基于改进型BP神经网络的手部动作识别  

Hand-motion recognition based on improved BP neural network

作　　者：尤波[1] 李忠杰 黄玲[1] YOU Bo;LI Zhongjie;HUANG Ling(School of Automation,Harbin University of Science and Technology,Harbin 150080,China)

机构地区：[1]哈尔滨理工大学自动化学院,黑龙江哈尔滨150080

出　　处：《智能系统学报》2018年第5期848-854,共7页CAAI Transactions on Intelligent Systems

基　　金：国家"863"计划重大项目(2009AA043803)

摘　　要：对手部动作进行模式识别,首先将采集到的肌电信号进行降噪处理,选择时域分析法中的方差算法对采集信号进行特征提取。将特征信号进行归一化处理,实验发现普通BP神经网络分类器出现学习速率慢,泛化能力较差,不同动作识别准确率差别较大等问题。针对以上问题,提出了一种改进型BP神经网络,将神经网络输入数据进行人工升维处理,并对网络学习速率慢的原因进行理论推导,然后引入交叉熵代价函数并对其进行正则化处理,以提高网络的泛化能力以及网络的识别准确率。实验结果表明,改进型BP神经网络的学习速率、泛化能力以及动作分类的准确率均优于普通网络,识别准确率平均为94.34%。To achieve accurate pattern recognition of hand motions,in this study,we first denoised collected electromyogram(EMG)signals,and then used a variance algorithm in the time domain to extract features from the collected signals.After normalizing the characteristic signal,in the experiment,we found that the general BP neural network classifier has a slow learning rate,poor generalization ability,and large differences in its accuracy of recognizing diverse motions.To address the above problems,we propose an improved BP neural network that processes its input data by artificially increasing the dimensions.It then theoretically determines the reason for the slow network learning rate,and introduces a cross-entropy cost function to regularize it,thereby improving the network's generalization ability and increasing its reaction speed.Experimental results show that the improved BP neural network has a better learning speed,generalization ability,and accuracy in hand motion classification than the ordinary neural network,with an average recognition accuracy of 94.34%.

关 键 词：BP神经网络 SEMG信号 交叉熵 手部动作识别 特征提取 正则化 机器学习 模式识别 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]