Research On A New CSI Indoor Positioning Method Based On Convolutional Neural Network

Location-based services have greatly facilitated our lives.Relevant research has been valued by researchers at home and abroad.In recent years,localization systems have begun to utilize channel state information(CSI).CSI contains rich multipath information and is more stable.How to make full use of the CSI to establish an indoor localization system with higher positioning accuracy,lower computational cost and stronger environmental adaptability has extremely high research value.In previous CSI-based localization systems,combining deep learning techniques has become a trend.The use of images transformed from CSI data and Convolutional Neural Network(CNN)to establish a localization system is a very innovative way,but the challenges and opportunities coexist,mainly lie in:(1)Whether the features extracted from CSI are stable and reliable(2)How to ensure that the network can adapt to complex environments and ensure high recognition capabilities.(3)At the stage of position estimation,the system can flexibly select suitable candidate points according to different scenarios.Based on the above background,in order to improve the positioning accuracy and reduce the computational complexity,this paper proposes a new indoor localization system based on convolutional network.The main research contents of this paper are summarized as follows:1.First to fourth central moments of CIR amplitude distribution are utilized to increase the data dimension and improve data stability by calibrating the phase.2.The new construction of multi-dimensional image is proposed to convert multi-dimensional CSI features into images.The mapping between location and fingerprint is done using a convolutional network.The new construction is smaller in size but contains richer data features,reducing computational time and improving the classification capabilities of convolutional network systems.3.This paper analyzes the design requirements of CSI-based localization systems and determines system structure and process design.In order to solve network underfitting,we expand the training set by using the mobile robot with low speed controlled by Arduino to collect CSI data in an area.We collect diverse data to build a training set for convolutional networks.4.The multi-image localization algorithm based on spectral clustering is proposed.In this paper,the spectral clustering in graph theory is introduced into an indoor positioning system,and a method of dynamically selecting k values is established to improve localization accuracy and efficiency.This paper establishes a localization system,and collects measured data to verify the feasibility of the proposed method.