Adaptive denoising method for electronic stethoscope based on time-domain fully convolutional deep neural network

Abstract

The present invention relates to signal noise reduction, and in particular to an adaptive noise reduction method for an electronic stethoscope based on a time-domain full convolutional deep neural network. The electronic stethoscope collects a cardiopulmonary sound signal with background sound interference through a main acquisition channel, and simultaneously collects the background sound signal through a remote acquisition channel, and inputs the signal into an original signal buffer space. The main acquisition channel and the remote acquisition channel in the original signal buffer space are respectively normalized to obtain a normalized cardiopulmonary sound signal s with background sound interference and a background sound signal x. The cardiopulmonary sound signal s with background sound interference and the background sound signal x are input into an ANC system including an adaptive filter and an adaptive algorithm to obtain an adaptive filter output y and a system error e. The adaptive filter output y and the system error e are input into a trained full convolutional deep neural network. The technical solution provided by the present invention can effectively overcome the defect that the environmental noise interference cannot be effectively eliminated in the process of collecting cardiopulmonary sound signals.

Description

Technical Field

[0001] The present invention relates to signal noise reduction, and in particular to an adaptive noise reduction method for an electronic stethoscope based on a time-domain fully convolutional deep neural network. Background Art

[0002] As a non-invasive auxiliary diagnostic instrument, the electronic stethoscope can easily collect human heart and lung sound signals, including heart sounds and lung sounds. Doctors determine the patient's condition by analyzing the information contained in these heart and lung sounds. Although electronic stethoscopes are widely used, they still have many problems, such as the gap between the body and the capacitive earphones, the different positions of the stethoscope capacitive earphones, and the susceptibility to environmental noise, which will seriously affect the auscultation judgment results.

[0003] Signals such as heart sounds and lung sounds are extremely weak. To ensure that clear signals can be heard, the amplification factor of electronic stethoscopes is often set very high, which makes the auscultation signal extremely susceptible to environmental noise. However, environmental noise is unknown and difficult to model, which makes it difficult for some traditional algorithms to achieve optimal results in the field of heart and lung sound noise reduction. To address this problem, the prior art proposes an adaptive noise cancellation (ANC) method that can restore the target signal to a certain extent. It usually uses a remote acquisition channel to record background sound, the mixed sound data collected by the main channel and the background sound interference data collected by the remote acquisition channel, and trains a set of finite impulse response (FIR) filters through an adaptive algorithm to eliminate background sound interference.

[0004] The invention patent CN105496447B authorized and announced on February 5, 2019 discloses an electronic stethoscope with active noise reduction and auxiliary diagnosis functions, which uses a dual-channel ANC algorithm to improve the auscultation effect. However, the traditional ANC technology processing has the problem of over-elimination, which makes the signal changes caused by the pathological factors that doctors are concerned about disappear, and when the external background sound changes, the traditional ANC technology needs to go through a long iteration process to converge, which seriously affects the real-time performance of the noise reduction processing.

[0005] In addition, dual-channel spectral subtraction is also a commonly used cardiopulmonary sound noise reduction algorithm. The dual-channel multi-band spectral subtraction method proposed by Emmanouilidou et al. (Adaptive Noise Suppression of Pediatric Lung Auscultations With Real Applications to Noisy Clinical Settings in Developing Countries. IEEE Transactions on Biomedical Engineering, vol. 62, no. 9, 2015, pp. 2279-2288) in 2015 eliminates environmental noise in the process of lung sound collection. It can dynamically eliminate background noise interference while ensuring the integrity of the lung sound signal. However, it can only eliminate the interference of additive noise and has poor processing effect on convolution or nonlinear noise. Summary of the invention

[0006] 1. Technical issues to be solved

[0007] In view of the above-mentioned shortcomings of the prior art, the present invention provides an adaptive noise reduction method for an electronic stethoscope based on a time-domain fully convolutional deep neural network, which can effectively overcome the defect of the prior art that it is unable to effectively eliminate environmental noise interference during the process of collecting cardiopulmonary sound signals.

