An ECG signal reconstruction method based on AE-GAN with enhanced decoding

Abstract

The present invention relates to the technical field of physiological signal processing and analysis, and discloses an ECG signal reconstruction method based on AE-GAN with enhanced decoding. An AE-GAN hybrid neural network model is constructed, wherein the AE-GAN hybrid neural network model includes a generator and a discriminator. The generator includes an encoder and a decoder, and the discriminator includes a first fully convolutional layer, a fully connected layer, and a Sigmoid activation function layer. ECG signals containing noise with different signal-to-noise ratios are input into the AE-GAN hybrid neural network model, and reconstructed noise-free ECG signals are output after processing. While the GAN network is used to capture the global features of the ECG, the AE network convolution operation is used to capture the local temporal features of the ECG, and an attention mechanism is added to avoid the loss of channel information of the network during the encoding process. A generator loss function is designed that integrates a least squares function following the Pearson chi-square divergence, a distance function, and a local maximum difference function to overcome the gradient diffusion problem during the GAN network training process. The method has higher accuracy and stability in reconstructing ECG signals.

Description

Technical Field

[0001] The present invention relates to the technical field of physiological signal processing and analysis, and in particular to an ECG signal reconstruction method based on AE-GAN (Electrode Generative Adversarial Network) enhanced decoding. Background Art

[0002] Electrocardiogram (ECG) is the reflection of the heart's electrical activity on the surface of the human body and is the gold standard for clinical assessment of heart and cardiovascular health status. With the increasing popularity of portable monitoring devices, the collection and analysis of ECG in daily behavioral environments has become a research hotspot. However, long-term ECG recordings are often affected by factors such as acquisition equipment, human activity, and electrode contact, which results in the ECG collected by the device containing various noises and interferences (mainly including baseline drift noise, myoelectric interference, electrode interference, etc.). Even more, the noise may overwhelm the ECG signal, making it impossible for the monitoring device to correctly obtain the physiological state. Therefore, the development of relevant algorithms to remove ECG noise and reconstruct a pure ECG is an important preliminary work for intelligent diagnosis of cardiovascular diseases. In response to the above problems, the present invention provides an ECG signal reconstruction method of AE-GAN with enhanced decoding.

[0003] Common ECG noise artifacts include baseline wander (BW) caused by respiration (frequency range: 0.05-2Hz), power frequency interference (50Hz or 60Hz), muscle artifacts (MA) caused by muscle movement (frequency range: 5-2000Hz), and electrode motion (EM) caused by changes in electrode contact during exercise. Researchers have devoted considerable effort to removing these noise artifacts and reconstructing a pure ECG signal, and have achieved considerable success. Existing ECG signal noise reduction algorithms are mainly divided into traditional noise reduction methods and deep learning-based noise reduction algorithms. Traditional ECG signal filtering involves analog and digital filter designs. Analog filters are generally ineffective and are typically implemented as hardware peripherals with amplifier circuits. High-precision filtering at the control end is achieved by digital filters. Common digital filters include FIR / IIR filters, wavelet transform (WT) filters, empirical mode decomposition (EMD) filters, adaptive filters, and improved algorithms based on these theories. Because ECG signal energy is primarily concentrated in the 0-100 Hz range, which overlaps with the aforementioned noise, FIR / IIR filters are ineffective and prone to group delay. WT decomposes the ECG signal into different frequency bands. It is generally believed that noise is distributed in the high-frequency portion of the wavelet decomposition, while the ECG signal is primarily distributed in the low-frequency portion. Therefore, reconstructing the ECG signal after soft or hard thresholding the high-frequency coefficients of the wavelet decomposition can remove noise. The effectiveness of this method depends heavily on the choice of wavelet mother function and threshold. The EMD method decomposes the ECG signal into multiple intrinsic mode functions (IMFs), then removes the mode function with the most noise, and reconstructs the ECG signal from the remaining IMFs. However, high-frequency noise often aliases into multiple IMFs and intersects with the QRS complex, leading to modal aliasing and even QRS distortion. To mitigate the effects of modal aliasing, integrated empirical mode decomposition, variational mode decomposition, and their improved algorithms have been widely introduced in the field of ECG denoising. Since ECG noise is complex, the signal itself has large individual and pathological differences, and is severely affected by human motion, accurate reconstruction of a pure ECG often requires the coordinated use of the above-mentioned multiple methods.

