A method and system for denoising electrocardiogram signals based on asymmetric convolutional diffusion network

Abstract

The present application relates to a kind of electrocardiosignal denoising method, system based on asymmetric convolution diffusion network, belong to electrocardiosignal denoising technical field.First, the local receptive field of one-dimensional convolution layer is used to adapt to feature type data;Second, asymmetric convolution network architecture is used, and the feature map in encoder is combined with the feature map in decoder by jump connection.Simplify upsampling is used instead of transpose convolution to reduce noise introduction, and upsampling maintains the authenticity of original signal in denoising process by amplifying repeated signal.Finally, filter is added after jump connection, which helps to restore the original form and characteristics of signal, to improve the quality of decoder reconstructed signal.The present application is different for the interference received by electrocardiosignal in different frequency bands, and the proposed asymmetric convolution network can remove specific noise, avoid error in reconstruction process, and realize efficient electrocardiosignal denoising.

Description

TECHNICAL FIELD

[0001] The present application relates to an electrocardiosignal denoising method and system based on an asymmetric convolution diffusion network, and belongs to the technical field of electrocardiosignal denoising. BACKGROUND

[0002] The electrocardiosignal (ECG) is a physiological signal that records the electrical activity of the heart and is widely used in the diagnosis of various diseases such as arrhythmia, coronary artery disease, myocardial infarction, and some other mental disorders. As a key indicator of heart health, the accuracy of the electrocardiosignal is crucial. However, the electrocardiosignal is a very weak bioelectric signal, and the acquisition process is easily disturbed by external interference, especially local interference. For example, power line interference, electrode contact noise, electromyographic noise, and baseline drift. These noises not only reduce the quality of the signal, but also can lead to misdiagnosis and missed diagnosis. High-quality signals are the basis for research, and the quality of the signal directly affects the effectiveness of subsequent analysis. Therefore, by effectively removing noise, improving the quality of the signal, it can greatly help the subsequent research work.

[0003] In previous studies, traditional methods are often used for noise removal, without in-depth research on the characteristics of electrocardiosignals and the design of corresponding models for denoising. First, in terms of noise, during the acquisition process, patients often have local intense disturbances such as turning over and limb movement, which severely damage the signal. Second, in terms of feature extraction, different frequency bands of the signal are affected differently. In the low-frequency band, it is affected by baseline drift, and in the high-frequency band, it is mainly affected by electromyographic and power frequency interference. These disturbances have different forms of expression and different removal methods, but existing methods do not separate the important features for extraction and preservation according to this characteristic. Finally, in terms of the signal, the real signal is often very weak, making it difficult to identify.

[0004] To address the above problems, the present application proposes an electrocardiosignal denoising method and system based on an asymmetric convolution diffusion network, and designs an asymmetric convolution network Asy_UNet. SUMMARY

[0005] The present application provides an electrocardiosignal denoising method and system based on an asymmetric convolution diffusion network to improve the ability of electrocardiosignal denoising.

[0006] The technical solution of the present application is as follows: In a first aspect, the present application provides an electrocardiosignal denoising method based on an asymmetric convolution diffusion network, and the specific steps of the method are as follows:

[0007] Step 1, ECG signal preprocessing, including standardization and format conversion;

[0008] Step 2: Gradually add different levels of random Gaussian white noise to the preprocessed ECG signal until the original signal is covered by the noise;

[0009] Step 3: Back propagation. Taking the original signal as the target function, the model trains the noisy ECG signal and the corresponding clean signal. In this process, the features of the ECG signal are extracted, the ECG signal is reconstructed based on the extracted features, and finally the denoised ECG signal is generated.

[0010] As a further solution of the present invention, the process in Step 2 is a diffusion process. The diffusion process simulates the degradation process of data by gradually adding noise to the data. This process is described as a series of discrete steps, each step adding a certain amount of noise until the data becomes almost completely noisy. Assume that there is an initial data x0. This data is clear at the beginning of the diffusion process. In each step t, the data x0 is added to the data x0. t-1 Add noise to create new data x t , expressed as:

[0011]

[0012] where ∈ is random noise sampled from a standard normal distribution, x t-1 is the data state of the previous step, α t is a predefined noise level coefficient that determines how much noise is added at each step, α t is chosen to ensure a gradual increase in noise, ultimately making x T That is, the data in the last step is close to pure noise; the entire diffusion process consists of multiple time steps. In the initial step, the noise added is relatively small, but as the steps proceed, the noise gradually increases.

