阅读说明：本技术 心电信号噪声处理方法 (Electrocardiosignal noise processing method ) 是由 王丽荣 朱文亮 邱励燊 蔡文强 王朵朵 俞杰 张淼 于 2021-01-15 设计创作，主要内容包括：本发明公开了一种心电信号噪声处理方法,包括以下步骤：使用人工合成的数据集训练轻量深度学习网络得到训练完成的轻量深度学习网络,测试后端应用算法得到对应的预设熵阈值；将采集到的心电信号数据分段并输入训练完成的训练轻量深度学习网络,分类得到含噪声部分的信号段数据；计算含噪声部分的信号段数据的样本熵,与预设熵阈值比较并将大于预设熵阈值的信号段数据去除,得到去噪后的心电信号数据。本发明通过轻量深度学习网络直接对段信号进行分类,避免了人工提取特征的弊端；通过样本熵对分段信号去噪,有效降低肌电干扰、电极运动干扰噪声对诊断系统造成的误警,提高后端应用算法在心电信号中含有噪声时的准确性。(The invention discloses an electrocardiosignal noise processing method, which comprises the following steps: training a light-weight deep learning network by using a artificially synthesized data set to obtain a trained light-weight deep learning network, and testing a back-end application algorithm to obtain a corresponding preset entropy threshold; segmenting the collected electrocardiosignal data, inputting the segmented electrocardiosignal data into a trained lightweight deep learning network, and classifying to obtain signal segment data containing a noise part; and calculating the sample entropy of the signal segment data of the noise-containing part, comparing the sample entropy with a preset entropy threshold value, and removing the signal segment data larger than the preset entropy threshold value to obtain the denoised electrocardiosignal data. According to the method, the segment signals are directly classified through the lightweight deep learning network, so that the defect of manual feature extraction is overcome; the segmented signals are denoised through the sample entropy, false alarms of the diagnosis system caused by electromyographic interference and electrode movement interference noise are effectively reduced, and the accuracy of the rear-end application algorithm when the electrocardiosignal contains noise is improved.)

1. An electrocardiosignal noise processing method is characterized in that: the method comprises the following steps:

training a lightweight deep learning network by using an artificially synthesized data set to obtain the trained lightweight deep learning network, and testing a rear-end application algorithm by using the artificially synthesized data set to obtain a preset entropy threshold corresponding to the rear-end application algorithm;

segmenting the collected electrocardiosignal data, inputting the segmented electrocardiosignal data into a trained lightweight deep learning network, and classifying to obtain signal segment data containing a noise part;

and calculating the sample entropy of the signal segment data of the noise-containing part, comparing the sample entropy with a preset entropy threshold corresponding to a rear-end application algorithm, and removing the signal segment data larger than the preset entropy threshold to obtain the denoised electrocardiosignal data.

2. The method for processing electrocardiosignal noise according to claim 1, characterized in that: the light weight deep learning network is a double-branch segmentation network which comprises a learning down-sampling stage, a global feature extraction stage, a feature fusion stage and a classifier stage,

the learning down-sampling stage is used for extracting shallow features, the global feature extraction stage is used for extracting signal global semantics, the feature fusion stage is used for fusing features of different fine granularities and different levels in the electrocardiosignal, and the classifier stage is used for outputting the category of each sampling point in the signal.

3. The method for processing electrocardiosignal noise according to claim 2, characterized in that: the learning down-sampling stage comprises a two-dimensional convolution layer and two depth separable convolution structure layers with different channel numbers.

4. The method for processing noise of electrocardiographic signals according to claim 3, characterized in that: the two-dimensional convolution layer in the study downsampling stage and the depth separable convolution structure layer rear end that two channel numbers are different are all equipped with batch standard layer, the activation function that adopts in the depth separable convolution structure layer that two-dimensional convolution layer and two channel numbers are different is Leaky ReLU, use Depthwise convolution and Pointwise convolution simultaneously in the depth separable convolution structure layer for can distinguish mutually with the depth separable convolution.

5. The method for processing electrocardiosignal noise according to claim 2, characterized in that: the global feature extraction stage comprises a bottleneck layer and a pyramid pooling layer, wherein the bottleneck layer is used for reducing network parameters and maintaining network accuracy, and the pyramid pooling layer is used for extracting features in the multiple scale signals.

6. The method for processing electrocardiosignal noise according to claim 2, characterized in that: the feature fusion phase comprises a first layer, a second layer and an addition layer, wherein the first layer comprises an upper sampling layer, a depth separable convolution structure layer and a two-dimensional convolution layer, the second layer is a two-dimensional convolution layer, the addition layer is used for adding output results of the first layer and the second layer, and the addition layer is a batch normalization layer.

