CN113935378A - ECG noise reduction method based on antagonistic depth full convolution network - Google Patents

ECG noise reduction method based on antagonistic depth full convolution network Download PDF

Info

Publication number
CN113935378A
CN113935378A CN202111196251.8A CN202111196251A CN113935378A CN 113935378 A CN113935378 A CN 113935378A CN 202111196251 A CN202111196251 A CN 202111196251A CN 113935378 A CN113935378 A CN 113935378A
Authority
CN
China
Prior art keywords
convolution
noise
signal
layer
parts
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202111196251.8A
Other languages
Chinese (zh)
Inventor
刘瑞霞
侯彦荣
舒明雷
陈长芳
周书旺
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Qilu University of Technology
Shandong Institute of Artificial Intelligence
Original Assignee
Qilu University of Technology
Shandong Institute of Artificial Intelligence
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Qilu University of Technology, Shandong Institute of Artificial Intelligence filed Critical Qilu University of Technology
Priority to CN202111196251.8A priority Critical patent/CN113935378A/en
Publication of CN113935378A publication Critical patent/CN113935378A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/02—Preprocessing
    • G06F2218/04—Denoising
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Abstract

An ECG noise reduction method based on an antagonistic depth full convolution network is characterized in that a convolution block is adopted in a model to further retain details, expansion convolution is adopted to reduce the complexity of calculation, and batch normalization is adopted for each layer to obtain better gradient flow for rapid convergence. And finally, the discriminator is utilized to better learn optimization. The design can embody a noise reduction signal with low complexity and high precision while keeping details.

