Electrocardiosignal artifact elimination method and system based on dynamic channel weighting

Abstract

The invention discloses an electrocardiosignal artifact elimination method and system based on dynamic channel weighting, and the method comprises the steps: a data collection and preprocessing stage: collecting original ECG physiological signal data, and building a data set suitable for subsequent model training; a network model construction stage: designing a network architecture with a dynamic channel weighting mechanism for electrocardiosignal motion artifact confrontation elimination; a model training stage: training the model by using the constructed data set; in the practical application stage, ECG signals to be processed are input into the trained model, artifact filtering is achieved through multi-stage feature processing, and purified ECG signals are output. The system comprises a data set construction unit, a model construction unit, a model training unit and an artifact elimination unit. By using the method and the device, artifacts in the electrocardiosignals can be effectively eliminated. The method can be widely applied to the field of signal processing.

Description

Technical Field

[0001] The present invention relates to the field of signal processing, and in particular to a method and system for eliminating electrocardiogram (ECG) signal artifacts based on dynamic channel weighting. Background Art

[0002] Dynamic and continuous monitoring of ECG signals is of great value in early warning of cardiovascular diseases, diagnosis of neurological diseases, and health management. With the rapid development of flexible electronic technology and wearable devices, medical monitoring scenarios are extending from clinical environments to daily life scenarios. However, motion artifacts caused by daily human activities (such as walking and gestures) seriously interfere with signal quality. Their spectrum highly overlaps with the target physiological signal, making it difficult for traditional filtering methods to distinguish between effective signals and noise. According to industry reports, signal distortion caused by motion artifacts has become the primary factor reducing the data availability of wearable medical devices, directly affecting the reliability of core functions such as atrial fibrillation screening and sleep apnea monitoring.

[0003] Traditional motion artifact elimination technologies are mostly implemented using adaptive filtering methods. Adaptive filtering uses reference signals such as accelerometers to estimate noise, but is limited by linear assumptions and has limited effect on eliminating complex nonlinear coupling interference in actual scenarios (such as dynamic changes in electrode contact impedance). Summary of the Invention

[0004] In view of this, in order to solve the technical problem that existing artifact removal methods are limited by linear assumptions and cannot complete the nonlinear coupling separation of non-stationary motion artifacts and physiological signals, in a first aspect, the present invention proposes an ECG signal artifact removal method based on dynamic channel weighting, the method comprising the following steps:

[0005] In the data acquisition and preprocessing stage, raw ECG physiological signal data is collected and a data set suitable for subsequent model training is established;

[0006] During the network model construction phase, a network architecture with a dynamic channel weighting mechanism was designed to combat motion artifacts in ECG signals.

[0007] In the model training phase, the model is trained using the constructed dataset;

[0008] In the actual application stage, the ECG signal to be processed is input into the trained model, and artifacts are filtered out through multi-level feature processing to output the purified ECG signal.

[0009] The model includes a U-shaped architecture generator and a discriminator. The U-shaped architecture generator is used to perform multi-level downsampling feature extraction, and a dynamic parameter freezing mechanism triggered by the QRS wave position is embedded in the upsampling stage. The time domain branch of the discriminator based on time-frequency domain input is used to capture local waveform distortion features, and the frequency domain branch is used to analyze amplitude anomalies to achieve cross-domain collaborative identification of nonlinear characteristics of motion artifacts.

[0010] In a second aspect, the present invention further proposes an electrocardiogram signal artifact removal system based on dynamic channel weighting, the system comprising:

[0011] A data set construction unit, used to collect raw ECG physiological signal data and establish a data set suitable for subsequent model training;

[0012] Model building unit, designing a network architecture with a dynamic channel weighting mechanism for anti-motion artifact elimination of ECG signals;

[0013] A model training unit, used to train the model using the constructed data set;

[0014] The artifact removal unit is used to input the ECG signal to be processed into the trained model, filter out artifacts through multi-level feature processing, and output the purified ECG signal.