7. The method for processing noise of an electrocardiographic signal according to claim 6, characterized in that: the depth separable convolution structure layer in the first layer, the two-dimensional convolution layer in the first layer and the two-dimensional convolution layer in the second layer are all provided with batch standardization layers at the rear ends, and the activation functions adopted in the depth separable convolution structure layer in the first layer and the addition layer are both Leaky ReLU.

8. The method for processing electrocardiosignal noise according to claim 2, characterized in that: the classifier stage includes a plurality of depth separable convolutional structural layers of different scales, a two-dimensional convolutional layer, and an upsampling layer.

9. The method for processing noise of electrocardiographic signals according to claim 8, characterized in that: the multi-scale deep separable convolution structure layers and the two-dimensional convolution layer in the classifier stage are provided with batch standardization layers at the rear ends, and the activation functions adopted in the multi-scale deep separable convolution structure layers, the two-dimensional convolution layers and the up-sampling layers in the classifier stage are all Leaky ReLU.

10. The method for processing noise of an electrocardiographic signal according to any one of claims 1 to 9, characterized in that: the specific method for calculating the sample entropy of the signal segment data containing the noise part comprises the following steps:

forming a discrete time sequence with the length of N into a vector sequence X with the length of m and the dimension of N-m +1_m(i) Wherein m and r are hyper-parameters of sample entropy, and determine the size and similarity threshold of the same element to be searched;

vector X_m(i) And X_m(j) Distance d [ X ] between_m(i)，X_m(j)](j is more than or equal to 1 and less than or equal to N-m, j is not equal to i) is the absolute value of the maximum difference value of the two corresponding elements, and the calculation formula is d [ X ]_m(i)，X_m(j)]＝max_{k＝0，...，m-1}(|x(i+k)-x(j+k)|)；

For X_m(i) Calculating d [ X ]_m(i)，X_m(j)]R is not more than r, (1 is not less than j and not more than N-m, j is not equal to i) in_m(j) Number of (B)_i，B_iIs calculated by the formulaCalculate two ordersProbability B of column matching m points under similarity tolerance r^m(r) is calculated by the formula

Vector sequence X_m(i) Increasing the length of (c) to m +1 to obtain X_m+1(i) Calculating d [ X ]_m+1(i)，X_m+1(j)]R is not more than r, (1 is not less than j and not more than N-m, j is not equal to i) in_m+1(j) Number of (A)_i，A^m(r) is calculated byCalculate the probability A that two sequences match m +1 points^m(r) is calculated by the formula

The calculation formula of the sample entropy isSample entropy of N is finite

Technical Field

The invention relates to the technical field of electrocardiosignal processing, in particular to an electrocardiosignal noise processing method.

Background

Arrhythmia usually has the characteristics of instantaneity, paroxysmal and sometimes no symptoms, and brings challenges to diagnosis, and electrocardiosignal analysis is one of effective means for judging heart diseases. The existing electrocardiosignal acquisition equipment can only intermittently acquire electrocardiosignals, heart diseases with sporadic characteristics can not be noticed timely, and therefore a patient is generally required to be equipped with a wearable electrocardiosignal monitoring device to dynamically monitor the electrocardiosignals. However, the dynamic monitoring device also introduces more noise while monitoring the electrocardiosignals in real time. At present, the study of clinical electrocardiosignal diagnosis is endless, but serious signal interference caused by noise still causes many false alarms, so that the alarm of a guardian to the system becomes numb, the phenomenon of alarm fatigue occurs, and the alarm is finally ignored. The dynamic electrocardiosignal mainly comprises various interferences such as baseline drift, power frequency interference, electromyographic interference, electrode movement interference and the like, wherein the baseline drift and the power frequency interference have mature elimination algorithms, and the electrode movement interference and the electromyographic interference are still difficult to filter. The parts with larger electrode movement interference and myoelectric interference are removed, the performance index of the rear-end application algorithm in a noise environment is improved, and the false alarm rate can be reduced.

