Research On ECG Signal Denoising Method Based On Deep Autoencoder

Currently,cardiovascular disease has been referred to as the ‘leading cause of death worldwide’ posing a significant threat to human health and life.Electrocardiogram(ECG),as a key factor for evaluating heart disease,has important research significance for the prevention and treatment of cardiovascular disease.However,ECG is easily interfered by various types of noises,leading to the loss of medical value of the ECG signal,thus affecting the diagnosis of doctors.Therefore,this thesis conducts in-depth research on the denoising problem of ECG signals,and the main research contents are as follows:(1)Aiming at the problem that existing ECG signal denoising methods cannot eliminate multiple types of noises and the loss of detailed feature information,an ECG denoising method based on variational autoencoder and masked convolutional network is proposed.The method uses the idea of variational inference to learn the effective latent variable representation of ECG signals and uses the KL divergence to measure the information content encoded in the latent variables,effectively capturing the global features of the ECG signal.To address the issue of loss of detailed feature information during the denoising process,a masked convolution network module with strong generality is proposed.Unlike the independence assumption of the latent variable model,this module fully utilizes the correlation between the ECG signal sample points by applying mask operations in different regions,thus retaining more detailed feature information of the ECG signal.Additionally,the module uses residual blocks to improve the efficiency of feature extraction and is applicable to most deep learning-based detailed feature processing problems.Based on the complementary feature information extracted by the two modules,the method effectively removes single noise and mixed noise in ECG signals.(2)Since the existing ECG denoising algorithms do not utilize the correlation between ECG waveforms,a denoising method for electrocardiogram(ECG)signals is proposed based on stacked autoencoders and long short-term memory(LSTM)networks.To optimize the training time of the LSTM network,which suffers from a large number of parameters,the proposed method first utilizes stacked autoencoders to extract highlevel features from the input data layer by layer and optimizes the network parameters based on supervised learning.Then,the LSTM network captures the long-term dependencies between signal cycles and the masked convolution module models the interaction between the sample points of the ECG signal,which captures more local detail features avoids while avoiding the problem of gradient disappearance.Finally,ablation experiments were conducted to verify the complementary advantages of different modules.The proposed model’s loss function integrates the reconstruction loss,crossentropy loss,and distance loss function,which not only considers the differences between denoised and clean signals but also considers the interdependence between ECG signal sequences,achieving accurate and efficient ECG signal denoising.The two models proposed in this thesis are verified on the MIT-BIH arrhythmia database for single noise and mixed noise and compared with other ECG signal denoising methods.The results demonstrate that both methods achieved outstanding noise reduction performance in terms of signal-to-noise ratio(SNR)and root mean square error(RMSE)indicators.