[0008] (II) Technical solution

[0009] To achieve the above objectives, the present invention is implemented through the following technical solutions:

[0010] The electronic stethoscope adaptive noise reduction method based on time domain full convolution deep neural network includes the following steps:

[0011] S1, the electronic stethoscope collects the cardiopulmonary sound signal with background sound interference through the main acquisition channel, and simultaneously collects the background sound signal through the remote acquisition channel, and inputs it into the original signal buffer space;

[0012] S2, normalizing the main acquisition channel and the remote acquisition channel in the original signal buffer space respectively to obtain the normalized cardiopulmonary sound signal s and the background sound signal x with background sound interference;

[0013] S3, inputting the cardiopulmonary sound signal s with background sound interference and the background sound signal x into an ANC system including an adaptive filter and an adaptive algorithm, and obtaining an adaptive filter output y and a system error e;

[0014] S4: Input the adaptive filter output y and the system error e into the trained fully convolutional deep neural network to obtain a clean heart and lung sound signal

[0015] Preferably, after the input into the original signal buffer space, when the buffer space data accumulates to a set time length T, S2 is executed, otherwise S1 is repeatedly executed;

[0016] After normalizing the main acquisition channel and the remote acquisition channel in the original signal buffer space respectively, clearing the buffer space data, and synchronously executing S1;

[0017] The clean heart and lung sound signal is obtained Afterwards, when the cache space data accumulates to the next set time length T, it returns to S2 to continue execution.

[0018] Preferably, the adaptive filter in S3 is an M-order FIR transverse filter, and the specific filtering process is:

[0019] y(k)=w T (k)x(k), where k is the number of iterations, w(k)=[w0(k),w1(k),...,w M-1 (k)] represents the filter tap weight vector, x(k) = [x(k), x(k-1), ..., x(k-M+1)] represents iterating through the background sound signal x to generate the filter tap input vector;

[0020] Subtract each sampling point s(k) in the cardiopulmonary sound signal s with background sound interference from y(k) after iteration to obtain the system error e(k) of each sampling point;

[0021] After the iteration is completed, y(k) is saved to obtain the adaptive filter output y, and the system error e(k) of each sampling point is saved to obtain the system error e.

[0022] Preferably, the adaptive algorithm in S3 is a normalized minimum mean square error NLMS algorithm, which is used to update the filter tap weight vector w(k), and the updating process is:

[0023]

[0024] Among them, μ0 is the convergence factor, and its value range is (0,1], and ζ is a very small positive number used to limit the step size and avoid the denominator being 0.

[0025] Preferably, the fully convolutional deep neural network in S4 includes an encoder, an eliminator and a decoder, wherein the network parameters are obtained by training with a labeled data training set.

[0026] Preferably, the encoder is composed of I one-dimensional convolutional layers, wherein the number of convolution kernels of the first one-dimensional convolutional layer is N, the length of the convolution kernel is K, the convolution step is K / 2, and the remaining I-1 one-dimensional convolutional layers are stacked dilated convolutional layers, and the dilation factor of each dilated convolutional layer in the stacking order is d=1,2,...,2 I-1 , and each dilated convolutional layer output is connected in series with a parameter-corrected linear unit PReLU, the adaptive filter output y and the system error e are respectively input into two parallel encoders to obtain their respective encoding features

[0027] The eliminator includes a temporal convolutional network module for extracting multi-level features, an interaction module at the input end, and a mask generation module at the output end. The eliminator receives the encoded features Given the mask M for the gain feature, the gain encoding feature F is obtained by matrix dot multiplication G , the calculation process is

[0028] The decoder is composed of I one-dimensional deconvolution layers, wherein the first I-1 one-dimensional deconvolution layers are stacked dilated deconvolution layers, and the dilation factor of each dilated deconvolution layer in the stacking order is d=2 I-1 ,2 I-2 ,...,1, and each expansion deconvolution layer output end is connected in series with a parameter correction linear unit PReLU, the number of convolution kernels of the I-th one-dimensional deconvolution layer is N, the length of the convolution kernel is K, the convolution step is K / 2, and the gain encoding feature F G Input to the decoder to get clean heart and lung sound signal

[0029] Preferably, the temporal convolutional network module includes R repeated layers, which receive the output of the interaction module and give a feature output F i TCN , i=1,2,...,R;