Description

ECG noise reduction method based on antagonistic depth full convolution network
Technical Field
The invention relates to the technical field of electrocardiosignal processing, in particular to an ECG noise reduction method based on a antagonism depth full convolution network.
Background
Electrocardiosignals are an indispensable way for identifying heart abnormalities of human bodies, and as shown in figure 1, the electrocardiosignals mainly comprise P waves, QRS complex waves, T waves and U waves, and the abnormal conditions of all wave bands represent different body diseases respectively. The normal electrocardiosignal frequency range is in between, and the normal electrocardiosignal frequency range is concentrated in between.
The electrocardiosignal is used as an important index for identifying the heart abnormality of a human body, and the common problem is that some unnecessary and even interfering noises are introduced in the electrocardiosignal acquisition process. Particularly, with the popularization and application of wearable devices in recent years, noise, especially motion artifacts, are introduced in the acquisition process, and partial electrocardiosignals are distorted, so that the accuracy of the acquired electrocardiosignals is reduced. Therefore, the elimination of these noises in order to analyze whether the signal is abnormal is the first step and also the key step. In order to obtain clean and accurate electrocardiosignals, noise reduction becomes the core of an electrocardiosignal processing task. At the beginning of research, there are many researches on an electrocardiosignal denoising method based on a traditional method, such as empirical mode decomposition, a wavelet technology, a filtering technology, sparse representation and the like. However, the existing electrocardiosignal noise reduction methods have advantages and disadvantages respectively. A common disadvantage is that the generalization capability is not good under different noise backgrounds. In recent years, with the rapid development of deep learning knowledge in the directions of image processing, voice recognition and the like, the application of deep learning knowledge to the field of electrocardiosignal processing makes a major breakthrough, such as a full convolution noise automatic compression encoder, a stack noise automatic encoder, a deep recurrent neural network and the like. It exhibits good generalization capability, especially for various noise conditions, however, it is not perfect. The problem is that although the method has better generalization capability, the noise reduction effect of the designed network structure is superior to the traditional noise reduction effect, but the method still shows low signal-to-noise ratio and high error in the deep learning technology, and electrocardiosignal distortion and important information loss are caused.
Disclosure of Invention
In order to overcome the defects of the technology, the invention provides an ECG processing method which utilizes a time-frequency domain loss function and utilizes a discriminator network in a generation countermeasure network to optimize parameters of a deep convolution network so as to eliminate all ECG noise and avoid ECG distortion as much as possible.
The technical scheme adopted by the invention for overcoming the technical problems is as follows:
an ECG noise reduction method based on a antagonism depth full convolution network comprises the following steps:
a) selecting noises numbered BW, MA and EM in an MIT-BIH noise pressure test database, intercepting, selecting a noise signal of a channel I in intercepted BW noise and cutting the noise signal into N parts and M parts, wherein the lengths of the cut noise signals of the N parts and the M parts are both L, selecting a noise signal of the channel I in intercepted MA noise and cutting the noise signal into the N parts and the M parts, wherein the lengths of the cut noise signals of the N parts and the M parts are both L, selecting a noise signal of the channel I in intercepted EM noise and cutting the noise signal into the N parts and the M parts, and wherein the lengths of the cut noise signals of the N parts and the M parts are both L;
b) selecting 48 records of an MIT-BIH arrhythmia database, cutting each selected signal into N clean signals with the length of L, selecting 105 records of a QT database, and cutting each selected signal into M clean signals with the length of L;
c) normalizing the noise signals cut in the step a), normalizing the clean signals cut in the step b), respectively injecting N parts of noise signals and M parts of noise signals in the normalized BW noise into the normalized N parts of clean signals and M parts of clean signals, respectively injecting N parts of noise signals and M parts of noise signals in the normalized MA noise into the normalized N parts of clean signals and M parts of clean signals, respectively injecting the N parts of noise signals and M parts of noise signals in the normalized EM noise into the normalized N parts of clean signals and M parts of clean signals, and obtaining signals needing noise reduction;
d) dividing noise signals cut in BW, MA and EM into a training set, a verification set and a test set respectively, dividing clean signals cut in the MIT-BIH arrhythmia database into the training set, the verification set and the test set, and dividing clean signals cut in the QT database into the training set, the verification set and the test set;
e) establishing a noise reduction network model in a neural network model by the formula x ═ v + n, wherein
Figure BDA0003302972970000021
j-1 is represented as a noise signal in BW noise, j-2 is represented as a noise signal in MA noise, j-3 is represented as a noise signal in EM noise,
Figure BDA0003302972970000022
to add the ith noisy signal of type j,
Figure BDA0003302972970000023
vifor the (i) th clean signal to be,
Figure BDA0003302972970000024
Figure BDA0003302972970000025
in order to add the ith noise signal with the noise type j, distinguishing the original signal and the noise-reduced signal by utilizing a classification task in a discriminator network in a neural network model;
f) training a neural network model by taking a noisy signal x as input and a clean signal v as a target and obtaining a formula
Figure BDA0003302972970000031
Calculating a total parameter theta of the neural network model, wherein
Figure BDA0003302972970000032
For the ith signal subjected to noise reduction and obtained by the noise reduction network model, arg is an optimal parameter when the loss function takes the minimum value, lambda is a balance term, and lambda takes the value of 10-3，
Figure BDA0003302972970000033
