CN112704503B - Electrocardiosignal noise processing method - Google Patents

Abstract

The invention discloses an electrocardiographic signal noise processing method, comprising the following steps: using artificially synthesized data sets to train a lightweight deep learning network to obtain a trained lightweight deep learning network, and testing a back-end application algorithm to obtain corresponding preset entropy Threshold; segment the collected ECG signal data and input it into the trained lightweight deep learning network, and classify the signal segment data containing noise; calculate the sample entropy of the signal segment data containing noise, and the preset entropy The threshold value is compared and the signal segment data larger than the preset entropy threshold value is removed to obtain the denoised ECG signal data. The invention directly classifies the segment signals through a lightweight deep learning network, avoiding the drawbacks of manual feature extraction; denoising the segment signals through sample entropy, effectively reducing the false alarm caused by the electromyography interference and electrode motion interference noise to the diagnosis system , to improve the accuracy of the back-end application algorithm when the ECG signal contains noise.

Description

Electrocardiographic signal noise processing method

technical field

The invention relates to the technical field of electrocardiographic signal processing, in particular to a method for processing electrocardiographic signal noise.

Background technique

Cardiac arrhythmias are usually transient, paroxysmal, and sometimes asymptomatic, which brings challenges to diagnosis. ECG signal analysis is one of the effective methods for diagnosing cardiac diseases. Existing ECG signal acquisition equipment can only collect intermittently, and some sporadic cardiac diseases are often not noticed in time. Therefore, patients usually need to be equipped with a wearable ECG monitoring device to dynamically monitor ECG signals. However, the dynamic monitoring equipment also introduces more noise while monitoring the ECG signal in real time. At present, there are endless methods to study the diagnosis of clinical ECG signals, but severe signal interference caused by noise will still cause many false alarms, which makes the monitoring personnel become numb to the alarm of the system, resulting in the phenomenon of "alarm fatigue", and finally ignore the alarm. The ambulatory ECG signals mainly include baseline drift, power frequency interference, EMG interference, electrode motion interference and other interferences. Among them, baseline drift and power frequency interference already have relatively mature elimination algorithms, while electrode motion interference and EMG interference still remain. Difficult to filter out. Remove the parts with large electrode motion interference and EMG interference, improve the performance index of the back-end application algorithm in a noisy environment, and reduce the false alarm rate.

In recent years, many studies have made great contributions to the work of evaluating ECG quality signals. In the 2011 CinC Challenge [1], several SQI indicators (Signal Quality Index, speech quality index) were used to evaluate the segment, and a segment of ECG signal was divided into various quality levels. Since then, similar studies have gradually increased. Typical signal quality indicators include bSQI, tSQI, iSQI, aSQI, pSQI, sSQI, kSQI, basSQI, etc., which are collectively referred to as SQIs [2, 3]. Different combinations of SQIs have appeared in different studies, NeginYaghmaie et al. [4] proposed a new signal quality indicator dSQI, and combined with four SQIs, using support vector machine to clean the normal ECG signal and abnormal ECG The signals were classified separately from noisy ECG signals with 96.9% and 96.3% accuracy, respectively. Zhao, Z et al. [5] quantified the membership function of SQIs by combining qSQI, kSIQ, pSQI, and basSQI, and then combined Cauchy distribution, rectangular distribution and trapezoidal distribution to establish a fuzzy vector, and then select a bounded operator to carry out Fuzzy synthesis, which utilizes weighted membership functions for evaluation and classification, achieves 94.67% accuracy in both high-quality and low-quality binary classification tasks. Zhang, Y et al. [6] proposed the Lempel-Ziv complexity as an ECG signal quality evaluation index. Subsequently, Liu, C.Y. et al. [7] combined typical SQIs with sample entropy, fuzzy measure entropy, and Lempel-Ziv complexity, and used the SVM classifier to classify the signal quality into 5 levels, and achieved good results. In addition to using SQIs to classify the signal quality, Satija, U. et al. [8] decomposed the ECG signal through CEEMD, and extracted the features of different noises in the intrinsic mode decomposition function IMFs to realize the localization and classification of noise. Moeyersons, J. et al. [9] segmented the ECG signal into 5-second signal segments and extracted its ACF, then extracted features from the ACF, and classified the signal quality into 5 levels through the RUSBoost classifier. In addition, Zhang, Q. et al. [10] converted the time-spectral signal into a picture with a resolution of 257 × 63, used multiple cascaded CNNs as classifiers, and divided the ECG signal into 5 grades, in the public database. The accuracy rate reached 92.7%. Finally, Satija, U. et al. [11] made a very good summary of related research before 2017, which is of great significance to the evaluation of ECG signal quality.