In recent years, many studies have made a great contribution to the work of evaluating quality signals of electrocardio. In the year 2011 of CinC challenge [1], there are several SQI indicators (Signal Quality indicators) used to evaluate segments and classify a segment of cardiac Signal into various Quality grades, and similar research is gradually increasing thereafter. Typical signal quality indicators include bSQI, tSQI, iSQI, aSQI, pSQI, sSQI, kSQI, basSQI, etc., collectively referred to as SQIs [2, 3 ]. Different SQIs combinations appear in different researches, NeginYaghmaie et al [4] propose a new signal quality index dSQI, combine with four SQIs, and use a support vector machine to classify clean normal electrocardiosignals and abnormal electrocardiosignals with noise respectively, wherein the accuracy rates are respectively 96.9% and 96.3%. Zhao, Z et al [5] combined qSQI, kSIQ, pSQI, basSQI, then combined with cauchy distribution, rectangular distribution and trapezoidal distribution, quantized the membership functions of sqi, established fuzzy vectors, then selected bounded operators for fuzzy synthesis, evaluated and classified by the weighted membership functions, and obtained 94.67% accuracy in the high-quality and low-quality two-classification tasks. Zhang, Y et al [6] propose Lempel-Ziv complexity as an ECG signal quality evaluation index. Subsequently, Liu, c.y. et al [7] combined typical sqiis with sample entropy, fuzzy measure entropy, Lempel-Ziv complexity, and used an SVM classifier to divide the signal quality into 5 levels, which achieved a fair effect. In addition to using SQIs to classify signal quality, Satija, U.S. et al [8] decompose electrocardiosignals through CEEMD, extract features of different noises in IMFs, and achieve noise localization and classification. Moeyerson, j. et al [9] split the ecg signal into 5-second signal segments and extract its ACF, extract features from the ACF, and divide the signal quality into 5 levels by a RUSBoost classifier. In addition, Zhang, q. et al [10] convert the time-frequency spectral signals into pictures with a resolution of 257 × 63, use multiple cascaded CNNs as classifiers, divide the electrocardiographic signals into 5 levels, and achieve an accuracy of 92.7% in the public database. Finally, Satija, U.S. et al [11] performed a very good summary of relevant studies before 2017, and were of great significance to the electrocardiosignal quality assessment work.

Meanwhile, part of research results can identify and remove the part with poor quality of the electrocardiosignal, and a good effect is achieved. Mico, P. et al [12] obtained 97% sensitivity and 16% false detection rate in the task of filtering MA signals, based on sample entropy, using its sensitive characteristic to noise irregularities, verified by MIT-BIH database using a sliding window calculation method. Satija, U.S. et al [13] first decompose electrocardiosignals using wavelets, then perform feature extraction on the decomposed different frequency band coefficients, and finally use a plurality of empirical thresholds to achieve the positioning and classification of noise and obtain good results. Bashar, s.k. et al [14] respectively look for noise features in the time and frequency domains of the cardiac signal, use empirical thresholds to achieve more detailed identification of the portions of poor cardiac signal quality, and reduce 94% false positives in atrial fibrillation detection.

However, the existing signal quality indicators (sqs) are features of artificial feature selection, and in any combination of the sqs, the problems of feature redundancy or insufficient features may exist; moreover, the extraction accuracy of some signal quality indexes is related to an extraction algorithm of the indexes, such as a pSQI accuracy-dependent R-wave positioning algorithm, so that the accuracy of the electrocardiosignal quality classification depends on the extraction algorithm of the signal quality indexes. Moreover, in the existing research, the basis for dividing the signal quality grade comes from subjective judgment of a annotator, not from a back-end application algorithm, so that the signal grade is not necessarily suitable for various back-end application algorithms. Meanwhile, in the real electrocardiosignal, the gradual change phenomena of the contents of several common noises, including electrode motion interference and myoelectricity interference, can occur, which leads to the hesitation psychology of the same annotator in the process of signal level annotation. Although some researches have established labeling rules for signal level labeling work in a hesitation state, the hesitation phenomenon cannot be avoided in practice, and the problem of ambiguous labels from the same annotator is difficult to eliminate.

The present invention references are as follows:

[1]Silva,I.,G.B.Moody,and L.Celi,Improving the Quality of ECGs Collected Using Mobile Phones:The PhysioNet/Computing in Cardiology Challenge 2011.2011Computing in Cardiology,2011.38:p.273-276.

[2]Behar,J.,et al.,A single channel ECG quality metric,in Computing in Cardiology.2012.p.381-384.

[3]Clifford,G.D.,et al.,Signal quality indices and data fusion for determining clinical acceptability ofelectrocardiograms.PhysiolMeas,2012.33(9):p.1419-33.

[4]Yaghmaie,N.,et al.,Dynamic signal quality index for electrocardiograms.PhysiolMeas,2018.39(10):p.105008.

