Research On Intelligent Wireless Sensing Based On Deep Learning

Nowadays,with the rapid development of Internet of Things,the demand for wireless data traffic is increasing rapidly,and a variety of wireless signals,such as,WiFi,Bluetooth,mm Wave,etc.are full of people's working and living space.Beyond communication,the wireless signals can also be used to sense the status of surrounding objects,such as position,velocity,gesture,gait,vital signs,etc.Different status of target with their reflection,scattering and refraction will bring in different influence on signals.To realize different sensing tasks,wireless sensing technology reveals and analyzes the influence pattern of target towards surrounding wireless signals,and with its advantages of not carrying any hardware device,easy deployment,not causing privacy leakage and high security,leading to a broad application prospect in the smart home,human-computer interaction,security aid and so on.Deep learning has achieved great success in computer vision in recent years.Some existing studies have proved that wireless sensing technology can benefit from it,which can make the traditional wireless network evolve to intelligent wireless network with target status sensing ability,and make it a brand-new step towards ubiquitous sensing.However,at present,there are still some challenges of deep learning based wireless sensing technology.Such as,the well-trained sensing model in a certain scenario has poor adaptability and its cost of retraining is high in a new working scenario,and the training samples are insufficient,which leads to model degradation and a decline on sending performance.When encountering a sensing task in new working scenarios,deep learning based wireless sensing systems usually require a large number of wireless signal data in the new scenario to participate in the retraining of the deep network,which reduces the efficiency of wireless sensing model.Thus,this paper proposes a meta learning based wireless sensing method,which trains the deep learning-based sensing model to learn the transferability knowledge hidden in wireless signal samples,and evaluate the distance between the features of wireless samples automatically,so as to quickly adapt the task of wireless sensing with small samples in new working scenarios,which effectively alleviates the problem of model degradation with less training effort,and improves the applicability of the sensing model constructed by training scenario in a new working application scenario.Due to the stringent condition of wireless signal sample collection,obtaining standard wireless signal data is usually time-consuming and labor-intensive.When the training samples are insufficient,it is difficult for deep learning-based wireless sensing system to learn essential characteristics of wireless signal data,and the performance of the sensing system will decrease.Thus,this paper proposes a generative adversarial-based wireless sensing method,and use modified Generative Adversarial Networks to synthesize virtual radio images,which optimize the structure of adversarial network and strength the constraint conditions to maximize the acquisition of the rich information contained in the wireless data sample and realize wireless data augmentation with small training sample so that improve the generalization ability of the deep learning based-wireless sensing model with a sample collection cost.This paper has designed human sensing experiments including sign language recognition,letter recognition and human gesture recognition,to evaluate the meta learning-based wireless sensing algorithm and the generative adversarial-based wireless sensing algorithm.Experimental results show that the proposed deep learning-based wireless sensing method effectively can improve the sensing adaptability of the trained deep wireless sensing model,augment the wireless dataset with a small experimental sample collection cost,and construct a high-performance wireless sensing system.