Research And Implementation Of ECG Signal Classification Algorithm Based On Multiscale Convolutional Neural Network

With the rapid development of artificial intelligence,more and more researchers are applying machine learning and deep learning to the classification of ECG signals,and have made relevant progress in the diagnosis of clinical abnormal heart rate,but still have some problems and challenges.Due to the interference of noise during the electrocardiogram signal acquisition process,So,feature extraction of traditional machine learning has poor performance,like in low classification accuracy that cannot meet clinical me dical requirements.Traditional machine learning has poor generalization ability,low effi ciency,and is unable to handle current complex scenarios,and with the development of medical equipment,electrocardiogram signals have changed from single-lead to multi-le ad data,resulting in excessive computational complexity and complex model training in traditio-nal convolutional networks,and because of the collection cost of ECG dataset is too high,the number of sample is unbalanced,resulting in the low classification accuracy of traditional convolutional networks,and network is easy to classify normal heart rates with large sample numbers while ignoring more important abnormal heart rates.So,this thesis proposes a deep convolutional neural network to address these issues and respectively achieves an accuracy of 99.6% and 95% on the MIT-BIH dataset and Hefei Open Source dataset.The main contributions and innovative points are summarized as f ollows:(1)This thesis proposes a structure called Multi-scale Fusion Convolutional Block,which performs parallel feature extraction using multiple convolution layers,max pooling layers,and min pooling layers.Finally,the outputs are stacked in the channel dimension to enable the network to have receptive fields of different sizes.Additionally,cause different convolution branches have different feature extraction capabilities,SQUEEZEAND-EXCITATION BLOCKS(SE)self-attention mechanisms are used to allow the network to autonomously learn the importance of different convolution branches.Experimental results show that this structure has excellent feature extraction capabilities.(2)This thesis proposes a novel pooling pyramid structure for the peak and valley characteristics of electrocardiogram data.The structure takes the nodes of the previous l ayer as inputs and outputs passing through the max pooling and mini pooling layers,this process is repeated iteratively until reaching the last layer of nodes.The proposed structure enhances the extraction of peak and valley features of electrocardiogram signals while reducing data dimensionality.To prevent network degradation,a residual channel shortcut structure is introduced,where shallow electrocardiogram data undergoes one-dim ensional convolution and is directly stacked with deep outputs on the channel dimension to ensure that the network does not lose shallow information while preventing overfitting.Furthermore,to tackle the issue of imbalanced electrocardiogram data,data augment ation is performed in the data preprocessing stage,and in loss function design,CB focal loss is employed to balance the large gap between positive and negative samples and reduce the impact of data imbalance.(3)This thesis conducts research on electrocardiogram signal classification using the Vit-Transformer network based on the recurrent neural network structure,and performs experimental comparisons by fusing convolutional networks and Vit-Transformer.