Research On WiFi Sensing Technology Based On Deep Learning

Accurate perception of the physical world has a huge demand for applications in future intelligent interaction scenarios such as Human-Computer Interaction,Virtual Reality,and the Metaverse.WiFi sensing is an intelligent interaction technology that uses ubiquitous WiFi wireless signals to achieve scene recognition,which can realize indoor positioning,behavior recognition,person tracking and emotion computing,etc.With the characteristics of low cost,easy deployment,and non-visual distance,it brings great convenience to people’s daily life and work and has received wide attention in recent years.Traditional WiFi sensing techniques usually implement different sensing tasks through Machine Learning(ML)or building mathematical or physical models between human activity and signals,but these methods are limited by the efficiency of manual feature extraction or the matching between the model and the actual task and are less generalizable to complex sensing tasks.In recent years,WiFi sensing technology based on Deep Learning(DL)has received much attention.DL technology uses deep neural networks to automatically extract available information from WiFi signals,avoiding the drawbacks of traditional algorithms that manually extract features or modeling,and enabling efficient sensing of the environment and people.However,existing DL-based WiFi sensing techniques usually only use Convolutional Neural Networks(CNN)to extract local features in action space,or use Long Short-Term Memory(LSTM)networks and Gate Recurrent Unit(GRU)networks to extract temporal features in signals,to realize the sensing and recognition of human activity features.Although the classical LSTM and GRU structures can capture the long-term dependencies in the time series data,they only consider the connection between the past information of human activities and ignore the comprehensive consideration of the future information,and cannot comprehensively extract the temporal features of the action data.In addition,these existing schemes generally directly apply CNN or LSTM,GRU,and other modules to carve the feature information of the perceived action in spatial dimension or temporal dimension and do not comprehensively extract the temporal and spatial features of human activity data,resulting in low accuracy of recognition of complex human activities.With the above research background,this thesis conducts a series of research on WiFi sensing technology based on the DL algorithm:(1)An WiFi sensing platform was built and small-scale respiratory and large-scale human activity datasets were constructed.Based on the Intel 5300 NIC and dedicated data acquisition tool CSI Tool,the constructed platform was used to collect small-scale respiratory and large-scale human action data respectively.Among them,four types of respiratory actions with a total of 160 minutes of respiratory data and seven types of large-scale human actions with a total of 2450 action data were included.(2)For small-scale respiratory pattern sensing,a recognition method based on the Bidirectional Long Short-Term Memory(BiLSTM)neural network is proposed.The proposed method utilizes the bidirectional feature extraction capability of BiLSTM to achieve the extraction of temporal correlation features of respiration data,to effectively classify different respiration patterns.Experimental results show that the proposed method achieves 98% recognition accuracy for four respiration patterns,which is 20%higher than the traditional ML-based method and 1.5% higher than the classical LSTM algorithm.Meanwhile,the proposed method utilizes the ubiquitous WiFi signals in indoor and outdoor environments to achieve respiration sensing,which greatly reduces the application cost compared to the sensing methods through dedicated medical devices and wearable sensors.(3)For large-scale human activity sensing,CNN-Attention-BiLSTM(CNN-ABiLSTM)and CNN-Attention-BiGRU(CNN-ABiGRU)based recognition methods are proposed,respectively.The proposed method addresses the problems of insufficient spatial feature extraction capability of LSTM and GRU and poor modeling capability of CNN for temporal data,combining the advantages of CNN spatial feature extraction and the advantages of temporal data modeling of BiLSTM network and Bidirectional Gate Recurrent Unit(BiGRU)network,supplemented with attention mechanism,to extract rich action spatiotemporal features.The experimental results show that CNN-ABiLSTM achieves 96% recognition accuracy for seven large-scale human actions,while CNN-ABiGRU achieves 97% recognition accuracy due to the simplified gating mechanism of GRU,and CNN-ABiGRU improves recognition accuracy and reduces time complexity at the same time.Compared with the classical DL algorithm,the recognition accuracy of both network architectures is improved by1%-20%.