Research On Intelligent Wireless Sensing Of Human Behavior Based On Channel State Information

Currently,there is an increasing demand for the Internet of Things,and many industries need to intelligently control existing devices.Wi-Fi sensing technology provides a new method for achieving intelligent control.We transmit Wi Fi signals to the sensed environment,and the receiver collects signals that carry environmental information after reflection,scattering,and multipath transmission.After complex signal processing,we search for environmental features and identify people,objects,and human actions in the environment.We take indoor human behavior as the research object and adopt more sophisticated Channel State Information(CSI)to achieve efficient human motion perception.And we explore the impact of different experimental scenarios on the model to achieve the universality of the same model in different indoor scenarios.The main research content of this thesis is as follows:(1)To solve the problem of performance degradation of wireless sensing models for different research objects in Wi-Fi environments,we propose a human independent motion recognition model,Wi Pg,which is constructed from convolutional neural networks and generated adversarial networks.Model training and testing were conducted on CSI data from 14 yoga movements of 10 experimental personnel with different body types,and we ultimately obtained recognition results unrelated to the body type of the personnel.The experimental results show that the average accuracy rate of Wi Pg for recognizing 14 yoga movements can reach 92.7%,realizing a high-precision and personnel independent yoga movement recognition method.(2)To solve the problem of unstable recognition performance caused by experimental scene changes in Wi-Fi motion recognition methods,we propose a scenarioindependent CSI information fusion sharing model CSI-F,combining Convolutional Neural Network(CNN)and Gated Recurrent Unit(GRU).To prove the effectiveness of CSI-F,we conducted experiments on data sets collected in a real environment,and utilized antenna diversity to eliminate random phase shifts and ignore static interference.Then,we selected Hampel filter,principal component analysis,and Butterworth filter to eliminate the impact of noise beyond the action information.On this basis,we use window slicing and multi task information fusion models to share and overlay the same action CSI,further focusing on the impact of actions on CSI data,thereby achieving highprecision,scene-independent recognition of CSI-F and environmental mobility of the model.Experiments fully demonstrate the effectiveness and feasibility of CSI-F.(3)To further explore the universality of the model in the scene and personnel size,we propose a depth domain adaptation method,Wi-CAL.Firstly,noise is filtered from the collected data,and the data is expanded using dynamic time warping(DTW)center averaging(DBA)based on a weighted form.Secondly,the CSI data features that best represent the corresponding human motion are calculated using a feature weighting algorithm.Finally,we use CNN to automatically extract action CSI features,and combine softmax classifier with Correlation Alignment(CORAL)loss to obtain a classification model suitable for multiple scenarios.Experiments on public datasets and the expanded multi-scene datasets collected in this thesis show that Wi-CAL has stable and good recognition performance in the case of random changes in scenes and personnel.