processed by the vector calculation unit 307 can also be used for calculation in the hidden layer 2 (232).

The instruction fetch buffer 309 connected to the controller 304 is used to store instructions used by the controller 304;

The unified memory 306, the input memory 301, the weight memory 302, and the fetch memory 309 are all On-Chip memories. The external memory is independent of the NPU hardware architecture.

Among them, the calculation of each layer in the noise reduction autoencoder shown in FIG. 5 can be executed by the arithmetic circuit 303 or the vector calculation unit 307.

The methods involved in the embodiments of the present application are described in detail below.

Example one:

FIG. 7A is a schematic flowchart of a method for training a noise-reducing autoencoder according to Embodiment 1 of the present invention, and FIG. 7B is a schematic explanatory diagram of a method for training a noise-reducing autoencoder according to Embodiment 1 of the present invention. The method can be specifically executed by the training device 120 shown in FIG. 3. Optionally, steps S702-S706 in the method may also be pre-executed by other functional modules before the training device 120, that is, the data of the original samples received or obtained from the database 130 is preprocessed to obtain training samples Then, the training device executes S708 and S710 through the training samples to train the noise-reducing autoencoder. Optionally, the method may be processed by the CPU, or may be processed by the CPU and a processor suitable for neural network calculations (the neural network processor 30 shown in FIG. 6) together to process the neural network processor shown in FIG. 6 30. This application is not restricted. The method may include some or all of the following steps:

S702: Superimpose the noise-free ECG signal and the EMG noise signal to obtain a synthesized ECG signal, where the signal-to-noise ratio (SNR) of the noise-free ECG signal is not less than the first Threshold, noise-free ECG signal includes M beat ECG signal, each beat ECG signal contains a QRS complex, H is an integer greater than 1.

In an implementation of the present application, the noise-free ECG signal may be an ECG signal collected by an ECG acquisition device when the person is in a static state. The noise-free ECG signal may include the H-beat ECG signal, which is a series of sampling points sorted by time, and the value of the sampling point represents the intensity of the electrical signal on the surface of the organism when the sampling point is collected. It should be understood that the noise-free ECG signal in the embodiments of the present application refers to the ECG signal that contains no or almost no EMG noise, but it is not excluded that the noise-free ECG signal includes such as power frequency noise, baseline drift or other noise, etc. . Specifically, the signal-to-noise ratio of the noise-free ECG signal is not less than the first threshold, and the first threshold can be a fixed value greater than 10db, such as 10db, 15db, 20db or 40db, etc., or it can be based on the power of the noise-free ECG signal It is determined, for example, that the first threshold is such that the signal-to-noise ratio of the noise-free ECG signal is equal to the second threshold. The signal-to-noise ratio may specifically be the ratio of the power of the noise-free ECG signal to the power of myoelectric noise, or the ratio of the power of the noise-free ECG signal to the power of the noise.

Specifically, the noise-free ECG signal and the EMG noise signal have the same sampling frequency and length. One way to generate a synthetic ECG signal may be to superimpose the EMG noise signal and the noise-free ECG signal, that is, to synthesize The amplitude of the ECG signal is the sum of the amplitude of the noise-free ECG signal and the amplitude of the EMG noise signal, or the weighted sum of the amplitude of the noise-free ECG signal and the amplitude of the EMG noise signal, also There may be other superposition methods, which are not limited in the embodiment of the present application. In specific calculations, the value of the k-th sampling point in the synthesized ECG signal is equal to the sum of the value of the k-th sampling point in the noise-free ECG signal and the value of the k-th sampling point in the noise signal.

In order to distinguish each sample in this application, a sample composed of a synthetic ECG signal and a noise-free ECG signal corresponding to the synthetic ECG signal is called an original sample, and multiple original samples constitute an original sample set. The noise signals used in the synthesized ECG signals superimposed on different original samples may be different or the same, and there is no limitation on this. It should be understood that the length of an ECG signal (such as a synthesized ECG signal or a noiseless ECG signal, etc.) refers to the duration of the ECG signal. In the embodiment of the present application, the length of the ECG signal may be 5-10 minutes or longer. Or shorter, not limited here. It should also be understood that the number of sampling points in the ECG signal is related to the sampling frequency and the length of the ECG signal. Two signals with the same length at the same sampling frequency have the same number of sampling points.

Optionally, the noise-free ECG signal may be preprocessed before superimposing and generating the synthesized ECG signal. The preprocessing process may include wavelet transform (wavelet transform, WT) to remove the central electrical signal of the noiseless ECG signal. Noise outside the signal band. It is also possible to perform wavelet transformation on the synthesized ECG signal after superposing and generating the synthesized ECG signal, which is not limited in the embodiment of the present application. It should be understood that the frequency band of the ECG signal is generally between 0.05-60Hz, and the frequency of myoelectric noise is mainly concentrated in 0.01-100Hz. Wavelet transform can be used to remove the noise signal outside the ECG signal band (0-0.05Hz and above 60H) In order to reduce the difficulty of the later noise reduction process, the specific implementation of wavelet transform is the prior art, which will not be repeated here. At this time, the synthetic ECG signal in the original sample may be the ECG signal processed by wavelet transform, and meanwhile, the noise-free ECG signal corresponding to the synthetic ECG signal may be the ECG signal processed by the wavelet transformation.

S704: Decompose the synthesized ECG signal into a reference ECG signal and a noisy residual ECG signal by using an average beat subtraction method.

The first implementation of S704 may specifically include the following steps:

S7041: Perform average processing on the W beat ECG signal in the target ECG signal to obtain the second average ECG signal, where the target ECG signal can be a synthetic ECG signal, a noiseless ECG signal or from the first user (That is, users who have been collected to obtain a noise-free ECG signal) historically collected ECG signals, etc., W≤H, W is a positive integer.

S7042: Substituting the second average ECG signal for A _j corresponding to the H-beat ECG signal in the synthesized ECG signal to obtain a reference ECG signal, where A _j represents the H-beat ECG signal centered on R _j and left and right. For the ECG signal in the Δt interval, R _j is the apex of the QRS complex in the W beat ECG signal, j=1, 2...H.

S7043: Remove the reference ECG signal from the synthesized ECG signal to obtain a noisy residual ECG signal.

It should be understood that the W beat ECG signal used to generate the second average ECG signal in S7041 can be derived from a synthetic ECG signal, a noise-free ECG signal, or from the first user (that is, a noise-free ECG signal that is collected) User) historically collected ECG signals, etc. The W-beat ECG signal may be a continuous W-beat ECG signal, or a discontinuous or partially continuous multi-beat ECG signal. Before the training device performs signal averaging on the W-beat ECG signal, it is necessary to ensure that the position of the W-beat ECG signal is aligned, and further, the aligned W-beat ECG signal is superimposed and averaged. The electrical signal is illustrated by taking the average processing of the W beat ECG signal to obtain the second average ECG signal as an example.

Second average ECG signal

among them,

Is the second average ECG signal

The value of the v-th sampling point in the W beat ECG signal, A _k (v) is the value of the v-th sampling point in the ECG signal in the Δt interval with R _k as the center in the W beat ECG signal, 1≤v≤ V,1≤k≤W, v,k are positive integers.

It should be understood that in the foregoing implementation manner 1, for different synthetic ECG signals, the second average electrocardiogram signals obtained through implementation manner 1 are different, and then the reference electrocardiogram signals obtained through the second average electrocardiogram signal in implementation manner 1 Also different.

It can be seen that for different synthetic ECG signals, the second average ECG signal obtained by the implementation method 1 is different, and the reference ECG signal obtained by the second average ECG signal in the implementation method 1 is also different. This implementation method 1 adaptively selects the reference ECG signal for different synthetic ECG signals, and the obtained reference ECG signal can extract the obvious characteristics of the synthetic ECG signal, so that the target noise reduction autoencoder obtained by training can be adapted Different ECG signals, thereby improving the noise reduction performance of the target noise reduction autoencoder.

Average ECG signal in the second

The second implementation of S704 may specifically include but is not limited to the following steps:

S7044: The training device can detect the R peak of each beat of the ECG signal in the W beat ECG signal in the target ECG signal (that is, the apex of the R wave), and the R peak is the highest energy value in the beat of the ECG signal Sampling point.

S7045: Perform average processing on the W beat ECG signal in the target ECG signal to obtain the second average ECG signal. The R peak of the second average ECG signal includes the ECG signal of Δt1 on the left side and the ECG signal of Δt2 on the right side. Signal, the target ECG signal can be a synthetic ECG signal, a noiseless ECG signal, or a historically collected ECG signal from the first user (that is, the user who has been collected to obtain the noiseless ECG signal), etc., W≤ H and W are positive integers.

S7046: Replace the second average ECG signal with A _j corresponding to the H-beat ECG signal in the synthesized ECG signal to obtain the reference ECG signal. A _j means that the H-beat ECG signal takes R _j as the reference and takes Δt1 from the left and takes the right The ECG signal in the interval obtained by Δt2, R _j is the apex of the QRS complex in the W beat ECG signal, j=1, 2...H.