[0004] With the rapid development of machine learning and neural network technologies, data-trained ECG denoising and reconstruction algorithms have been gradually proposed. These methods can be broadly categorized into two types: one uses machine learning training to improve, optimize, and update key parameters in traditional noise reduction models to achieve better denoising results; the other uses data-based training to decompose and reconstruct the ECG data using neural networks, seeking point-to-point corresponding data reconstruction models. For example, error backpropagation neural networks are combined with variational mode decomposition to remove BW noise; supervised deep factor models are constructed to remove linear superposition noise from ECG data; and autoencoders (AEs) are applied to ECG denoising, leveraging their compression coding and representation learning capabilities. The focus is on identifying the nonlinear relationship between noisy ECG data and feature vectors by designing appropriate encoding processes. Because generative adversarial neural networks (GANs), through the combined action of a generator and discriminator, can make the generated signal approach the pure signal in both global and local characteristics, some researchers have introduced them into the field of ECG denoising, achieving some promising results. However, single neural networks have certain limitations when dealing with the complex and diverse noise of ECG data. Among them, AE is prone to losing ECG detail features during the compression encoding process, which may lead to the loss of transient cardiac pathological information and cause an irreversible situation. In the application of GAN, the structure of the generator and discriminator greatly affects the final noise reduction performance. Therefore, researchers often improve the performance of the network by introducing residual neural networks, deep convolutional neural networks, conditional convolution, etc. into the generator and discriminator. However, the noise reduction algorithms based on deep learning have the following problems: (1) Some methods require ECG segmentation, the signal processing process is complicated, and the stability is not high; (2) ECG is rich in details and pathological information. Some noise reduction methods lose detailed information seriously while removing noise, and the accuracy needs to be improved. Summary of the Invention

[0005] In view of the shortcomings of the existing technology, the purpose of the present invention is to provide an ECG signal reconstruction method based on AE-GAN with enhanced decoding.

[0006] In order to achieve the above object, the present invention provides the following technical solutions:

[0007] An ECG signal reconstruction method based on AE-GAN with enhanced decoding is proposed. The AE-GAN hybrid neural network model is constructed. The AE-GAN hybrid neural network model includes a generator and a discriminator, wherein the generator includes an encoder and a decoder, and the discriminator includes a first fully convolutional layer, a fully connected layer, and a sigmoid activation function layer. ECG signals containing myoelectric interference, baseline drift noise, electrode interference, and mixed noise with different signal-to-noise ratios are input into the AE-GAN hybrid neural network model. After processing, a reconstructed noise-free ECG signal is output.

[0008] In the present invention, preferably, the generator is constructed by an AE network, and an attention mechanism is added to enable the AE network model to strengthen channel change information in the convolution operation.

[0009] In the present invention, preferably, a SE layer is added after each deconvolution layer of the decoder to capture the channel relationship.

[0010] In the present invention, preferably, the loss function used by the generator is:

[0011]

[0012]

[0013]

[0014] In the formula, the first term is the least squares loss function, L dist is a distance function used to characterize the global difference of data; L max is the maximum difference, which is used to characterize the local difference of data. Represents a noise signal The distribution of N represents the sample length, represents the denoised signal of sample point i, x i Represents the pure signal (i.e., label data) of sample point i, λ1 and λ2 are L dist and L max The weighting coefficients of are set to 0.7 and 0.2 respectively in the experiment.

[0015] In the present invention, preferably, the loss function used by the discriminator is:

[0016]

[0017] In the formula, the first term, x~p data (x) is the distribution of the real data x, D(x) is the result of the discriminator identifying the pure signal, and 1 is the label of the real sample (ie, the pure signal); the second item, Represents a noise signal The distribution of is the result of the discriminator identifying the generated sample, -1 is the label data of the false sample (i.e., the generated sample), and the optimization goal of this formula is to distinguish the samples generated by G by improving the resolution ability.

[0018] In the present invention, preferably, the AE-GAN hybrid neural network model is trained using pre-processed ECG data, and the pre-processing of the ECG data includes:

[0019] S1. Prepare training data by adding myoelectric interference, baseline drift noise, electrode noise, and mixed noise of different signal-to-noise ratios to the pure ECG signal to obtain noisy ECG signal samples;

[0020] S2, segmenting the noisy ECG signal into 1024 points per sample and normalizing the data amplitude;

[0021] S3. Divide all data into 95% training set and 5% test set.

