Research On Wireless Sensing Based On Machine Learning

Wireless sensing is a novel technology which could mine information such as identity information,gesture information and activity information from the ubiquitous wireless signal by collecting,processing,and analyzing it.Wireless sensing technology could achieve centimeter-level perception of targets without any additional equipment owing to the physical properties of wireless signal,and it has the advantages of easy deployment,high security,and non-invasion of human privacy and could improve quality of human life and production efficiency.With the popularization of wireless network devices,the available wireless signal data shows a blowout growth,and the data-driven wireless sensing technology based on machine learning has gradually developed.In the first place,the sensitivity of wireless signal to the information in the deployment environment allows wireless sensing based on machine learning technology to sense the subtle movements of the target.However,wireless signal is too sensitive,and various noise is involved and it is difficult to extract features with high performance.We propose a wireless sensing algorithm based on intrinsic feature which separates the fluctuation trend related to target information from the original wireless signal to construct the eigen-signal,and extracts features with good characterization ability to achieve high-performance sensing of target information.In addition,the construction of wireless sensing models based on machine learning often requires the participation of a large amount of wireless signal data,and when a new sensing problem is encountered,large amounts of wireless signal data is required to be collected.We propose a sensing algorithm based on meta-learning,which makes the knowledge learned by the sensing model achieve not only high-performance sensing but also good transferability.The algorithm could realize the reconstruction of the wireless sensing model with a small amount of training data,which saves the data efforts and improves the efficiency of solving wireless sensing problems.Finally,the sensing model often faces changes of working environment in practical applications,and changes in the working environment will cause the features extracted by the wireless sensing model to shift,which will cause the wireless sensing model to make incorrect classification for the target data near the decision boundary.In response to this problem,we propose a scenario-adaptive wireless sensing algorithm,which utilizes dragging distribution strategy and dragging center strategy to align the feature distribution of source scenario and target scenario,and achieve cross-scenario wireless sensing in an unsupervised manner.We design gait identity experiments,gesture recognition experiments and activity recognition experiments to evaluate the performance of wireless sensing algorithm based on intrinsic feature,wireless sensing algorithm based on meta-learning and scenario-adaptive wireless sensing algorithm,and the results show that the three algorithms proposed in this paper could extract wireless signal feature the high-performance,rebuild a sensing model with small training effort and achieve the unsupervised cross-scenario wireless sensing efficiently.