[0013] As a further solution of the present invention, in Step 3, the back propagation starts from the noise data x obtained in Step 2. T , which is the result of the forward diffusion process; at each time step t, the goal of the model is to predict the t The noise ∈ t , and use this prediction to reconstruct x t-1 , achieved by the following steps:

[0014] Noise prediction: Using a neural network f θ (x t, t) to predict the noise ∈ t , the noise that the network adds to the data at the current step is expressed as: ∈ t =f θ (x t, t);

[0015] Data reconstruction: Use the predicted noise to reconstruct the data x of the previous step t-1 ,σ t is a predefined noise level system;

[0016]

[0017] This iterative process iterates from t=T to t=0. In each step, the model uses the predicted noise to remove the noise in the current data and gradually reconstructs a clearer data state. After multiple steps of iteration, the model finally reconstructs clear data close to the original data x0.

[0018] As a further solution of the present invention, Step 3 includes:

[0019] (1) Use a one-dimensional convolution layer to slide the convolution kernel on the time axis to capture the local features of the ECG signal;

[0020] (2) An asymmetric convolutional network architecture is used, and the jump connection combines the feature maps in the encoder with the feature maps in the decoder; the asymmetric structure optimizes this process by selectively fusing features;

[0021] (3) Use upsampling instead of transposed convolution to reduce the introduction of noise. Upsampling maintains the authenticity of the original signal during the denoising process by amplifying the repeated signal;

[0022] (4) Adding a filter after the jump connection helps to restore the original shape and characteristics of the signal, thereby improving the quality of the decoder's reconstructed signal.

[0023] As a further solution of the present invention, said (2) includes: adjusting the convolution kernel size, filter shape, channel number allocation and upsampling step size of the network according to the characteristics of the ECG signal and the denoising requirements;

[0024] In the encoder, the number of channels is gradually increased to capture deep features, and in the decoder, the number of channels is gradually reduced to merge features and reduce noise. The asymmetric convolutional network architecture optimizes the feature fusion process by selectively fusing the features of the encoder and decoder, and optimizes the spatial resolution of the signal by adjusting the upsampling step size.

[0025] By adjusting the number of feature channels at different levels, the asymmetric convolutional network architecture can simultaneously learn multi-scale features; using a different number of channels at each stage of the asymmetric convolutional network architecture introduces a deep supervision mechanism; this indicates that each level of the network can be trained to minimize the loss of the corresponding level of the asymmetric convolutional network architecture, thereby improving the overall denoising performance.

[0026] As a further scheme of the present application, the (3) comprises:

[0027] Specifically, the nearest neighbor interpolation is used instead of the replacement convolution, and the nearest feature value of the original is copied to the new place, the nearest neighbor interpolation provides a more direct method to restore the size of the feature, so that the feature matches the original input; the up-sampling is completed by the convolution operation, wherein the weights of the convolution kernel are set as the interpolation mode, and the method allows the network to learn the optimal interpolation weight in the training process.

[0028] As a further scheme of the present application, the (4) comprises: in the training process, the parameters of the filter are adjusted according to the loss function to reduce the difference between the reconstructed signal and the original signal.

[0029] In another aspect, the present application provides an electrocardiosignal denoising system based on an asymmetric convolution diffusion network, comprising a module for executing the method of the first aspect.

[0030] The present application has the beneficial effects that: the present application improves the electrocardiosignal denoising capability, and each convolution layer in the model corresponding to the method of the present application uses one-dimensional convolution instead of two-dimensional convolution, which is more suitable for processing one-dimensional electrocardiosignal data; different levels correspond to different channel numbers to remove different frequency band noises; in addition, the model uses a filter in the connection part to optimize signal processing; the up-sampling process is also specially designed to ensure the accuracy of signal reconstruction.