[5]Zhao,Z.and Y.Zhang,SQI Quality Evaluation Mechanism of Single-Lead ECG Signal Based on Simple Heuristic Fusion and Fuzzy Comprehensive Evaluation.Front Physiol,2018.9:p.727.

[6]Zhang,Y.,et al.,Using Lempel-Ziv Complexity to Assess ECG Signal Quality.J Med Biol Eng,2016.36(5):p.625-634.

[7]Liu,C.Y.,et al.,Signal Quality Assessment and Lightweight QRS Detection for Wearable ECG SmartVest System.Ieee Internet ofThings Journal,2019.6(2):p.1363-1374.

[8]Satija,U.,B.Ramkumar,and M.Manikandan,Automated ECG Noise Detection and Classification System for Unsupervised Healthcare Monitoring.IEEE Journal ofBiomedical and Health Informatics,2017.PP.

[9]Moeyersons,J.,et al.,Artefact detection and quality assessment of ambulatory ECG signals.Comput Methods Programs Biomed,2019.182:p.105050.

[10]Zhang,Q.,L.Fu,and L.Gu,A Cascaded Convolutional Neural Network for Assessing Signal Quality of Dynamic ECG.Comput Math Methods Med,2019.2019:p.7095137.

[11]Satija,U.,B.Ramkumar,and M.S.Manikandan,A Review of Signal Processing Techniques for Electrocardiogram Signal Quality Assessment.IEEE Rev Biomed Eng,2018.11:p.36-52.

[12]Mico,P.,et al.,Automatic segmentation of long-term ECG signals corrupted with broadband noise based on sample entropy.Comput Methods Programs Biomed,2010.98(2):p.118-29.

[13]Satija,U.,B.Ramkumar,and M.S.Manikandan,An automated ECG signal quality assessment method for unsupervised diagnostic systems.Biocybernetics and Biomedical Engineering,2018.38(1):p.54-70.

[14]Bashar,S.K.,et al.,Noise Detection in Electrocardiogram Signals for Intensive Care Unit Patients.IEEE Access,2019.7:p.88357-88368.

disclosure of Invention

The invention aims to provide an electrocardiosignal noise processing method which can effectively reduce false alarms of electromyographic interference and electrode movement interference noise on a diagnosis system and improve the accuracy in a noise environment.

In order to solve the technical problem, the invention provides an electrocardiosignal noise processing method, which comprises the following steps:

training a lightweight deep learning network by using an artificially synthesized data set to obtain the trained lightweight deep learning network, and testing a rear-end application algorithm by using the artificially synthesized data set to obtain a preset entropy threshold corresponding to the rear-end application algorithm;

segmenting the collected electrocardiosignal data, inputting the segmented electrocardiosignal data into a trained lightweight deep learning network, and classifying to obtain signal segment data containing a noise part;

and calculating the sample entropy of the signal segment data of the noise-containing part, comparing the sample entropy with a preset entropy threshold corresponding to a rear-end application algorithm, and removing the signal segment data larger than the preset entropy threshold to obtain the denoised electrocardiosignal data.

Further, the light-weight deep learning network is a dual-branch segmentation network, the dual-branch segmentation network comprises a learning down-sampling stage, a global feature extraction stage, a feature fusion stage and a classifier stage,

the learning down-sampling stage is used for extracting shallow features, the global feature extraction stage is used for extracting signal global semantics, the feature fusion stage is used for fusing features of different fine granularities and different levels in the electrocardiosignal, and the classifier stage is used for outputting the category of each sampling point in the signal.

Further, the learning down-sampling stage includes a two-dimensional convolution layer and two depth separable convolution structure layers with different channel numbers.

Furthermore, the two-dimensional convolution layer in the learning down-sampling stage and the rear end of the depth separable convolution structure layer with different two channel numbers are both provided with batch standardization layers, the activation functions adopted in the two-dimensional convolution layer and the depth separable convolution structure layer with different two channel numbers are both Leaky ReLU, and Depthwise convolution and Pointwise convolution are simultaneously used in the depth separable convolution structure layer and are used for being distinguished from the depth separable convolution.

Further, the global feature extraction stage includes a bottleneck layer for reducing network parameters while maintaining network accuracy, and a pyramid pooling layer for extracting features in the multiple scale signals.

Further, the feature fusion stage includes a first layer, a second layer and an addition layer, the first layer includes an upsampling layer, a depth separable convolution structure layer and a two-dimensional convolution layer, the second layer is a two-dimensional convolution layer, the addition layer is used for adding output results of the first layer and the second layer, and the addition layer is a batch normalization layer.