[0030] The number of interaction modules is the same as the number of repeated layers. The interaction module first uses the global layer normalization gLN to encode the features Normalized and input into two parallel 1×1 convolutional layers to reduce the feature dimension. The outputs of the two 1×1 convolutional layers are concatenated according to the channel dimension and then subjected to a global layer normalization gLN operation. The subsequent 1×1 convolutional layers, parameterized corrected linear units PReLU and global layer normalization gLN are connected in series to integrate the concatenated features to obtain the local mask M. j , the local mask M j Integrate with the main branch input to form the output of the interactive module The integration process is where j = 0, 1, ..., R-1,

[0031] The mask generation module is composed of a parameterized corrected linear unit PreLU, a 1×1 convolutional layer, a global layer normalization gLN, and a Sigmoid function connected in series. Its input is the feature output F of the R repeated layers in the temporal convolutional network module. i TCN The output is the mask M used to gain features.

[0032] Preferably, the repetitive layer is composed of M stacked one-dimensional convolution blocks, the input end of each one-dimensional convolution block includes two convolution layers, one of which is a 1×1 convolution layer, and the other is an expanded convolution layer, the output end of each convolution layer is sequentially connected in series with a parameter-corrected linear unit PreLU and a global layer normalization gLN, and the output end of each one-dimensional convolution block includes two parallel 1×1 convolution layers;

[0033] Through the two convolutional layers at the input end, two parallel 1×1 convolutional layers at the output end give two outputs of the one-dimensional convolution block, namely the residual output and the skip connection output. The input of the subsequent one-dimensional convolution block is the residual output of the previous one-dimensional convolution block. The skip connection outputs of M stacked one-dimensional convolution blocks are superimposed to generate the feature output F i TCN ;

[0034] Among them, the dilated convolutional layers in the M stacked one-dimensional convolutional blocks have dilation factors of d=1,2,...,2 in the stacking order. M-1 , the input of the first one-dimensional convolutional block in the repeated layer is the output of the interactive module

[0035] Preferably, the training process of the fully convolutional deep neural network in S4 is:

[0036] A large number of pure heart and lung sound signals are mixed with background noise data to generate a data training set. The data training set needs to be input into the ANC system first to obtain the adaptive filter output and system error as the input of the network. The pure heart and lung sound signal is used as the target value, and the network is trained based on the neural network back propagation method to obtain the network parameters.

[0037] The network parameters are updated once the network is trained until the final cost function value stabilizes near a certain value, then the training can be stopped and the current network parameters and network structure information can be saved;

[0038] Among them, the cost function used in the training process of the network is the negative scale-invariant signal-to-noise ratio SI-SNR, which is defined as in s target are the output and target value of the network respectively.

[0039] (III) Beneficial effects

[0040] Compared with the prior art, the adaptive denoising method for electronic stethoscope based on time-domain fully convolutional deep neural network provided by the present invention combines the traditional adaptive filtering method with the deep neural network, realizes the adaptive elimination of background sound interference of the electronic stethoscope, and does not need to use high-latency Fourier transform. The background sound interference can be effectively eliminated only by using the time domain signal, which provides strong technical support for the diagnosis of heart sounds and lung sounds. In addition, the present invention does not need to make any assumptions about the relationship between the source of the heart and lung sound signals and the background noise, and has a very wide range of applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention, and for ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0042] Figure 1 It is a schematic diagram of the process of the present invention;

[0043] Figure 2 It is a schematic diagram of the structure of the time-domain fully convolutional deep neural network in the present invention;

[0044] Figure 3 It is a structural schematic diagram of the interaction module in the present invention;

[0045] Figure 4 Schematic diagram of the structure of a one-dimensional convolution block of a repeated layer in the present invention;

[0046] Figure 5 (a) is a time-frequency diagram of heart sounds with background sound interference input in an embodiment of the present invention;

[0047] Figure 5 (b) is a time-frequency diagram of the corresponding background sound input in an embodiment of the present invention;

[0048] Figure 5 (c) is a heart sound time-frequency diagram outputted in an embodiment of the present invention after background sound interference is eliminated;

[0049] Figure 6 (a) is a time-frequency diagram of lung sounds with background sound interference input in an embodiment of the present invention;

[0050] Figure 6 (b) is a time-frequency diagram of the corresponding background sound input in an embodiment of the present invention;

[0051] Figure 6 (c) is a time-frequency diagram of lung sounds output after eliminating background sound interference in an embodiment of the present invention. DETAILED DESCRIPTION