Is composed of
Figure BDA0003302972970000034
Fast Fourier transform of (V)iIs v isiThe fast fourier transform of (a) the fast fourier transform,
Figure BDA0003302972970000035
by the formula
Figure BDA0003302972970000036
Computing a minimization loss function l in a discriminator networkDIn the formula
Figure BDA0003302972970000037
For the discriminator network to determine the probability that the signal will come from a clean signal rather than a noise-reduced signal, D (v)i) A probability of identifying a clean signal for the discriminator network;
g) and d) carrying out noise reduction treatment on the noise signals in the test set in the step d) through the neural network model trained in the step f).
Preferably, in step a), N is 600, M is 300, and L is 512.
Preferably, in step b), N is 600, M is 300, and L is 512.
Preferably, the noise signals cut in BW, MA and EM are divided into a training set, a validation set and a test set in step d) by setting the signal-to-noise ratios to 0dB, 1.25dB and 5dB, respectively.
Further, in the step d), the number of samples of the training set is selected to be 23040, the number of samples of the verification set is selected to be 2880, the number of samples of the test set is selected to be 2880, the number of samples of the training set is selected to be 25200, the number of samples of the verification set is selected to be 3150, and the number of samples of the test set is selected to be 3150 in the clean signal cut by the MIT-BIH arrhythmia database.
Further, in the 1 st convolutional layer of the noise reduction network model in the step e), the input signal is changed into a signal with a channel of 56 for 1 channel, then the signal is input into the 2 nd convolutional block to be convolved to obtain a feature map T1, the feature map T1 is transmitted to the 3 rd downsampling layer, the feature map T1 is convolved by using a method of striding convolution instead of downsampling, the result after the convolution operation is used as the 4 th convolutional block to be convolved by the convolutional block to obtain a feature map T2, and the output feature map is outputInputting T2 into the 5 th down-sampling layer, performing convolution and dimensionality reduction by using a down-sampling method, inputting the data after the convolution and dimensionality reduction into the 6 th convolution block, performing convolution operation by using a convolution block to obtain a feature map T3, inputting the feature map T3 into the 7 th down-sampling layer, performing convolution and dimensionality reduction by using a down-sampling method, inputting the data after the convolution into the 8 th convolution block, performing convolution operation by using a convolution block to obtain a feature map T4, inputting the feature map T4 into the 9 th down-sampling layer, performing convolution operation by using a down-sampling method, inputting the result after the convolution operation into the 10 th convolution block, performing convolution operation by using a convolution block to obtain a feature map T5, inputting the feature map T5 into the 11 th up-sampling layer, performing deconvolution processing by using stride convolution instead of up-sampling to obtain a feature map T6, inputting the feature map T6 and the feature map T4 into the 12 th convolution block, performing convolution operation by using a convolution block, inputting a convolution operation result into a 13 th upper sampling layer, performing deconvolution processing by using an upper sampling method to obtain a characteristic diagram T7, inputting the characteristic diagram T7 and the characteristic diagram T3 into a 14 th convolution block, performing convolution operation by using the convolution block, inputting a convolution operation result into a 15 th upper sampling layer, performing deconvolution processing by using the upper sampling method to obtain a characteristic diagram T8, inputting the characteristic diagram T8 and the characteristic diagram T2 into a 16 th convolution block, performing convolution operation by using the convolution block, inputting a convolution operation result into a 17 th upper sampling layer, performing deconvolution processing by using the upper sampling method to obtain a residual signal characteristic diagram T9, inputting the characteristic diagram T9 into an 18 th convolution layer, performing convolution operation by using residual connection with the 1 st convolution layer to obtain a noise signal subjected to a noise reduction network model
Figure BDA0003302972970000041
By the formula
Figure BDA0003302972970000042
Calculating noise reduced signal
Figure BDA0003302972970000043
Further, the discriminator network in the step e) adopts two full convolution layers and three full connection layers, the size of the convolution kernel in the first full convolution layer is 1 multiplied by 1, and the step length is 1The padding is (0,0) to change the signal with the channel number of 1 into the signal with the channel number of 56, the size of a convolution kernel in the second full convolution layer is 1 multiplied by 1, the step length is 1, the padding is (0,0) to change the signal with the channel number of 56 into the signal with the channel number of 1, each full convolution layer is provided with a Batch Norm layer and a Relu activation layer output, the output of the second full convolution layer is used as the input of three full connection layers, the first two full connection layers output 150 units and Relu activation, and the last full connection layer output is 1 unit and uses a Sigmoid activation function to execute the binary task.
Further comprising selecting the MIT-BIH arrhythmia database with the number: 103. 105, 116, 213, 219, 230 with the QT database numbered: sel123, sel302, sel820, sel16420, sel 0106, sel32, sel14046 signals are expressed by the formula
Figure BDA0003302972970000044
Calculating cosine similarity between signals
Figure BDA0003302972970000045
In the formula
Figure BDA0003302972970000046
Is a noise-reduced signal.