At the same time, some research results can identify and eliminate the parts with poor ECG signal quality, and have achieved good results. Based on the sample entropy, Mico, P. et al. [12] used the method of sliding window calculation based on the sample entropy and its sensitivity to noise irregularity, and verified by the MIT-BIH database, and obtained 97% in the task of filtering out the MA signal. sensitivity and 16% false detection rate. Satija, U. et al. [13] first used wavelet to decompose the ECG signal, then extracted features from the decomposed coefficients of different frequency bands, and finally used multiple empirical thresholds to locate and classify noise, and achieved good results. . Bashar, S.K. et al. [14] searched for noise features in the time domain and frequency domain of the ECG signal respectively, and realized a more detailed identification of the poor quality part of the ECG signal by using the empirical threshold, and reduced 94% in the detection of atrial fibrillation. % false positives.

However, the existing signal quality indicators (SQIs) are the features of artificial feature selection, and no matter how they are combined, there may be redundant or insufficient features; and the extraction accuracy of some signal quality indicators is different from the extraction algorithm of the indicator. For example, the accuracy of pSQI depends on the R-wave localization algorithm, which leads to the accuracy of the ECG signal quality classification depending on the extraction algorithm of the signal quality index. Moreover, the basis for dividing the signal quality level in the existing research comes from the subjective judgment of the annotator, not from the back-end application algorithm, so the signal level is not necessarily suitable for a variety of back-end application algorithms. At the same time, in the real ECG signal, the content of several common noises will appear gradual change, including electrode motion interference and EMG interference, which causes the same annotator to be hesitant in the process of signal level labeling. Although some studies have formulated labeling rules for signal level labeling in hesitant state, in practice the phenomenon of hesitation is unavoidable, and the problem of ambiguous labels from the same labeler is difficult to eliminate.

References of the present invention 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 of electrocardiograms. Physiol Meas, 2012.33(9):p.1419-33.

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

[5] Zhao, Z. and Y. Zhang, SQI Quality Evaluation Mechanism of Single-LeadECG 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 SignalQuality. J Med Biol Eng, 2016.36(5):p.625-634.

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

[8] Satija, U., B. Ramkumar, and M. Manikandan, Automated ECG NoiseDetection and Classification System for Unsupervised Healthcare Monitoring. IEEE Journal of Biomedical 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 SignalProcessing 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 signalquality 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.

SUMMARY OF THE INVENTION

The technical problem to be solved by the present invention is to provide an electrocardiographic signal noise processing method that can effectively reduce the false alarm caused by the electromyographic interference and electrode motion interference noise to the diagnosis system, and improve the accuracy in the noise environment.

In order to solve the above-mentioned technical problems, the present invention provides a method for processing ECG signal noise, comprising the following steps:

Use the artificially synthesized data set to train the lightweight deep learning network to obtain the trained lightweight deep learning network, and use the artificially synthesized data set to test the back-end application algorithm to obtain the preset entropy threshold corresponding to the back-end application algorithm;

Segment the collected ECG signal data and input it into the trained lightweight deep learning network, and classify the signal segment data containing the noise part;

Calculate the sample entropy of the signal segment data containing noise, compare with the preset entropy threshold corresponding to the back-end application algorithm, and remove the signal segment data larger than the preset entropy threshold to obtain denoised ECG signal data.