[0015] Based on the above scheme, the present invention provides a method and system for removing ECG signal artifacts based on dynamic channel weighting, which performs targeted suppression on noise characteristics of different frequency bands through multi-scale signal fusion and hierarchical processing mechanism. The model simultaneously analyzes high-frequency transient interference, medium-frequency motion artifacts and low-frequency drift, significantly improving signal quality in dynamic scenarios. The time-frequency dual-stream discrimination module further identifies residual noise from two dimensions: waveform coherence and spectral distribution, ensuring that the denoised signal is close to the real physiological characteristics in both the time domain and the frequency domain. The present invention designs multiple protection strategies for clinical core features such as QRS complex and ST segment in ECG signals. Through real-time waveform detection and dynamic gradient control, key areas are actively locked and parameter updates are suspended during the denoising process to avoid waveform distortion or over-smoothing problems common in traditional algorithms. In addition, the attention mechanism guides the model to prioritize signal components related to pathological diagnosis (such as P wave morphology and T wave slope) to ensure the accuracy of the output signal. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a flowchart of the steps of a method for eliminating ECG signal artifacts based on dynamic channel weighting according to the present invention;

[0017] Figure 2 This is a structural block diagram of the signal motion artifact elimination generator model of the present invention;

[0018] Figure 3This is a block diagram of the structure of the up- and down-sampling convolution block of the present invention;

[0019] Figure 4 This is a structural block diagram of the discriminator of the present invention.

[0020] Figure 5 This is a structural block diagram of the time domain convolution block of the present invention;

[0021] Figure 6 This is a structural block diagram of an electrocardiogram signal artifact elimination system based on dynamic channel weighting according to the present invention; DETAILED DESCRIPTION

[0022] In recent years, deep learning has shown potential in the field of signal denoising. Supervised learning methods can capture complex noise patterns by training models with "clean-noisy" paired data, but the cost of obtaining high-quality labeled data is high, and models trained with static data sets are difficult to adapt to individual differences and the diversity of movement patterns. Generative adversarial networks (GANs) alleviate the data dependency problem through unsupervised learning, but traditional frameworks have defects such as over-smoothing of generated signals and distortion of key physiological waveforms (such as QRS complexes). At the same time, the contradiction between model complexity and resource constraints of embedded devices is prominent: medical-grade algorithms must meet the rigid requirements of low latency and low power consumption, while existing deep learning models generally have computational redundancy, making it difficult to achieve real-time processing on wearable hardware.

[0023] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0024] It should be noted that, for ease of description, only the parts related to the invention are shown in the drawings. In the absence of conflict, the embodiments and features in the embodiments of this application can be combined with each other.

[0025] It should be understood that the terms "system," "device," "unit," and / or "module" used in this application are a method for distinguishing different components, elements, parts, portions, or assemblies at different levels. However, if other terms can achieve the same purpose, the terms may be replaced by other expressions.

[0026] As used in this application and the claims, unless the context clearly indicates an exception, the terms "a," "an," "an," and / or "the" are not intended to refer to the singular and may include the plural, unless the context clearly indicates otherwise. Generally speaking, the terms "comprises" and "include" only indicate the inclusion of the steps and elements specifically identified, and these steps and elements do not constitute an exclusive list. A method or apparatus may also include other steps or elements. The phrase "comprises a..." does not preclude the presence of additional identical elements in the process, method, product, or apparatus that includes the elements.

[0027] In the description of the embodiments of this application, "plurality" refers to two or more than two. The terms "first" and "second" below are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of the technical features indicated. Therefore, features specified as "first" or "second" may explicitly or implicitly include one or more of the features.

[0028] In addition, flow charts are used in this application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, the steps may be processed in reverse order or simultaneously. Furthermore, other operations may be added to these processes, or one or more operations may be removed from these processes.

[0029] Reference Figure 1 , which is a flow chart of an optional example of a method for removing ECG signal artifacts based on dynamic channel weighting proposed by the present invention. The method can be applied to a computer device. The artifact removal method proposed in this embodiment may include but is not limited to the following steps:

[0030] Step S1, acquiring ECG signals and constructing a data set;

[0031] Step S2: constructing an ECG signal motion artifact countermeasure elimination network based on dynamic channel weighting to obtain an elimination model;

[0032] Step S3: training the elimination model based on the ECG signal dataset;

[0033] Step S4: Input the ECG signal to be measured into the trained elimination model, and complete the artifact elimination through multi-level feature extraction and feature fusion combined with dynamic parameter freezing.