[0031] The present application proposes an asymmetric convolution network aiming at the different interference of electrocardiosignal in different frequency bands, which can remove specific noise and avoid errors in the reconstruction process, and realizes efficient electrocardiosignal denoising. BRIEF DESCRIPTION OF DRAWINGS

[0032] Figure 1 The electrocardiosignal denoising flowchart of the present application;

[0033] Figure 2 The network model architecture of the present application;

[0034] Figure 3 The visualization diagram of adding different levels of noise to the original signal;

[0035] Figure 4 The signal time domain analysis diagram before and after denoising of the present application;

[0036] Figure 5 The frequency spectrum comparison analysis diagram of the signal before and after denoising of the present application. DETAILED DESCRIPTION

[0037] Example 1: as Figures 1-5As shown, an electrocardiogram signal denoising method based on asymmetric convolution diffusion network, the specific steps of the method are as follows: the specific steps of the method are as follows:

[0038] Step1, ECG signal preprocessing, including standardization and format conversion, to adapt to the input requirements of the network;

[0039] Step2, gradually add 0.1-0.5 level random Gaussian white noise to the preprocessed ECG signal until the original signal is covered by noise;

[0040] The process in Step2 is a diffusion process, and the diffusion process is to simulate the degradation process of data by gradually increasing noise to the data, which is described as a series of discrete steps, each step increases a certain amount of noise until the data is almost completely noisy; suppose there is an initial data x0, this data is clear at the beginning of the diffusion process, in each step t, to the data x t-1 Add noise to create new data x t , denoted as:

[0041]

[0042] Where, ∈ is a random noise sampled from a standard normal distribution, x t-1 is the data state of the previous step, α t is a predefined noise level coefficient, which determines how much noise is added in each step, α t is selected to ensure gradual increase of noise, and finally x T , that is, the data of the last step, is close to pure noise; the entire diffusion process contains multiple time steps, in the initial step, the addition of noise is relatively small, but as the step proceeds, the noise gradually increases.

[0043] Step3, back propagation, taking the original signal as the objective function, the model trains the noisy ECG signal and the corresponding clean signal; in this process, the features of the electrocardiogram signal are extracted, and the electrocardiogram signal is reconstructed according to the extracted features, and finally the denoised electrocardiogram signal is generated.

[0044] As a further scheme of the present application, in Step3, back propagation starts from the noise data x T obtained in Step2, which is the result of the forward diffusion process, and these data are almost completely noisy and cannot see any information of the original data; in each time step t, the goal of the model is to predict the noise ∈ t added to x t , and use this prediction to reconstruct x t-1 , which is achieved by the following steps:

[0045] Noise prediction: Using a neural network f θ (x t, t) to predict the noise ∈ t , the noise that the network adds to the data at the current step is expressed as: ∈ t =f θ (x t, t);

[0046] Data reconstruction: Use the predicted noise to reconstruct the data x of the previous step t-1 ,σ t is a predefined noise level system;

[0047]

[0048] This iterative process runs from t = T to t = 0. At each step, the model uses the predicted noise to remove noise from the current data and gradually reconstructs a clearer data state. After multiple iterations, the model ultimately reconstructs clear data that is close to the original data x0. This reconstructed data is usually close to the original signal, indicating that the model has successfully recovered useful information from the noisy data.

[0049] The denoised ECG signal is then compared with the original signal and the corresponding error is calculated to quantify the denoising effect. The denoising effect is mainly evaluated using signal-to-noise ratio (SNR), mean square error (MSE), signal distortion (SD), and coefficient of determination (R2).

[0050] The signal-to-noise ratio (SNR) is an important indicator for measuring signal quality. It is used to describe the ratio between the signal power and the background noise power. Its calculation formula is:

[0051]

[0052] Mean Squared Error (MSE) is an important indicator for measuring the accuracy of a model's predictive ability. It is calculated by taking the square of the difference between the predicted value and the actual value and averaging these squared differences. The calculation formula is:

[0053]

[0054] Signal distortion refers to the degree of deformation or quality degradation of a signal compared to the original signal during transmission or processing.

[0055]

[0056] The coefficient of determination (R-squared) is the ratio of the variance explained by the regression model to the total variance. Its calculation formula is:

[0057]

[0058] The trained model can effectively remove various noises in ECG signals, such as power line noise, baseline drift, and myoelectric interference, while maintaining the important features and waveform of the signal.