S7047: Remove the reference ECG signal to be noise-reduced from the synthesized ECG signal to obtain the noise-containing residual ECG signal.

A specific implementation of S7045 can be to select the ECG signal in the interval obtained by taking the R peak as the reference to the left and right taking Δt2 for each beat of the W beat ECG signal to obtain W ECG signals Fragment. Furthermore, the W ECG signal segments are averaged to obtain a second average ECG signal. Among them, each ECG signal segment includes a QRS complex, and the R peak position of all ECG signal segments has the same distance from the starting position of the ECG signal segment where it is located, that is, W ECG signal segments have the same R peak The number of sampling points on both sides are the same to ensure that the W ECG signal segments are aligned.

In another implementation, the selected W ECG signal segments of length L may also include P waves, QRS complexes, and T waves; or include P waves, QRS complexes, and so on.

Second average ECG signal

among them,

Is the second average ECG signal

at this time,

Average ECG signal

The value of the v-th sampling point in can be expressed as:

among them,

Average ECG signal

Further, select the average ECG signal

It should be understood that the first implementation described above is a special case of the second implementation. The longer the length of the synthetic ECG signal in the original sample, the larger N will be, and the second average ECG signal will be more accurate when the route is determined based on a large number.

It should also be understood that signal averaging can reduce the noise of the ECG signal. By averaging the synthesized ECG signal or the noise-free ECG signal, the second average ECG signal is a one-beat ECG signal after noise reduction processing. , The reference ECG signal generated by the second average ECG signal can be considered to contain no EMG noise, and the reference ECG signal obtained above is generated for a specific synthetic ECG signal, which can more accurately represent the synthetic ECG The distinctive feature of the signal.

Further, the synthetic ECG signal is obtained by synthesizing the noise-free ECG signal and the noise signal. Compared with the second average ECG signal obtained by averaging the synthesized ECG signal, the first average ECG signal obtained by averaging the noise-free ECG signal The two-average ECG signal has less EMG noise, and more EMG noise is retained in the noisy residual ECG signal, so that the trained target noise reduction autoencoder can learn to deal with this part of EMG noise The noise reduction function.

The specific method of forming the reference ECG signal in S7042 or S7046 is as follows:

It should be understood that the R peak position of the reference ECG signal obtained by the average beat subtraction is the same as the R peak position of the synthesized ECG signal. At this time, the reference ECG signal has the same R peak position and R-R interval as the synthetic ECG signal.

The embodiment of the present application takes a specific implementation of S7046 as an example for description. It should be understood that the above-mentioned first implementation is a special case of the second implementation, and the specific implementation of S7042 can refer to the implementation of S7046, which will not be repeated here. Please also refer to the schematic diagram of the principle of calculating the reference ECG signal shown in FIG. 7C. A specific implementation of S7046 may include but is not limited to the following steps:

S70461: Detect the R peak position of the synthetic ECG signal and the R peak position of the second average ECG signal.

It should be understood that the R peak detection in the ECG signal is a prior art, and will not be repeated here.

S70462: Obtain H signal segments according to the second average ECG signal, the R peak position in the synthesized ECG signal, and the R peak position of the second average ECG signal, and the H signal segments are the same as the second average ECG signal And the R peak position of the h-th signal segment in the H signal segments is equal to the h-th R peak position in the synthesized ECG signal, h is a positive integer, and h is not greater than the total R peak in the synthesized ECG signal Number. Fig. 7C takes H=3 as an example for illustration.

Specifically, taking the R peak position as a reference, the second average ECG signal is aligned with each beat of the composite ECG signal, and the second average ECG signal aligned with the R peak position in the composite ECG signal is called It should be understood that, relative to the second average ECG signal, the position of each sampling point in the signal segment moves as a whole, but its waveform remains unchanged.

It should also be understood that the interval between any two adjacent R peaks in the synthetic ECG signal (ie, the RR interval) may be different, and the length of the second average ECG signal may be greater than the length of an RR interval, or may be less than the length of an RR interval. length. At this time, two adjacent signal segments may partially overlap or have a certain distance.

S70463: Generate a reference ECG signal according to the H signal segments, where the value of the sampling point of the reference ECG signal at the third position is the value of the sampling point of the third signal segment at the third position and the value of the fourth signal segment at the third position. The average of the value of the sampling point at the position, the third position is the position where there are multiple sampling points in the H signal fragments, the third signal fragment and the fourth signal fragment are the H signal fragments with the sampling point at the third position Two signal segments; the value of the sampling point of the reference ECG signal at the fourth position is obtained by interpolating the values of the two closest sampling points to the fourth position in the multiple signal segments, and the third position is H signals Where the segments overlap, the fourth position is the position between H signal segments.

Among them, the interpolation algorithm is an existing technology, and will not be repeated here.

It should be understood that in another implementation of S70463, the value of the sampling point of the reference ECG signal at the third position may be the value of the sampling point of the third signal segment at the third position or the value of the sampling point of the fourth signal segment at the third position. The value of the sampling point at the location. The value of the sampling point of the reference ECG signal at the fourth position can also be set to 0, which is not limited here.

S7043 and S7047 remove the reference ECG signal from the synthesized ECG signal to obtain the noisy residual ECG signal.

It should be understood that the synthesized ECG signal, the noise-free ECG signal, the reference ECG signal, the noise-containing residual ECG signal, and the noise-free residual ECG signal all include the same number of sampling points. Remove the reference ECG signal from the synthetic ECG signal, that is, the value of the sample point in the synthetic ECG signal is subtracted from the value of the sample point in the reference ECG signal, that is, the residual ECG signal with noise The value of the xth sampling point is equal to the difference between the value of the xth sampling point in the synthetic ECG signal and the value of the xth sampling point in the reference ECG signal.

S706: Remove the reference ECG signal from the noise-free ECG signal to obtain a noise-free residual ECG signal.

In the same way, removing the reference ECG signal from the noise-free ECG signal means that the value of the sample point in the noise-free ECG signal is subtracted from the value of the sample point in the reference ECG signal, that is, the noise-free ECG signal The value of the xth sample point in the remaining ECG signal is equal to the difference between the value of the xth sample point in the noise-free ECG signal and the value of the xth sample point in the reference ECG signal. Among them, x is the index of the sampling point, x is a positive integer, and x is not greater than the total number of sampling points in the synthesized ECG signal.

In the implementation manner (1) of the embodiment of this application:

Synthetic ECG signal X _{q The} noisy residual ECG signal obtained after removing the reference ECG signal

The length meets the length requirement of the input data of the noise reduction autoencoder. At this time, the noisy residual ECG signal

And noiseless residual ECG signal

The formed samples are called training samples, among which, the synthetic ECG signal X _q is obtained by superimposing the noise-free ECG signal Y _q with the EMG noise signal, and the noise-free residual ECG signal

It is the signal obtained after the noise-free ECG signal Y _{q is} removed from the reference ECG signal. The original sample corresponds to the training sample one to one. Multiple training samples constitute a training sample set. This training sample set is used to train the noise reduction autoencoder.

At this point, a training sample in the training sample set can be expressed as

Among them, q is the index of the original sample in the original sample set, which is also the index of the training sample in the training sample set in this implementation; the training sample set includes Q training samples, q and Q are positive integers, and q≤Q.

Optionally, the noisy residual ECG signal in the training sample may be obtained by normalizing the noisy residual ECG signal obtained by decomposition, and the normalization method may be maximum-minimum normalization , Or other normalization methods, here, take the maximum and minimum value normalization as an example to illustrate:

among them,

Is the noisy residual ECG signal in the training sample,

To decompose the noisy residual ECG signal,

Is the minimum signal strength of the noisy residual ECG signal in the training sample,

Is the maximum signal strength in the noisy residual ECG signal in the training sample.

In the same way, the noise-free residual ECG signal in the training sample is obtained by normalizing the noise-free residual ECG signal obtained by decomposition. Take the maximum and minimum value normalization as an example to illustrate:

among them,

Is the noise-free residual ECG signal in the training sample,

In order to decompose the noiseless residual ECG signal,

Is the minimum signal strength of the noise-free residual ECG signal in the training sample,

Is the maximum signal strength of the noise-free residual ECG signal in the training sample.

At this point, a training sample in the training sample set can be expressed as

It should be understood that normalization is not a necessary step for generating a training sample set. The training device may also preprocess the training samples before inputting the training samples to the denoising autoencoder. The preprocessing process may include normalization. Operations can also include other operations, which are not limited here.

In the implementation manner (2) of the embodiment of this application:

The length of is far greater than the length of the input data required by the noise reduction autoencoder. At this point, the noisy residual ECG signal in the original sample can be

And noiseless residual ECG signal

Carry out cutting, and cut the noisy residual ECG signal slices

Meet the length of the input data required by the noise reduction autoencoder. Similarly, for the noiseless residual ECG signal

Perform cutting to obtain noise-free residual ECG signal slices

Among them, the synthesized ECG signal X _q is obtained by superimposing the noise-free ECG signal Y _q with the EMG noise signal, and the noise-free residual ECG signal

It is the signal obtained by removing the reference ECG signal from the noise-free ECG signal Y _q , and e is the noise-free residual ECG signal slice obtained from Q noise-free residual ECG signal slices. Index, e and E are positive integers, e≤E. It should be understood that noisy residual ECG signal

Multiple slices of noisy residual ECG signal and noiseless residual ECG signal obtained by cutting

The multiple noiseless residual ECG signals obtained by cutting correspond one to one.

