| With the rapid growth of the demand for location-based services in the indoor environment,WiFi signal fingerprint positioning(hereinafter referred to as fingerprint positioning),as a mainstream indoor positioning technology,is receiving more and more attention by more and more researchers due to its simple system equipment and high positioning accuracy.The quality and quantity of the fingerprint database have a crucial impact on the entire fingerprint positioning system.Many current studies still use Received Signal Strength(RSS)to build a fingerprint database.This fingerprint database has low precision and large granularity of area division,which cannot meet the needs of high-precision positioning.The accuracy of fingerprint positioning depends on the density of reference points.The higher the density,the better the positioning effect.The construction of the fingerprint database requires a lot of time and human resources to collect signal information from many reference points.How to reduce the cost of information collection and designing an efficient indoor positioning algorithm is an urgent problem that needs to be solved at present.In this paper,Channel State Information(CSI)of Massive Multiple Input Multiple Output(Ma MIMO)system is used as the original information for constructing the Wi-Fi signal fingerprint database.Compared with RSS,CSI has richer signal feature expression capabilities and can provide detailed information at the subcarrier level.In this paper,Fourier operation is used to extract the time domain and frequency domain information contained in CSI,and combined with the multiple signal classification(MUSIC)algorithm to obtain the angle of arrival(AOA)of the signal to jointly build a fingerprint data.The fingerprint database construction method can express the characteristics of the signal to the greatest extent and improve the expression ability of the fingerprint database.In order to solve the problem of time-consuming and laborious fingerprint collection,this paper proposes a new generative adversarial network model(DS-WGAN).In order to adapt to the CSI fingerprint database with a large amount of information,this model draws on the construction method of Dense Net based on the Wasserstein GAN(WGAN),and replaces the network of the generation model and the discriminant model with the Dense Net module.Experiments show that the model can generate a large amount of data similar to the initial fingerprint distribution in a short period of time,and the diversity among the data is high.In addition,a deep learning-based fingerprint localization algorithm MS-CNN is proposed in this paper.Based on the Inception network,the network uses multi-scale convolution to extract the features of CSI information from different receptive fields,at the same time,a one-dimensional convolution kernel is used to extract the antenna and sub-carrier characteristics of CSI respectively,and using skip connections to reduce error propagation during network training.Experiments show that this method is significantly improved compared with other indoor positioning algorithms,and the mean square positioning error can reach up to 49 mm. |