[0034] In some feasible embodiments, step S1 specifically includes:

[0035] S1.1. Acquire ECG signal;

[0036] In this embodiment, an electrocardiogram device is used to collect ECG signals, wherein the collected ECG signals include normal and arrhythmia ECG signals, and the data sampling frequency is 360 Hz.

[0037] S1.2. Apply median filtering to eliminate low-frequency baseline drift caused by breathing or electrode movement, and avoid overall waveform offset;

[0038] S1.3. Use a low-pass filter to remove high-frequency components such as myoelectric noise, retaining the high-frequency details of the QRS complex;

[0039] S1.4. Cut the continuous signal into 5-second non-overlapping segments (corresponding to 5000 points at 1000 Hz sampling). Calculate the mean (μ) and standard deviation (σ) of each segment independently and perform z-score normalization:

[0040]

[0041] Limit the normalized value range to [-3,3] to avoid interference from extreme noise points;

[0042] S1.5. Multi-channel data generation and concatenation, including: 1) Original signal: retain the original sampling rate; 2) Downsampled signal: downsample to 250Hz and then linearly interpolate to align the length, emphasizing the low- and medium-frequency components; 3) Envelope signal: full-wave rectify the original signal (absolute value), then obtain the envelope through a 5Hz low-pass filter to characterize the baseline drift (0.1-5Hz). The above three channels of data are concatenated along the channel dimension to form a 3×5000 input tensor, X3=[x origin ,x d ,], covering different frequency band characteristics;

[0043] S1.6. Synchronously acquire clean ECG signals at rest, generate reference labels through wavelet threshold denoising combined with template matching, and use the dynamic time warping (DTW) algorithm to ensure phase alignment between the noisy signal and the clean label waveforms.

[0044] S1.7. Dynamically add real motion artifacts to the clean signal, with the signal-to-noise ratio (SNR) randomly set from -10dB to 10dB to simulate interference under different activity intensities; locally stretch / compress the signal (amplitude ±5%) to enhance the model's adaptability to heart rate fluctuations, and flip the signal polarity with a 10% probability to improve robustness to poor lead contact.

[0045] S1.8. After the above processing, a data set is generated.

[0046] The ECG signal of this embodiment is a 1D sequence with a total of 650,000 sampling points, which is divided into 130 segments on average, with each segment having 5,000 sample points. Each segment serves as the input of the model, and the output is a set of 5,000×1 waveforms.

[0047] In some possible embodiments, such as Figure 2 As shown in Figure 1, the elimination model includes a generator and a discriminator. The U-shaped architecture generator consists of an encoder and a decoder. A jump connection is used between the encoder and the decoder to fuse the low-level features of the encoder with the high-level features of the decoder.

[0048] The encoder and decoder of the generator both use 4-layer network layers; each layer of the encoder's 4-layer network uses depth-wise separable convolution; each layer of the decoder's 4-layer network uses transposed convolution; the input of each layer of the decoder is the fusion of the output features of the previous layer and the corresponding jump connection output features.

[0049] The ECG signal X is used as the input of the depth separation convolution of the first layer of the encoder. The convolution output feature of the first layer of the network is Then the features They are respectively used as the input of the separable convolution of the second network layer and the first skip connection; they are processed in sequence by multiple separable convolutions and a dynamic channel attention module.

[0050] For the i-th layer encoder:

[0051]

[0052] Where DC represents the depth convolution operation, PC represents the point-by-point convolution operation, and F i-1 Output feature of the previous layer, dimension is (B,C in ,L), B is the batch size, L is the sequence length; is the depth convolution kernel of the c-th channel, size (k, 1), k is the convolution kernel size; is the c-th channel bias term; is a point-by-point convolution kernel with a size of (C in ,C out ). The mathematical expression of dynamic channel attention is as follows:

[0053]

[0054] a c =σ(W2·δ(W1·z c +b1)+b2),W1∈R C×C / r , W2∈R C / r×C

[0055]

[0056] Among them, z c represents the global average pooling result of the c-th channel, δ is the Relu activation function, σ is the Sigmoid function, and the output channel weight α c ∈[0,1].