Further, a batch normalization layer is arranged at the rear end of each of the depth separable convolution structure layer in the first layer, the two-dimensional convolution layer in the first layer and the two-dimensional convolution layer in the second layer, and activation functions adopted in the depth separable convolution structure layer in the first layer and the addition layer are both Leaky ReLU.

Further, the classifier stage includes a plurality of depth separable convolutional structure layers of different scales, a two-dimensional convolutional layer, and an upsampling layer.

Furthermore, a batch normalization layer is arranged at the rear end of each of the depth separable convolution structure layers and the two-dimensional convolution layer with different scales in the classifier stage, and all the activation functions adopted in the depth separable convolution structure layers, the two-dimensional convolution layer and the up-sampling layer with different scales in the classifier stage are Leaky ReLU.

Further, the method for calculating the sample entropy of the signal segment data of the noise-containing part specifically includes:

forming discrete time sequence with length N into length m, dimensionVector sequence X with degree of N-m +1_m(i) Wherein m and r are hyper-parameters of sample entropy, and determine the size and similarity threshold of the same element to be searched;

vector X_m(i) And X_m(j) Distance d [ X ] between_m(i)，X_m(j)](j is more than or equal to 1 and less than or equal to N-m, j is not equal to i) is the absolute value of the maximum difference value of the two corresponding elements, and the calculation formula is d [ X ]_m(i)，X_m(j)]＝max_{k＝0，...，m-1}(|X(i+k)-x(j+k)|)；

For X_m(i) Calculating d [ X ]_m(i)，X_m(j)]R is not more than r, (1 is not less than j and not more than N-m, j is not equal to i) in_m(j) Number of (B)_i，B_iIs calculated by the formulaCalculate the probability B that two sequences match m points with a similarity tolerance r^m(r) is calculated by the formula

Vector sequence X_m(i) Increasing the length of (c) to m +1 to obtain X_m+1(i) Calculating d [ X ]_m+1(i)，X_m+1(j)]R is not more than r, (1 is not less than j and not more than N-m, j is not equal to i) in_m+1(j) Number of (A)_i，A^m(r) is calculated byCalculate the probability A that two sequences match m +1 points^m(r) is calculated by the formula

The calculation formula of the sample entropy isSample entropy of N is finite

The invention has the beneficial effects that:

(1) the existing Signal Quality Indexes (SQIs) are the characteristics of manual characteristic selection, and the problems of characteristic redundancy or insufficient characteristics exist in manual selection; meanwhile, there are cases where the accuracy of extraction of the signal quality index is related to the extraction algorithm of the index. For a given section of electrocardiosignal, the traditional method is to calculate the kurtosis, the skewness and the R wave interval goodness of fit of the section of electrocardiosignal, and then classify the section of electrocardiosignal according to the kurtosis, the skewness and the R wave goodness of fit. In the process, the skewness calculation, the kurtosis calculation and the R wave interval goodness of fit belong to artificial extraction features, wherein the R wave interval goodness of fit depends on the accuracy of two or more R wave extraction algorithms, and if the R wave extraction accuracy is poor, the feature fails, so that the final classification result of the signal section has deviation indirectly. In the invention, the lightweight deep learning network is used for directly classifying the segment signals, thereby avoiding the defect of manually extracting the characteristics; meanwhile, the light-weight deep learning network belongs to a network light-weight (small calculated amount) convolutional neural network, and is suitable for processing electrocardiosignals in real time.

(2) The basis for dividing signal quality grades in the existing research comes from subjective judgment of a annotator, not from a back-end application algorithm, and the signal grades are not necessarily suitable for various back-end application algorithms. According to the invention, the quantized value of the segmented signal is obtained through the sample entropy, and if the segmented signal exceeds a preset threshold value, the segmented signal is removed. The preset threshold is obtained by continuous testing, the accuracy of the algorithm applied by the back end of the artificially synthesized signal test is used, and the threshold is continuously adjusted until a proper preset threshold is obtained; meanwhile, aiming at different test back-end application algorithms, different corresponding preset thresholds are obtained through adjustment. The preset threshold value can effectively filter signals with excessive noise, can ensure the performance of back-end application algorithms such as an R wave detection algorithm, a heartbeat classification algorithm and the like, improves the accuracy in a noise environment, is suitable for various back-end application algorithms, and is very flexible.