Furthermore, the 2 nd, 4 th, 6 th, 8 th, 10 th, 12 th, 14 th and 16 th convolution blocks all use the channel number of 56, the convolution kernel size is 3 × 1, the step size is 1, the expansion rate scaled is 3, the padding is (0,3), each convolution block comprises four convolution layers, a Batch Norm layer is configured after the four convolution layers to normalize the output signal data, and the normalized signal data is output by using a linear activation function Relu; the 3 rd down-sampling layer, the 5 th down-sampling layer, the 7 th down-sampling layer and the 9 th down-sampling layer use convolution kernel with the size of 2 multiplied by 1, the step length is 2, padding is (0,0), each down-sampling layer is provided with a Batch Norm layer to normalize the output signal data, and the normalized output signal data is output by using a linear activation function Relu; the convolution kernel size of the 11 th upsampling layer, the 13 th upsampling layer, the 15 th upsampling layer and the 17 th upsampling layer is 2 multiplied by 1, the step length is 2, padding is (0,0), each downsampling layer is provided with a Batch Norm layer to normalize output signal data, and the normalized output signal data are output by using a linear activation function Relu; the 1 st convolutional layer and the 18 th convolutional layer change the signal dimension from 1 to 56 and 56 to 1 using convolutional kernel size of 1 × 1, step size of 2, padding of (0, 0).
The invention has the beneficial effects that: by the antagonism electrocardiosignal noise reduction method based on the deep convolution network, the convolution block is adopted in the model to further retain details, the expansion convolution is adopted to reduce the complexity of calculation, and batch normalization is adopted in each layer to obtain better gradient flow and fast convergence. And finally, the discriminator is utilized to better learn optimization. The design can embody a noise reduction signal with low complexity and high precision while keeping details.
Drawings
FIG. 1 is a diagram of the neural network architecture of the present invention.
Detailed Description
The invention is further described below with reference to fig. 1.
An ECG noise reduction method based on a antagonism depth full convolution network comprises the following steps:
a) selecting noises numbered BW, MA and EM in an MIT-BIH noise pressure test database, intercepting, selecting a noise signal of a channel I in intercepted BW noise and cutting the noise signal into N parts and M parts, wherein the lengths of the cut noise signals of the N parts and the M parts are both L, selecting a noise signal of the channel I in intercepted MA noise and cutting the noise signal into the N parts and the M parts, wherein the lengths of the cut noise signals of the N parts and the M parts are both L, selecting a noise signal of the channel I in intercepted EM noise and cutting the noise signal into the N parts and the M parts, and wherein the lengths of the cut noise signals of the N parts and the M parts are both L.
b) Selecting 48 records in the MIT-BIH arrhythmia database, cutting each selected signal into N clean signals with the length of L, selecting 105 records in the QT database, and cutting each selected signal into M clean signals with the length of L.
c) Normalizing the noise signal cut in the step a), normalizing the clean signal cut in the step b), respectively injecting N parts of noise signals and M parts of noise signals in the normalized BW noise into the normalized N parts of clean signals and M parts of clean signals, respectively injecting N parts of noise signals and M parts of noise signals in the normalized MA noise into the normalized N parts of clean signals and M parts of clean signals, respectively injecting the N parts of noise signals and M parts of noise signals in the normalized EM noise into the normalized N parts of clean signals and M parts of clean signals, and obtaining the signal needing noise reduction.
d) The method comprises the steps of dividing noise signals cut in BW, MA and EM into a training set, a verification set and a test set respectively, dividing clean signals cut in an MIT-BIH arrhythmia database into the training set, the verification set and the test set, and dividing clean signals cut in a QT database into the training set, the verification set and the test set.
e) Establishing a noise reduction network model in a neural network model by the formula x ═ v + n, wherein
Figure BDA0003302972970000061
j-1 is represented as a noise signal in BW noise, j-2 is represented as a noise signal in MA noise, j-3 is represented as a noise signal in EM noise,
Figure BDA0003302972970000062
to add the ith noisy signal of type j,
Figure BDA0003302972970000063
vifor the (i) th clean signal to be,
Figure BDA0003302972970000064
Figure BDA0003302972970000065
in order to add the ith noise signal with the noise type j, the original signal and the noise-reduced signal are identified by utilizing a classification task in an identifier network in a neural network model.
f) With noisy signal x as input, clean signalv is a target training neural network model, and the neural network model is obtained through a formula
Figure BDA0003302972970000066
Calculating a total parameter theta of the neural network model, wherein
Figure BDA0003302972970000067
For the ith signal subjected to noise reduction and obtained by the noise reduction network model, arg is an optimal parameter when the loss function takes the minimum value, lambda is a balance term, and lambda takes the value of 10-3，
Figure BDA0003302972970000068
Is composed of
Figure BDA0003302972970000069
Fast Fourier transform of (V)iIs v isiThe fast fourier transform of (a) the fast fourier transform,
Figure BDA00033029729700000610
by the formula
Figure BDA0003302972970000071
Computing a minimization loss function l in a discriminator networkDIn the formula
Figure BDA0003302972970000072
For the discriminator network to determine the probability that the signal will come from a clean signal rather than a noise-reduced signal, D (v)i) Probability of identifying a clean signal for the discriminator network.
g) And d) carrying out noise reduction treatment on the noise signals in the test set in the step d) through the neural network model trained in the step f).
By the antagonism electrocardiosignal noise reduction method based on the deep convolution network, the convolution block is adopted in the model to further retain details, the expansion convolution is adopted to reduce the complexity of calculation, and batch normalization is adopted in each layer to obtain better gradient flow and fast convergence. And finally, the discriminator is utilized to better learn optimization. The design can embody a noise reduction signal with low complexity and high precision while keeping details.
Example 1:
preferably, in step a), N is 600, M is 300, and L is 512. In step b), the value of N is 600, the value of M is 300, and the value of L is 512. In the step d), noise signals cut in BW, MA and EM are divided into a training set, a verification set and a test set respectively by setting signal-to-noise ratios to be 0dB, 1.25dB and 5 dB.
Example 2:
