Research On ECG Signal Processing Technology Based On Sparse Characteristics And Deep Learning

Electrocardiogram(ECG)is the safe,effective and quick method for diagnosing heart diseases.At the same time,ECG is an important indicator for the detection,classification and treatment of heart diseases.Therefore,removing noise in ECG efficiently and accurately is of great significance to assisted cardiac diagnosis and treatment.In this context,combined with the sparse characteristics of the signal,the neural network based method will be used to deeply study the noise reduction of the ECG signal.The main innovations of this article will be shown in the following three aspects.(1)Aiming at the problem of poor applicability of traditional filter-based noise reduction methods to signals,a new ECG signal noise reduction method is proposed based on the combination of Generative Adversarial Network(GAN)and Deep Residual Network(DRN).Adding DRN to GAN greatly reduces the possibility of gradient explosion or dispersion of GAN during training,thereby improving the stability of network training.GAN itself has a strong ability to learn the difference(between the noisy signal and the original clean signal),and can better extract the "noise characteristics" to efficiently and accurately complete the noise reduction.Experiments show that the method proposed in this paper can remove three common noises and a variety of mixed noises,and it shows excellent applicability compared with general noise reduction networks.The MIT-BIH noise pressure database is used to verify the proposed algorithm,and the RMSE is 0.0102 and the SNR is 40.8526 db.The performance of this method is compared with that of multiple noise reduction algorithms.The results show that the proposed method based on the combination of GAN and DRN achieves excellent noise reduction results.(2)Aiming at complex optimization problems that often need to be solved in sparse signal processing,this paper proposes an ECG sparse noise reduction method based on deep expansion network.By utilizing the deep unfolding network,the iterations in the neural network are replaced by the iterations in the sparse optimization problem.The parameter solving problem in the optimization problem is transformed into the weight in the neural network to learn,which avoid to solve the complex optimization problem in the process of sparse noise reduction.At the same time,the use of deep learning overcomes the limitations of the artificial setting of the feature space in the traditional noise reduction methods.Experimental results show that the minimum SNR of this method 25.78 db,and this also validates the correctness of the theoretical model.(3)Because the local correlation and global correlation are not fully considered in the existing noise reduction methods,and this will cause serious distortion of the denoised signal,a new loss function is proposed which can better retain the key information of the signal after noise reduction.In this new loss function,the local difference and the overall difference between the denoised signal and the original clean signal are taken into account.Therefore,the local maximum difference function and the overall difference are added to the original generator network loss function.The value function is used to better capture the local and global features of the signal.Experiments have verified that the denoised signal is highly consistent with the original clean signal,and the key information of the original signal is retained to a great extent.