Research On ECG Signal Denoising Algorithm Based On Deep Neural Network

As an invisible killer,the high incidence and high disability rate of cardiovascular disease bring a heavy economic and psychological burden to society,families,and individual patients.Therefore,the prevention,monitoring,and post-treatment of cardiovascular diseases are particularly important.The Electrocardiogram(ECG)is widely used in the diagnosis of heart disease.However,in real situations,ECG signal recordings are often polluted by different noises,which cause distortion in the signal band and affect the final diagnostic results.Therefore,it is of great importance to propose an effective ECG signal noise reduction algorithm.Over many years,although significant advances have taken place in the processing and analysis of ECG signal,ECG signal denoising still faces many challenges,such as:(1)It is difficult to further improve the denoising performance and remove a single type of noise when using traditional models for ECG signal denoising.(2)Although learning-based models achieve the desired results,most neural networks,as a kind of black box,are mainly dependent on human experience in their architecture design and do not have strong interpretability.Based on the above problems,this thesis proposes two ECG signal denoising algorithms for study,whose innovations and contributions are as follows:(1)Aiming at the problems that traditional denoising models remove a single type of noise and have insufficient generalization ability,this thesis proposes an ECG denoising model based on the adversarial denoising convolutional neural network.Firstly,an adversarial denoising convolutional neural network model is used to learn the residual signal in the noisy signal to obtain the denoised signal.Secondly,the discriminator network is used to classify the denoised signal and the original signal,and is fed back to the adversarial denoising convolutional neural network model for parameter adjustment.In addition,a new minimized time-frequency domain loss function was designed to train the network model as a way to capture non-linear features and waveform details,and also to ensure the stability of the network training process.The denoising performance of the proposed model is verified by dividing three different types of datasets on the MIT-BIH database and the QT database.Experimental results show that the method has certain advantages compared with existing models.(2)The traditional ECG denoising model has the problem of removing a single type of noise and the lack of interpretability of deep neural networks.This thesis proposes a deep neural network denoising model based on the sparse representation algorithm for ECG,incorporating of traditional methods into the design of neural networks,and explores the connection between the sparse representation algorithm and neural networks.Firstly,the ECG signal is modeled using the sparse representation algorithm.Secondly,the half quadratic splitting algorithm is used to transform the optimization problem into two sub-problems.Based on the design and solution of the sub-problems,a new neural network architecture is proposed and the designed neural network architecture has a clear and meaningful interpretation of its parameters and data flow.In addition,a new weight distribution module is designed to take hyperparameters adaptively using the correlation between ECG signal data,which greatly improves the efficiency of hyperparameter selection.To verify the validity of the proposed model,four different denoising models and corresponding different data pre-processing techniques are used to make comparisons.After extensive experimental validation and simulation studies,it is shown that the framework has good denoising performance.