in the step d), selecting 23040 samples of the training set from the clean signals cut by the MIT-BIH arrhythmia database, 2880 samples of the verification set, 2880 samples of the test set, 25200 samples of the training set from the clean signals cut by the QT database, 3150 samples of the verification set and 3150 samples of the test set.
Example 3:
the 1 st convolution layer of the noise reduction network model in the step e) converts an input signal into a signal with a channel passage of 56 and inputs the signal into a 2 nd convolution block to carry out convolution operation to obtain a characteristic diagram T1, the characteristic diagram T1 is conveyed to a 3 rd down-sampling layer, the characteristic diagram T1 is subjected to convolution operation by using a stride convolution instead of a down-sampling method, the result after the convolution operation is used as a 4 th convolution block to carry out convolution operation by using the convolution block to obtain a characteristic diagram T2, the output characteristic diagram T2 is input into a 5 th down-sampling layer to carry out convolution dimensionality reduction by using a down-sampling method, the data after the convolution dimensionality reduction processing is input into a 6 th convolution block to obtain a characteristic diagram T3 by using the convolution block, the characteristic diagram T3 is input into a 7 th down-sampling layer to carry out convolution dimensionality reduction by using the down-sampling method, the data after the convolution is input into an 8 th convolution block to obtain a characteristic diagram T4, inputting the feature map T4 into the 9 th downsampling layer, performing convolution operation by adopting a downsampling method, inputting the result of the convolution operation into the 10 th convolution block, performing convolution operation by adopting the convolution block to obtain a feature map T5, inputting the feature map T5 into the 11 th upsampling layer, performing deconvolution processing by adopting a method of striding convolution instead of upsampling to obtain a feature map T6, inputting the feature map T6 and the feature map T4 into the 12 th convolution block together, and performing convolution operation by adopting a method of striding convolution and replacing upsamplingPerforming convolution operation by using a convolution block, inputting a convolution operation result into a 13 th upper sampling layer, performing deconvolution processing by using an upper sampling method to obtain a characteristic diagram T7, inputting the characteristic diagram T7 and the characteristic diagram T3 into a 14 th convolution block, performing convolution operation by using the convolution block, inputting a convolution operation result into a 15 th upper sampling layer, performing deconvolution processing by using the upper sampling method to obtain a characteristic diagram T8, inputting the characteristic diagram T8 and the characteristic diagram T2 into a 16 th convolution block, performing convolution operation by using the convolution block, inputting a convolution operation result into a 17 th upper sampling layer, performing deconvolution processing by using the upper sampling method to obtain a residual signal characteristic diagram T9, inputting the characteristic diagram T9 into an 18 th convolution layer, performing convolution operation by using residual connection with the 1 st convolution layer to obtain a noise signal subjected to a noise reduction network model
Figure BDA0003302972970000081
By the formula
Figure BDA0003302972970000082
Calculating noise reduced signal
Figure BDA0003302972970000083
Example 4:
in the step e), the discriminator network adopts two full convolution layers and three full connection layers, the size of a convolution kernel in the first full convolution layer is 1 multiplied by 1, the step length is 1, padding is (0,0) to change a signal with the channel number of 1 into a signal with the channel number of 56, the size of a convolution kernel in the second full convolution layer is 1 multiplied by 1, the step length is 1, padding is (0,0) to change a signal with the channel number of 56 into a signal with the channel number of 1, each layer of the full convolution layers is provided with a Batch m layer and a Relu active layer output, the output of the second full convolution layer is used as the input of the three full connection layers, the first two full connection layers output 150 units and Norlu activation, and the last full connection layer output 1 unit and uses a Sigmoid activation function to execute the binary task.
Example 5:
further comprising selecting the MIT-BIH arrhythmia database with the number: 103. 105, 116, 213, 219. 230 with the QT database numbered: sel123, sel302, sel820, sel16420, sel 0106, sel32, sel14046 signals are expressed by the formula
Figure BDA0003302972970000084
Calculating cosine similarity between signals
Figure BDA0003302972970000091
In the formula
Figure BDA0003302972970000092
Is a noise-reduced signal.
Example 6:
the 2 nd, 4 th, 6 th, 8 th, 10 th, 12 th, 14 th and 16 th convolution blocks all use the channel number of 56, the convolution kernel size is 3 × 1, the step size is 1, the expansion rate is 3, the padding is (0,3), each convolution block comprises four convolution layers, a Batch Norm layer is configured behind the four convolution layers to normalize the output signal data, and the normalized signal data is output by using a linear activation function Relu; the 3 rd down-sampling layer, the 5 th down-sampling layer, the 7 th down-sampling layer and the 9 th down-sampling layer use convolution kernel with the size of 2 multiplied by 1, the step length is 2, padding is (0,0), each down-sampling layer is provided with a Batch Norm layer to normalize the output signal data, and the normalized output signal data is output by using a linear activation function Relu; the convolution kernel size of the 11 th upsampling layer, the 13 th upsampling layer, the 15 th upsampling layer and the 17 th upsampling layer is 2 multiplied by 1, the step length is 2, padding is (0,0), each downsampling layer is provided with a Batch Norm layer to normalize output signal data, and the normalized output signal data are output by using a linear activation function Relu; the 1 st convolutional layer and the 18 th convolutional layer change the signal dimension from 1 to 56 and 56 to 1 using convolutional kernel size of 1 × 1, step size of 2, padding of (0, 0).
Finally, it should be noted that: although the present invention has been described in detail with reference to the foregoing embodiments, it will be apparent to those skilled in the art that changes may be made in the embodiments and/or equivalents thereof without departing from the spirit and scope of the invention. Any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims (9)

