Research On Denoising Method Of ECG Signal Based On Autoencoder-Conditional Generative Adversarial Network

Nowadays,cardiovascular disease has become one of the most serious diseases causing human casualties,which has brought great medical burden to China.Electrocardiogram(ECG)is a non-invasive test to diagnose the health status of human heart,which will not cut the skin or enter any of the body spaces.Its detection method is simple and quick which’s results have high reliability and accuracy.However,in the actual process of dynamic ECG acquisition,the central electrical signal is often polluted by noise,which affects the quality of the signal.Therefore.noise reduction is an important signal processing work before signal recognition and disease diagnosis.The main content of this paper is as follows:(1)The research status of electrocardiogram signal denoising is analyzed.By dividing the noise reduction methods into traditional methods and deep learning-based methods,the characteristics and shortcomings of the existing methods are analyzed.Through literature research,it is found that existing denoising methods of ECG signal have some problems such as poor signal detail retention ability and weak generalization ability after denoising.And some of them need to divide the input signal according to the heart beat.(2)In this work,Generative Adversarial Networks and Autoencoder network are demonstrated and analyzed in principle and characteristics.In order to solve the shortcomings of the existing denoising methods of ECG signals,this work proposes a denoising method of ECG signal base on Autoencoder-Generative Adversarial Networks.By using the Generated Adversarial Network structure and referring to the time domain characteristics of ECG signals,this method can antagonistically learn the distribution characteristics of noise in noisy signals and suppress it,so as to accurately recover the details of micro-waveform in ECG signals The convolutional layer is used to effectively retain the spatial information of all the adjacent regions of the output samples,which eliminates the restriction of dividing the input data according to the heart beat before noise reduction.(3)An ECG noise reduction method based on improved AE-CGAN was proposed.Considering that there are potential vector inputs in the original Generative Adversarial Networks and various noise reduction methods,and the unstable input is difficult to have a positive impact in the noise reduction process,we use the network structure mentioned above and the potential vector input mode of SEGAN to build two models with potential vector input and without potential vector input.The SNR and RMSE of denoising signal produced by the two models after denoising the ECG signal under the influence of the same noise intensity are measured.The experimental results show that the average SNR increased by 3.88dB and the average RMSE decreased to 0.0037 after the random vector input was removed.Through data discussion and analysis,the negative effect of potential vector input on noise reduction process is found.Considering the similarity between the full convolutional autoencoder generator and the denoising autoencoder in this method,and the influence of the encoding capacity in the denoising autoencoder on the denoising effect,five models with different encoding capacity are built by adjusting the number of layers with half the depth in the convoluted feature data.The SNR and RMSE of denoising signals of five models under the same training set,test set and generalization verification set were measured.Through the analysis and discussion of the average SNR,it is found that the SNR of the test set and the generalization verification set are positively correlated with the coding capacity in a certain range,and the SNR is negatively correlated with the coding capacity after the certain range.That is to say,there is an optimal coding capacity in this experiment.Based on the experimental results,the best coding capacity in this experiment is selected to optimize the network structure of the proposed method.(4)A variety of noise reduction experiments were carried out on the optimized model and the results were compared with those of the existing noise reduction methods.The experimental results show that the denoising method of ECG signal with Generative Adversarial Networks structure based on full convolutional autoencoder has the highest average SNR after denoising the ECG signals which are respectively affected by three kinds of single noises.The average SNR under seven kinds of noise conditions is 41.47dB,which is 25.05%higher than that of the best performing Adversarial method at present,and the improvement value reaches 8.31dB.Experimental results show that the proposed method has better noise reduction effect than the existing methods on the basis of the fact that there is no need to divide the input data into heart beats and the micro-waveform features can be retained.