Research On ECG Signal Denoising Method Based On Generative Adversarial Network

ECG is an important non-invasive technology for detecting and diagnosing heart disease.However,the frequency and amplitude of ECG signals are low,and they are very sensitive to various instruments and biological interference,so they are usually polluted by various noises.The existing noise will cause the waveform change of the ECG signal,which causes great inconvenience for professionals to identify cardiac pathology.At present,many algorithms have been developed for medical signal denoising under complex environmental conditions,but there are still problems such as weak model generalization ability or signal distortion after denoising.Therefore,this thesis is based on the deep neural network and uses the method of Generative Adversarial Network(GAN)to conduct in-depth research on ECG signal denoising.The following is the embodiment of the innovation of this thesis.(1)This thesis proposes an ECG signal denoising method based on GAN.Compared with the traditional denoising model,this method can effectively remove multiple noise types in the signal and it has better generalization ability and noise reduction performance.The method uses the encoder structure of the Transformer as the generator of GAN to generate denoising signals and introduces the attention mechanism to visualize the importance of different parts of the signal to help the model better understand the signal characteristics.The discriminator of GAN introduces a multi-scale Inception structure,which improves the performance of deep neural networks by using convolution kernels of different scales to capture features of different scales.The experimental results exhibit that this method can better remove three common single noises and mixed noises in ECG signals.For single noise,the average Signal-to-Noise Ratio of the denoised ECG signal can reach 29.64 d B,and the corresponding Root Mean Square Error is 0.0106.For mixed signals,the average Signal-to-Noise Ratio of the denoised ECG signal is up to 28.07 d B,and the corresponding Root Mean Square Error is 0.0134.(2)This thesis proposes a two-stage denoising method based on the Half Instance Normalization network,which alleviates the distortion of the ECG waveform caused by denoising.Each stage of the method is composed of the U-Net network.In the first stage,the Instance Norm layer is used to standardize each sample independently,which better preserves the features of each sample.In the second stage,Cross-Stage Feature Fusion(CSFF)is used to connect the two stages of the network,thereby simplifying the flow of information and making the network optimization process more stable.In addition,aiming at the problem that the existing denoising methods do not consider the local and global characteristics of the signal,a regular term loss function(gradient difference loss)is proposed innovatively.The loss function takes the gradient of the signal into account and suppresses the interference factors in the denoised signal by ensuring that the gradient of the denoised signal and the clean signal in the same interval is as equal as possible,thereby accelerating the convergence speed of the model.The final experimental results verify the effectiveness and robustness of the method.The two noise reduction methods proposed in this thesis are tested on the MIT-BIH database.The experimental results show that both model frameworks can achieve excellent ECG signal-denoising effects.