1. An ECG noise reduction method based on a antagonism depth full convolution network is characterized by comprising the following steps:
a) selecting noises numbered BW, MA and EM in an MIT-BIH noise pressure test database, intercepting, selecting a noise signal of a channel I in intercepted BW noise and cutting the noise signal into N parts and M parts, wherein the lengths of the cut noise signals of the N parts and the M parts are both L, selecting a noise signal of the channel I in intercepted MA noise and cutting the noise signal into the N parts and the M parts, wherein the lengths of the cut noise signals of the N parts and the M parts are both L, selecting a noise signal of the channel I in intercepted EM noise and cutting the noise signal into the N parts and the M parts, and wherein the lengths of the cut noise signals of the N parts and the M parts are both L;
b) selecting 48 records of an MIT-BIH arrhythmia database, cutting each selected signal into N clean signals with the length of L, selecting 105 records of a QT database, and cutting each selected signal into M clean signals with the length of L;
c) normalizing the noise signals cut in the step a), normalizing the clean signals cut in the step b), respectively injecting N parts of noise signals and M parts of noise signals in the normalized BW noise into the normalized N parts of clean signals and M parts of clean signals, respectively injecting N parts of noise signals and M parts of noise signals in the normalized MA noise into the normalized N parts of clean signals and M parts of clean signals, respectively injecting the N parts of noise signals and M parts of noise signals in the normalized EM noise into the normalized N parts of clean signals and M parts of clean signals, and obtaining signals needing noise reduction;
d) dividing noise signals cut in BW, MA and EM into a training set, a verification set and a test set respectively, dividing clean signals cut in the MIT-BIH arrhythmia database into the training set, the verification set and the test set, and dividing clean signals cut in the QT database into the training set, the verification set and the test set;
e) establishing a noise reduction network model in a neural network model by the formula x ═ v + n, wherein
Figure FDA0003302972960000011
j-1 is represented as a noise signal in BW noise, j-2 is represented as a noise signal in MA noise, j-3 is represented as a noise signal in EM noise,
Figure FDA0003302972960000012
to add the ith noisy signal of type j,
Figure FDA0003302972960000013
vifor the (i) th clean signal to be,
Figure FDA0003302972960000014
Figure FDA0003302972960000015
in order to add the ith noise signal with the noise type j, distinguishing the original signal and the noise-reduced signal by utilizing a classification task in a discriminator network in a neural network model;
f) training a neural network model by taking a noisy signal x as input and a clean signal v as a target and obtaining a formula
Figure FDA0003302972960000021
Calculating a total parameter theta of the neural network model, wherein
Figure FDA0003302972960000022
For the ith signal subjected to noise reduction and obtained by the noise reduction network model, arg is an optimal parameter when the loss function takes the minimum value, lambda is a balance term, and lambda takes the value of 10-3，
Figure FDA0003302972960000023
Is composed of
Figure FDA0003302972960000024
Fast Fourier transform of (V)iIs v isiThe fast fourier transform of (a) the fast fourier transform,
Figure FDA0003302972960000025
by the formula
Figure FDA0003302972960000026
Computing a minimization loss function l in a discriminator networkDIn the formula
Figure FDA0003302972960000027
For the discriminator network to determine the probability that the signal will come from a clean signal rather than a noise-reduced signal, D (v)i) A probability of identifying a clean signal for the discriminator network;
g) and d) carrying out noise reduction treatment on the noise signals in the test set in the step d) through the neural network model trained in the step f).
2. The ECG noise reduction method based on the antagonistic deep full convolution network of claim 1, characterized in that: in step a), the value of N is 600, the value of M is 300, and the value of L is 512.
3. The ECG noise reduction method based on the antagonistic deep full convolution network of claim 1, characterized in that: in step b), the value of N is 600, the value of M is 300, and the value of L is 512.
4. The ECG noise reduction method based on the antagonistic deep full convolution network of claim 1, characterized in that: in the step d), noise signals cut in BW, MA and EM are divided into a training set, a verification set and a test set respectively by setting signal-to-noise ratios to be 0dB, 1.25dB and 5 dB.
5. The ECG noise reduction method based on the antagonistic deep full convolution network of claim 1, characterized in that: in the step d), selecting 23040 samples of the training set from the clean signals cut by the MIT-BIH arrhythmia database, 2880 samples of the verification set, 2880 samples of the test set, 25200 samples of the training set from the clean signals cut by the QT database, 3150 samples of the verification set and 3150 samples of the test set.
