基于GASF及神经网络的多周期脉象信号识别分类研究  

Research on Multi-Cycle Pulse Diagnosis Signal in TCM Classification Based on GASF and Neural Networks

作　　者：刘轩吉 刘光浚 邓威 郝龙辉 王维广 陈占春 LIU Xuanji;LIU Guangjun;DENG Wei;HAO Longhui;WANG Weiguang;CHEN Zhanchun(College of Mechanical and Vehicle Engineering,Taiyuan University of Technology,Taiyuan 030024,Taiyuan,China;China Academy of Chinese Medical Sciences,Beijing University of Chinese Medicine,Beijing 100029,China)

机构地区：[1]太原理工大学机械与运载工程学院,山西太原030024 [2]北京中医药大学中医学院,北京100029

出　　处：《中华中医药学刊》2024年第10期22-27,I0003,I0004,共8页Chinese Archives of Traditional Chinese Medicine

基　　金：国家科技部科技基础资源调查专项(2022FY102000,2022FY102002);中央高校基本科研业务费专项(2022-JYB-JBZR-011);山西浙大新材料与化工研究院项目(2022SX-TD021)。

摘　　要：目的旨在解决一维时序的脉象信号特征提取阶段参数量不够以及一维信号转化为二维序列图像时逆运算缺失和数据与时序关系模糊等问题。方法提出了基于无分段聚合近似(PAA)的格拉姆和角场(GASF)及深度学习网络模型相结合的一维脉象信号多周期数据分类方法。首先通过GASF编码将一维脉象信号转换为二维时序图像,然后输入神经网络(TCPNet)进行训练并分类。设置了4273段同样长度的多周期脉象信号作为输入数据集。结果研究发现使用无分段聚合近似的格拉姆角场处理的网络准确率不低于89%。模型最高准确率达到93.61%,精确度为93.63%,F1分数为93.60%,召回率为93.61%。结论基于文章方法建立的脉象分类模型准确率明显提高,力证了分类方法的有效性,也为脉象信号的分类问题提供了新的思路和方法。Objective To solve the problems of insufficient parameter quantity in the feature extraction stage of pulse signals in one-dimensional time series as well as missing inverse operations and blurry relationship between data and time series when converting one-dimensional signals into two-dimensional sequence images.Methods It proposed a one-dimensional pulse signal multi period data classification method that combined Gram and angle field(GASF)based on non segmented aggregation approximation(PAA)and deep learning network models.Firstly,the one-dimensional pulse signal was converted into a two-dimensional temporal image through GASF encoding and then input into a neural network(TCPNet)for training and classification.It set 4273 segments of multi cycle pulse signals of the same length as the input dataset.Results This study found that the network accuracy using non segmented aggregation approximation for GASF processing was not less than 89%.The highest accuracy of the model reached 93.61%,with an accuracy of 93.63%,an F1 score of 93.60%and a recall rate of 93.61%.Conclusions The accuracy of the pulse classification model established based on the method of this paper is significantly improved,which forcefully proves the effectiveness of the classification method of this paper,and also provides a new idea and method for the classification problem of pulse signals.

关 键 词：多周期脉象信号 脉象识别分类 GASF 卷积神经网络 残差块 

分 类 号：R241.1[医药卫生—中医诊断学]