[0059] This paper designs an asymmetric convolutional network Asy_UNet. The main functions are: (1) using the local receptive field of the one-dimensional convolutional layer to adapt to the characteristic type of data (ECG signal). In the context of ECG signal denoising, the local receptive field focuses on the short-term fluctuations of the signal, which has great advantages for identifying and removing local, short-term noise such as noise generated by stray heartbeats, electrode friction, changes in sleeping posture, etc. (2) adjusting the number of feature channels at different levels can more effectively encode different aspects of the signal and model capacity. At the primary level, fewer channels are used to capture the basic waveform structure, while at deeper levels, more channels can focus on removing high-frequency noise. (3) adjustments are made between the encoder and decoder to optimize the processing of ECG signals through asymmetric convolutional layers and filters to better preserve the important features of the signal. (4) simple upsampling is used instead of transposed convolution to reduce noise introduction. Upsampling maintains the authenticity of the original signal during the denoising process by amplifying repeated signals.

[0060] As a further solution of the present invention, Step 3 includes:

[0061] (1) Use a one-dimensional convolution layer to slide the convolution kernel on the time axis to capture the local features of the ECG signal;

[0062] ECG signals are time-series data, whose key features unfold along the time axis. Key ECG signal information, such as the P wave, QRS complex, and T wave, all vary along the time axis. Therefore, ECG signal processing and analysis primarily focus on local features and variations along the time dimension.

[0063] One-dimensional convolution is a convolution operation designed for time series data. It applies the convolution kernel only in a single dimension (time dimension). By sliding the convolution kernel on the time axis, one-dimensional convolution can accurately capture these local temporal features without distracting attention to unnecessary spatial features like two-dimensional convolution. The expression is:

[0064]

[0065] Where x is the input ECG signal sequence, h is the convolution kernel, and n is the current sliding position. Convolution operation is essentially a filtering process that can enhance useful components in the signal while suppressing noise through the weights of the convolution kernel. The one-dimensional convolution designed by the application can capture the local features of the ECG signal by sliding the convolution kernel on the time series.

[0066] (2) An asymmetric convolutional network architecture is adopted, and the skip connection combines the feature maps in the encoder with those in the decoder. The asymmetric structure optimizes this process by selectively fusing features.

[0067] The Asy_UNet proposed in the application improves the asymmetric convolutional network architecture, and the asymmetric structure can optimize this process by selectively fusing features. In order to retain more detailed information in the reconstruction process. This can be expressed as:

[0068] concatenate(u,v)=(u,v)

[0069] Where u and v are features from the encoder and decoder respectively. The optimization principle is to adjust the convolution kernel size, filter shape, channel number allocation and up-sampling of the network according to the characteristics of the ECG signal and the noise removal requirements.

[0070] In this way, the model can more accurately capture and retain the key features of the ECG signal, while selectively suppressing noise of specific frequencies.

[0071] In the encoder, the gradually increasing number of channels is used to capture deep features, and in the decoder, the gradually decreasing number of channels helps to merge features and reduce noise. In addition, the asymmetric convolutional network architecture also optimizes the feature fusion process by selectively fusing the features of the encoder and decoder, and optimizes the spatial resolution of the signal by adjusting the up-sampling step. This optimization not only improves the sensitivity and accuracy of the model to ECG signals, but also enhances the robustness of noise removal, achieving higher performance in the ECG signal denoising task.

[0072] ECG signals contain information at multiple time scales, such as short-term heart rate variability and long-term rhythm patterns. By adjusting the number of feature channels at different levels, the asymmetric convolutional network architecture can simultaneously learn these multi-scale features, better understanding the overall structure of the signal. Using different channel numbers at each stage of the asymmetric convolutional network architecture can introduce a deep supervision mechanism. This indicates that each level of the network can be trained to minimize the loss of the corresponding level of the asymmetric convolutional network architecture, thereby improving the overall denoising performance. In the decoder, by reducing the number of channels and combining the features in the encoder, the network can effectively fuse the features. This fusion helps to reconstruct a clean and detailed ECG signal. The advantage of this improvement is that by adjusting the number of feature channels at different levels, the network can create a feature hierarchy, so that the shallow layer captures local and detailed information, while the deep layer integrates these information and learns more abstract representations, because noise and signal features will be exhibited at different scales.