At this time, a slice of noisy residual ECG signal obtained by cutting

And the noisy residual ECG signal slice

Corresponding noise-free residual ECG signal slice

The formed samples are called training samples. E training samples form the training sample set. At this time, one training sample in the training sample set can be expressed as

Among them, e is also called the index of the training sample in the training sample set. At this time, one original sample can correspond to multiple training samples. Multiple training samples constitute a training sample set. The training sample set is used to train the noise reduction autoencoder to obtain the target noise reduction autoencoder.

It should be noted that the above steps S702-S706 may also be executed by other equipment before the training device, or may be executed by the training device.

The training device can generate multiple training samples. The multiple training samples can be divided into a training sample set and a test sample set. The training device can use the training sample set to train the noise-reducing autoencoder, where the noisy residual ECG signal is Training input, the noise-free residual ECG signal corresponding to the noisy residual ECG signal is the training label, and finally the target noise reduction autoencoder is obtained. The training device can use the test sample set to test the trained target denoising autoencoder to evaluate the robustness and generalization ability of the target denoising autoencoder. The embodiment of the present application takes the training sample set described in the foregoing implementation (1) as an example for illustration. It should be understood that the training samples in step S708 may also be in the form in the foregoing implementation (2) or other forms, which will not be repeated here. Specifically, the specific implementation of the training device using the training samples to train the noise reduction autoencoder may include the following steps:

S708: Input the noisy residual ECG signal into the noise reduction autoencoder to obtain the predicted residual ECG signal.

For the network structure of the noise reduction autoencoder, refer to the description in FIG. 2 above, and details are not described in the embodiment of the present application.

At this time, the denoising autoencoder is the initialized neural network or the denoising autoencoder updated during the training process.

The embodiment of this application takes the training sample as

As an example, the residual ECG signal with noise

Input to the noise reduction autoencoder, and the noise reduction autoencoder

Process to get predicted residual ECG signal

S710: Update the parameters of the noise reduction autoencoder according to the error between the predicted residual ECG signal and the noise-free residual ECG signal to obtain the target noise reduction autoencoder.

Among them, one training sample may be used in one training process of the noise reduction autoencoder, or multiple training samples or all training samples may be used, which is not limited in the embodiment of the present application. The target denoising autoencoder is a denoising autoencoder trained through the training sample set.

The embodiment of this application takes Q training samples as an example, where U is a positive integer not greater than Q, and the training device can determine Q according to the error between the predicted residual ECG signal and the noise-free residual ECG signal of each training sample For the loss corresponding to each training sample, according to the loss corresponding to the Q training sample, the parameters of the denoising autoencoder are updated through the optimization algorithm, so that the loss becomes smaller and smaller. The optimization algorithm may be gradient descent, Adam algorithm or other optimization algorithms, which are not limited in the embodiment of the present application.

In the embodiment of the present application, the loss function is used to calculate the loss corresponding to Q training samples. The loss function can be the error between the predicted residual ECG signal and the noise-free residual ECG signal, where the predicted residual ECG signal and the noise-free residual ECG signal The error of the electrical signal can be the mean absolute error (MAE), mean squared error (MSE), or root mean squared error between the predicted residual ECG signal and the noise-free residual ECG signal. Error, RMSE), etc., can also be used to predict the cross entropy of the residual ECG signal and the noise-free residual ECG signal, and can also have other forms, which are not limited in this application.

For example, in the embodiment of the present application, the loss function L can be expressed by predicting the average absolute error between the residual ECG signal and the noise-free residual ECG signal, then:

The above method does not directly use the synthetic ECG signal for the training of the noise reduction autoencoder, but first removes the reference ECG signal including the R peak position and other obvious features from the synthetic ECG signal, and the reference ECG signal remains The obvious characteristics of the synthesized ECG signal, such as the R peak position, the average amplitude of the R peak, etc., are used to avoid the distortion of the ECG signal; the noisy residual ECG signal obtained after the synthetic ECG signal is removed from the reference ECG signal As the input of the noise reduction autoencoder, compared with the prior art where the entire ECG signal is used as the input of the autoencoder, the noise reduction autoencoder in the embodiment of the present application only needs to extract the encoding representation of the detailed information of the ECG signal. There is no need to obtain the encoding representation of the entire ECG signal, which reduces the difficulty of training the denoising autoencoder, and in turn, enables the trained target denoising autoencoder to better extract the details of the noisy residual ECG signal Feature, thereby improving the noise reduction performance of the obtained target noise reduction autoencoder.

Embodiment two:

FIG. 7D is a schematic flowchart of another method for training a noise reduction autoencoder according to Embodiment 2 of the present invention. The method can be specifically executed by the training device 120 shown in FIG. 3. Optionally, steps S712-S716 in the method may also be pre-executed by other functional modules before the training device 120, that is, the data of the original samples received or obtained from the database 130 is preprocessed to obtain training samples Then, the training device executes S718 and S720 through the training samples to train the noise reduction autoencoder. Optionally, the method may be processed by the CPU, or may be processed by the CPU and a processor suitable for neural network calculations (the neural network processor 30 shown in FIG. 6) together to process the neural network processor shown in FIG. 6 30. This application is not restricted. The method may include some or all of the following steps:

S712: Superimpose the noise-free ECG signal and the EMG noise signal to obtain a synthesized ECG signal, where the signal-to-noise ratio (SNR) of the noise-free ECG signal is not less than the first Threshold, the noise-free ECG signal may include a multi-beat ECG signal H-beat ECG signal, each beat ECG signal contains a QRS complex, and H is an integer greater than 1.

S714: Decompose the noise-free ECG signal into a reference ECG signal and a noise-free residual ECG signal by using an average beat subtraction method.

For the specific implementation method of S714, please refer to the related description in S704 using average beat subtraction to decompose the noise-free ECG signal into the reference ECG signal and the noise-free residual ECG signal. Similarly, the first implementation of S714 is specific:

The first implementation of S714 may specifically include the following steps:

S7141: Perform average processing on the W beat ECG signal in the target ECG signal to obtain a second average ECG signal, where the target ECG signal can be a synthetic ECG signal, a noiseless ECG signal or from the first user (That is, users who have been collected to obtain a noise-free ECG signal) historically collected ECG signals, etc., W≤H, W is a positive integer.

S7142: Replace the second average ECG signal with A _j corresponding to the H-beat ECG signal in the noise-free ECG signal to obtain the reference ECG signal. A _j represents the H-beat ECG signal with Δt as the center of R _j. For the ECG signal in the interval, R _j is the apex of the QRS complex in the W beat ECG signal, j=1, 2...H.

S7143: Remove the reference ECG signal from the noise-free ECG signal to obtain the noise-free residual ECG signal.

It should be understood that the first implementation of S714 is the same as the first implementation of S704 in the embodiment. For details, please refer to a related description of the foregoing embodiment, which will not be repeated in this embodiment.

The second implementation of S714 may specifically include but is not limited to the following steps:

S7144: The training device can detect the R peak of each beat of the ECG signal in the W beat ECG signal in the target ECG signal (that is, the apex of the R wave), and the R peak is the maximum energy value in the beat of the ECG signal Sampling point.

S7145: Perform average processing on the W beat ECG signal in the target ECG signal to obtain the second average ECG signal. The R peak of the second average ECG signal includes the ECG signal of Δt1 on the left side and the ECG signal of Δt2 on the right side. Signal, the target ECG signal can be a synthetic ECG signal, a noiseless ECG signal, or a historically collected ECG signal from the first user (that is, the user who has been collected to obtain the noiseless ECG signal), etc., W≤ H and W are positive integers.

S7146: Replace the second average ECG signal with A _j corresponding to the H-beat ECG signal in the noise-free ECG signal to obtain the reference ECG signal. A _j means that the H-beat ECG signal takes R _j as the reference left and Δt1 right Take the ECG signal in the interval obtained by Δt2, R _j is the apex of the QRS complex in the W beat ECG signal, j=1, 2...H.

S7147: Remove the reference ECG signal from the noise-free ECG signal to obtain the noise-free residual ECG signal.

It should be understood that the second implementation of S714 is the same as the second implementation of S704 in the embodiment. For details, please refer to a related description of the foregoing embodiment, which will not be repeated in this embodiment.

S716: Remove the reference ECG signal from the synthesized ECG signal to obtain a noisy residual ECG signal.

S718: Input the noisy residual ECG signal into the noise reduction autoencoder to obtain the predicted residual ECG signal.

S720: Update the parameters of the noise reduction autoencoder according to the error between the predicted residual ECG signal and the noise-free residual ECG signal to obtain the target noise reduction autoencoder.