Further, the lightweight deep learning network is a dual-branch segmentation network, and the dual-branch segmentation network includes a learning downsampling stage, a global feature extraction stage, a feature fusion stage, and a classifier stage,

The learning downsampling stage is used to extract shallow features, the global feature extraction stage is used to extract the global semantics of the signal, and the feature fusion stage is used to fuse features of different fine-grained and different levels in the ECG signal, so the The classifier stage described above is used to output the class of each sample point in the signal.

Further, the learning downsampling stage includes a two-dimensional convolutional layer and two depthwise separable convolutional structure layers with different numbers of channels.

Further, the two-dimensional convolutional layer in the learning downsampling stage and the two depthwise separable convolutional structure layers with different numbers of channels are provided with a batch normalization layer at the back end, and the two-dimensional convolutional layer and the two channels The activation functions used in different depthwise separable convolutional structure layers are all Leaky ReLU. In the depthwise separable convolutional structure layer, both Depthwise convolution and Pointwise convolution are used to distinguish it from depthwise separable convolution. .

Further, the global feature extraction stage includes a bottleneck layer and a pyramid pooling layer, the bottleneck layer is used to reduce network parameters while maintaining network accuracy, and the pyramid pooling layer is used to extract multiple scale signals. feature.

Further, the feature fusion stage includes a first layer, a second layer and an addition layer, the first layer includes an upsampling layer, a depthwise separable convolutional structure layer and a two-dimensional convolutional layer, the second layer is a two-dimensional convolutional layer, the addition layer is used to add the output results of the first layer and the second layer, and the addition layer is a batch normalization layer.

Further, the depthwise separable convolutional structure layer in the first layer, the two-dimensional convolutional layer in the first layer, and the backend of the two-dimensional convolutional layer in the second layer are all provided with batch normalization. layer, the depthwise separable convolution structure layer in the first layer and the activation function used in the addition layer are both Leaky ReLU.

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

Further, a batch normalization layer is provided at the back end of a plurality of depth-separable convolutional layers of different scales and a two-dimensional convolutional layer in the classifier stage, and a depth of a plurality of different scales in the classifier stage is separable. The activation functions used in the convolutional structure layer, the two-dimensional convolutional layer and the upsampling layer are all Leaky ReLU.

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

A discrete time series of length N is composed of a vector sequence X _m (i) of length m and dimension N-m+1, where m, r are hyperparameters of sample entropy, which determine the size and similarity of the search for the same element threshold;

The distance d[X _m (i), X _m (j)] (1≤j≤Nm, j≠i) between the vectors X _m (i) and X _m (j) is the maximum difference between the two corresponding elements The absolute value of , the calculation formula is d[X _m (i), X _m (j)]=max _{k=0, ..., m-1} (|X(i+k)-x(j+k)| );

计算两个序列在相似容限r下匹配m个点的概率B^m(r)，计算公式为

For X _m (i), calculate the number B _i , B _i of X _m (j) in d[X _m (i), X _m (j)]≤r, (1≤j≤Nm, j≠i) The calculation formula is

Calculate the probability B ^m (r) that the two sequences match m points under the similarity tolerance r, the calculation formula is

计算两个序列匹配m+1个点的概率A^m(r)，计算公式为

Increase the length of the vector sequence X _m (i) to m+1 to get X _m+1 (i), calculate d[X _m+1 (i), X _m+1 (j)]≤r, (1≤j ≤Nm, j≠i) the number A _i of X _m+1 (j), the calculation formula of A ^m (r) is:

Calculate the probability A ^m (r) that the two sequences match m+1 points, the calculation formula is

N为有限值时样本熵为

The formula for calculating sample entropy is

When N is a finite value, the sample entropy is

Beneficial effects of the present invention:

(1) The existing signal quality indicators (SQIs) are the features of manual feature selection, and there are problems of redundant or insufficient features in manual selection; at the same time, there are situations in which the extraction accuracy of signal quality indicators is related to the extraction algorithm of the indicators. For a given piece of ECG signal, the traditional method is to first find the kurtosis, skewness, and R-wave interval fit of the signal, and then use the kurtosis, skewness, and R-wave fit results for this segment of the signal. sort. In this process, seeking skewness, kurtosis and R wave interval fit are all features of artificial extraction, in which R wave interval fit depends on the accuracy of two or more R wave extraction algorithms. Poor extraction accuracy will lead to the failure of this feature, which indirectly leads to a deviation in the final classification result of this signal. In the present invention, a lightweight deep learning network is used to directly classify segment signals, avoiding the disadvantages of manually extracting features; at the same time, the lightweight deep learning network belongs to a kind of convolutional neural network with lightweight network (small amount of calculation), which is suitable for real-time Process the ECG signal.

(2) The basis for dividing the signal quality level in the existing research comes from the subjective judgment of the annotator, not from the back-end application algorithm, and the signal level is not necessarily suitable for a variety of back-end application algorithms. In the present invention, the quantized value of the segmented signal is obtained through the sample entropy, and if the segmented signal exceeds a preset threshold, the segmented signal is removed. The preset threshold is obtained through continuous testing. The accuracy of the back-end application algorithm is tested by artificially synthesized signals, and the threshold is continuously adjusted until a suitable preset threshold is obtained. the corresponding preset threshold. The preset threshold can not only effectively filter out the signal with too much noise, but also ensure the performance of the R-wave detection algorithm, heartbeat classification algorithm and other back-end application algorithms, improve the accuracy in noisy environments, and adapt to a variety of back-end application algorithms. flexible.

(3) In view of the problem that labeling hesitancy will lead to ambiguous labels, the data used for network training in the present invention is artificially synthesized and does not require manual intervention, so there are no ambiguous labels or problems due to phenomena. At the same time, the threshold value is more delicate than the level, which avoids the situation of discarding too many signals according to the level, and can effectively reduce the false alarm caused by the electromyography interference and electrode motion interference noise to the diagnosis system.

The above description is only an overview of the technical solution of the present invention. In order to understand the technical means of the present invention more clearly, and implement it according to the content of the description, the preferred embodiments of the present invention are described in detail below with the accompanying drawings.

Description of drawings

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

FIG. 2 is a schematic structural diagram of a dual-branch splitting network in the present invention.

Detailed ways

The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand the present invention and implement it, but the embodiments are not intended to limit the present invention.

In the description of the present invention, the term "comprising" is intended to cover non-exclusive inclusion, such as a process, method, system, product or device comprising a series of steps or elements, not limited to the listed steps or elements but rather Steps or units not listed are optionally included, or other steps or units inherent to these processes, methods, products, or devices are optionally included.

The back-end application takes the R-peak wave detection algorithm as an example. The processing flow of the ECG signal containing noise in this embodiment is shown in Figure 1, including the following steps:

Step 1: Use the artificially synthesized data set to train the lightweight deep learning network to obtain the trained lightweight deep learning network, and use the artificially synthesized data set to test the back-end application algorithm to obtain the preset entropy threshold corresponding to the back-end application algorithm. The artificially synthesized data set is generated by the superposition of clean ECG signal and noise. The clean ECG signal and noise are partially controllable. Using the artificially synthesized data set to train a lightweight deep learning network and test the back-end application algorithm can obtain good performance. The training of the completed lightweight deep learning network and a suitable preset entropy threshold. The preset entropy threshold of the R-peak wave detection algorithm obtained by testing in this embodiment is 2.0.

