Research On Classification Model And Algorithm Of 12 Lead Electrocardiogram Signal Based On Deep Learning

Electrocardiogram(ECG)signal is kind of weak and complex biological signal,which frequency and amplitude vary with time,with non-linear and non-stationarity and its combination relation and structure characteristics are more complex,its classification processing also is always significant in the fields of signal analysis and biomedical research.Feature extraction is the critical factor of pattern classification and detection,the stronger the feature representation ability is,the better the performance of pattern classification will be.However,Low frequency,weak signal,baseline drift and most existing methods only focus on digital feature extraction,ignoring the information of signals time dimension,integrity and objectivity,locality and global factors.The classification detection faces a huge challenge.The research framework and contents of this paper are summarized as follows.To solve the problem of nonlinear time-varying signal classification,the existing RNN serial information processing mode is analyzed.A recurrent neural network model of parallel computing structure for time-varying signal classification is designed,which consists of a multi-channel time-varying signal input layer,a parallel processing structure unit,a signal feature fusion layer and a Softmax classifier.In these parallel processing units,the RNN expands the existing RNN serial processing mode for multi-channel time-varying signals into parallel mode,the input signal for each channel corresponds to a deep recurrent neural network,and feature extraction and attribute association of single-channel signals were performed to achieve parallel processing of all-channel signals.In the signal feature fusion layer,feature vectors from each channel signal were integrated to constitute a comprehensive feature matrix.On that basis,the Softmax function was employed as a classifier for multi-channel signals.In terms of mechanism,the model can realize the independent extraction of single channel signal features,the fusion of each channel signal features,and the signal classification based on the comprehensive feature matrix,so as to maintain the feature combination relationship of multi-channel signals.The existing serial mode of RNN multi-channel signal processing is improved to reduce the loss of structural feature information and improve the algorithm efficiency.In particular,the resolution of signal samples with similar distribution characteristics is significantly improved,which verifies the effectiveness of the model and algorithm.For the recognition of 12-lead multi-channel time-varying signals with similar feature distribution,analyzing the problems of deep network degradation,deep network can’t be optimized well,a deep convolutional sequence residual network model is proposed,which is composed of multi-channel signal input layer,convolutional time series residual block,fully connected layer,typical feature embedding layer and classifier.In convolution sequence block residual,stacking the convolution block with spatial feature mapping function and the recurrent block with time domain feature extraction function according to the way of residual connection to form the convolution time sequence residual block,the convolution time sequence residual block has spatiotemporal characteristics on the characteristics of time-varying signals,so as to obtain more detailed characteristics of the characteristic signals.In typical feature embedding layer,typical samples will be constructed to expand the dataset to assist the identification in view of the small sample data volume and unbalanced sample distribution.Specifically,the typical sample features are acquired by using the classic beat segmentation algorithm and the time series sliding window algorithm,and then the matched beat samples are screened according to the screening rules to form the typical sample data set,and the new sample dataset is formed by parallel fusion with the features extracted by the neural network.On this basis,the Softmax classifier is used to realize multi-channel signal classification To solve the problem that the model lacks the representation of spatial-temporal characteristic signals and the fusion of global and local features,the convolution time sequence residual block is constructed to acquire the spatial-temporal characteristics of the signals,so as to realize the fusion of global and local features and reduce the loss of time domain features.The effectiveness of the model and related algorithms is verified by experiments.