It should be noted that the foregoing S718 may refer to the related description of S708 in Embodiment 1, and the foregoing S720 may refer to the related description of S710 in Embodiment 1, which will not be repeated here.

Example three:

FIG. 8A is a schematic flowchart of an ECG signal noise reduction method according to Embodiment 2 of the present invention, and FIG. 8B is a schematic explanatory diagram of an ECG signal noise reduction method according to Embodiment 2 of the present invention. The target noise reduction autoencoder obtained by training in the first embodiment realizes the noise reduction of the ECG signal to be denoised. The method may be specifically executed by the execution device 110 shown in FIG. 3, and the ECG signal to be denoised in the method may be the input data given by the user device 140 shown in FIG. The preprocessing module 113 can be used to execute the signal superimposing module 114 in the execution device 110 in S802-S804 of the method 800 to execute S808 in the method 800, and the calculation module 111 in the execution device 110 can Used to execute the S806. Optionally, the method 800 may be processed by a CPU, or a CPU and a processor suitable for neural network calculation (for example, the neural network processor 30 shown in FIG. 6), which is not limited in this application.

The ECG signal noise reduction method 800 may include but is not limited to some or all of the following processes:

S802: Obtain the ECG signal to be noise-reduced.

The user equipment may send the ECG signal to be noise-reduced to the execution device, requesting the execution device to reduce the noise of the ECG signal to be noise-reduced. The execution device may also collect the user's ECG signal in real time through the ECG device, and the ECG signal of the preset length collected in real time is the ECG signal to be noise-reduced. Among them, the preset length is the length required by the noise reduction autoencoder for input data.

Optionally, before S804, the ECG signal to be denoised can be preprocessed. The preprocessing can include but is not limited to operations such as wavelet transform. At this time, the ECG signal to be denoised in the following steps is the preprocessed ECG signal. Noise reduction ECG signal. The execution device can perform wavelet transform on the ECG signal to be denoised to remove the noise outside the frequency band of the central electrical signal of the ECG signal to be denoised, so as to reduce the difficulty of the later noise reduction process. The specific implementation of wavelet transform is the existing technology, here No longer.

In a possible implementation, the execution device may be a wearable device or terminal equipped with an ECG sensor, such as a smart bracelet, smart watch, etc. At this time, a specific implementation of S802 may be: The sensor collects the analog ECG signal on the surface of the user's skin; further, the analog ECG signal is processed through the digital-to-analog conversion module to obtain a digital ECG signal to be denoised.

In another possible implementation, the execution device may be a server or a terminal, etc. In this case, a specific implementation of S802 may be: the execution device receives the ECG signal to be noise-reduced sent by the ECG acquisition device. Among them, the ECG acquisition device may be a wearable device or a terminal equipped with an ECG sensor.

For example, a wearable device, such as a smart watch, collects the ECG signal of the wearer, and the ECG signal is the ECG signal to be noise-reduced. The smart watch can send the signal to the terminal (such as a mobile phone) bound to it through Bluetooth. Noise reduction ECG signal, the mobile phone receives the ECG signal to be noise reduction. After the mobile phone receives the ECG signal to be noise-reduced, it can perform noise reduction processing on the ECG signal to be noise-reduced; it can also send the ECG signal to be noise-reduced to the server, and the server implements the ECG signal to be noise-reduced Perform noise reduction processing.

S804: Decompose the ECG signal to be denoised into the reference ECG signal to be denoised and the remaining ECG signal to be denoised by using the average beat subtraction method.

Similar to the first implementation of S704 in Embodiment 1, the first implementation of S804 may specifically include the following steps:

S8041: Perform averaging processing on the N-beat ECG signal in the ECG signal to be noise-reduced to obtain the first average ECG signal, where N is less than M, and N is a positive integer.

S8042: Substituting the first average ECG signal for the B _j corresponding to the M-beat ECG signal in the ECG signal to be noise-reduced to obtain the reference ECG signal to be noise-reduced, and B _j represents that R _j is used in the M-beat ECG signal Is the ECG signal in the Δt interval of the center, R _j is the apex of the QRS complex in the M-beat ECG signal, j=1, 2...M.

S8043: Remove the reference ECG signal to be denoised from the ECG signal to be denoised to obtain the remaining ECG signal to be denoised.

It should be understood that the N-beat ECG signal used to generate the first average ECG signal in S8041 may be a continuous N-beat ECG signal in the ECG signal to be denoised, or it may be a discontinuous or partially continuous N-beat ECG signal. electric signal. Before the execution device performs signal averaging on the N-beat ECG signal, the N-beat ECG signal needs to be aligned, and then the aligned multi-beat ECG signal is superimposed and averaged, and the specific implementation is the same as that described in S704 above. The specific implementation principles of the two average electrocardiogram signals are similar, and can refer to the related descriptions in the foregoing implementation 1 and implementation 2, which will not be repeated here.

Corresponding to the first implementation of S704 in the first embodiment above, in the first implementation of S804, implementation 1 of the first average ECG signal may be:

Second average ECG signal

among them,

Further, the first average ECG signal and the ECG signal in the Δt interval may include V sampling points, where V is a positive integer greater than 1, and the signal average may be obtained by the formula:

among them,

Is the first average ECG signal

In the v-th sampling point values, B _i (v) about the center from each Sign v-th sampling point in the ECG interval Δt to ECG R _i is a value for N, 1≤v≤ V,1≤i≤N, v,i are positive integers.

It can be seen that for different ECG signals to be noise-reduced, the first average ECG signals obtained through implementation 4 are different, and the reference ECG signals obtained through the first average ECG signal in implementation 1 are also different. This implementation manner 4 adaptively selects the reference ECG signal for different ECG signals to be denoised, and the obtained reference ECG signal can more accurately extract the obvious features of the ECG signal to be denoised, and then the denoised ECG signal The ECG signal can better retain the obvious features of the ECG signal to be noise-reduced, reduce the distortion of the ECG signal after noise reduction, and improve the quality of the ECG signal after noise reduction.

Implementation mode 2 of the first average ECG signal may be:

First average ECG signal

The second implementation of S804 may specifically include but not limited to the following steps:

S8044: Detect the R peak of each beat of the ECG signal in the N-beat ECG signal in the ECG signal to be noise-reduced (that is, the apex of the R wave), and the R peak is the sample with the largest energy value in one beat of the ECG signal point.

S8045: Perform averaging processing on the N-beat ECG signal in the ECG signal to be noise-reduced to obtain the first average ECG signal. The R peak of the first average ECG signal includes the ECG signal of Δt3 on the left side and the heart signal of Δt4 on the right side. Electric signal, N≤M, N is a positive integer.

S8046: Substituting the first average ECG signal for the B _j corresponding to the M-beat ECG signal in the ECG signal to be noise-reduced, to obtain the reference ECG signal to be noise-reduced, B _j represents the M-beat ECG signal based on R _j Take the ECG signal in the interval obtained by taking Δt3 on the left and taking Δt4 on the right. R _j is the apex of the QRS complex in the N-beat ECG signal, j = 1, 2...M.

S8047: Remove the reference ECG signal to be noise-reduced from the ECG signal to be noise-reduced to obtain the noise-containing residual ECG signal.

A specific implementation of S8045 can be to select the ECG signal in the interval obtained by taking the R peak as the reference to the left and right taking Δt4 for each beat of the N-beat ECG signal to obtain N ECG signals Fragment. Furthermore, the N ECG signal segments are averaged to obtain the first average ECG signal. Among them, each ECG signal segment includes a QRS complex, and the R peak position of all ECG signal segments has the same distance from the starting position of the ECG signal segment where it is located, that is, N ECG signal segments have the same R peak The number of sampling points on both sides are the same to ensure alignment of N ECG signal segments.

In an implementation of the embodiment of the present application, Δt1 may be equal to Δt3 in the first embodiment, and Δt2 may be equal to Δt4 in the first embodiment. In an implementation of the embodiment of the present application, Δt1 in the first embodiment may not be equal to Δt3, and Δt2 in the first embodiment may not be equal to Δt4, which is not limited.

For the specific implementation of the foregoing S8045, refer to the specific implementation of S7045 in the foregoing embodiment 1, which will not be repeated here.

The specific method for forming the reference ECG signal to be noise-reduced in S8042 or S8046 is as follows:

The embodiment of the present application takes a specific implementation of S8046 as an example for illustration. It should be understood that the above-mentioned first implementation is a special case of the second implementation. For the specific implementation of S8042, refer to the implementation of S8046, which will not be repeated here.

The principle of calculating the reference ECG signal is the same as in the first embodiment above. Please also refer to the schematic diagram of the principle of calculating the reference ECG signal to be denoised shown in Figure 8C. A specific implementation of S8046 may include but is not limited to the following step:

S80461: Detect the R peak position of the ECG signal to be noise-reduced and the R peak position of the first average ECG signal.