[0057] In this embodiment, the transposed convolution block of the 4th layer of the decoder is based on the output feature of the 4th network layer of the encoder. And the output features of the dynamic channel attention module The fusion is taken as input; then the convolution of the 3rd layer of the decoder is based on the output features of the 4th layer of the decoder And the output features of the 3rd b encoder As input, and so on, it is finally processed by the transposed convolution block of the first layer of the decoder and then output. The decoder decodes the spatial and temporal features of the ECG signal through convolution blocks and skip connections; and generates the ECG waveform through the Tanh function.

[0058] The decoder uses the encoded features and the details of the jump connection to gradually reconstruct the signal. The calculation formula is as follows: For the i-th layer decoder:

[0059]

[0060] Among them, CT represents the transposed convolution operation, Represents the transposed convolution kernel, with a size of (k,C in ,C out ),F i+1 Represents the output of the previous layer decoder.

[0061] The mathematical expression of the skip connection feature fusion process can be expressed as:

[0062]

[0063] in, is the weighted feature of the DCAM of the same-layer coding path, The convolution kernel is fused for skip connection, and the size is (1,2C,C). The signal reconstruction output expression is:

[0064]

[0065] Where W out is the output convolution kernel, with a size of (1, C, 1), is the denoised ECG signal with a dimension of (B, 1, L). The loss function constraint is:

[0066]

[0067] Where y is the clean ECG signal and λ is the frequency domain loss weight.

[0068] In this embodiment, Figure 3 As shown in the figure, the encoder consists of a depthwise separable convolutional layer and a LeakyReLU activation layer. The convolutional layer is used to extract time domain features from the ECG signal and capture local patterns; the LeakyReLU activation layer introduces nonlinear features to enhance the model's expressiveness. The decoder consists of four decoding layers and a QRS freeze layer, each of which consists of a transposed convolutional layer and a ReLU activation layer. The transposed convolutional layer implements the upsampling operation, and its core function is to gradually restore the low-resolution feature map to the high-resolution space. The functions of the other layers are consistent with those of the corresponding layers of the decoder.

[0069] In this embodiment, Figure 4 As shown, the discriminator consists of a time domain discriminant branch and a frequency domain discriminant branch. The time domain discriminant branch consists of four convolution blocks, a global pooling layer, and a fully connected layer. Each convolution block consists of a 1D convolution layer and spectral normalization. The frequency domain discriminant branch consists of one STFT layer, two complex convolution layers, one global average pooling layer, and a fully connected layer. The feature extraction mathematical expression of the time domain discriminant branch is as follows:

[0070]

[0071] Where f represents the activation function LeakyReLU, is the feature of the previous layer, and the first layer input is the original signal y or the generated signal is the time domain convolution kernel of the i-th layer, size (k,C in ,C out ); α is the negative slope coefficient of LeakyReLU. The time domain discrimination output is:

[0072]

[0073] in represents the final convolution kernel, with a size of (1, C, 1), σ is the Sigmoid function, and outputs the time domain authenticity probability ∈ [0, 1].

[0074] The calculation process of its frequency domain discrimination branch can be expressed as follows:

[0075] 1) Short-time Fourier transform

[0076] X=STFT(x),X∈C N×T

[0077]

[0078] The mathematical expression of STFT is

[0079]

[0080] Where X[n,k] is a complex number, representing the time-frequency distribution of the signal at time frame n and frequency k; x[m] represents the input time domain signal (discrete sampling points) with a length of L; w[mn] is a sliding window function with a window length of M; n is the time frame index, n∈[0,N-1], k is the frequency index, k∈[0,M / 2], R is the frame shift. In the discriminator, for the input signal x∈R L :

[0081] X real =Re(X[n,k]),X imag =Im(X[n,k])

[0082] Concatenate into a 2-channel tensor:

[0083] F stft =[X real ,X imag ]∈R 2×N×K

[0084] in, is the number of valid frequency points.

[0085] 2) Frequency domain convolution processing

[0086]

[0087] in is the frequency domain convolution kernel, size is (k f ,k t ,C in ,C out ),k f =3×3. The frequency domain discrimination output is:

[0088]

[0089] 3) Adaptive output

[0090] D(x)=βD t (x)+(1-β)D freq (x)

[0091] Where β is a learnable fusion parameter.