(3) Aiming at the problem that the labeling hesitation psychology can cause ambiguous labels, the data used by the network training is artificially synthesized without manual intervention, so that the problem of ambiguous labels or phenomena does not exist. Meanwhile, compared with the grade, the threshold value is finer and smoother, the condition that too many signals are discarded when the signals are discarded according to the grade is avoided, and false alarm caused by electromyographic interference and electrode movement interference noise on a diagnosis system can be effectively reduced.

The foregoing description is only an overview of the technical solutions of the present invention, and in order to make the technical solutions of the present invention more clearly understood and to implement them in accordance with the contents of the description, the following detailed description is given with reference to the preferred embodiments of the present invention and the accompanying drawings.

Drawings

Fig. 1 is a schematic diagram of a noise signal processing flow in the embodiment of the present invention.

Fig. 2 is a schematic structural diagram of a two-branch split network according to the present invention.

Detailed Description

The present invention is further described below in conjunction with the following figures and specific examples so that those skilled in the art may better understand the present invention and practice it, but the examples are not intended to limit the present invention.

In the description of the present invention, the term "comprises/comprising" is intended to cover a non-exclusive inclusion, such that a process, method, system, article, or apparatus that comprises a list of steps or elements is not limited to the listed steps or elements but may alternatively include other steps or elements not expressly listed or inherent to such process, method, article, or apparatus.

The back-end application takes an R-peak wave detection algorithm as an example, and the processing flow of the electrocardiosignal containing noise in the embodiment is shown in fig. 1, and includes the following steps:

step 1: training the light-weight deep learning network by using the artificially synthesized data set to obtain the trained light-weight deep learning network, and testing the back-end application algorithm by using the artificially synthesized data set to obtain a preset entropy threshold corresponding to the back-end application algorithm. The artificially synthesized data set is generated by overlapping clean electrocardiosignals and noise, the clean electrocardiosignals and the noise are partially controllable, and the artificially synthesized data set is used for training the light-weight deep learning network and testing the back-end application algorithm, so that the light-weight deep learning network with good performance after training and a proper preset entropy threshold value can be obtained. The preset entropy threshold of the R-peak wave detection algorithm obtained by the test in this embodiment is 2.0.

Step 2: the collected electrocardiosignal data (electrocardiosignal damaged by electrode motion interference or electromyographic interference) are segmented and input into a trained lightweight deep learning network, and signal segment data of a noise-containing part is obtained by classification. In this embodiment, the acquired electrocardiographic signal data is divided into signals of one segment every 10s, and through light-weight network classification, the portion containing noise in the signal segment data is marked as 1, and the portion without noise is marked as 0.

The lightweight deep learning network is a dual-branch segmentation network, and the dual-branch segmentation network shown in fig. 2 includes four stages, namely a learning down-sampling stage, a global feature extraction stage, a feature fusion stage and a classifier stage.

The learning down-sampling stage is used for extracting shallow layer features, and comprises three layers of a 16-channel two-dimensional Convolution layer (Conv2D), a 24-channel depth Separable Convolution structure layer (DSConv) and a 32-channel depth Separable Convolution structure layer, wherein a Batch Normalization layer (BN, similar to common data Normalization, is a way of unifying dispersed data and is also a method for optimizing a neural network) is used at the rear end of each layer to improve the optimization speed of the network. In the 16-channel two-dimensional Convolution layer (Conv2D), the 24-channel depth Separable Convolution structure layer (DSConv) and the 32-channel depth Separable Convolution structure layer, the adopted activation function is LeakyReLU (the linear unit function with leakage correction is a variant of ReLu activation function, the output of the function has a very small gradient to negative input; because the derivative is always not zero, the occurrence of silent neurons can be reduced, learning based on the gradient is allowed, and the problem that the neurons cannot learn after the Relu function enters a negative interval is solved); the reason why the leakyreu is used to replace the ReLU is that a negative value exists in the electrocardiographic signal, and the ReLU (Rectified Linear Unit, also called a modified Linear Unit, is a commonly used activation function in an artificial neural network, and usually refers to a nonlinear function represented by a ramp function and a variant thereof) directly loses the signal characteristics of the negative value part. And in the 24-channel depth separable convolution structure layer (DSConv) and the 32-channel depth separable convolution structure layer (DSConv), Depthwise convolution and Pointwise convolution are simultaneously used for distinguishing from the depth separable convolution, and an activation function exists between the Depthwise convolution and the Pointwise convolution.

The specific structure of the learning down-sampling phase is shown in table 1, where BN denotes the batch normalization layer, f denotes the activation function, and all the activation functions f are leakage ReLU. "/BN" indicates that the output result passes through the BN layer; "/f" indicates that the output result passes through the activation function f; "/BN/f" indicates that the output result passes through the BN layer, and the obtained result passes through the activation function f.