Step 2: Segment the collected ECG signal data (the ECG signal disrupted by electrode motion or EMG interference) and input it into the trained lightweight deep learning network, and classify the signal segment data containing noise. In this embodiment, the collected ECG signal data is divided into a signal segment of every 10s. After light network classification, the part of the signal segment data that contains noise is marked as 1, and the part without noise is marked as 0.

The lightweight deep learning network is a dual-branch segmentation network. The dual-branch segmentation network shown in Figure 2 includes four parts: a learning downsampling stage, a global feature extraction stage, a feature fusion stage, and a classifier stage.

The learning downsampling stage is used to extract shallow features, including a 16-channel two-dimensional convolutional layer (Conv2D), a 24-channel depthwise separable convolutional structure layer (Depthwise Separable Convolution, DSConv) and a 32-channel depthwise separable layer. The convolutional structure has three layers, and the back end of each layer uses a batch normalization layer (Batch Normalization, BN, similar to ordinary data standardization, is a method of unifying scattered data, and also a method of optimizing neural networks) to improve Optimized speed of the network. In the 16-channel two-dimensional convolutional layer (Conv2D), the 24-channel depthwise separable convolutional structure layer (Depthwise Separable Convolution, DSConv), and the 32-channel depthwise separable convolutional structure layer, the activation function used is LeakyReLU (The linear unit function with leak correction is a variant of the ReLu activation function, whose output has a small slope to negative inputs; since the derivative is always non-zero, it reduces the occurrence of silent neurons, allowing gradient-based learning, This solves the problem that the neurons do not learn after the Relu function enters the negative interval); the use of LeakyReLU instead of ReLU is due to the existence of negative values in the ECG signal, ReLU (Rectified Linear Unit, linear rectification function, also known as modified linear unit, is a The activation function commonly used in artificial neural networks, usually referring to the nonlinear function represented by the ramp function and its variants, will directly lose the signal features of the negative part. In the 24-channel depthwise separable convolutional structure layer (DSConv) and the 32-channel depthwise separable convolutional structure layer (DSConv), Depthwise convolution and Pointwise convolution are used at the same time, which are used to compare with the depthwise separable convolution. The difference, there is an activation function between Depthwise convolution and Pointwise convolution.

The specific structure of the learning downsampling stage is shown in Table 1, where BN represents the batch normalization layer, f represents the activation function, and all the activation functions f are Leaky ReLU. "/BN" means that the output result passes through the BN layer; "/f" means that the output result passes through the activation function f; "/BN/f" means 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 stride 3600×1×1 Conv2D/BN/f 16 8 2 1800×1×16 DSConv/BN/f twenty four 4 2 900×1×24 DSConv/BN/f 32 2 2

Table 1 Structure table of learning downsampling stage

The global feature extraction stage is used to extract the global semantics of the signal, including a 32-channel bottleneck layer (Bottleneck, Bo, generally used in a network with higher depth. Two 1X1 filters are included to reduce and increase the feature dimension, respectively, The number of parameters can be reduced to reduce the amount of calculation, and data training and feature extraction can be performed more effectively and intuitively after dimensionality reduction) and a 64-channel pyramid pooling layer (a traditional network architecture model, in which the convolution A layer of pyramid pooling (Pyramid pooling, Py) is added between the layer and the fully connected layer, which can solve the situation of different input image sizes), and the bottleneck layer is used to reduce network parameters while maintaining network accuracy. One of the keys to achieve network lightweight; some studies believe that ECG signals also have multiple scales, so the pyramid pooling layer is used to extract features from multiple scale signals, which is conducive to extracting global semantic information.