S80462: Obtain M signal segments according to the first average ECG signal, the R peak position in the ECG signal to be noise-reduced, and the R peak position of the first average ECG signal, and the M signal segments have the same value as the first average ECG signal. The signal has the same waveform, and the R peak position of the h-th signal segment in the M signal segments is equal to the h-th R peak position in the ECG signal to be denoised, h is a positive integer, and h is not greater than the ECG to be denoised The total number of R peaks in the signal. Fig. 8C takes M=3 as an example for illustration.

Specifically, based on the R peak position, the first average ECG signal is aligned with each beat of the ECG signal to be denoised, and the first average ECG signal is aligned with the R peak position in the ECG signal to be denoised. The average ECG signal is called a signal segment. It should be understood that, relative to the first average ECG signal, the position of each sampling point in the signal segment moves as a whole, but its waveform remains unchanged.

It should also be understood that the interval between any two adjacent R peaks in the ECG signal to be noise-reduced (ie, the RR interval) may be different, and the length of the first average ECG signal may be greater than the length of one RR interval, or may be less than one RR interval. The length of the interval. At this time, two adjacent signal segments may partially overlap or have a certain distance.

S80463: Generate a reference ECG signal to be denoised according to the M signal segments, where the value of the sampling point of the reference ECG signal to be denoised at the first position is the value of the sampling point of the first signal segment at the first position and the The average value of the sampling points of the two signal segments at the first position. The first position is the position where there are multiple sampling points in the M signal segments. The first signal segment and the second signal segment are the M signal segments in the first position. Two signal segments with sampling points at the position; the value of the sampling point of the reference ECG signal to be denoised at the second position is interpolated based on the value of the two most adjacent sampling points in the M signal segments It is obtained that the first position is the position including the overlap on the M signal segments, and the second position is the position between the M signal segments.

It should be understood that in another implementation of S80463, the value of the sampling point of the reference ECG signal to be denoised at the first position may be the value of the sampling point of the first signal segment at the first position or the second signal segment The value of the sampling point at the first position. The value of the sampling point of the reference ECG signal to be denoised at the second position can also be set to 0, which is not limited here.

It should be understood that the method of generating the reference ECG signal to be denoised used by the trained target denoising autoencoder may be consistent with the method of generating the reference ECG signal to be denoised.

A specific implementation of S8043 or S8047 can be:

It should be understood that the ECG signal to be denoised, the reference ECG signal to be denoised, and the remaining ECG signal to be denoised all include the same number of sampling points. The reference ECG signal to be denoised is removed from the ECG signal to be denoised, that is, the value of the sample point in the ECG signal to be denoised is subtracted from the value of the sample point in the reference ECG signal to be denoised. That is, the value of the yth sampling point in the remaining ECG signal to be denoised is equal to the value of the yth sampling point in the ECG signal to be denoised and the value of the yth sampling point in the reference ECG signal to be denoised difference. Among them, y is the index of the sampling point, y is a positive integer, and y is not greater than the total number of sampling points in the synthesized ECG signal.

Optionally, after obtaining the remaining ECG signal to be denoised, a normalization operation may be performed on the remaining ECG signal to be denoised. Whether the normalization operation is required and the target noise reduction autoencoder's requirements for input data are determined. When the noisy residual ECG signal in the training sample of the target noise reduction autoencoder is normalized data, the residual ECG signal to be denoised is input to the target noise reduction autoencoder. The normalization operation is performed on the remaining ECG signal to be denoised; otherwise, the normalization operation is not required to be performed on the remaining ECG signal to be denoised, which will not be repeated in this embodiment of the application.

S806: Input the residual ECG signal to be noise-reduced into the target noise-reduction autoencoder to obtain the residual ECG signal after noise reduction.

Among them, the target denoising autoencoder is a target denoising autoencoder trained by the training method of the denoising autoencoder described in the first embodiment or the second embodiment, and has the function of denoising the residual ECG signal For details, please refer to the related description in the above-mentioned Embodiment 1 or Embodiment 2, which will not be repeated here.

S808: Superimpose the reference ECG signal to be noise-reduced and the remaining ECG signal after noise reduction to obtain the ECG signal after noise reduction. The ECG signal after noise reduction is the ECG signal to be noise-reduced after noise reduction processing After the ECG signal.

Specifically, the ECG signal obtained by superimposing the reference ECG signal to be noise-reduced and the residual ECG signal after noise reduction, that is, the value of the sampling point in the reference ECG signal and the value of the sampling point in the residual ECG signal after noise reduction The value corresponds to the addition, that is, the value of the zth sample point in the denoised ECG signal is equal to the value of the zth sample point in the reference ECG signal to be denoised and the residual ECG signal after noise reduction The value of the zth sampling point in, z is the index of the sampling point, and z is not greater than the total number of sampling points in the ECG signal to be denoised.

Perform the above method, extract from the ECG signal to be denoised the reference ECG signal to be denoised including obvious features such as the R peak position of the ECG signal to be denoised, and use the target denoising autoencoder to treat the denoised ECG signal After the signal is removed from the reference ECG signal, the remaining ECG signal to be de-noised is denoised, avoiding the noise reduction processing of the target denoising autoencoder on the obvious features of the de-noising ECG signal, so that the reference ECG signal to be denoised is The residual ECG signal after noise is superimposed to obtain the noise-reduced ECG signal, which can better retain the R peak position in the ECG signal to be noise-reduced, and reduce the distortion of the ECG signal after noise reduction.

Please refer to the schematic illustration of the denoising result of the ECG signal of the target denoising autoencoder shown in FIG. 8D. 8D shows the ECG signal to be denoised, the ideal ECG signal (that is, the ECG signal that the ECG signal to be denoised is expected to obtain through noise reduction processing), and the convolutional autoencoder in the prior art The denoised ECG signal obtained after the noise reduction process is performed on the noisy ECG signal, the denoised ECG signal is reduced by the target denoising autoencoder in the embodiment of the present application and the ECG signal denoising method provided in the present application The denoised ECG signal obtained after noise processing. It can be seen from FIG. 8D that, compared with the prior art method for reducing the noise of the ECG signal, the method of the embodiment of the present application better extracts the detailed features of the ECG signal to be denoised, and reduces the denoised ECG signal. Distortion, improve noise reduction performance.

It can be understood that the first embodiment is the training stage of the denoising autoencoder (the stage performed by the training device 120 as shown in FIG. 4), and the specific training adopts any one of the possibilities based on the first embodiment and the first embodiment. The noise reduction autoencoder provided in the implementation manner is performed; and the second embodiment can be understood as the application stage of the target noise reduction autoencoder obtained by training (the stage executed by the execution device 110 as shown in FIG. 4). It can be embodied as adopting the target noise reduction autoencoder trained in the first embodiment, and obtaining the output signal according to the input residual ECG signal to be denoised, that is, the residual ECG signal after noise reduction in the second embodiment. Finally, the remaining ECG signal after noise reduction and the reference ECG signal to be noise-reduced are superimposed to obtain the noise-reduced ECG signal.

The following describes the devices involved in the embodiments of the present application with reference to the drawings.

Fig. 9A is a schematic block diagram of a training device for noise reduction and self-encoding in an embodiment of the present invention. The noise reduction self-encoding training device 90 shown in FIG. 9A (the device 90 may specifically be the training device 120 of FIG. 4), which may include:

The superimposing unit 901 is used to superimpose the noise-free ECG signal and the EMG noise signal to obtain a synthetic ECG signal. The noise-free ECG signal includes an H-beat ECG signal, and each beat of the ECG signal includes a QRS complex, H is a positive integer greater than 1, and the signal-to-noise ratio of the noise-free ECG signal is not less than the first threshold;

A decomposition unit 902, configured to decompose the synthesized ECG signal into a reference ECG signal and a noisy residual ECG signal by using average beat subtraction;

The removing unit 903 is configured to remove the reference ECG signal from the noise-free ECG signal to obtain a noise-free residual ECG signal;

The training unit 904 is configured to train the noise-reducing autoencoder according to the noise-containing residual ECG signal and the noise-free residual ECG signal corresponding to the noise-containing residual ECG signal, wherein the noise-containing residual ECG signal is Training input, and the noise-free residual ECG signal corresponding to the noisy residual ECG signal is a training label.

In a possible implementation, the decomposition unit 902 is specifically configured to:

Perform average processing on the W beat ECG signal in the synthesized ECG signal to obtain a second average ECG signal, W is less than H, and W is a positive integer;

Substituting the second average ECG signal for A _j corresponding to the H-beat ECG signal in the synthesized ECG signal to obtain the reference ECG signal, A _j represents the H-beat ECG signal where R _j is The left and right sides of the center each take the ECG signal in the Δt interval, R _j is the apex of the QRS complex in the W beat ECG signal, j=1, 2...H;

The reference ECG signal is removed from the synthetic ECG signal to obtain the noisy residual ECG signal.

In a possible implementation, the decomposing unit 902 performing the averaging processing of the W beat ECG signal in the synthesized ECG signal to obtain the second average ECG signal specifically includes:

among them,

Represents the second average ECG signal, _Ak represents the ECG signal in the W-beat ECG signal that takes R _k as the center and takes the left and right Δt intervals, and R _k is the QRS wave in the W-beat ECG signal For the vertices of the group, k = 1, 2...W.