[0052] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0053] Adaptive noise reduction method for electronic stethoscope based on time domain fully convolutional deep neural network, such as Figure 1 As shown, the following steps are included:

[0054] S1, the electronic stethoscope collects the cardiopulmonary sound signal with background sound interference through the main acquisition channel, and collects the background sound signal through the remote acquisition channel at the same time, and inputs it into the original signal buffer space. When the buffer space data accumulates to the set time length T, execute S2, otherwise repeat S1;

[0055] S2, respectively normalize the main acquisition channel and the remote acquisition channel in the original signal buffer space to obtain the normalized cardiopulmonary sound signal s and the background sound signal x with background sound interference, clear the buffer space data, and synchronously execute S1;

[0056] S3, inputting the cardiopulmonary sound signal s with background sound interference and the background sound signal x into an ANC system including an adaptive filter and an adaptive algorithm, and obtaining an adaptive filter output y and a system error e;

[0057] S4: Input the adaptive filter output y and the system error e into the trained fully convolutional deep neural network to obtain a clean heart and lung sound signal When the cache space data accumulates to the next set time length T, return to S2 to continue execution.

[0058] In the technical solution of the present application, the adaptive filter is an M (M=4) order FIR transverse filter, and the specific filtering process is:

[0059] y(k)=w T (k)x(k), where k is the number of iterations (k = 16000), w(k) = [w0(k), w1(k), ..., w M-1(k)] represents the filter tap weight vector, x(k) = [x(k), x(k-1), ..., x(k-M+1)] represents iterating through the background sound signal x to generate the filter tap input vector;

[0060] Subtract each sampling point s(k) in the cardiopulmonary sound signal s with background sound interference from y(k) after iteration to obtain the system error e(k) of each sampling point;

[0061] After the iteration is completed, y(k) is saved to obtain the adaptive filter output y, and the system error e(k) of each sampling point is saved to obtain the system error e.

[0062] In the technical solution of the present application, the adaptive algorithm is the normalized minimum mean square error NLMS algorithm, which is used to update the filter tap weight vector w(k), and the updating process is:

[0063]

[0064] Among them, μ0 is the convergence factor (μ0=0.0001), the value range is (0,1], ζ is a very small positive number (ζ=0.0001), which is used to limit the step size to be too large and avoid the denominator being 0.

[0065] In the technical solution of this application, Figure 2 As shown, the fully convolutional deep neural network includes an encoder, an eliminator and a decoder, wherein the network parameters are trained by a labeled data training set.

[0066] ① The encoder consists of I (I = 5) one-dimensional convolutional layers, where the number of convolution kernels in the first one-dimensional convolutional layer is N (N = 256), the length of the convolution kernel is K (K = 16), the convolution step is K / 2, and the remaining I-1 one-dimensional convolutional layers are stacked dilated convolutional layers (D-conv). The dilation factor of each dilated convolutional layer in the stacking order is d = 1,2 ,...,2 I-1 , and each dilated convolutional layer output is connected in series with a parameter-corrected linear unit PReLU, the adaptive filter output y and the system error e are respectively input into two parallel encoders to obtain their respective encoding features

[0067] ②The eliminator includes a temporal convolutional network module for extracting multi-level features, an interaction module at the input end, and a mask generation module at the output end. The eliminator receives the encoded features Given the mask M for the gain feature, the gain encoding feature F is obtained by matrix dot multiplication G , the calculation process is

[0068] 1) The temporal convolutional network module includes R (R = 4) repeated layers, which receive the output of the interaction module and give the feature output F i TCN , i=1,2,...,R.

[0069] The repetitive layer consists of M (M=8) stacked one-dimensional convolution blocks (1-D Conv), as Figure 4 As shown in the figure, the input end of each one-dimensional convolution block contains two convolution layers, one of which is a 1×1 convolution layer (1×1Conv), and the other is an expanded convolution layer. The output end of each convolution layer is sequentially connected in series with a parameter-rectified linear unit PreLU and a global layer normalization gLN. The output end of each one-dimensional convolution block contains two parallel 1×1 convolution layers.

[0070] Through the two convolutional layers at the input end, two parallel 1×1 convolutional layers at the output end give two outputs of the one-dimensional convolution block, namely the residual output and the skip connection output. The input of the subsequent one-dimensional convolution block is the residual output of the previous one-dimensional convolution block. The skip connection outputs of M stacked one-dimensional convolution blocks are superimposed to generate the feature output F i TCN .