[0092] The encoder of the generator compresses multi-scale input into low-dimensional features and filters key information through the attention mechanism. The decoder uses the encoded features and the details of the jump connection to gradually reconstruct the signal; the discriminator based on time-frequency domain data input generates fake samples. The two are optimized through adversarial training to achieve a balance between noise suppression and waveform fidelity.

[0093] This embodiment significantly reduces model complexity and computational overhead by employing depthwise separable convolution and dynamic computation optimization techniques. The U-shaped structure achieves efficient feature reuse through skip connections, avoiding redundant computation.

[0094] In some feasible embodiments, step S3 specifically includes:

[0095] S3.1. Construct a three-channel input (original signal, downsampled intermediate frequency signal, and enveloped low frequency signal), which is concatenated into a 3 × 5000 tensor after segmented z-score normalization and dynamic truncation. The labels are the clean ECG signals annotated by experts to ensure input alignment and noise distribution adaptation for model learning.

[0096] S3.2. Initialize the generator (multi-scale encoding and decoding + dynamic attention) and the discriminator (time-frequency dual-stream identification), define the adversarial loss, cycle consistency loss and clinical feature constraint loss, set the dynamic weight (λ adv =0.5 increments, λ C =1.0,λ cycle =1.0 attenuation) balances confrontation and fidelity goals;

[0097] S3.3. Phased training strategy

[0098] S3.3.1. Warm-up (1-10 rounds): Train only the generator, freeze the discriminator, and strengthen the basic reconstruction capability;

[0099] S3.3.2, Confrontation (11-180 rounds): Alternately optimize G / D and gradually increase λ adv To 0.8, activate QRS freeze to protect critical waveforms;

[0100] S3.3.3, fine-tuning (181-200 rounds): Freeze the generator, sharpen the discriminator's recognition of frequency domain details, and reduce the learning rate to 1e-5.

[0101] This embodiment combines an improved adversarial training framework with a progressive learning strategy to effectively balance the optimization rhythm of the generator and the discriminator, thereby alleviating the mode collapse problem in traditional adversarial training.

[0102] In some feasible embodiments, in step S3, during the training process, the loss function of the elimination model is composed of adversarial loss, cycle consistency loss, and clinical feature constraint loss, where:

[0103] The adversarial loss formula is:

[0104]

[0105] The cycle consistency loss is:

[0106] L C =|D(x)-G d (G(D(x)))|1

[0107] Among them G d Represents the auxiliary noise generator, which is used to constrain the denoising signal to not lose the original information.

[0108] The frequency domain matching loss is:

[0109]

[0110] The clinical feature constrained loss is:

[0111]

[0112] Where λ1 and λ2 are balance weights, QRS pos (y) is the QRS wave position detection function, ST slope Represents the ST segment slope calculation function.

[0113] The total loss function is:

[0114] L total =λ adv L adv +λ C L C +λ freq L freq +λ Clin L clin

[0115] where λ adv ,λ C ,λ freq and λ Clin is the weight coefficient of each loss.

[0116] This embodiment enhances the model's adaptability to unknown noise types and individual signal differences by introducing multi-dimensional consistency constraints.

[0117] In some feasible embodiments, step S4 specifically includes:

[0118] S4.1, encoder downsampling operation:

[0119] S4.1.1, the one-dimensional ECG signal is used as the input of the convolutional block of the first layer of the encoder;

[0120] S4.1.2. The convolution layer of the first layer of the encoder first performs a convolution operation (the convolution kernel is 5×1), and then passes through the activation function LeakyReLU to output the feature f ec1 ;

[0121] S4.1.3, the convolutional block of the second layer of the encoder will f ec1 As input, the convolution operation is first performed (the convolution kernel is k2×1), and then the activation function LeakyReLU is used to output the feature f ec2 ;

[0122] S4.1.4, the convolutional block of the third layer of the encoder will f ec2 As input, the convolution operation is first performed (the convolution kernel is k3×1), and then the activation function LeakyReLU is used to output the feature f ec3 ;

[0123] S4.1.5, the convolutional block of the 4th layer of the encoder will f ec3 As input, the convolution operation is first performed (the convolution kernel is k4×1), and then the activation function LeakyReLU is used to output the feature f ec4 ;