For the specific implementation of each unit in the embodiment of the present application, reference may be made to the related description in the foregoing embodiment 1, which will not be repeated here.

Fig. 9B is a schematic block diagram of a training device for noise reduction and self-encoding in an embodiment of the present invention. The training device 92 for noise reduction and self-encoding shown in FIG. 9B (the device 92 may specifically be the training device 120 of FIG. 4), which may include:

The superimposing unit 921 is used to superimpose the noise-free ECG signal and the EMG noise signal to obtain a synthesized ECG signal. The noise-free ECG signal includes an H-beat ECG signal, and each beat of the ECG signal includes a QRS complex, H is a positive integer greater than 1, and the signal-to-noise ratio of the noise-free ECG signal is not less than the first threshold;

The decomposition unit 922 is configured to decompose the noise-free ECG signal into a reference ECG signal and a noise-free residual ECG signal by using average beat subtraction;

The removing unit 923 is configured to remove the reference ECG signal from the synthesized ECG signal to obtain a noisy residual ECG signal;

The training unit 924 is configured to train the noise-reducing autoencoder according to the noise-containing residual ECG signal and the noise-free residual ECG signal corresponding to the noise-containing residual ECG signal, wherein the noise-containing residual ECG signal is Training input, and the noise-free residual ECG signal corresponding to the noisy residual ECG signal is a training label.

For the specific implementation of each unit in the embodiment of the present application, reference may be made to the related description in the second embodiment above, which will not be repeated here.

FIG. 10 is a schematic block diagram of an ECG signal noise reduction device provided by an embodiment of the present invention. The ECG signal noise reduction device 1000 shown in FIG. 10 (the device 1000 may specifically be the execution device 110 of FIG. 4), which may include :

The obtaining unit 1001 is configured to obtain an ECG signal to be noise-reduced, where the ECG signal to be noise-reduced includes M beats of the ECG signal, each beat of the ECG signal includes a QRS complex, and M is a positive integer greater than 1;

The first decomposition unit 1002 uses average beat subtraction to decompose the ECG signal to be denoised into the reference ECG signal to be denoised and the remaining ECG signal to be denoised;

The noise reduction unit 1003 is configured to input the residual ECG signal to be noise-reduced into the target noise reduction autoencoder to obtain the residual ECG signal after noise reduction;

The superimposing unit 1004 is configured to superimpose the reference ECG signal to be noise-reduced and the remaining ECG signal after noise reduction to obtain the ECG signal after noise reduction.

In a possible implementation, the first decomposition unit 1002 is specifically configured to:

Perform averaging processing on the N beats of the ECG signal in the ECG signal to be noise-reduced to obtain the first average ECG signal, where N is less than M, and N is a positive integer;

Substituting the first average ECG signal for the B _j corresponding to the M-beat ECG signal in the ECG signal to be noise-reduced to obtain the reference ECG signal to be noise-reduced, B _j represents the M-beat ECG signal Take R _j as the center and left and right to take the ECG signals in the Δt interval, R _j is the apex of the QRS complex in the M beat ECG signal, j=1, 2...M;

The reference ECG signal to be denoised is removed from the ECG signal to be denoised to obtain the remaining ECG signal to be denoised.

In a possible implementation, the first decomposition unit 1002 performing the averaging processing on the N-beat ECG signal in the ECG signal to be noise-reduced to obtain the first average ECG signal specifically includes:

among them,

Indicates the first average ECG, B _i represents the N ECG beat to about the center R _i in the ECG signal from each interval Δt, R _i beat from the ECG QRS wave of the N The vertices of the group, i = 1, 2...N.

Optionally, the device 1000 may further include some or all of the units in the noise reduction and self-encoding training device 90, 92 shown in FIG. 9A or FIG. 9B, and details are not described herein again.

For the specific implementation of each unit in the embodiment of the present application, reference may be made to the related description in the third embodiment above, which will not be repeated here.

FIG. 11 is a schematic diagram of the hardware structure of a training device for a noise reduction autoencoder provided by an embodiment of the present application. The training device 1100 of the noise reduction autoencoder shown in FIG. 11 (the device 1100 may specifically be a computer device) includes a memory 1101, a processor 1102, a communication interface 1103, and a bus 1104. Among them, the memory 1101, the processor 1102, and the communication interface 1103 implement communication connections between each other through the bus 1104.

The memory 1101 may be a read only memory (Read Only Memory, ROM), a static storage device, a dynamic storage device, or a random access memory (Random Access Memory, RAM). The memory 1101 may store a program. When the program stored in the memory 1101 is executed by the processor 1102, the processor 1102 and the communication interface 1103 are used to execute the training method of the noise reduction autoencoder described in Embodiment 1 or Embodiment 2 of this application The various steps.

The processor 1102 may adopt a general-purpose central processing unit (Central Processing Unit, CPU), microprocessor, application specific integrated circuit (Application Specific Integrated Circuit, ASIC), graphics processing unit (graphics processing unit, GPU), or one or more The integrated circuit is used to execute related programs to realize the functions required by the units in the training device of the noise reduction autoencoder in this embodiment of the application, or to execute the noise reduction self in the first or second embodiment of the method of this application. Encoder training method.

The processor 1102 may also be an integrated circuit chip with signal processing capability. In the implementation process, each step of the training method of the noise reduction autoencoder of this application can be completed by the integrated logic circuit of the hardware in the processor 1102 or instructions in the form of software. The aforementioned processor 1102 may also be a general-purpose processor, a digital signal processor (Digital Signal Processing, DSP), an application specific integrated circuit (ASIC), a ready-made programmable gate array (Field Programmable Gate Array, FPGA) or other programmable logic devices , Discrete gates or transistor logic devices, discrete hardware components. The methods, steps, and logical block diagrams disclosed in the embodiments of the present application can be implemented or executed. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor or the like. The steps of the method disclosed in the embodiments of the present application may be directly embodied as being executed and completed by a hardware decoding processor, or executed and completed by a combination of hardware and software modules in the decoding processor. The software module can be located in a mature storage medium in the field such as random access memory, flash memory, read-only memory, programmable read-only memory, or electrically erasable programmable memory, registers. The storage medium is located in the memory 1101, and the processor 1102 reads the information in the memory 1101, and combines its hardware to complete the functions required by the units included in the training device for the noise reduction autoencoder in this embodiment of the application, or execute the method of the application The training method of the noise reduction autoencoder of the first embodiment or the second embodiment.

The communication interface 1103 uses a transceiver device such as but not limited to a transceiver to implement communication between the device 1100 and other devices or communication networks. For example, training data (such as the training samples described in Embodiment 1 of the present application) can be obtained through the communication interface 1103.

The bus 1104 may include a path for transferring information between various components of the device 1100 (for example, the memory 1101, the processor 1102, and the communication interface 1103).

It should be understood that the superimposing units 901 and 921, the decomposing

units

902 and 922, the removing

units

903 and 923, and the

training units

904 and 924 in the training device 90 or 92 of the noise reduction autoencoder may be equivalent to the processor 1102.

For the specific implementation of each of the foregoing functional devices, reference may be made to the related description in the foregoing embodiment 1, and details are not repeated in the embodiment of the present application.

FIG. 12 is a schematic diagram of the hardware structure of an electrocardiographic signal noise reduction device provided by an embodiment of the present application. The electrocardiographic signal noise reduction device 1200 shown in FIG. 12 (the device 1200 may specifically be a computer device) includes a memory 1201, a processor 1202, an electrocardiographic sensor 1203, a communication interface 1204, and a bus 1205. Among them, the memory 1201, the processor 1202, and the communication interface 1204 implement communication connections between each other through the bus 1205.

The memory 1201 may be a read only memory (Read Only Memory, ROM), a static storage device, a dynamic storage device, or a random access memory (Random Access Memory, RAM). The memory 1201 may store a program. When the program stored in the memory 1201 is executed by the processor 1202, the processor 1202 and the communication interface 1204 are used to execute the steps of the method for reducing the noise of the ECG signal in the third embodiment of the present application.

The processor 1202 may adopt a general-purpose central processing unit (Central Processing Unit, CPU), a microprocessor, an application specific integrated circuit (Application Specific Integrated Circuit, ASIC), a graphics processing unit (graphics processing unit, GPU), or one or more The integrated circuit is used to execute related programs to realize the functions required by the units in the electrocardiographic signal noise reduction device 1200 in this embodiment of the present application, or to implement the electrocardiographic signal noise reduction method in the third method embodiment of this application.

