Research On Classification Algorithm Of ECG Signal Based On Deep Transfer Learning

In recent years,cardiovascular disease is one of the important diseases that threaten human life,and its morbidity and mortality have increased year by year.Most of the early symptoms of cardiovascular disease are accompanied by arrhythmia.It is of great medical value and social significance to timely and accurately classify abnormal ECG signal categories.ECG signal classification algorithms currently have two problems.Most classification methods are based on ECG signal data classification.There is little research on the classification of ECG signal time-domain waveforms and transform domain data sets,and the accuracy of single signal classification is relatively high.Low,and the time complexity of the classification algorithm model is high and the amount of operation parameters is large.In view of the above problems,this paper presents a data feature analysis algorithm based on time-frequency transform,and a classification discriminant algorithm based on deep convolutional network combined with transfer learning,which obtains better classification performance in small data samples and reduces time Complexity and amount of calculation data.Main work as follows:1)Give a signal preprocessing method based on wavelet decomposition.First,Daubechies wavelet arid Symlet wavelet transform are used to perform denoising preprocessing on the original ECG signal,and high-frequency noise is filtered by wavelet decomposition.Experimental simulation comparison uses Daubechies wavelet to achieve the highest signal-to-noise ratio when the decomposition layer is 6.Secondly,the QRS wave is detected according to the information marked in the MIT-BIH database file,and the heart beat is divided to complete the construction of the time-domain waveform data set of the ECG signal.In the time-domain data set,the characteristics of abnormal signals are mainly selected from the perspective of QRS wave morphology to distinguish six types of ECG signals.2)Give a signal characteristic analysis method based on time-frequency transform.Aiming at the problem of low accuracy of single feature recognition of ECG signals,the methods based on short-time Fourier transform and continuous wavelet transform are selected to complete the time-frequency transform of ECG signal data and construct frequency domain and wavelet domain with sufficient effective features Data sets to train deep migration networks.In the time-frequency domain data set,feature vectors are mainly extracted from two aspects:amplitude spectrum analysis and power spectrum estimation,including three eigenvalues:spectrum mean,spectrum standard deviation,and power spectrum standard deviation,and then distinguish normal signals from arrhythmia signals.3)Design a deep migration learning network based on ECG signal classification and recognition.Aiming at the problem of the need to train a large amount of data and the high model complexity in the classification method,transfer learning is introduced.Using the characteristics of model-based transfer learning,VGG-16 is improved as a pre-training network to automatically extract ECG signal features.The complexity of the network algorithm.Experiments show that the introduction of transfer learning shortens the training time and reduces the amount of computational complexity.And experiments are conducted on the three designed data sets to verify the effectiveness of the improved algorithm given in this paper.