After repeated tests, adding multiple bottleneck layers did not improve the results, so it is believed that the combination of the bottleneck layer and the pyramid pooling layer has fully extracted the signal features. The specific structure of the global feature extraction stage is shown in Table 2:

Table 2 Structure table of global feature extraction stage

The feature fusion stage is used to fuse features of different fine-grained and different levels in the ECG signal, the feature fusion stage includes a first layer, a second layer and an addition layer, and the first layer includes an upsampling layer (Upsample, Up), a 64-channel depthwise separable convolutional structure layer (DSConv), and a 64-channel two-dimensional convolutional layer (Conv2D), the second layer includes a 64-channel two-dimensional convolutional layer, the phase The addition layer is used to add the output results of the first layer and the second layer, and the addition layer is a batch normalization layer (BN). The 64-channel depthwise separable convolutional structure layer in the first layer, the 64-channel two-dimensional convolutional layer in the first layer, and the 64-channel two-dimensional convolutional layer in the second layer are all set at the back end. There is a batch normalization layer, the 64-channel depthwise separable convolution structure layer in the first layer and the activation function used in the summation layer are both LeakyReLU.

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

Table 3 Structure table of feature fusion stage

The output of the input signal segment data after passing through the structure in the learning downsampling stage in Table 1 is divided into two parts, one part passes through the bottleneck layer and the pyramid pooling layer in the global feature extraction stage in Table 2 as the input of the feature fusion stage, and passes through successively. The upsampling layer, the two-dimensional convolutional layer and the depthwise separable convolutional structure layer in the first layer of Table 3; the other part is directly input from the learning downsampling stage to the feature fusion stage, and passes through the depthwise separable volume of the second layer of Table 3. The output of the two parts is added in the batch normalization layer of Table 3, and the added structure passes through the activation function f as the input of the classifier stage.

The classifier stage is used to output the category of each sampling point in the signal, and the classifier stage includes a plurality of depthwise separable convolutional structure layers with different scales (DSConv) and a 1-channel two-dimensional convolutional layer (Conv2D) And the upsampling layer (Upsample). It has been found through many experiments that the segmentation accuracy of multi-layer DSConv with different scales is higher than that of multi-layer fixed-scale DSConv, which is called a multi-scale serial depthseparable convolution module (Multiscale Serial Depth Separable Convolution Module (Multiscale Serial Depth Separable Convolution Module) DSConv Module, Mu). In this embodiment, the multi-scale serial depthwise separable convolution module includes four depth separable volumes of different scales: DSConv with 64 channels, DSConv with 32 channels, DSConv with 16 channels, and DSConv with 8 channels. In the classifier stage, multiple depth-separable convolutional structure layers of different scales and one-channel two-dimensional convolutional layers are provided with batch normalization layers at the back end. The activation functions used in the depthwise separable convolutional structure layer, 1-channel 2D convolutional layer and upsampling layer are all Leaky ReLU. The specific structure of the classifier stage is shown in Table 4:

network input size layer name Upsampling multiple number of channels Convolution size Convolution stride 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 Structure table of classifier stages

Step 3: Calculate the sample entropy of the signal segment data containing noise, compare with the preset entropy threshold corresponding to the back-end application algorithm, and remove the signal segment data larger than the preset entropy threshold to obtain denoised ECG signal data. In this embodiment, the sample entropy of the signal segment data of the noise-containing part is calculated to be 2.5, which is greater than the preset sample entropy threshold of 2.0. Therefore, this part of the signal segment data is removed, and the remaining signal segment data is used as the R-peak wave detection algorithm input of.

This method uses the sample entropy to quantify the noise-corrupted part, and can get a clean signal; at the same time, the sample entropy is a nonlinear measure to estimate the regularity of the time series, which can find similar patterns with the same length in the time series. The smaller the frequency and possibility of occurrence, the greater the entropy value of the sequence. With the decrease of the signal-to-noise ratio, the entropy value increases, and the sample entropy and noise power are closely related; It will be affected by the clean signal energy, so the entropy value of the signal part that is not disturbed by electrode movement and EMG is significantly different from that of the disturbed part. Therefore, using the sample entropy to quantize and denoise the noise-corrupted part can obtain a clean signal, which can ensure that the input signal of the back-end application algorithm is clean, thereby improving the performance of the back-end application algorithm.