The processor 1202 may also be an integrated circuit chip with signal processing capability. In the implementation process, the various steps of the method for reducing the noise of the electrocardiogram signal of the present application can be completed by the integrated logic circuit of hardware in the processor 1202 or instructions in the form of software. The aforementioned processor 1202 may also be a general-purpose processor, a digital signal processor (Digital Signal Processing, DSP), an application specific integrated circuit (ASIC), an off-the-shelf programmable gate array (Field Programmable Gate Array, FPGA) or other programmable logic devices , Discrete gates or transistor logic devices, discrete hardware components. The methods, steps, and logical block diagrams disclosed in the embodiments of the present application can be implemented or executed. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor or the like. The steps of the method disclosed in the embodiments of the present application may be directly embodied as being executed and completed by a hardware decoding processor, or executed and completed by a combination of hardware and software modules in the decoding processor. The software module can be located in a mature storage medium in the field such as random access memory, flash memory, read-only memory, programmable read-only memory, or electrically erasable programmable memory, registers. The storage medium is located in the memory 1201, and the processor 1202 reads the information in the memory 1201, and combines its hardware to complete the functions required by the units included in the electrocardiographic signal noise reduction device of the embodiment of the present application, or execute the method embodiment of the present application The noise reduction method of the ECG signal.

The ECG sensor 1203 includes two electrodes and an analog-to-digital conversion module. When the ECG sensor 1203 is working, the two electrodes decompose and contact the skin surface of different parts of the user to collect the user's ECG signal. The digital-to-analog conversion module is used to collect The obtained ECG signal is converted into a digitized ECG signal, which is the ECG signal to be noise-reduced in this embodiment of the application.

It should be understood that the ECG sensor 1203 is not a necessary part of the ECG signal noise reduction device 1200. In an implementation scenario, the ECG signal noise reduction device 1200 is specifically a wearable device such as a smart bracelet or a smart watch, and the ECG signal noise reduction device 1200 may include the ECG sensor 1203. In another implementation scenario, the ECG signal noise reduction device 1200 is specifically a mobile phone or server, etc. The ECG signal noise reduction device 1200 may not include the ECG sensor 1203, and the ECG signal noise reduction device 1200 can receive ECG collection The ECG signal to be noise-reduced sent by the device may be a wearable device such as a smart bracelet or a smart watch equipped with an ECG sensor. It should be understood that the embodiments of the present application may also include other application scenarios.

The communication interface 1204 uses a transceiving device such as but not limited to a transceiver to implement communication between the device 1200 and other devices or a communication network. For example, training data (such as the ECG signal to be noise-reduced as described in the second embodiment of the present application) can be obtained through the communication interface 1204.

The bus 1205 may include a path for transferring information between various components of the device 1200 (for example, the memory 1201, the processor 1202, and the communication interface 1204).

It should be understood that the acquisition unit 1001 in the ECG signal noise reduction device 1000 is equivalent to the communication interface 1204 or the ECG sensor 1203 in the ECG signal noise reduction device 1200; the first decomposition unit 1002 in the ECG signal noise reduction device 1000 The noise reduction unit 1003 and the superposition unit 1004 may be equivalent to the processor 1202.

For the specific implementation of the above-mentioned functional units, reference may be made to the relevant description in the above-mentioned embodiment 3, and details are not repeated in this embodiment of the application.

It should be noted that although the devices 1100 and 1200 shown in FIG. 11 and FIG. 12 only show a memory, a processor, and a communication interface, in the specific implementation process, those skilled in the art should understand that the devices 1100 and 1200 also include implementations. Other devices necessary for normal operation. At the same time, according to specific needs, those skilled in the art should understand that the devices 1100 and 1200 may also include hardware devices that implement other additional functions. In addition, those skilled in the art should understand that the apparatuses 1100 and 1200 may also only include the necessary devices for implementing the embodiments of the present application, and not necessarily all the devices shown in FIG. 11 or FIG. 12.

It can be understood that the device 1100 is equivalent to the training device 120 in FIG. 4, and the device 1200 is equivalent to the execution device 110 in FIG. 4. A person of ordinary skill in the art may be aware that the units and algorithm steps of the examples described in the embodiments disclosed in this document can be implemented by electronic hardware or a combination of computer software and electronic hardware. Whether these functions are executed by hardware or software depends on the specific application and design constraint conditions of the technical solution. Professionals and technicians can use different methods for each specific application to implement the described functions, but such implementation should not be considered beyond the scope of this application.

Those skilled in the art can clearly understand that, for the convenience and conciseness of description, the specific working process of the above-described system, device, and unit can refer to the corresponding process in the foregoing method embodiment, which will not be repeated here.

In the several embodiments provided in this application, it should be understood that the disclosed system, device, and method may be implemented in other ways. For example, the device embodiments described above are only illustrative. For example, the division of the units is only a logical function division, and there may be other divisions in actual implementation, for example, multiple units or components can be combined or It can be integrated into another system, or some features can be ignored or not implemented. In addition, the displayed or discussed mutual coupling or direct coupling or communication connection may be indirect coupling or communication connection through some interfaces, devices or units, and may be in electrical, mechanical or other forms.

The units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the objectives of the solutions of the embodiments.

In addition, the functional units in each embodiment of the present application may be integrated into one processing unit, or each unit may exist alone physically, or two or more units may be integrated into one unit.

If the function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer readable storage medium. Based on this understanding, the technical solution of this application essentially or the part that contributes to the existing technology or the part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including Several instructions are used to make a computer device (which may be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage media include: U disk, mobile hard disk, read-only memory (Read-Only Memory, ROM), random access memory (Random Access Memory, RAM), magnetic disk or optical disk and other media that can store program code .

The above are only specific implementations of this application, but the protection scope of this application is not limited to this. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed in this application. Should be covered within the scope of protection of this application. Therefore, the protection scope of this application should be subject to the protection scope of the claims.

Claims

An ECG signal noise reduction method, which is characterized in that it comprises:

Acquire an ECG signal to be denoised, where the ECG signal to be denoised includes M beats of the ECG signal, each beat of the ECG signal includes a QRS complex, and M is a positive integer greater than 1;

Decompose the ECG signal to be denoised into a reference ECG signal to be denoised and the remaining ECG signal to be denoised by using average beat subtraction;

Input the residual ECG signal to be noise-reduced into the target noise-reduction autoencoder to obtain the residual ECG signal after noise reduction;

The reference ECG signal to be noise-reduced and the remaining ECG signal after noise reduction are superimposed to obtain a noise-reduced ECG signal.
The method according to claim 1, characterized in that the use of average beat subtraction (average beat subtraction) to decompose the ECG signal to be denoised into a reference ECG signal to be denoised and a residual ECG signal to be denoised include:

Perform averaging processing on the N beats of the ECG signal in the ECG signal to be noise-reduced to obtain the first average ECG signal, where N is less than M, and N is a positive integer;

Substituting the first average ECG signal for the B j corresponding to the M-beat ECG signal in the ECG signal to be noise-reduced to obtain the reference ECG signal to be noise-reduced, B j represents the M-beat ECG signal Take R j as the center and left and right to take the ECG signals in the Δt interval, R j is the apex of the QRS complex in the M beat ECG signal, j=1, 2...M;

The reference ECG signal to be denoised is removed from the ECG signal to be denoised to obtain the remaining ECG signal to be denoised.
The method according to claim 2, wherein the averaging processing of the N-beat ECG signal in the ECG signal to be noise-reduced to obtain the first average ECG signal specifically comprises:

among them,
Indicates the first average ECG, B i represents the N ECG beat to about the center R i in the ECG signal from each interval Δt, R i beat from the ECG QRS wave of the N The vertices of the group, i = 1, 2...N.
The method according to any one of claims 1-3, characterized in that, before inputting the residual ECG signal to be noise-reduced into a target noise-reducing autoencoder, and obtaining the residual ECG signal after noise reduction, The method also includes:

The noise-free ECG signal and the EMG noise signal are superimposed to obtain a synthetic ECG signal. The noise-free ECG signal includes an H-beat ECG signal, and each ECG signal includes a QRS complex, and H is a positive value greater than 1. An integer, the signal-to-noise ratio of the noise-free ECG signal is not less than the first threshold;

Decomposing the synthetic ECG signal into a reference ECG signal and a noisy residual ECG signal by using average beat subtraction (average beat subtraction);

Removing the reference ECG signal from the noise-free ECG signal to obtain a noise-free residual ECG signal;

Train the noise-reducing autoencoder according to the noisy residual ECG signal and the noise-free residual ECG signal to obtain the target noise-reducing autoencoder, wherein the noisy residual ECG signal is a training input, so The noise-free residual ECG signal is a training label.
The method according to claim 4, wherein the decomposing the synthesized ECG signal into a reference ECG signal and a noisy residual ECG signal using average beat subtraction (average beat subtraction) comprises:

Perform average processing on the W beat ECG signal in the synthetic ECG signal to obtain a second average ECG signal, W is less than H, and W is a positive integer; the second average ECG signal replaces the synthetic ECG signal The A j corresponding to the H-beat ECG signal is obtained, and the reference ECG signal is obtained. A j represents the ECG signal in the interval of Δt between the left and right sides of the H-beat ECG signal with R j as the center, and R j is the W beat the apex of QRS complex in ECG signal, j = 1, 2……H;