[0022] In the present invention, preferably, the encoder includes 6 second fully convolutional layers, the decoder includes 6 deconvolutional layers and 6 SE layers, the deconvolutional layers and SE layers are alternately arranged, and the second fully convolutional layers are jump-connected to the deconvolutional layers.

[0023] In the present invention, preferably, a layer of PReLU activation function is added after each second full convolutional layer and deconvolution layer.

[0024] In the present invention, preferably, the discriminator includes 5 first fully convolutional layers, 1 fully connected layer and 1 Sigmoid activation function layer, the convolution kernel sizes of the first fully convolutional layers are 16, 16, 16, 16, 8, respectively (the step sizes are 2, 4, 4, 4, 1, respectively), and each of the first fully convolutional layers is connected to a layer of LeakyReLU activation function.

[0025] In the present invention, preferably, the generator and the discriminator both use the Adam optimizer, with an initial learning rate of 0.0001 and a batch size of 64.

[0026] Compared with the prior art, the present invention has the following beneficial effects:

[0027] The method of the present invention uses the GAN network to capture the global features of ECG, while adopting the AE network convolution operation to capture the local temporal features of ECG, and avoids the channel information loss that may occur in the network during the encoding process by adding an attention mechanism. The design integrates the generator loss function that follows the least square function of the Pearson chi-square divergence, the distance function and the local maximum difference function to overcome the gradient diffusion problem in the GAN network training process; it has higher accuracy and stability in reconstructing ECG signals. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 This is a schematic diagram of the structure of the AE-GAN hybrid neural network model constructed by the ECG signal reconstruction method based on enhanced decoding AE-GAN described in the present invention.

[0029] Figure 2 This is a schematic diagram of the generator structure of the present invention.

[0030] Figure 3 Schematic diagram of the discriminator structure of the present invention.

[0031] Figure 4 Schematic diagram of the SE layer structure of the present invention.

[0032] Figure 5 This is a graph showing the results of removing MA noise using the ECG signal reconstruction method based on enhanced decoding AE-GAN described in the present invention.

[0033] Figure 6 This is a diagram showing the BW noise removal results of the ECG signal reconstruction method based on enhanced decoding AE-GAN described in the present invention.

[0034] Figure 7 This is a diagram showing the EM noise removal results of the ECG signal reconstruction method based on enhanced decoding AE-GAN described in the present invention.

[0035] Figure 8 This is a graph showing the removal results of MA+BW+EM mixed noise using the ECG signal reconstruction method based on enhanced decoding AE-GAN described in the present invention. DETAILED DESCRIPTION

[0036] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only 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 making creative efforts are within the scope of protection of the present invention.

[0037] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which this invention pertains. The terms used in this specification of the present invention are for the purpose of describing specific embodiments only and are not intended to limit the present invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0038] Please also see Figures 1 to 3 One embodiment of the present invention provides an ECG signal reconstruction method based on AE-GAN with enhanced decoding. Based on the GAN model, the generator of the GAN model is combined with an improved AE structure. The structure captures channel change information by adding an SE structure in the decoding process, thereby reducing the loss of ECG detail information in the AE process. The design of the GAN discriminator is based on a fully convolutional structure. In view of the fact that the traditional GAN ​​training based on the cross entropy loss function is prone to gradient diffusion, this method designs a loss function that integrates the least squares function following the Pearson chi-square divergence, the distance function and the local maximum difference function. Loss function, according to a large number of experiments, this method has higher accuracy and stability in reconstructing ECG signals. This method specifically includes constructing an AE-GAN hybrid neural network model, which includes a generator and a discriminator. The generator includes an encoder and a decoder, and the discriminator includes a first fully convolutional layer, a fully connected layer, and a Sigmoid activation function layer. ECG signals containing electromyographic interference, baseline drift noise, electrode interference, and mixed noise with different signal-to-noise ratios are input into the AE-GAN hybrid neural network model, and the reconstructed noise-free ECG signal is output after processing.

[0039] Specifically, the generator is constructed by the AE network, and an attention mechanism is added to enable the AE network model to strengthen channel change information in the convolution operation, thereby improving the accuracy of ECG signal reconstruction.