[0124] S4.1.6, dynamic channel attention module will f ec4 As input, the output feature f aout ;

[0125] S4.2, decoder upsampling

[0126] S4.2.1. Feature f aout and feature f e4 Add the features f by element e4a , f e4a As the input of the 4th layer convolution block of the decoder, the output feature f d4 ;

[0127] S4.2.2, Feature f e3 and feature f d4 Add the features f by element e3d4 , f e3d4 As the input of the decoder's third convolutional block, the output feature f d3 ;

[0128] S4.2.3, Feature f e2 and feature f d3 Add the features f by element e2d3 , f e2d3 As the input of the second convolution block of the decoder, after the activation function outputs the feature f d2 ;

[0129] S4.2.4, Feature fe1 and feature f d2 Add the features f by element e1d2 , f e1d2 As the input of the first convolutional block of the decoder, the denoised ECG data is output after passing through the activation function;

[0130] S4.3. Discriminator Decision

[0131] S4.3.1 The denoised ECG data is input into the first layer of the time domain discriminant branch and the frequency domain discriminant branch of the discriminator, and the output features f are respectively d1t , f d1f ;

[0132] S4.3.2 Feature f d1t , f d1f Input the second layer network of the time domain discrimination branch and the frequency domain discrimination branch respectively, and output the feature features f d2t , f d2f ;

[0133] S4.3.3 Feature f d2t , f d2f Input the third layer network of the time domain discrimination branch and the frequency domain discrimination branch respectively, and output the feature features f d3t , f d3f ;

[0134] S4.3.4 Feature f d3t , f d3f Input the 4th layer network of the time domain discrimination branch and the frequency domain discrimination branch respectively, and output the feature features f d4t , f d4f ;

[0135] S4.3.5 Feature f d4t Input the 5th layer network of the time domain discrimination branch and output the feature features f d5t ;

[0136] S4.3.6 Feature f d5t and f d4f Fusion, calculation of comprehensive scores, and decision on whether the artifact-free ECG signal passes.

[0137] In some feasible embodiments, the following further comprises:

[0138] Validation of the elimination model;

[0139] The training and test sets were constructed using noise signals from the MIT-BIH Arrhythmia Database (AHDB) and the MIT-BIH Noise Stress Test Database (NSTDB). Cosine similarity (CS), output signal-to-noise ratio (SNR), and root mean square error (RMSE) were used as evaluation indicators for verification. The experimental results show that the method of this embodiment performs well in the task of removing motion artifacts from ECG signals, as shown in Table 1.

[0140] Table 1 Comparison of denoising performance of different methods under various input signal-to-noise ratios

[0141]

[0142] like Figure 6 As shown, a system for removing electrocardiogram signal artifacts based on dynamic channel weighting includes:

[0143] A data set construction unit, configured to execute step S1;

[0144] A model building unit, configured to execute step S2;

[0145] A model training unit, configured to execute step S3;

[0146] The artifact removal unit is configured to execute step S4.

[0147] The contents of the above method embodiments are all applicable to the present system embodiments. The functions specifically implemented by the present system embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0148] An electrocardiogram signal artifact removal device based on dynamic channel weighting:

[0149] at least one processor;

[0150] at least one memory for storing at least one program;

[0151] When the at least one program is executed by the at least one processor, the at least one processor implements the above-mentioned method for eliminating ECG signal artifacts based on dynamic channel weighting.

[0152] The contents of the above method embodiments are all applicable to the present device embodiments. The functions specifically implemented by the present device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0153] A storage medium stores instructions executable by a processor, wherein the instructions executable by the processor are used to implement the above-mentioned method for eliminating electrocardiogram signal artifacts based on dynamic channel weighting when executed by the processor.

[0154] The contents of the above method embodiments are all applicable to the present storage medium embodiment. The functions specifically implemented by the present storage medium embodiment are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0155] The above is a specific description of the preferred implementation of the present invention, but the invention is not limited to the embodiments. Those skilled in the art can make various equivalent modifications or substitutions without violating the spirit of the present invention. These equivalent modifications or substitutions are all included in the scope defined by the claims of this application.

