姚嘉伟  蔡延光

（广东工业大学自动化学院，广东 广州 510006）

摘要：为提高心电图身份识别过程中神经网络的训练效率及识别正确率，提出一种基于多核卷积和多头自注意力的心电图身份识别方法。首先，利用多个大小不同的卷积核对预处理后的单个心拍进行特征提取；然后，采用多头自注意力模块加强卷积通道中全局特征的提取效果；接着，将特征通道一分为二进行半实例归一化，使多头自注意力模块能够关注归一化前后的特征，提升神经网络的收敛速度；最后，将多核多头自注意力模块进行ResNet残差连接。该方法在QT数据集上经过20个epoch训练，实现了94.92%的识别正确率。利用ResNet进行对比实验的结果表明，该方法能够有效地提升神经网络的训练效率及识别正确率。

关键词：多核卷积；多头自注意力机制；半实例归一化；心电图；身份识别

中图分类号：TP391          文献标志码：A            文章编号：1674-2605(2023)05-0005-06

DOI：10.3969/j.issn.1674-2605.2023.05.005

ECG Identity Recognition Method Based on Multi-kernel Convolution and Multi-head Self-attention 

YAO Jiawei  CAI Yanguang

(College of Automation, Guangdong University of Technology, Guangzhou 510006, China)

Abstract: To improve the training efficiency and recognition accuracy of neural networks in the process of electrocardiogram identity recognition, a electrocardiogram identity recognition method based on multi-kernel convolution and multi-head self-attention is proposed. Firstly, feature extraction is performed on a preprocessed single heartbeat using multiple convolution checks of different sizes; Then, a multi-head self-attention module is used to enhance the extraction effect of global features in the convolutional channel; Next, the feature channels are divided into two for semi instance normalization, enabling the multi-head self-attention module to focus on the features before and after normalization, thereby improving the convergence speed of the neural network; Finally, connect the multi-kernel and multi-head self-attention module to ResNet residual. This method achieved a recognition accuracy of 94.92% after 20 epochs of training on the QT dataset. The results of comparative experiments using ResNet show that this method can effectively improve the training efficiency and recognition accuracy of the neural network.

Keywords: multi-kernel convolution; multi-head self-attention mechanism; semi instance normalization; electrocardiogram; identity recognition