[0073] (3) Using upsampling instead of transposed convolution to reduce noise introduction, upsampling maintains the authenticity of the original signal in the denoising process by amplifying the repeated signal;

[0074] The Step3 includes:

[0075] Specifically, nearest neighbor interpolation is used instead of transposed convolution, which copies the nearest feature values from the original to the new location. This is different from transposed convolution, which requires a fixed replacement rule to be defined in advance. Nearest neighbor interpolation provides a more direct method to restore the size of the feature, making the feature match the original input; upsampling is completed through convolution operation, where the weights of the convolution kernel are set to the interpolation mode. This method allows the network to learn the optimal interpolation weight during training. The interpolation can be represented as:

[0076] upsampling(x(n))=x(n*f)

[0077] where f is the upsampling factor. The upsampling layer can adjust the reconstruction strategy through learning, so the network can adjust the way of upsampling according to the characteristics of ECG signals and noise patterns, thereby better restoring the details and quality of the signal. This means that the network can decide how to most effectively upsample based on the data itself, rather than relying on a fixed rule. Transposed convolution requires a replacement rule to be defined in advance, which to some extent limits the flexibility and adaptability of the model, and increases the computational cost. The advantage of using nearest neighbor interpolation instead of transposed convolution is its directness and flexibility, which allows the model to more accurately and adaptively restore the spatial dimensions of ECG signals, amplify the values of the repeated original signal, prevent overfitting, and achieve better results in the denoising task.

[0078] (4) After the jump connection, the filter is added to help restore the original shape and characteristics of the signal, thereby improving the quality of the reconstructed signal of the decoder.

[0079] During the training process, the parameters of the filter are adjusted according to the loss function to reduce the difference between the reconstructed signal and the original signal, thereby improving the denoising performance. Therefore, after the jump connection, the filter can be used to enhance the key features of the ECG signal while suppressing the unimportant noise components. The parameters of the filter can be optimized during the training process, so that the model can learn the optimal denoising strategy. This end-to-end optimization ensures that the filter and the entire network structure work together to achieve the best denoising effect. By precisely adjusting the frequency components of the signal, the filter helps to restore the original shape and characteristics of the signal, thereby improving the usability of the denoised signal.

[0080] In another aspect, the present application provides an ECG signal denoising system based on an asymmetric convolution diffusion network, comprising a module for executing the method of the first aspect described above.

[0081] The asymmetric convolution network model in an ECG signal denoising system based on an asymmetric convolution diffusion network can specifically include:

[0082] Encoder: The encoder part gradually extracts features and reduces the spatial dimension of the features through a series of convolutional layers and pooling layers. The main components of the encoder are as follows:

[0083] Input layer: receives the original ECG signal, which has a shape of (None, 6000, 1), where 6000 represents the signal length per minute and 1 represents the number of signal channels. This is specially designed for network processing one-dimensional ECG time series data.

[0084] Convolutional layer: the first convolutional layer Conv1D has 16 convolutional kernels with a size of 3, and the output shape is (None, 6000, 16), which is used to capture the primary features in the input signal. The ReLU activation function ensures non-linear feature activation; the second convolutional layer Conv1D_1 has 32 convolutional kernels to extract more complex features, and again uses the ReLU activation function and'same' padding, with an output shape of (None, 6000, 32).

[0085] Pooling Layer: The first max pooling layer Max_pooling1D reduces the spatial dimension of the features by half to (None, 3000, 16), from 6000 points to 3000 points. This not only reduces the computational load but also enhances the model's insensitivity to small changes. The second max pooling layer Max_pooling1D_1 further reduces the spatial dimension of the features by half to (None, 1500, 32), further abstracting and compressing the data to prepare for high-level feature extraction.

[0086] Encoder: By gradually reducing the size of the features, the encoder increases the depth of the features, which helps the model capture abstract features in the ECG signal.

[0087] Bridge: The bridge layer is the connecting part between the encoder and the decoder, which continues to extract deeper features in the signal.

[0088] Bridge Conv1D Layer: Conv3 layer is the core of the bridge part, which uses 64 filters to further extract features. The features at this level have a high level of abstraction, providing a foundation for signal denoising and feature reconstruction.

[0089] Decoder: The decoder part gradually restores the spatial resolution of the features through upsampling layers and convolutional layers, and reconstructs the detailed information of the signal.

[0090] Upsampling Layer: The first upsampling layer Up_sampling1D increases the size of the features to (None, 3000, 64) to prepare for detail reconstruction; the second upsampling layer Up_sampling1D_1 further increases the size to (None, 6000, 32).