6. The ECG noise reduction method based on the antagonistic deep full convolution network of claim 1, characterized in that: the 1 st convolution layer of the noise reduction network model in the step e) converts an input signal into a signal with a channel passage of 56 and inputs the signal into a 2 nd convolution block to carry out convolution operation to obtain a characteristic diagram T1, the characteristic diagram T1 is conveyed to a 3 rd down-sampling layer, the characteristic diagram T1 is subjected to convolution operation by using a stride convolution instead of a down-sampling method, the result after the convolution operation is used as a 4 th convolution block to carry out convolution operation by using the convolution block to obtain a characteristic diagram T2, the output characteristic diagram T2 is input into a 5 th down-sampling layer to carry out convolution dimensionality reduction by using a down-sampling method, the data after the convolution dimensionality reduction processing is input into a 6 th convolution block to obtain a characteristic diagram T3 by using the convolution block, the characteristic diagram T3 is input into a 7 th down-sampling layer to carry out convolution dimensionality reduction by using the down-sampling method, the data after the convolution is input into an 8 th convolution block to obtain a characteristic diagram T4, inputting a feature map T4 into a 9 th downsampling layer, performing convolution operation by adopting a downsampling method, inputting the result after the convolution operation into a 10 th convolution block, performing convolution operation by adopting a convolution block to obtain a feature map T5, inputting a feature map T5 into an 11 th upsampling layer, performing deconvolution processing by adopting a step-by-step convolution instead of upsampling to obtain a feature map T6, inputting the feature map T6 and the feature map T4 into a 12 th convolution block, performing convolution operation by adopting the convolution block, inputting the result of the convolution operation into a 13 th upsampling layer, performing deconvolution processing by adopting an up sampling method to obtain a feature map T7, inputting the feature map T7 and the feature map T3 into a 14 th convolution block, performing convolution operation by adopting the convolution block, inputting the result of the convolution operation into a 15 th upsampling layer, and performing deconvolution processing by adopting an up sampling methodProcessing to obtain a characteristic diagram T8, inputting the characteristic diagram T8 and the characteristic diagram T2 into a 16 th convolution block, performing convolution operation by adopting the convolution block, inputting the convolution operation result into a 17 th upsampling layer, performing deconvolution processing by adopting an upper sampling method to obtain a residual signal characteristic diagram T9, inputting the characteristic diagram T9 into an 18 th convolution layer, performing convolution operation by connecting the characteristic diagram T9 with a 1 st convolution layer through residual errors to obtain a noise signal subjected to a noise reduction network model
Figure FDA0003302972960000031
By the formula
Figure FDA0003302972960000032
Calculating noise reduced signal
Figure FDA0003302972960000033
7. The ECG noise reduction method based on the antagonistic deep full convolution network of claim 1, characterized in that: in the step e), the discriminator network adopts two full convolution layers and three full connection layers, the size of a convolution kernel in the first full convolution layer is 1 multiplied by 1, the step length is 1, padding is (0,0) to change a signal with the channel number of 1 into a signal with the channel number of 56, the size of a convolution kernel in the second full convolution layer is 1 multiplied by 1, the step length is 1, padding is (0,0) to change a signal with the channel number of 56 into a signal with the channel number of 1, each layer of the full convolution layers is provided with a Batch m layer and a Relu active layer output, the output of the second full convolution layer is used as the input of the three full connection layers, the first two full connection layers output 150 units and Norlu activation, and the last full connection layer output 1 unit and uses a Sigmoid activation function to execute the binary task.
8. The ECG noise reduction method based on the antagonistic deep full convolution network of claim 1, characterized in that: further comprising selecting the MIT-BIH arrhythmia database with the number: 103. 105, 116, 213, 219, 230 with the QT database numbered: sel123, sel302, sel820, sel16420,sel 0106, sel32, sel14046 signals by formula
Figure FDA0003302972960000041
Calculating cosine similarity between signals
Figure FDA0003302972960000042
In the formula
Figure FDA0003302972960000043
Is a noise-reduced signal.
9. The method of ECG noise reduction based on a antagonistic deep full convolution network of claim 6, characterized in that: the 2 nd, 4 th, 6 th, 8 th, 10 th, 12 th, 14 th and 16 th convolution blocks all use the channel number of 56, the convolution kernel size is 3 × 1, the step size is 1, the expansion rate is 3, the padding is (0,3), each convolution block comprises four convolution layers, a Batch Norm layer is configured behind the four convolution layers to normalize the output signal data, and the normalized signal data is output by using a linear activation function Relu; the 3 rd down-sampling layer, the 5 th down-sampling layer, the 7 th down-sampling layer and the 9 th down-sampling layer use convolution kernel with the size of 2 multiplied by 1, the step length is 2, padding is (0,0), each down-sampling layer is provided with a Batch Norm layer to normalize the output signal data, and the normalized output signal data is output by using a linear activation function Relu; the convolution kernel size of the 11 th upsampling layer, the 13 th upsampling layer, the 15 th upsampling layer and the 17 th upsampling layer is 2 multiplied by 1, the step length is 2, padding is (0,0), each downsampling layer is provided with a Batch Norm layer to normalize output signal data, and the normalized output signal data are output by using a linear activation function Relu; the 1 st convolutional layer and the 18 th convolutional layer change the signal dimension from 1 to 56 and 56 to 1 using convolutional kernel size of 1 × 1, step size of 2, padding of (0, 0).
CN202111196251.8A 2021-10-14 2021-10-14 ECG noise reduction method based on antagonistic depth full convolution network Pending CN113935378A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111196251.8A CN113935378A (en) 2021-10-14 2021-10-14 ECG noise reduction method based on antagonistic depth full convolution network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111196251.8A CN113935378A (en) 2021-10-14 2021-10-14 ECG noise reduction method based on antagonistic depth full convolution network