[0040] Specifically, a SE layer is added after each deconvolution layer of the decoder. The SE layer is a SE (Squeeze-and-Excitation, SE) structure to capture the channel relationship, thereby overcoming the information loss caused by channel compression during the decoding process, such as Figure 4 As shown, the SE structure input x is L, the number of channels is C, F tr It is a Transformation structure, which is implemented using a full convolutional network. tr Operation, the output length is L2, the number of channels is C2, the SE structure implementation steps mainly include two operations Squeeze and Excitation. Squeeze uses the global average pooling operation to compress the spatial information into the channel, the output length is 1, the number of channels is C2, as shown in the following formula:

[0041]

[0042] Excitation is used to capture the relationship between channels. First, a fully connected layer is used to compress the channel to C2 / rate, and then another fully connected layer is used to restore the channel to C2. The activation functions of the two fully connected layers are respectively selected as ReLU functions:

[0043]

[0044] And the Sigmoid function:

[0045]

[0046] The output s after the Excitation operation is a C2-dimensional vector.

[0047] s=F ex (z,W)=σ(g(z,W))=σ(W2δ(W1z))

[0048] Each value corresponds to the weight of the channel. The input feature u is multiplied by the corresponding weight to screen the importance of the feature, giving a high weight to important features and a low weight to useless features.

[0049]

[0050] In this embodiment, the network loss function is designed: to address the gradient diffusion problem caused by the cross entropy loss function of the traditional GAN ​​network, a new loss function integrating the least squares loss function, the distance function, and the maximum local difference function is designed in combination with ECG features to improve the stability of ECG reconstruction. The loss function used by the generator is:

[0051]

[0052]

[0053]

[0054] In the formula, the first term is the least squares loss function, L dist is a distance function used to characterize the global difference of data; L max is the maximum difference, which is used to characterize the local difference of data. Represents a noise signal The distribution of N represents the sample length, represents the denoised signal of sample point i, x i Represents the pure signal (i.e., label data) of sample point i, λ1 and λ2 are L dist and L maxThe weighting coefficients of are set to 0.7 and 0.2 respectively in the experiment;

[0055] The loss function used by the discriminator is:

[0056]

[0057] In the formula, the first term, x~p data () is the distribution of the real data x, D(x) is the result of the discriminator identifying the pure signal, and 1 is the label of the real sample (ie, the pure signal); the second item, Represents a noise signal The distribution of is the result of the discriminator identifying the generated sample, -1 is the label data of the false sample (i.e., the generated sample), and the optimization goal of this formula is to distinguish the samples generated by G by improving the resolution ability.

[0058] Specifically, the AE-GAN hybrid neural network model is trained using preprocessed ECG data. The preprocessing of ECG data includes:

[0059] S1. Prepare training data by adding myoelectric interference (MA), baseline wander noise (BW), electrode noise (EM), and a mixture of these three noises at different signal-to-noise ratios (-1dB, 3dB, 7dB) to the pure ECG signal to obtain noisy ECG signal samples. The pure ECG signal comes from the public dataset MIT-BIH Arrhythmia Database.

[0060] S2, segmenting the noisy ECG signal into 1024 points per sample and normalizing the data amplitude;

[0061] S3. Divide all data into 95% training set and 5% test set.

[0062] Specifically, the preprocessed ECG data is used to train the constructed AE-GAN hybrid neural network model, wherein the noisy ECG signal of the training set is input into the AE-GAN hybrid neural network model for training, and Adam is used to dynamically adjust the learning rate so that the gradient descent is iteratively updated along the negative gradient direction, thereby updating all the weights of the network until the loss function is minimized. The iteration ends and the model training ends. Then, the noisy ECG signal of the test set is input into the trained AE-GAN hybrid neural network model, and the denoised ECG signal is obtained at the output end. The output is compared with the pure ECG signal in the corresponding test set, that is, the performance indicators signal-to-noise ratio improvement (SNRimp), root mean square error (RMSE) and percentage root mean square difference (PRD) are calculated to evaluate the effectiveness of the denoising algorithm;

[0063] Where SNRimp represents the difference in signal-to-noise ratio between the noise-reduced signal and the original pure signal.

[0064] SNR imp =SNR out -SNR in

[0065]

[0066]

[0067] SNR imp The higher the value of , the better the noise reduction performance of the corresponding algorithm;

[0068] The RMSE is used to measure the error between the noise reduction signal and the original pure signal.

[0069]

[0070] The smaller the RMSE value, the smaller the error between the noise-reduced signal and the original pure signal;

[0071] Among them, PRD is used to measure the similarity between the reconstructed signal and the target signal.

[0072]

[0073] The smaller the PRD value is, the higher the reconstruction similarity of the corresponding denoising method is; in the above formula, x i represents a pure ECG signal, represents a noisy ECG signal, represents the ECG signal after noise reduction, and N is the signal length in one calculation cycle.

