Key Technology Research On WiFi-based Indoor Contactless Intelligent Sensing

WiFi signals offer numerous advantages,including contactless sensing,no invasion of privacy,no requirements for lighting and line of sight,low cost,and universality,making them the ideal medium for indoor contactless intelligent sensing applications.Consequently,WiFi-based indoor contactless intelligent sensing technology has garnered widespread attention from both academia and industry,and has been applied in various fields such as smart homes,healthcare,security surveillance,as well as augmented and virtual reality.However,the technological boundaries of WiFi-based indoor contactless intelligent sensing research are stretching towards meeting more complex sensing requirements and are facing several practical application issues:(1)With respect to sensing models,there has been an evolution from simple models trained based on manual features to deep learning models that automatically extract features,reflecting a trend of "deepening of sensing models".However,it is difficult to apply deep learning driven by multi-class data training to one-class classification model.(2)With respect to sensing tasks,there has been an evolution from single sensing tasks to composite sensing tasks in specific contexts,reflecting a trend of "compounding of sensing tasks".However,they are limited to simple combinations of different single sensing tasks,which can easily lead to some sensing defects;(3)With respect to sensing scenarios,there has been an evolution from in-domain sensing within specific scenario to cross-domain sensing adaptable to a variety of scenarios,reflecting a trend of "cross-domain sensing scenarios".However,cross-domain models have limitations in the use of multiple transceivers or target domain data,and are also limited to cross-one-factor sensing,leading to difficulty in cross-muitiple-factor sensing.To address the above practicality issues,this thesis explores from three dimensions:sensing models,sensing tasks,and sensing scenarios.Specifically,the main research contents of this thesis are as follows:(1)In response to the key issue of "difficulty in applying deep learning to one-class classification models" in the "deepening of sensing models",this thesis investigates "wireless one-class classification model based on deep autoencoders",abbreviated as CSI-OC.CSI-OC innovatively utilizes "wireless feature extraction method based on deep autoencoder" and "wireless feature clustering learning mechanism based on hypersphere" to apply deep learning methods to wireless one-class classification models,in order to achieve the research goal of "improving sensing performance".Extensive validation experiments and performance analysis on two real-world wireless datasets demonstrate that,compared with the state-of-theart methods,CSI-OC improves the positive-class recognition rate by 3.72% and the negativeclass recognition rate by 5.78%,with a processing delay of only one-seventh or even less.(2)In response to the key issue of "sensing defects caused by simple combination of composite sensing tasks" in the "compounding of sensing tasks",this thesis takes the contactless continuous authentication as an application example to investigate "optimization method for contactless continuous authentication based on contextual features",abbreviated as WiPT.WiPT innovatively utilizes "low-computational two-step user state continuous detection mechanism" and "behavioral biometric implicit authentication method based on posture transition",combined with contextual features,to optimize the sensing methods and supplement the feasibility of existing contactless continuous authentication,in order to achieve the research goal of "solving sensing defects".Extensive validation experiments and performance analysis based on three real-world scenarios demonstrate that,compared with the state-of-the-art WiFi-based contactless continuous authentication system,WiPT condenses regular 5-minute interval authentication to a continuous 2-second monitoring,reduces the total processing delay within 5 minutes to just over a tenth of its original time,improves the authentication accuracy by 4.49%,and defense accuracy by 4.78%.(3)In response to the key issue of "substantial usage restrictions on cross-domain model" in the "cross-domain sensing scenarios",this thesis investigates "wireless domain generalization model based on style randomization",abbreviated as WiSR.WiSR innovatively utilizes "domain style quantification mechanism based on subcarrier differences" and "domain invariant feature learning method based on style randomization" to achieve domain generalization sensing without the need for multiple transceivers and target domain data,in order to achieve the research goal of "simplifying sensing conditions".Extensive validation experiments and performance analysis on three public real-world wireless datasets from various WiFi devices demonstrate that,compared with the state-of-the-art generic domain generalization methods and wireless cross-domain methods,WiSR offers superior universality and improves average cross-domain accuracy by 2.26% and 3.79%,respectively.(4)In response to the key issue of "difficulty in cross-multiple-factor sensing" in the "cross-domain sensing scenarios",this thesis investigates "cross-multiple-factor sensing method based on wireless data augmentation",abbreviated as WiSGP.WiSGP innovatively utilizes "data augmentation method based on subdomain-guided perturbations" and "domain generalization training mechanism based on subdomain data augmentation" to achieve crossmultiple-factor sensing under domain generalization settings without the need for multiple transceivers and target domain data,in order to achieve the research goal of "overcoming sensing difficulty".Extensive validation experiments and performance analysis on three public real-world wireless datasets from various WiFi devices demonstrate that,compared with the state-of-the-art generic domain generalization methods and wireless cross-domain methods,WiSGP offers superior universality and portability,and improves average crossdomain accuracy by 1.67% and 5.09%,respectively.The above research contents address the key issues faced by WiFi-based indoor contactless intelligent sensing technology from three dimensions and four perspectives,and have important research significance.The resolution of these issues will have a huge driving effect on the commercial development of intelligent sensing technologies.