Publications (1)

Publication Number Publication Date
CN113935378A true CN113935378A (en) 2022-01-14

Family

ID=79279269

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111196251.8A Pending CN113935378A (en) 2021-10-14 2021-10-14 ECG noise reduction method based on antagonistic depth full convolution network

Country Status (1)

Country Link
CN (1) CN113935378A (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114781446A (en) * 2022-04-11 2022-07-22 山东省人工智能研究院 Electrocardiosignal noise reduction method based on HIN network and gradient difference loss
CN115470827A (en) * 2022-09-23 2022-12-13 山东省人工智能研究院 Antagonistic electrocardiosignal noise reduction method based on self-supervision learning and twin network

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
SE9302436D0 (en) * 1993-07-16 1993-07-16 Siemens-Elema Ab PROCEDURE AND DEVICE TO INCREASE SIGNAL NOISE CONTROLS WITH ECG SIGNALS
CN111046824A (en) * 2019-12-19 2020-04-21 上海交通大学 Time series signal efficient denoising and high-precision reconstruction modeling method and system
CN111680785A (en) * 2020-05-29 2020-09-18 山东省人工智能研究院 ECG signal processing method based on sparse characteristic and antagonistic neural network combination
CN111814656A (en) * 2020-07-02 2020-10-23 山东省人工智能研究院 Electrocardiosignal noise reduction method based on countermeasure generation network
CN112487914A (en) * 2020-11-25 2021-03-12 山东省人工智能研究院 ECG noise reduction method based on deep convolution generation countermeasure network
US20220070150A1 (en) * 2020-09-02 2022-03-03 Sap Se Privacy-enhanced data stream collection

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
SE9302436D0 (en) * 1993-07-16 1993-07-16 Siemens-Elema Ab PROCEDURE AND DEVICE TO INCREASE SIGNAL NOISE CONTROLS WITH ECG SIGNALS
CN111046824A (en) * 2019-12-19 2020-04-21 上海交通大学 Time series signal efficient denoising and high-precision reconstruction modeling method and system
CN111680785A (en) * 2020-05-29 2020-09-18 山东省人工智能研究院 ECG signal processing method based on sparse characteristic and antagonistic neural network combination
CN111814656A (en) * 2020-07-02 2020-10-23 山东省人工智能研究院 Electrocardiosignal noise reduction method based on countermeasure generation network
US20220070150A1 (en) * 2020-09-02 2022-03-03 Sap Se Privacy-enhanced data stream collection
CN112487914A (en) * 2020-11-25 2021-03-12 山东省人工智能研究院 ECG noise reduction method based on deep convolution generation countermeasure network

Non-Patent Citations (4)

* Cited by examiner, † Cited by third party
Title
CORNELIU T ARSENE: "Complex Deep Learning ModelHuman Heart ECG signalss for Denoising of", 《ARXIV》 *
MEENU RANI 等: "ECG Signal Analysis Using S-Domain", 《INTERNATIONAL JOURNAL OF RESEARCH IN ELECTRONICS AND COMMUNICATION TECHNOLOGY (IJRECT 2015)》 *
孙瀚: "基于多模态生物电信号人机交互技术研究", 《中国优秀博士学位论文全文数据库 信息科技辑》 *
陈健 等: "基于卷积自编码神经网络的心电信号降噪", 《计算机工程与应用》 *

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114781446A (en) * 2022-04-11 2022-07-22 山东省人工智能研究院 Electrocardiosignal noise reduction method based on HIN network and gradient difference loss
CN114781446B (en) * 2022-04-11 2022-12-09 山东省人工智能研究院 Electrocardiosignal noise reduction method based on HIN network and gradient difference loss
CN115470827A (en) * 2022-09-23 2022-12-13 山东省人工智能研究院 Antagonistic electrocardiosignal noise reduction method based on self-supervision learning and twin network

Similar Documents

Publication Publication Date Title
Zhou et al. Resnext and res2net structures for speaker verification
CN109841226B (en) Single-channel real-time noise reduction method based on convolution recurrent neural network
CN113935378A (en) ECG noise reduction method based on antagonistic depth full convolution network
CN111291614B (en) Child epileptic syndrome classification method based on migration learning multimode decision fusion
CN111933124B (en) Keyword detection method capable of supporting self-defined awakening words
CN110751636B (en) Fundus image retinal arteriosclerosis detection method based on improved coding and decoding network
CN112735460B (en) Beam forming method and system based on time-frequency masking value estimation
CN110349593A (en) The method and system of semanteme based on waveform Time-Frequency Analysis and the dual identification of vocal print
CN113191178B (en) Underwater sound target identification method based on auditory perception feature deep learning
CN111598894B (en) Retina blood vessel image segmentation system based on global information convolution neural network
CN111524530A (en) Voice noise reduction method based on expansion causal convolution
CN114283829B (en) Voice enhancement method based on dynamic gating convolution circulation network
CN113723171A (en) Electroencephalogram signal denoising method based on residual error generation countermeasure network
CN114129171B (en) Electrocardiosignal noise reduction method based on improved residual error dense network
CN114492522B (en) Automatic modulation classification method based on improved stacked hourglass neural network
CN114595728A (en) Signal denoising method based on self-supervision learning
CN117174105A (en) Speech noise reduction and dereverberation method based on improved deep convolutional network
CN116405100B (en) Distortion signal restoration method based on priori knowledge
Li et al. Deeplabv3+ vision transformer for visual bird sound denoising
CN110811591A (en) Heart failure grading method based on heart rate variability
CN113940638B (en) Pulse wave signal identification and classification method based on frequency domain dual-feature fusion
AU2021104217A4 (en) A system and method for optimum wavelet basis function selection for ecg arrhythmia denoising using artificial intelligence
CN114781446B (en) Electrocardiosignal noise reduction method based on HIN network and gradient difference loss
CN113707172A (en) Single-channel voice separation method, system and computer equipment of sparse orthogonal network
CN116631429B (en) Voice and video processing method and system based on VOLTE call

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
RJ01 Rejection of invention patent application after publication

Application publication date: 20220114

RJ01 Rejection of invention patent application after publication