The reference ECG signal is removed from the synthetic ECG signal to obtain the noisy residual ECG signal.
The method according to claim 5, wherein the averaging processing of the W beat ECG signal in the synthesized ECG signal to obtain the second average ECG signal specifically comprises:

among them,
Represents the second average ECG signal, Ak represents the ECG signal in the W-beat ECG signal that takes R k as the center and takes the left and right Δt intervals, and R k is the QRS wave in the W-beat ECG signal For the vertices of the group, k = 1, 2...W.
A method for training a noise reduction autoencoder, which is characterized in that it includes:

The noise-free ECG signal and the EMG noise signal are superimposed to obtain a synthetic ECG signal. The noise-free ECG signal includes an H-beat ECG signal, and each ECG signal includes a QRS complex, and H is a positive value greater than 1. An integer, the signal-to-noise ratio of the noise-free ECG signal is not less than the first threshold;

Decomposing the synthetic ECG signal into a reference ECG signal and a noisy residual ECG signal by using average beat subtraction (average beat subtraction);

Removing the reference ECG signal from the noise-free ECG signal to obtain a noise-free residual ECG signal;

The noise-reducing autoencoder is trained according to the noisy residual ECG signal and the noise-free residual ECG signal, wherein the noisy residual ECG signal is a training input, and the noise-free residual ECG signal is a training label .
8. The method according to claim 7, wherein the decomposing the synthesized ECG signal into a reference ECG signal and a noisy residual ECG signal by using average beat subtraction (average beat subtraction) comprises:

Perform average processing on the W beat ECG signal in the synthesized ECG signal to obtain a second average ECG signal, W is less than H, and W is a positive integer;

Substituting the second average ECG signal for A j corresponding to the H-beat ECG signal in the synthesized ECG signal to obtain the reference ECG signal, A j represents the H-beat ECG signal where R j is The left and right sides of the center each take the ECG signal in the Δt interval, R j is the apex of the QRS complex in the W beat ECG signal, j=1, 2...H;

The reference ECG signal is removed from the synthetic ECG signal to obtain the noisy residual ECG signal.
The method according to claim 8, wherein the averaging processing of the W beat ECG signal in the synthesized ECG signal to obtain the second average ECG signal specifically comprises:

among them,
Represents the second average ECG signal, Ak represents the ECG signal in the W-beat ECG signal that takes R k as the center and takes the left and right Δt intervals, and R k is the QRS wave in the W-beat ECG signal For the vertices of the group, k = 1, 2...W.
An ECG signal noise reduction device, which is characterized by comprising:

The acquiring unit is configured to acquire the ECG signal to be denoised, the ECG signal to be denoised includes M beats of the ECG signal, each beat of the ECG signal includes a QRS complex, and M is a positive integer greater than 1;

The first decomposition unit uses an average beat subtraction to decompose the ECG signal to be denoised into a reference ECG signal to be denoised and a residual ECG signal to be denoised;

A noise reduction unit, configured to input the residual ECG signal to be noise-reduced into the target noise reduction autoencoder to obtain the residual ECG signal after noise reduction;

The superimposing unit is used to superimpose the reference ECG signal to be noise-reduced and the remaining ECG signal after noise reduction to obtain the ECG signal after noise reduction.
The device according to claim 10, wherein the first decomposition unit is specifically configured to:

Perform averaging processing on the N beats of the ECG signal in the ECG signal to be noise-reduced to obtain the first average ECG signal, where N is less than M, and N is a positive integer;

Substituting the first average ECG signal for the B j corresponding to the M-beat ECG signal in the ECG signal to be noise-reduced to obtain the reference ECG signal to be noise-reduced, B j represents the M-beat ECG signal Take R j as the center and left and right to take the ECG signals in the Δt interval, R j is the apex of the QRS complex in the M beat ECG signal, j=1, 2...M;

The reference ECG signal to be denoised is removed from the ECG signal to be denoised to obtain the remaining ECG signal to be denoised.
The device according to claim 11, wherein the first decomposition unit performs the averaging processing of the N-beat ECG signal in the ECG signal to be noise-reduced, and obtaining the first average ECG signal specifically comprises :

among them,
Indicates the first average ECG, B i represents the N ECG beat to about the center R i in the ECG signal from each interval Δt, R i beat from the ECG QRS wave of the N The vertices of the group, i = 1, 2...N.
The device according to any one of claims 10-12, wherein the noise reduction unit executes the input of the remaining ECG signal to be noise-reduced into a target noise-reduction autoencoder to obtain the remaining noise-reduction Before the ECG signal, the device further includes:

The synthesis unit is used to superimpose the noise-free ECG signal and the EMG noise signal to obtain a synthesized ECG signal. The noise-free ECG signal includes an H beat ECG signal, and each beat ECG signal includes a QRS complex, H Is a positive integer greater than 1, and the signal-to-noise ratio of the noise-free ECG signal is not less than the first threshold;

The second decomposition unit is configured to decompose the synthetic ECG signal into a reference ECG signal and a noisy residual ECG signal by using average beat subtraction;

A removing unit for removing the reference ECG signal from the noise-free ECG signal to obtain a noise-free residual ECG signal;

The training unit is configured to train a noise-reducing autoencoder according to the noise-containing residual ECG signal and the noise-free residual ECG signal to obtain the target noise-reducing autoencoder, wherein the noise-containing residual ECG signal It is a training input, and the noise-free residual ECG signal is a training label.
The device according to claim 13, wherein the second decomposition unit is specifically configured to:

Perform average processing on the W beat ECG signal in the synthetic ECG signal to obtain a second average ECG signal, W is less than H, and W is a positive integer; the second average ECG signal replaces the synthetic ECG signal The A j corresponding to the H-beat ECG signal is obtained, and the reference ECG signal is obtained. A j represents the ECG signal in the interval of Δt between the left and right sides of the H-beat ECG signal with R j as the center, and R j is the W beat the apex of QRS complex in ECG signal, j = 1, 2……H;

The reference ECG signal is removed from the synthetic ECG signal to obtain the noisy residual ECG signal.
The device according to claim 14, wherein the second decomposition unit performing the averaging processing on the W beat ECG signal in the synthesized ECG signal to obtain the second average ECG signal specifically comprises:

among them,
Represents the second average ECG signal, Ak represents the ECG signal in the W-beat ECG signal that takes R k as the center and takes the left and right Δt intervals, and R k is the QRS wave in the W-beat ECG signal The vertices of the group, k = 1, 2...W.
A training device for a noise reduction autoencoder, which is characterized in that it comprises:

The superimposing unit is used to superimpose the noise-free ECG signal and the EMG noise signal to obtain a synthetic ECG signal. The noise-free ECG signal includes an H beat ECG signal, and each beat ECG signal includes a QRS complex, H Is a positive integer greater than 1, and the signal-to-noise ratio of the noise-free ECG signal is not less than the first threshold;

A decomposition unit, configured to decompose the synthesized ECG signal into a reference ECG signal and a noisy residual ECG signal by using average beat subtraction (average beat subtraction);

A removing unit for removing the reference ECG signal from the noise-free ECG signal to obtain a noise-free residual ECG signal;

The training unit is configured to train the noise-reducing autoencoder according to the noisy residual ECG signal and the noise-free residual ECG signal, wherein the noisy residual ECG signal is a training input, and the noise-free residual ECG signal The electrical signal is a training label.
The device according to claim 16, wherein the decomposition unit is specifically configured to:

Perform average processing on the W beat ECG signal in the synthesized ECG signal to obtain a second average ECG signal, W is less than H, and W is a positive integer;

Substituting the second average ECG signal for A j corresponding to the H-beat ECG signal in the synthesized ECG signal to obtain the reference ECG signal, where A j represents the H-beat ECG signal where R j is The left and right sides of the center each take the ECG signal in the Δt interval, R j is the apex of the QRS complex in the W beat ECG signal, j=1, 2...H;

The reference ECG signal is removed from the synthetic ECG signal to obtain the noisy residual ECG signal.
18. The device according to claim 17, wherein the decomposition unit performing the averaging processing on the W beat ECG signal in the synthesized ECG signal to obtain the second average ECG signal specifically comprises:

among them,
Represents the second average ECG signal, Ak represents the ECG signal in the W-beat ECG signal that takes R k as the center and takes the left and right Δt intervals, and R k is the QRS wave in the W-beat ECG signal The vertices of the group, k = 1, 2...W.
An electrocardiographic signal noise reduction device, characterized by comprising: a memory and a processor, the memory is used for a program, the processor executes the program stored in the memory, and when the program stored in the memory is executed, the The processor is used to execute the ECG signal noise reduction method according to any one of claims 1-6.
A training device for a noise reduction autoencoder, which is characterized by comprising: a memory and a processor, the memory is used for a program, the processor executes the program stored in the memory, and when the program stored in the memory is executed, The processor is configured to execute the method for training a noise reduction autoencoder according to any one of claims 7-9.
A computer-readable storage medium, wherein the computer-readable medium stores program code for device execution, and the program code includes the ECG signal noise reduction method according to any one of claims 1-6 .
A computer-readable storage medium, wherein the computer-readable medium stores program code for device execution, and the program code includes the noise reduction autoencoder according to any one of claims 7-9 Training method.