In this embodiment, removing the signal segment data greater than the preset entropy threshold is deleting the signal segment data greater than the preset entropy threshold or zeroing the signal segment data greater than the preset entropy threshold. The specific method for calculating the sample entropy of the signal segment data containing the noise part is:

Step 3-1: The segmented signal data of the discrete time series X(n)={x(1), x(2), . . . , x(N)} of length N is composed of length m and dimension N-m+1 vector sequence X _m (i), X _m (i) = [x(1), x(i+1), ..., x(i-m+1)], 1≤i ≤N-m+1, where m and r are hyperparameters of sample entropy, which determine the size and similarity threshold of searching for the same element; in this embodiment, m is 2, and r is 0.25.

Step 3-2: Define the distance d[X _m (i), X _m (j)] (1≤j≤Nm, j≠i) between the vectors X _m (i) and X _m (j) as both The absolute value of the maximum difference in the corresponding element, the calculation formula: d[X _m (i), X _m (j)]=max _k=0,...,m-1 (|x(i+k)-x (j+k)|); where X _m (i) is the sample set, i represents the ith sample set, N represents the number of vectors in the sample set, and k is the offset of the ordinal number

定义B^m(r)为两个序列在相似容限r下匹配m个点的概率，计算公式：

Step 3-3: For X _m (i), calculate the number of X _m (j) in d[X _m (i), X _m (j)]≤r, (1≤j≤Nm, j≠i) B _i , the calculation formula of B _i :

Define B ^m (r) as the probability that two sequences match m points under the similarity tolerance r, and the calculation formula is:

定义A^m(r)为两个序列匹配m+1个点的概率，计算公式：

Step 3-4: Increase the length of the vector sequence X _m (i) to m+1 to obtain X _m+1 (i), calculate d[X _m+1 (i), X _m+1 (j)]≤r , (1≤j≤Nm, j≠i) the number A _i of X _m+1 (j), the calculation formula of A ^m (r):

Define A ^m (r) as the probability that two sequences match m+1 points, and the calculation formula is:

当N为有限值时，样本熵为：

Step 3-5: Calculation formula of sample entropy:

When N is a finite value, the sample entropy is:

The denoised ECG signal data obtained in step 3 is input into the R-peak wave detection algorithm applied at the back end. Compared with the traditional method, the accuracy rate of the denoised ECG signal data of the present invention is increased by 40%. It is increased to 98%, which further illustrates the beneficial effect of the present invention.

The above-mentioned embodiments are only preferred embodiments for fully illustrating the present invention, and the protection scope of the present invention is not limited thereto. Equivalent substitutions or transformations made by those skilled in the art on the basis of the present invention are all within the protection scope of the present invention. The protection scope of the present invention is subject to the claims.

Claims (9)

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; the light-weight deep learning network is a double-branch segmentation network, the double-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 electrocardiosignals, and the classifier stage is used for outputting the category of each sampling point in the signals;

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

3. The method for processing electrocardiosignal noise according to claim 2, 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.

4. The method for processing electrocardiosignal noise according to claim 1, 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.

5. The method for processing electrocardiosignal noise according to claim 1, 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.

6. The method for processing noise of an electrocardiographic signal according to claim 5, 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.

7. The method for processing electrocardiosignal noise according to claim 1, 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.

8. The method for processing noise in an electrocardiographic signal according to claim 7, 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.

9. The method for processing noise of an electrocardiographic signal according to any one of claims 1 to 8, 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 _i Is calculated by the formula

Calculate 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 less than r, (1 is not less than j is not less than N-m, j is not equal to i) X _m+1 (j) Number of (A) _i ，A ^m (r) is calculated by

Calculate two sequence matches m +1Probability of a point A ^m (r) is calculated by the formula

The calculation formula of the sample entropy is

Sample entropy of N is finite

Images