[0091] Convolutional Layer: The following convolutional layers Conv1D_4 and Conv1D_5 continue to process the features, with outputs of (None, 3000, 32) and (None, 6000, 16) respectively; the final convolutional layer Conv1D_5 produces a single-channel output (None, 6000, 1), corresponding to the denoised ECG signal. The final output is a feature with the same size as the original signal.

[0092] Output Layer: The last layer is the output layer, which converts multi-channel features to single-channel denoised signals through a 1-filter convolution. The Sigmoid activation function is used to limit the output to between 0 and 1.

[0093] During model compilation, the Adam optimizer is used, which is an adaptive learning rate optimization algorithm that adjusts the learning rate to improve model performance. The loss function used is Mean Squared Error (MSE), which directly quantifies the difference between the denoised signal and the original signal.

[0094] The network described in the present application is as Figure 2 The model includes a backbone network and three functional improvements to enhance the ability of electrocardiosignal denoising. Each convolutional layer in the model uses one-dimensional convolution instead of two-dimensional convolution, which is more suitable for processing one-dimensional electrocardio data; different levels correspond to different channel numbers to remove different frequency band noises; in addition, the model uses filters in the connection part to optimize signal processing; the upsampling process is also specially designed to ensure the accuracy of signal reconstruction.

[0095] For Figure 4 The time domain analysis chart of the signal before and after denoising of the present application shows the time domain comparison between the original electrocardiogram (ECG) signal and the signal after denoising. From the waveform analysis, it can be seen that the blue line represents the original electrocardio signal, which has typical ECG waveform and sharp QRS complex, and the orange dotted line represents the denoised signal, which has much smaller amplitude than the original signal, but retains the basic shape and rhythm of the original signal. In terms of noise level, the noise existing in the original signal leads to large signal fluctuation, while the denoised signal is relatively smooth, which shows that the denoising process effectively eliminates noise without introducing too much distortion or changing the basic characteristics of the signal. Although the amplitude of the denoised signal is reduced, the important QRS complex and other features are still identifiable. From the signal amplitude change, it can be seen that the amplitude of the signal may be compressed during the denoising process, which is common in ECG signal processing, the purpose is to reduce the influence of noise on the signal, but without excessive compression, avoiding the loss of important information. In terms of frequency retention, the denoised signal is smoother, but the basic frequency characteristics of the signal remain unchanged, which means that the basic rhythm of the heartbeat is not distorted by the denoising process. During the entire time domain signal, the denoised signal shows good continuity without appearing broken or abnormal fluctuations, and the denoising algorithm shows smoothness and continuity in processing the signal without obvious distortion area, which shows the stability and consistency of the denoising algorithm.

[0096] For Figure 5For the spectral comparison of the original and denoised ECG signals, the figure shows the power spectral density (PSD) of the original and denoised ECG signals. The PSD reflects the energy distribution of the signals at different frequencies. From the frequency distribution, it can be seen that the two signals are very similar at the beginning of the low-frequency region, indicating that the denoising process preserves the low-frequency components of the ECG signal. These low-frequency components usually contain most of the energy and information of the ECG signal; in the mid-frequency feature (10 to 30 Hz) frequency range, the denoised signal closely follows the changes of the original signal, indicating that important ECG information is preserved; in the high-frequency component, the PSD of the denoised signal is significantly lower than that of the original signal, indicating that the denoising algorithm effectively removes the high-frequency noise component, which may come from power lines, equipment interference or electrode friction, etc. By analyzing the spectral smoothness, overall, the spectrum of the denoised signal is smoother than that of the original signal, which shows that the denoising process effectively removes random noise without introducing additional pseudo-frequencies or interference. The denoising algorithm effectively reduces high-frequency noise while preserving low-frequency ECG features. Overall, the denoising algorithm smooths the spectrum of the signal, which means that the noise is well suppressed. However, the orange dotted line also shows some irregular fluctuations, which may be due to the over-compression of some parts of the signal during the denoising process.

[0097] The specific embodiments of the present application are described in detail above with reference to the accompanying drawings, but the present application is not limited to the above-described embodiments, and various changes can be made within the knowledge of those skilled in the art without departing from the spirit of the present application.