Network input size	Layer name	Number of channels	Convolution size	Convolution step size
					3600×1×1	Conv2D/BN/f	16	8	2
1800×1×16	DSConv/BN/f	24	4	2
					900×1×24	DSConv/BN/f	32	2	2

TABLE 1 Structure Table for learning Down-sampling phase

The global feature extraction stage is used for extracting signal global semantics, and comprises a Bottleneck layer (Bottleneck, Bo, which is generally used in a network with a higher depth) of 32 channels, two 1X1 filters, which are respectively used for reducing and increasing feature dimensions, can reduce the number of parameters so as to reduce the calculated amount, and can more effectively and intuitively train data and extract features after dimension reduction, and a Pyramid pooling layer (which is a traditional network architecture model, wherein a Pyramid pooling layer (Py) is added between a convolution layer and a full connection layer so as to solve the problem of different sizes of input pictures) of 64 channels, wherein the Bottleneck layer is used for reducing network parameters and simultaneously keeping network accuracy, and is one of keys for realizing network light weight; the phenomenon that electrocardiosignals have multiple scales is considered in some researches, so that the pyramid pooling layer is used for extracting the features in the signals with the multiple scales and is beneficial to extracting global semantic information.

Through repeated tests, the result is not improved by adding a plurality of bottleneck layers, so that the combination of the bottleneck layers and the pyramid pooling layer is considered to fully extract the signal features, and the specific structure of the global feature extraction stage is shown in table 2:

TABLE 2 Structure Table of Global feature extraction stage

The characteristic fusion stage is used for fusing characteristics of different fine granularities and different levels in a electrocardiosignal, and comprises a first layer, a second layer and an addition layer, wherein the first layer comprises an upsampling layer (Usampled, Up), a 64-channel depth separable convolution structure layer (DSConv) and a 64-channel two-dimensional convolution layer (Conv2D), the second layer comprises a 64-channel two-dimensional convolution layer, the addition layer is used for adding output results of the first layer and the second layer, and the addition layer is a batch normalization layer (BN). The back ends of the depth separable convolution structure layers of the 64 channels in the first layer, the two-dimensional convolution layers of the 64 channels in the first layer and the two-dimensional convolution layers of the 64 channels in the second layer are respectively provided with a batch standardization layer, and the activation functions adopted in the depth separable convolution structure layers of the 64 channels in the first layer and the addition layer are both Leaky ReLU.

The specific structure of the feature fusion stage is shown in table 3:

TABLE 3 structural Table of the feature fusion phase

The output of input signal segment data after passing through the structure in the learning down-sampling stage in table 1 is divided into two parts, one part of the input signal segment data is used as the input of the feature fusion stage after passing through the bottleneck layer and the pyramid pooling layer in the global feature extraction stage in table 2, and the input signal segment data sequentially passes through the up-sampling layer, the two-dimensional convolution layer and the depth separable convolution structure layer in the first layer in table 3; the other part is directly input into the feature fusion stage from the learning down-sampling stage and passes through a depth separable convolution structure layer of a second layer in the table 3; the outputs of the two parts are summed in the batch normalization layer of table 3, with the summed structure passing through the activation function f as input to the classifier stage.

The classifier stage is for outputting a class for each sample point in the signal, the classifier stage comprising a plurality of depth separable convolutional structure layers (DSConv) of different scales, a 1-channel two-dimensional convolutional layer (Conv2D), and an upsampling layer (upsamplie). Through a plurality of experiments, it is found that the segmentation accuracy of the multi-layer DSConv with different scales is higher than that of the multi-layer DSConv with the fixed scale, which is called as a multi-scale Serial depth separable convolution Module (Mu) in the present invention, the multi-scale serial depth-separable convolution module includes four different-scale depth-separable convolution structure layers of 64-channel DSConv, 32-channel DSConv, 16-channel DSConv and 8-channel DSConv, the multiple different-scale depth-separable convolution structure layers in the classifier stage and the back end of the 1-channel two-dimensional convolution layer are both provided with a batch normalization layer, the activation functions employed in the depth separable convolution structure layers of the plurality of different scales, the two-dimensional convolution layer of the 1 channel, and the upsampling layer in the classifier stage are all Leaky ReLU, and the specific structure of the classifier stage is shown in Table 4:

network input size	Layer name	Multiple of up-sampling	Number of channels	Convolution size	Convolution step size
						450×1×64	DSConv/BN/f	-	64	8	1
450×1×64	DSConv/BN/f	-	32	16	1
						450×1×32	DSConv/BN/f	-	16	32	1
450×1×16	DSConv/BN/f	-	8	64	1
						450×1×8	Conv2D/BN/f	-	1	1	1
450×1×1	Upsample/f	8	-	-	-

