Research On Electrocardiograph Recognition Based On Improved Convolutional Neural Nnetwork

Cardiovascular disease(CVD)has now become the most common cause of death,accounting for more than 31 percent of deaths worldwide.As a non-invasive test,Electrocardiogram(ECG)is an import tool for detecting and diagnosing heart disease.The intelligent detection and diagnosis of ECG signals by computer aided diagnosis has become a research hostpot in the filed of intelligent diagnosis of CVDs.In the era of big data,the data scale is growing exponentially,and traditional machine learning algorithms need to consume a large number of computing resources to learn data characteristics,and lack of stability.In recent years,Convolutional Neural Network(CNN),as the representative of deep learning,has a good application in the fields of image recognition,image segmentation,signal processing and natural language processing,which enable the deep learning represented by CNN to have a broad application prospect in ECG signal automatic recognition and classification.Based on the above considerations,this dissertation,based on the CNN model and according to the data characteristics of different ECG signals,studies the application of automatic recognition of ECG signals,mainly including the following three aspects:First of all,for the characteristics of single lead ECG signal samples with high dimension and large number of training samples,which lead to high computational complexity,it is necessary to consider how to use small amount of computation to excavate deep features.Therefore,this dissertation proposes a convolutional neural network model with differential sequences(1D-CNN-DS),which uses a simple difference method,smooth nonlinear activation function and adaptive learning rate combined with loss function to solve the difficulties in single lead ECG signal recognition.Combined with experiments.It is proved that the 1D-CNN-DS proposed in this paper has good performance in single lead ECG signal recognition.Secondly,due to the influence of privacy or data collection and other reasons,the amount of ECG signals in different types are unbalanced,resulting in false positive or false negative problems in the identification process.To solve this problem,a new algorithm for the Multiply Stochastic Synthetic Minority Over-sampling Technique(MSSMOTE)combined with the improved Feature Selection Convolutional Neural Network(FS-CNN)is proposed.Experimental results show that the proposed algorithm is effective in ECG multi-factor data classification.Finally,considering that ECG signals have a difference in length and a loss of signal value in the acquisition process,and that the input dimension of CNN Network is fixed,which makes it impossible to accept signals of different lengths,it has a Recurrent problem that hinders ECG sequence recognition research.This dissertation proposes a model combining Independent Recurrent Neural Network(Ind RNN)with improved CNN.Ind RNN was used to identify and combine the ECG signal to the demand length and CNN was used for identification.Meanwhile,improved Fruit fly optimization algorithm(FOA)was used to optimize the parameters of the combinatorial network model for better performance.Experiments show that Ind RNN-CNN network has good performance in dealing with variable length ECG sequences,while FOA has certain effect in optimizing Ind RNN-CNN network.