[0071] The number of convolution kernels of the 1×1 convolution layer in the one-dimensional convolution block is 128, the number of convolution kernels of the dilated convolution layer is 256, the length of the convolution kernel is 3, the step size is 1, and the dilation factor of each dilated convolution layer in the M stacked one-dimensional convolution blocks is d=1,2,...,2 in the stacking order. M-1 , the input of the first one-dimensional convolutional block in the repeated layer is the output of the interactive module

[0072] 2) If Figure 3 As shown in Figure 2, the number of interaction modules is the same as the number of repeated layers. Each interaction module receives one main branch input and two auxiliary branch inputs, where the main branch input received by the first interaction module is the encoded feature. The main branch input received by the remaining three interaction modules is the feature output F of the first three repeated layers in the temporal convolutional network module. i TCN , and the two auxiliary branch inputs are encoded features

[0073] The interaction module first uses the global layer normalization gLN to encode features Normalized and input into two parallel 1×1 convolutional layers to reduce the feature dimension. The outputs of the two 1×1 convolutional layers (the number of convolution kernels of the 1×1 convolutional layer is 256, the 1×1 convolutional layer represents a one-dimensional convolutional layer with a convolution kernel length of 1 and a convolution step of 1) are concatenated according to the channel dimension and then subjected to a global layer normalization gLN operation. The subsequent 1×1 convolutional layers, parameterized corrected linear units PReLU and global layer normalization gLN are connected in series to integrate the concatenated features to obtain the local mask M. j , the local mask M j Integrate with the main branch input to form the output of the interactive module The integration process is where j = 0, 1, ..., R-1,

[0074] 3) The mask generation module is composed of a parameterized rectified linear unit PreLU, a 1×1 convolutional layer, a global layer normalization gLN, and a Sigmoid function connected in series. Its input is the feature output F of the R repeated layers in the temporal convolutional network module. i TCN The output is the mask M used to gain features.

[0075] ③ The decoder consists of I (I = 5) one-dimensional deconvolution layers, among which the first I-1 one-dimensional deconvolution layers are stacked dilated deconvolution layers, the number of convolution kernels is 256, the length of the convolution kernel is set to 3, the step size is set to 1, and the dilation factor of each dilated deconvolution layer in the stacking order is d = 2 I-1 ,2 I-2 ,…,1, and each expansion deconvolution layer output end is connected in series with a parameter correction linear unit PReLU, the number of convolution kernels of the I-th one-dimensional deconvolution layer is N (N = 256), the length of the convolution kernel is K (K = 16), the convolution step is K / 2, and the gain encoding feature F G Input to the decoder to get clean heart and lung sound signal

[0076] In the technical solution of this application, the training process of the fully convolutional deep neural network is:

[0077] A large number of pure heart and lung sound signals are mixed with background noise data to generate a data training set. The data training set needs to be input into the ANC system first to obtain the adaptive filter output and system error as the input of the network. The pure heart and lung sound signal is used as the target value, and the network is trained based on the neural network back propagation method to obtain the network parameters.

[0078] The network parameters are updated once the network is trained until the final cost function value stabilizes near a certain value. The training can then be stopped and the current network parameters and network structure information can be saved.

[0079] Among them, the cost function used in the training process of the network is the negative scale-invariant signal-to-noise ratio SI-SNR, which is defined as in s target are the output and target value of the network respectively.

[0080] The data used in this embodiment are all from public data sets, including the heart sound database Physionet (https: / / physionet.org / content / challenge-2016), the lung sound database ICBHI (https: / / bhichallenge.med.auth.gr), and the noise database DEMAND (https: / / zenodo.org / record / 1227121#.YNqNl5AzYdU). After excluding the records with severe murmurs and unsuccessful heart and lung sound records, a total of 161 heart sound records were selected from Physionet, each with a different duration, and 90 lung sound records were selected from ICBHI, each with a different duration. 141 heart sound records and 76 lung sound records were randomly selected to construct the training data set, and the remaining records were used for testing.