[0074] In this embodiment, the encoder includes 6 layers of second full convolutional layers, and the decoder includes 6 layers of deconvolutional layers and 6 layers of SE layers. The deconvolutional layers and SE layers are arranged alternately. The second full convolutional layers and the deconvolutional layers are jump-connected. The convolution kernels of the full convolutional layers and the deconvolutional layers are both 32. A PReLU activation function is added after each second full convolutional layer and deconvolution layer:

[0075]

[0076] The discriminator consists of 5 first fully convolutional layers, 1 fully connected layer and 1 Sigmoid activation function layer. The convolution kernel sizes of the first fully convolutional layer are 16, 16, 16, 16, and 8, respectively, with step sizes of 2, 4, 4, 4, and 1. Each first fully convolutional layer is connected to a LeakyReLU activation function:

[0077]

[0078] Where a is a very small slope;

[0079] The input of the discriminator is the denoised ECG signal or the clean ECG signal, with a size of 1024*1; both the generator and the discriminator use the Adam optimizer, with an initial learning rate of 0.0001 and a batch size of 64.

[0080] Working principle:

[0081] The overall steps of this method are:

[0082] 1. Prepare training data. The model outputs label data. The noisy ECG signal input to the model is obtained by adding myoelectric interference (MA) with different signal-to-noise ratios (-1dB, 3dB, 7dB), baseline drift noise (BW), electrode noise (EM), and a mixture of the above three noises to the pure ECG signal. Data is segmented into 1024 points for each sample of the noisy ECG signal and the data amplitude is normalized. The normalization formula is:

[0083]

[0084] Where x is the input signal, x min is the minimum value of 1024 sample point data, x max The maximum value of the 1024 sample point data, x norm The data is normalized; all data are divided into 95% training set and 5% test set;

[0085] 2. Build the model structure. The AE-GAN hybrid neural network model includes a generator and a discriminator. The generator includes a connected encoder and decoder. The encoder consists of 6 layers of second-level full convolutional layers. The size of each convolution kernel is 32, and the activation function of each convolution output is the PReLU function:

[0086]

[0087] Where b is the slope of learning, x is the neuron input, and this activation function can ensure that when the input is negative, the neuron output still exists, thus avoiding the death of a large number of neurons during training.

[0088] The decoder consists of 6 deconvolution layers and 6 SE layers. The deconvolution layers and SE layers are arranged alternately. The SE layers in the decoder implement channel information compensation during the decoding process to reduce the information loss caused by the reduction of the number of channels during the decoding process. The activation function of each deconvolution layer output is the PReLU function.

[0089] The discriminator consists of 5 layers of full convolution, one layer of full connection and a Sigmoid activation function layer, as shown in Figure 3 As shown in the figure, the convolution kernel sizes of each convolution layer are 16, 16, 16, 16, and 8, respectively, with strides of 2, 4, 4, 4, and 1, respectively. Each convolution layer is followed by a LeakyReLU activation function. The input of the discriminator is the denoised ECG signal or the clean ECG signal, with a size of 1024*1.

[0090] 3. Model training, estimating the generated data corresponding to the sample through the adversarial process, that is, the denoised ECG signal The generator (denoted as G) and discriminator (denoted as D) of the whole hybrid neural network model are trained simultaneously: G is trained based on the input data, i.e. the noisy ECG signal (denoted as ), random noise z is distributed through the encoder output data, and then the observation data is generated by the decoder based on the integrated channel enhancement The discriminator estimates The probability of coming from a pure sample x. The optimization goal of G is to maximize the error probability of the discriminator, making it unable to distinguish the generated data and pure samples x; the optimization goal of D is to distinguish the samples generated by G by improving its ability to distinguish. G and D improve their respective abilities through adversarial training. The overall objective function of this adversarial process is to solve the following maximum-minimization equation:

[0091]

[0092] Where Pdata is the distribution of the pure ECG signal x, and Pz is defined as the distribution of the generated data G(z). However, the above equation uses the cross-entropy loss function, which easily causes the generated data distributed on the decision boundary to be judged as true samples, resulting in the gradient diffusion problem of the generator update. Therefore, this method integrates a new loss function with the least squares loss function, the distance function, and the maximum local difference function to improve the stability of ECG reconstruction;

[0093] 4. Model application. First, the noisy ECG signal with a dimension of 1024*1 enters the encoder. After the encoder, the dimension of the ECG feature vector output by each convolution layer changes to 1024*32, 512*64, 256*128, 128*256, 64*512, and 32*1024 respectively. Then the 32*1024 ECG feature data enters the decoder, and the dimension of the output signal of each layer changes to 64*512, 128*256, 256*128, 512*64, 1024*32, and 1024*1. The decoder outputs the noise reduction signal. That is the reconstructed ECG signal, which can be input into the discriminator to identify the authenticity. The denoised signal gradually converges and stabilizes during the iterative training process.

