Research And Implementation Of Domain Adaptation Method In Passive Wireless Sensing

The development of wireless communication and sensing technology promotes the development of passive wireless sensing.According to the influence of human body and its motions on wireless propagations,sensing tasks such as gesture recognition,activity recognition and positioning,can be realized by pattern recognition via learning algorithms.However,wireless signals have a great dependence on the environment,and the change of people or surrounding environment will lead to the failure of the established sensing model.In order to solve this problem,ths existing schemes usually adopt semi-supervised or unsupervised domain adaptation methods,which need to collect unlabeled or partially labeled data in the new environment.But in practice,the data in the new environment cannot be obtained in advance.Therefore,a solution to automatically generalize to a new environment without new environment data is needed.To achieve this goal,this thesis proposes a generalization framework for environment independent action recognition that integrates sample generation and episodic training.The framework uses several source domain data to build the action recognition model,extracts the domain independent action features,and can generalize to the target domain without any new data.To verify the generalization and effectiveness of the proposed framework in the field of wireless sensing,the proposed framework was evaluated for action recognition based on Wi Fi and millimeter wave respectively,and achieved acceptable recognition accuracy.(1)This thesis proposes a domain generalization universal wireless sensing framework that integrates virtual data generation and episodic training.The framework consists of three parts: data acquisition and preprocessing,virtual data generation,domain independent feature extraction and classification.It can reduce the impact of various environmental changes,such as person changes and room changes,on recognition performance in passive wireless sensing.(2)Based on this framework,Wi Fi environment independent action recognition is realized.When action recognition is carried out with Wi Fi,the three modes of data including virtual amplitude,virtual phase and virtual Doppler spectrum are generated based on the real amplitude data,and then the environment-independent action features are extracted from the spatio-temporal dimensions through neural network episodic training,so as to realize zero-cost generalization to the new environment.Experimental results show that the correct rate of gesture recognition for new users and new environments is more than 80% when the target domain has no data involved in model training,which is higher than other methods.(3)Based on this framework,millimeter wave environment independent action recognition is realized.When action recognition is carried out with millimeter wave,the Range-Doppler is first compressed to obtain the compressed Doppler,and then the virtual data are generated based on the real compressed Doppler data.Afterwards,the Convolutional Block Attention Module(CBAM)and the Res Net network are used to extract the domain-independent action features,so as to achieve the domain generalization effect.The experimental results show that the accuracy of the proposed method for action recognition of new users,new locations and new environments is more than 94% when no data from the target domain is involved in the model training.Compared with other methods,the proposed method has higher results.In the end,an intelligent wireless action recognition system is implemented based on this framework.The system can achieve action recognition based on Wi Fi and Millilmeter wave,including wireless data acquisition,model training,action recongition,etc.