Claims

1. A method for denoising electrocardiogram signals based on an asymmetric convolutional diffusion network, characterized by: The specific steps of the method are as follows: Step 1: ECG signal preprocessing, including standardization and format conversion; Step 2: Gradually add different levels of random Gaussian white noise to the preprocessed ECG signal until the original signal is covered by the noise; Step 3: Backpropagation: Using the original signal as the target function, the model trains on the noisy ECG signal and the corresponding clean signal. In this process, the features of the ECG signal are extracted, the ECG signal is reconstructed based on the extracted features, and finally the denoised ECG signal is generated. In Step 3, back propagation starts with the noise data obtained in Step 2. , which is the result of the forward diffusion process; at each time step t, the goal of the model is to predict the Noise , and use this prediction to reconstruct , achieved by the following steps: Noise Prediction: Using a Neural Network ( ) to predict the noise , the noise that this network adds to the data at the current step is expressed as: ( ); Data reconstruction: Use the predicted noise to reconstruct the data of the previous step is a predefined noise level system; ; This iterative process iterates from t=T to t=0. In each step, the model uses the predicted noise to remove the noise in the current data and gradually reconstructs a clearer data state. After multiple steps of iteration, the model finally reconstructs a state close to the original data. Clear data; The Step 3 includes: (1) Use a one-dimensional convolution layer to slide the convolution kernel on the time axis to capture the local features of the ECG signal; (2) An asymmetric convolutional network architecture is used, and the jump connection combines the feature map in the encoder with the feature map in the decoder; the asymmetric structure optimizes this process by selectively fusing features; (3) Use upsampling instead of transposed convolution to reduce noise introduction. Upsampling maintains the authenticity of the original signal during the denoising process by amplifying the repeated signal; (4) Adding a filter after the jump connection helps to restore the original form and characteristics of the signal, thereby improving the quality of the decoder's reconstructed signal.

2. The ECG signal denoising method based on an asymmetric convolutional diffusion network according to claim 1, characterized in that: The process in Step 2 is a diffusion process. The diffusion process simulates the degradation process of data by gradually adding noise to the data. This process is described as a series of discrete steps, each step adding a certain amount of noise until the data becomes almost completely noise. Suppose there is an initial data of , this data is clear at the beginning of the diffusion process, and in each step t, Add noise to create new data , expressed as: ； in, is random noise sampled from a standard normal distribution, is the data state of the previous step, is a predefined noise level factor that determines how much noise is added at each step, is chosen to ensure a gradual increase in noise, ultimately making That is, the data in the last step is close to pure noise; the entire diffusion process consists of multiple time steps. In the initial step, the noise added is relatively small, but as the steps proceed, the noise gradually increases.

3. The ECG signal denoising method based on an asymmetric convolutional diffusion network according to claim 1, characterized in that: Said (2) includes: adjusting the network's convolution kernel size, filter shape, channel number allocation, and upsampling step size according to the characteristics of the ECG signal and the denoising requirements; In the encoder, the number of channels is gradually increased to capture deep features, and in the decoder, the number of channels is gradually reduced to merge features and reduce noise. The asymmetric convolutional network architecture optimizes the feature fusion process by selectively fusing the features of the encoder and decoder, and optimizes the spatial resolution of the signal by adjusting the upsampling step size. By adjusting the number of feature channels at different levels, the asymmetric convolutional network architecture can simultaneously learn multi-scale features; using a different number of channels at each stage of the asymmetric convolutional network architecture introduces a deep supervision mechanism; this indicates that each level of the network can be trained to minimize the loss of the corresponding level of the asymmetric convolutional network architecture, thereby improving the overall denoising performance.

4. The ECG signal denoising method based on an asymmetric convolutional diffusion network according to claim 1, characterized in that: Said (3) includes: Specifically, nearest neighbor interpolation is used instead of permutation convolution to copy the nearest feature value in the original to a new location. Nearest neighbor interpolation provides a more direct way to restore the size of the feature so that the feature matches the original input; upsampling is done through convolution operation, in which the weights of the convolution kernel are set to interpolation mode. This method allows the network to learn the best interpolation weights during training.

5. The ECG signal denoising method based on an asymmetric convolutional diffusion network according to claim 1, characterized in that: Said (4) includes: during the training process, the parameters of the filter are adjusted according to the loss function to reduce the difference between the reconstructed signal and the original signal.

6. An electrocardiogram signal denoising system based on an asymmetric convolutional diffusion network, characterized in that: The method comprises a module for executing the method according to any one of claims 1 to 5.

Images