[0094] like Figure 5-Figure 8 Figure 2 shows the effectiveness of the proposed method for removing MA noise, BW noise, EM noise, and mixed MA+BW+EM noise at a -1dB level. The first row of the figure represents noisy ECG signals, the second row represents clean signals, and the third row represents noise-reduced signals. This figure demonstrates that the proposed method effectively removes noise while preserving both the overall and detailed ECG features.

[0095] The above description is a detailed description of the preferred embodiments of the present invention, but the embodiments are not intended to limit the scope of the patent application of the present invention. Any equivalent changes or modifications completed under the technical spirit suggested by the present invention should fall within the patent scope covered by the present invention.

Claims

1. An ECG signal reconstruction method based on AE-GAN with enhanced decoding, characterized in that: Constructing an AE-GAN hybrid neural network model, the AE-GAN hybrid neural network model including a generator and a discriminator, wherein the generator includes an encoder and a decoder, and the discriminator includes a first fully convolutional layer, a fully connected layer, and a sigmoid activation function layer, inputting ECG signals containing myoelectric interference, baseline drift noise, electrode interference, and mixed noise with different signal-to-noise ratios into the AE-GAN hybrid neural network model, and outputting a reconstructed noise-free ECG signal after processing; The loss function used by the generator is: Where the first term is the least squares loss function, is a distance function used to characterize the global difference of data; is the maximum difference, which is used to characterize the local difference of data. Represents the noise signal ( ), N represents the sample length, Representative sample points i The noise reduction signal, Representative sample points i The pure signal, and yes and The weighting coefficients of are set to 0.7 and 0.2 respectively in the experiment; The loss function used by the discriminator is: Where x~p data (x) is the distribution of the real data x, D(x) is the result of the discriminator identifying the pure signal; the second term, Represents the noise signal ( ), The result of generating samples for the discriminator to identify; The AE-GAN hybrid neural network model is trained using pre-processed ECG data. The pre-processing of ECG data includes: S1. Prepare training data by adding myoelectric interference, baseline drift noise, electrode noise and mixed noise of the three with different signal-to-noise ratios to the pure ECG signal to obtain noisy ECG signal samples; S2, segmenting the noisy ECG signal into 1024 points per sample and normalizing the data amplitude; S3. Divide all data into 95% training set and 5% test set.

2. The ECG signal reconstruction method based on AE-GAN with enhanced decoding according to claim 1, characterized in that: The generator is constructed by the AE network, and an attention mechanism is added to enable the AE network model to strengthen channel change information in the convolution operation.

3. The ECG signal reconstruction method based on AE-GAN with enhanced decoding according to claim 1, characterized in that: A SE layer is added after each deconvolution layer in the decoder to capture channel relationships.

4. The ECG signal reconstruction method based on AE-GAN with enhanced decoding according to claim 1, characterized in that: The encoder includes 6 second fully convolutional layers, the decoder includes 6 deconvolutional layers and 6 SE layers, the deconvolutional layers and SE layers are alternately arranged, and the second fully convolutional layers are jump-connected to the deconvolutional layers.

5. The ECG signal reconstruction method based on AE-GAN with enhanced decoding according to claim 4, characterized in that: A PReLU activation function is added after each second full convolutional layer and deconvolution layer.

6. The ECG signal reconstruction method based on AE-GAN with enhanced decoding according to claim 1, characterized in that: The discriminator includes 5 first fully convolutional layers, 1 fully connected layer and 1 Sigmoid activation function layer. The convolution kernel sizes of the first fully convolutional layers are 16, 16, 16, 16, and 8, respectively (with step sizes of 2, 4, 4, 4, and 1, respectively). Each of the first fully convolutional layers is connected to a layer of LeakyReLU activation function.

7. The ECG signal reconstruction method based on AE-GAN with enhanced decoding according to claim 1, characterized in that: Both the generator and the discriminator use the Adam optimizer with an initial learning rate of 0.0001 and a batch size of 64.

Images