Research On ECG Signal Denoising Technology Based On Residual Dense Network And U-Net

Recently,the incidence of cardiovascular disease in our country has been on the rise,mainly characterized by high recurrence rates,multiple complications and high mortality,which has brought a heavy burden to families and society.The electrocardiogram(ECG)can provide detailed information about the rhythm and function of the heart,and provide guidance for doctors to discover and track diseases,thereby ensuring the life and health of patients.However,the process of wearable ECG signals acquisition is susceptible to be contaminated by the environmental noise,which affects the accuracy of the waveforms.In addition,the existing denoising approaches cannot effectively remove various kinds of noises in ECG signals,which may lead to the loss of important waveform information.In view of this,the thesis will adopt deep learning methods such as residual dense network and U-Net to remove various noises in ECG signals.The main research contents of the thesis are as follows:(1)Aiming at the problem that existing denoising methods cannot effectively remove various noises in ECG signals and cannot fully extract the hierarchical features of ECG signals,a denoising method for ECG signals based on a multi-scale residual dense network is proposed.A dual-branch residual dense block composed of dense convolutional network and dilated convolutions is proposed,which realizes the adaptive extraction of local multi-scale features of ECG signals,and the local multi-scale features extracted by all modules can be adaptively fused and passed to all convolutional layers in subsequent dual-branch residual dense blocks,avoiding a large amount of feature stacking.Then,the local multi-scale features extracted by all dual-branch residual dense blocks are fused.Finally,based on residual learning,shallow features and multi-level features are fused,thereby realizing the extraction of multi-scale and multi-level features of ECG signals,and better preserve the waveform characteristics of the ECG signals.The proposed method is verified on the MIT-BIH database.Experiments show that this method can reduce the noises of ECG signals,while effectively retaining the detailed information of ECG signals,and the average signal-to-noise ratio of this method can reach 35.28 d B,which is significantly improved compared with the existing denoising methods.(2)A denoising method for ECG signals based on multi-scale residual dense U-Net is proposed.By integrating the dual-branch residual dense block into the up-sampling and down-sampling process of U-Net,the multi-scale features of ECG signals can be extracted under different receptive fields.Compared with the multi-scale residual dense network,this method needs a small number of parameters,and the model training is much faster.By reducing the size of the feature maps and increasing the feature dimensions of the hidden layer of the model,the down-sampling operation based on the dual-branch residual dense block speeds up the training of the model,and a better tradeoff can be realized between efficiency and effectiveness in exploiting the hierarchical features from all convolutional layers.In addition,by skip connection,the restored features from up-sampling are fused with the corresponding down-sampling features,and are transferred to the dual-branch residual dense block.This operation avoids the information loss that may be caused during down-sampling process,thereby capturing more accurate contextual and detailed information about the ECG signal.It has been verified by experiments that the waveforms obtained by the multi-scale residual dense U-Net are basically consistent with the waveforms of the clean signals,this method effectively retains the important waveform information of the ECG signals and has a good noise reduction performance.And the proposed method is superior to the existing denoising methods in terms of signal-to-noise ratio and root mean square error performance indicators.