A Study On Electrocardiogram Signal Denoising Method Based On Attention Mechanism And Generative Adversarial Network

12-lead electrocardiogram（ECG）is commonly used for assessing the heart’s health status by recording the electrical activity during the cardiac cycle.However,due to its high cost,complex operation,and inconvenience in carrying,it is not suitable for routine dynamic ECG monitoring.Holter ECG is often used for long-term ECG monitoring,but in complex dynamic environments,the collected ECG signals are easily affected by various types of noise,such as motion artifacts and baseline drift,resulting in poor data quality.Therefore,a set of single-lead ECG data collection system was designed in this paper with the aim to study noise reduction algorithm for ECG data acquisition,reduce motion artifact interference and have important implications for preventing cardiovascular diseases.The main works conducted in this paper include:（1）Designed a wearable ECG data collection system.The system consists of four parts:a power management module,an ECG acquisition module,a data processing module,and a delay switch module.It completes the collection of ECG signals and uses Bluetooth for real-time transmission.After testing and analyzing the collection system,the theoretical working time of the system was calculated during the power consumption test.It was found that the system can work continuously for 42 hours when fully charged,meeting the requirements of long-term monitoring.During signal collection tests in scenarios of sitting and slow walking,the ECG signals collected by the system met clinical diagnostic requirements,with clear QRS complex waves,P waves,and T waves.（2）Proposed an ECG denoising algorithm based on Variational Mode Decomposition（VMD）.VMD,Discrete Wavelet Transform（DWT）,and Non-Local Mean（NLM）are suitable for removing noise from a specific case of denoising ECG.This paper combines the advantages of these three algorithms by decomposing the noisy ECG signal into high-frequency and low-frequency components using VMD.The high-frequency components are filtered by DWT thresholding while the low-frequency components are filtered by NLM.The reconstructed signal can effectively denoise the ECG.Experimental results show that the proposed denoising method is significantly better than these comparative methods.Under Gaussian white noise with an input signal-to-noise ratio（SNR）of 5d B,the signal-to-noise ratio improvement(SNR_imp)of the denoised signal has increased by at least 20.92%.（3）Proposed a method for denoising electrocardiogram signals by combining the Efficient Channel Attention Network（ECA-Net）and Cycle Generative Adversarial Network（Cycle GAN）.To address noise such as motion artifacts and baseline drift that can occur during signal acquisition,a one-dimensional convolutional neural network was used to extract temporal features from ECG data.The ECA-Net module was embedded into the backbone network to highlight the key features of the ECG and suppress noise information,improving the denoising performance of the model.The loss function was optimized using the L1 norm and maximum difference function to capture global and local features of the signal and suppress maximum local errors,leading to better fitting of training data.On a open publicly available dataset,the SNR_imp after denoising single-type noise on average was 32.96%when the input SNR was 5d B,while that for mixed-type noise was 33.93%.The optimized denoising model has lower root-mean-square error（RMSE）and percentage root mean square difference（PRD）,indicating stronger generalization ability and denoising performance.The denoising model was used to remove motion artifacts from ECG signals collected by a monitoring system in dynamic motion environments.Experimental results show a significant improvement in the quality of the measured ECG signals after denoising.This paper presents a wearable ECG acquisition system and proposes an ECG denoising algorithm based on Variational Mode Decomposition,as well as an ECG denoising algorithm that combines the ECA-Net and Cycle GAN.The system and algorithm address the problems of inconvenient long-term dynamic ECG monitoring and inaccurate signal acquisition,which can meet the requirements for long-term ECG monitoring.The proposed system has important research significance for early prevention of heart disease and sudden cardiac arrest.