Claims

1. A method for removing electrocardiogram (ECG) signal artifacts based on dynamic channel weighting, characterized in that: The following steps are involved: Acquire ECG signals and construct datasets; Construct an ECG signal motion artifact elimination network based on dynamic channel weighting and obtain an elimination model; Training the elimination model based on the ECG signal data set to obtain a trained elimination model; The ECG signal to be measured is input into the trained elimination model, and artifact elimination is completed through multi-level feature extraction and feature fusion combined with dynamic parameter freezing.

2. The method for removing ECG signal artifacts based on dynamic channel weighting according to claim 1, characterized in that: The step of acquiring ECG signals and constructing a data set specifically includes: Acquire ECG signals; The ECG signal data is filtered, signal cut, multi-channel data generated and spliced, phase aligned, and artifact added to generate a data set.

3. The method for removing ECG signal artifacts based on dynamic channel weighting according to claim 1, characterized in that: The elimination model includes a U-shaped architecture generator and a discriminator, wherein: The U-shaped architecture generator includes an encoder and a decoder; The discriminator includes a time domain discrimination branch and a frequency domain discrimination branch.

4. The method for removing ECG signal artifacts based on dynamic channel weighting according to claim 3, characterized in that: Both the encoder and decoder consist of at least two layers of networks, where: For the i-th layer network of the encoder, the formula is expressed as follows: Among them, DC represents the depth convolution operation, PC represents the point-by-point convolution operation, and F i-1 Output features for the previous layer; is the depth convolution kernel of the c-th channel; is the c-th channel bias term; is the point-by-point convolution kernel; For the i-th layer network of the decoder, the formula is as follows: Among them, CT represents the transposed convolution operation, represents the convolution kernel, F i+1 Represents the output of the previous layer decoder.

5. The method for removing ECG signal artifacts based on dynamic channel weighting according to claim 3, characterized in that: The expression of feature extraction of the time domain discriminant branch is as follows: Among them, f represents the activation function LeakyReLU, is the feature of the previous layer; is the time domain convolution kernel of the i-th layer; α is the negative slope coefficient of LeakyReLU; The expression of the discriminant output of the time domain discriminant branch is as follows: in, represents the final convolution kernel weight, Represents the last layer of time domain feature map; σ is the Sigmoid function.

6. The method for removing ECG signal artifacts based on dynamic channel weighting according to claim 3, characterized in that: The frequency domain discrimination branch includes a trainable STFT and a complex convolutional network.

7. The method for removing ECG signal artifacts based on dynamic channel weighting according to claim 1, characterized in that: During the training process, the loss function of the elimination model includes adversarial loss, cycle consistency loss, frequency domain matching loss and clinical feature constraint loss.

8. The method for removing ECG signal artifacts based on dynamic channel weighting according to claim 3, characterized in that: The step of inputting the ECG signal to be measured into the trained elimination model, performing multi-level feature extraction and feature fusion, and combining dynamic parameter freezing to complete the artifact elimination step specifically includes: Input the ECG signal to be measured into the trained elimination model; Perform multi-level downsampling feature extraction based on the U-shaped architecture generator, and embed a dynamic parameter freezing mechanism triggered by the QRS wave position in the upsampling stage; The time domain branch of the discriminator is used to capture local waveform distortion features, and the frequency domain branch is used to analyze amplitude anomalies to obtain cross-domain artifact identification results. Artifact elimination is completed according to the artifact cross-domain identification result.

9. An electrocardiogram signal artifact removal system based on dynamic channel weighting, characterized in that: include: A data set construction unit, which acquires ECG signals and constructs a data set; A model building unit, which builds an ECG signal motion artifact elimination network based on dynamic channel weighting to obtain an elimination model; A model training unit, which trains the elimination model based on the ECG signal data set to obtain a trained elimination model; The artifact elimination unit inputs the ECG signal to be measured into the trained elimination model, and completes the artifact elimination through multi-level feature extraction and feature fusion combined with dynamic parameter freezing.

10. An electrocardiogram signal artifact removal device based on dynamic channel weighting, characterized in that: include: at least one processor; at least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements the method for eliminating electrocardiogram signal artifacts based on dynamic channel weighting as described in any one of claims 1 to 8.

Images