[0081] For the clean heart sound records selected in Physionet, the sampling rate is 2000Hz. At this time, the signal of each record is first upsampled to 8000Hz; for the clean lung sound records selected in ICBHI, the sampling rate is 44100Hz. At this time, the signal of each record is first downsampled to 8000Hz; the sampling rate of the noise records in DEMAND is 16000Hz. The noise signal of each record is first downsampled to 8000Hz.

[0082] Since the duration of noise recording is generally longer than that of heart and lung sound recording, this embodiment randomly extracts background sound signals from various noise recordings and mixes them with heart and lung sound recordings at a specified signal-to-noise ratio to obtain a simulated heart sound or lung sound recording with background sound interference. The signal-to-noise ratio is randomly selected from -5dB, 0dB, 5dB, 10dB, 15dB, and 20dB, and all simulated heart and lung sound recordings are traversed at an overlap rate of 50%. A heart and lung sound signal with a duration of 2 seconds and background sound interference is extracted, and the background sound signal and clean heart and lung sound signal are similarly extracted. Finally, 14,823 groups of training data and 2,578 groups of test data are obtained, each of which contains 2 seconds of heart and lung sound signals with background sound interference, 2 seconds of background sound signals, and 2 seconds of clean heart and lung sound signals.

[0083] For the heart sound signal processing results, such as Figure 5As shown in the figure, (a) is the input heart sound frequency diagram with background sound interference, (b) is the input corresponding background sound frequency diagram, and (c) is the output heart sound frequency diagram after eliminating background sound interference; for the lung sound signal processing results, as shown in Figure 6 As shown, (a) is the time-frequency diagram of lung sounds with background sound interference input in an embodiment of the present invention, (b) is the time-frequency diagram of the corresponding background sound input, and (c) is the time-frequency diagram of lung sounds output after eliminating the background sound interference.

[0084] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An adaptive noise reduction method for electronic stethoscope based on time-domain fully convolutional deep neural network, characterized by: The following steps are involved: S1, the electronic stethoscope collects the cardiopulmonary sound signal with background sound interference through the main acquisition channel, and simultaneously collects the background sound signal through the remote acquisition channel, and inputs it into the original signal buffer space; S2, normalizing the main acquisition channel and the remote acquisition channel in the original signal buffer space respectively to obtain the normalized cardiopulmonary sound signal s and the background sound signal x with background sound interference; S3, inputting the cardiopulmonary sound signal s with background sound interference and the background sound signal x into an adaptive noise cancellation ANC system including an adaptive filter and an adaptive algorithm, and obtaining an adaptive filter output y and a system error e; S4: Input the adaptive filter output y and the system error e into the trained fully convolutional deep neural network to obtain a clean heart and lung sound signal The fully convolutional deep neural network described in S4 includes an encoder, an eliminator, and a decoder, wherein the network parameters are trained by a labeled data training set; The encoder consists of I one-dimensional convolutional layers, where the number of convolution kernels of the first one-dimensional convolutional layer is N, the length of the convolution kernel is K, the convolution step is K / 2, and the remaining I-1 one-dimensional convolutional layers are stacked dilated convolutional layers, and the dilation factor of each dilated convolutional layer in the stacking order is d=1,2,...,2 I-1 , and each dilated convolutional layer output is connected in series with a parameter-corrected linear unit PReLU, the adaptive filter output y and the system error e are respectively input into two parallel encoders to obtain their respective encoding features The eliminator includes a temporal convolutional network module for extracting multi-level features, an interaction module at the input end, and a mask generation module at the output end. The eliminator receives the encoded features Given the mask M for the gain feature, the gain encoding feature F is obtained by matrix dot multiplication G The calculation process is The decoder is composed of I one-dimensional deconvolution layers, wherein the first I-1 one-dimensional deconvolution layers are stacked dilated deconvolution layers, and the dilation factor of each dilated deconvolution layer in the stacking order is d=2 I-1 ,2 I-2 ,...,1, and each expansion deconvolution layer output end is connected in series with a parameter correction linear unit PReLU, the number of convolution kernels of the I-th one-dimensional deconvolution layer is N, the length of the convolution kernel is K, the convolution step is K / 2, and the gain encoding feature F G Input to the decoder to get clean heart and lung sound signal The temporal convolutional network module includes R repeated layers, which receive the output of the interaction module and give a feature output F i TCN , i=1,2,...,R; The number of interaction modules is the same as the number of repeated layers. The interaction module first uses the global layer normalization gLN to encode the features Normalized and input into two parallel 1×1 convolutional layers to reduce the feature dimension. The outputs of the two 1×1 convolutional layers are concatenated according to the channel dimension and then subjected to a global layer normalization gLN operation. The subsequent 1×1 convolutional layers, parameterized corrected linear units PReLU and global layer normalization gLN are connected in series to integrate the concatenated features to obtain the local mask M. j , the local mask M j Integrate with the main branch input to form the output of the interactive module The integration process is where j = 0, 1, ..., R-1, The mask generation module is composed of a parameterized corrected linear unit PreLU, a 1×1 convolutional layer, a global layer normalization gLN, and a Sigmoid function connected in series. Its input is the feature output F of the R repeated layers in the temporal convolutional network module. i TCN The superposition of is the mask M used to gain features; The repetitive layer is composed of M stacked one-dimensional convolution blocks, and the input end of each one-dimensional convolution block includes two convolution layers, one of which is a 1×1 convolution layer and the other is an expanded convolution layer. The output end of each convolution layer is sequentially connected in series with a parameter-corrected linear unit PreLU and a global layer normalization gLN, and the output end of each one-dimensional convolution block includes two parallel 1×1 convolution layers; Through the two convolutional layers at the input end, two parallel 1×1 convolutional layers at the output end give two outputs of the one-dimensional convolution block, namely the residual output and the skip connection output. The input of the subsequent one-dimensional convolution block is the residual output of the previous one-dimensional convolution block. The skip connection outputs of M stacked one-dimensional convolution blocks are superimposed to generate the feature output F i TCN ; Among them, the dilated convolutional layers in the M stacked one-dimensional convolutional blocks have dilation factors of d=1,2,...,2 in the stacking order. M-1 , the input of the first one-dimensional convolutional block in the repeated layer is the output of the interactive module 2. The method for adaptive denoising of an electronic stethoscope based on a time-domain fully convolutional deep neural network according to claim 1 is characterized in that: After the input is into the original signal buffer space, when the buffer space data accumulates to the set time length T, execute S2, otherwise repeat S1; After normalizing the main acquisition channel and the remote acquisition channel in the original signal buffer space respectively, clearing the buffer space data, and synchronously executing S1; The clean heart and lung sound signal is obtained Afterwards, when the cache space data accumulates to the next set time length T, it returns to S2 to continue execution.

