AHNNet:融合注意力机制的行为识别混合神经网络模型  

AHNNet:Human Activity Recognition Based on Hybrid Neural Network Combining Attention Mechanism

作　　者：曹仰杰 李昊[1] 段鹏松[2] 王福超[2] 王超 CAO Yangjie;LI Hao;DUAN Pengsong;WANG Fuchao;WANG Chao(School of Information Engineering,Zhengzhou University,Zhengzhou 450000,China;School of Software,Zhengzhou University,Zhengzhou 450000,China)

机构地区：[1]郑州大学信息工程学院,郑州450000 [2]郑州大学软件学院,郑州450000

出　　处：《西安交通大学学报》2021年第5期123-132,共10页Journal of Xi'an Jiaotong University

基　　金：国家自然科学基金资助项目(61972092)。

摘　　要：针对Wi-Fi信号的行为感知研究中传统机器学习方法特征提取困难、深度学习方法特征提取方式单一,导致特征提取不充分、识别准确率不高等问题,提出融合注意力机制的人体行为识别混合神经网络模型AHNNet。在对信道状态信息影响因子分析的基础上,使用信道状态信息的振幅数据作为行为识别的基础数据;采用时间滑窗将长时间人体活动序列分割为短时间序列,构建样本数据,克服全局人体行为数据非实时、长度不固定的缺点;通过双向循环门控网络和时序卷积网络并行提取输入数据特征,充分挖掘数据潜在特征之间的关系;在双向循环门控网络中融合注意力机制以强化数据特征,进一步提高模型性能;将双向循环门控网络和时序卷积网络提取到的特征进行融合,增加特征的多样性;将融合特征输入到Softmax分类器进行分类,得到人体活动数据对应的行为。与长短期记忆网络、双向循环神经网络进行了对比,实验结果表明:在标准数据采集室数据集上,AHNNet的行为识别正确率达到97.15%,比未使用注意力机制的模型分类正确率提高1.81%;在公共数据集上,AHNNet的行为识别正确率比其他对比模型的提高至少0.65%,参数量下降47%;在不同环境下,AHNNet在卧室环境中的正确率为95.7%,比标准数据采集室中的下降1.45%。AHNNet具有良好的识别效果和鲁棒性,并且在复杂的居家环境中应具有应用价值。Activity wireless sensing is an important technology to implement health monitoring.Although the research on activity perception based on Wi-Fi has made good progress,there are still some problems,such as the difficulty of feature extraction in the traditional machine learning and the single feature extraction method in deep learning,which leads to insufficient feature extraction and low recognition accuracy.Here AHNNet,a hybrid neural network of human activity recognition integrating attention mechanism is proposed.Analyzing the influence factors of channel state information,the amplitude data of channel state information are used as the basic data for activity recognition,and the time sliding window is used to divide the long time human activity sequence into short time series to construct the sample data,which solves the difficulties of non-real-time and non-fixed length global human activity data.AHNNet parallelizes the bidirectional gated recurrent network and the temporal convolutional network to extract the features of the input data,so as to fully reveal the relationship between potential features of data.To further improve the performance of model recognition,AHNNet is combined with attention mechanism to strengthen the main features of data in bidirectional gated recurrent network.Then,the features extracted from the bidirectional gated recurrent network and the temporal convolutional network are fused to increase the diversity of the features,and the fused features are input into the Softmax classifier for classification to obtain the activity corresponding to human activity data.Experimental results demonstrate that AHNNet has stronger classification ability than the other models,and the average accuracy rate achieves 97.15%,and AHNNet has fewer parameters while maintaining high accuracy.The competing model demonstrates that the accuracy rate is 95.7%in the bedroom environment,1.45%lower than that in the standard data acquisition room.AHNNet has good recognition effect and robustness,especially in com

关 键 词：人体行为识别 信道状态信息 双向循环门控网络 时序卷积网络 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]