Research On WiFi Signal Indoor Localization Based On Neural Networks

With the advancement of wireless communication technology and the development of smart terminals,the demand for location-based services is also increasing,and the precise positioning of the target is the key to the realization of the above-mentioned various location-based services.WiFi devices are very popular in various indoor occasions and do not require additional device support,making WiFi-based indoor positioning systems a research hotspot.The commonly used research indicators of WiFi signals include Received signal Strength Indication(RSSI)and Channel State Information(CSI).Based on these two indicators,many positioning algorithms have been developed,which can be roughly divided into fingerprint identification positioning method and ranging positioning method.Among them,the fingerprint database positioning method needs to select a huge number of reference points in the environment in advance,collect its signal-related indicators(RSSI/CSI)and the corresponding plane coordinates and store them in the database,but the subsequent data maintenance and update is more complicated.A series of mathematical models need to be constructed in the ranging and localization method,and the indoor environment is complex,so it is difficult for the constructed model to express it accurately.The neural network can theoretically approximate any form of function model,so the neural network can be used to replace the complex model that needs to be constructed in the process of ranging and positioning.Based on this,this thesis implements an indoor positioning method based on WiFi channel state information combined with neural network.Firstly,the received signal is preprocessed.Transceiver equipment has hardware errors,which will inevitably cause amplitude or phase deviation in the process of signal generation,propagation and reception,so this thesis firstly performs linear calibration on it.The input to the neural network is then constructed using the calibrated data.The specific construction method is as follows: The type of CSI data obtained from the receiving antenna is a complex matrix.In this thesis,the real and imaginary parts of the original CSI complex matrix and its conjugate matrix are extracted and recombined to obtain a real matrix.On the basis of this real number matrix,this thesis designs a moving window interception method to construct a multi-channel input matrix.There is a certain mapping between the CSI of the received signal and its angle of arrival(AOA)and time of flight(TOF),so this thesis designs a multi-scale convolution kernel parallel neural network named MKCNN to simulate this mapping relationship.The AOA and TOF of the arriving signal are jointly estimated.This network is based on the Inception structure,and the convolution kernels of different scales can extract the AOA and TOF related features under different sizes of receptive fields from the input matrix.Then,the position probability map of the target is generated by using the joint estimation result to represent the probability of the target appearing at each point,and the position with the highest probability is taken as the pre-estimated position.In order to use of the location information of the past time,this thesis uses the LSTM network to predict the location of the current time according to the location of the past time,and the weighted average of the prediction result and the pre-estimation result is used to obtain the final location estimate,so as to achieve positioning tracking.In order to verify the effectiveness of the proposed method in this thesis,This thesis verifies the method in this thesis from multiple angles.The results show that in a complex indoor environment with 4 APs with a size of 16m×10m,the positioning accuracy can reach the decimeter level,and the positioning error is within 0.8m of coordinate data accounted for 90%.