3. The method for adaptive denoising of an electronic stethoscope based on a time-domain fully convolutional deep neural network according to claim 1 is characterized in that: The adaptive filter in S3 is an M-order FIR transversal filter, and the specific filtering process is: y(k)=w T (k)x(k), where k is the number of iterations, w(k)=[w0(k),w1(k),...,w M-1 (k)] represents the filter tap weight vector, x(k) = [x(k), x(k-1), ..., x(k-M+1)] represents iterating through the background sound signal x to generate the filter tap input vector; Subtract each sampling point s(k) in the cardiopulmonary sound signal s with background sound interference from y(k) after iteration to obtain the system error e(k) of each sampling point; After the iteration is completed, y(k) is saved to obtain the adaptive filter output y, and the system error e(k) of each sampling point is saved to obtain the system error e.

4. The method for adaptive denoising of an electronic stethoscope based on a time-domain fully convolutional deep neural network according to claim 1 is characterized in that: The adaptive algorithm described in S3 is the normalized minimum mean square error NLMS algorithm, which is used to update the filter tap weight vector w(k). The updating process is: Among them, μ0 is the convergence factor, and its value range is (0,1], and ζ is a very small positive number used to limit the step size and avoid the denominator being 0.

5. The method for adaptive denoising of an electronic stethoscope based on a time-domain fully convolutional deep neural network according to claim 1 is characterized in that: The training process of the fully convolutional deep neural network described in S4 is: A large number of pure heart and lung sound signals are mixed with background noise data to generate a data training set. The data training set needs to be input into the ANC system first to obtain the adaptive filter output and system error as the input of the network. The pure heart and lung sound signal is used as the target value, and the network is trained based on the neural network back propagation method to obtain the network parameters. The network parameters are updated once the network is trained until the final cost function value stabilizes near a certain value. The training can then be stopped and the current network parameters and network structure information can be saved.

Images