TABLE 4 structural Table of classifier stages

And step 3: and calculating the sample entropy of the signal segment data of the noise-containing part, comparing the sample entropy with a preset entropy threshold corresponding to a rear-end application algorithm, and removing the signal segment data larger than the preset entropy threshold to obtain the denoised electrocardiosignal data. In this embodiment, the sample entropy of the signal segment data of the calculated noise-containing part is 2.5, which is greater than the preset sample entropy threshold value 2.0, so that the signal segment data of the part is removed, and the remaining signal segment data is used as the input of the R-peak wave detection algorithm.

The method quantizes the part damaged by the noise by using the sample entropy to obtain a clean signal; meanwhile, the sample entropy is a nonlinear measurement for estimating the regularity of the time sequence, similar patterns with the same length can be searched in the time sequence, the frequency and the possibility of the patterns are smaller, the entropy value of the sequence is larger, the entropy value is increased along with the reduction of the signal-to-noise ratio, and the sample entropy is closely related to the noise power; in addition, because the sample entropy value is hardly influenced by clean signal energy, the entropy values of the signal part which is not subjected to electrode movement interference and myoelectricity interference and the interfered part are obviously different. Therefore, the noise of the part damaged by the noise is removed after quantification by using the sample entropy, a clean signal can be obtained, and the input signal of the back-end application algorithm can be ensured to be clean, so that the performance of the back-end application algorithm is improved.

In this embodiment, the removing of the signal segment data larger than the preset entropy threshold is to delete the signal segment data larger than the preset entropy threshold or to zero the signal segment data larger than the preset entropy threshold. The specific method for calculating the sample entropy of the signal segment data of the noise-containing part comprises the following steps:

step 3-1: segment signal data of a discrete time sequence X (N) of length N { X (1), X (2) }, X (N) }, is composed into a vector sequence X of length m and dimension N-m +1_m(i)，X_m(i)＝[x(1)，x(i+1)，...，x(i-m+1)]I is more than or equal to 1 and less than or equal to N-m +1, wherein m and r are hyper-parameters of sample entropy, and the size and similarity threshold value of the same searched element are determined; in this embodiment, m is 2 and r is 0.25.

Step 3-2: definition vector X_m(i) And X_m(j) In-line with the aboveDistance d [ X ] between_m(i)，X_m(j)](j is more than or equal to 1 and less than or equal to N-m, j is not equal to i) is the absolute value of the maximum difference value of the two corresponding elements, and the calculation formula is as follows: d [ X ]_m(i)，X_m(j)]＝max_{k＝0，...，m-1}(| x (i + k) -x (j + k) |); wherein X_m(i) Is a sample set, i represents the ith sample set, N represents the number of vectors in the sample set, and k is the offset of ordinal number

Step 3-3: for X_m(i) Calculating d [ X ]_m(i)，X_m(j)]R is not more than r, (1 is not less than j and not more than N-m, j is not equal to i) in_m(j) Number of (B)_i，B_iThe calculation formula of (2):definition B^m(r) is the probability that two sequences match m points with a similarity tolerance r, the formula is calculated:

step 3-4: vector sequence X_m(i) Increasing the length of (c) to m +1 to obtain X_m+1(i) Calculating d [ X ]_m+1(i)，X_m+1(j)]R is not more than r, (1 is not less than j and not more than N-m, j is not equal to i) in_m+1(j) Number of (A)_i，A^m(r) is the formula: definition A^m(r) is the probability that two sequences match m +1 points, the formula is calculated:

step 3-5: calculation formula of sample entropy:when N is a finite value, the sample entropy is:

inputting the denoised electrocardiosignal data obtained in the step (3) into an R-peak wave detection algorithm applied at the rear end, and compared with the traditional method, the accuracy of the denoised electrocardiosignal data is improved from 40% to 98%, so that the beneficial effect of the method is further illustrated.

The above-mentioned embodiments are merely preferred embodiments for fully illustrating the present invention, and the scope of the present invention is not limited thereto. The equivalent substitution or change made by the technical personnel in the technical field on the basis of the invention is all within the protection scope of the invention. The protection scope of the invention is subject to the claims.