| Indoor positioning has broad prospects for development in intelligent life.Fingerprint indoor positioning based on Wi-Fi has become one of the research hotspots in indoor positioning technical field due to its advantages of low cost and convenient deployment.The traditional Wi-Fi fingerprint positioning based on Receive Signal Strength(RSS)is limited by the instability of RSS,which makes the improvement of positioning accuracy a bottleneck.Channel State Information(CSI),as a finer-grained fingerprint feature,can be easily obtained and utilized by widely used wireless network cards.To further improve the accuracy of Wi-Fi fingerprint indoor positioning as well as reduce the complexity of positioning system deployment,utlize Wi-Fi wireless chanel CSI as fingerprints,a high-precision CSI fingerprint indoor positioning method was studied by adopting deep learning into indoor positioning and using deep learning to fully mine the fingerprint features of CSI.The main contribution is as fowllowing:(1)Aiming at the problems that RSS is easily affected by multipath signals,has instability and less signal characteristics,an experimental platform was built to collect the CSI of the Wi-Fi wireless channel,and further extracted the amplitude and phase data of the CSI.As the random distribution characteristics of the original phase data,phase calibration of linear transformation and phase difference were used to process the original phase.At the same time,to further improve the stability of the amplitude information,also adopted a method of constructing the amplitude difference.Through experimental analysis,the feasibility and effectiveness of indoor positioning using amplitude difference and phase difference as fingerprints are verified.(2)For the problems of poor convergence,low positioning accuracy and large network parameters of the Fully Connected(FC)neural network,designed and built a Convolutional Neural Networks(CNN)by combining the amplitude difference and phase difference of CSI as fingerprints for offline training.In the online stage,to make full use of the amplitude difference and phase difference,the amplitude and phase difference were weighted,and double sampling of fingerprints was used for online positioning estimation.To reduce the influence of noise information on positioning performance,a probabilistic index optimization positioning algorithm was designed.Finally,through experimental analysis,it is verified that the proposed positioning method can achieve high positioning accuracy.(3)In view of the large dimension of CSI fingerprint matrix and many data packets are required,a Convolutional Auto Encoder(CAE)was proposed to reduce the dimensionality of the CSI subcarrier amplitude difference and phase difference of a single data packet.By comparing the original CSI fingerprint and the fingerprint generated by CAE,it is verified that CAE has good fingerprint reconstruction ability,which reduces the number of data packets required for CSI training data,and therefore greatly optimize the number of training data and test data.Finally,with almost no loss of positioning accuracy,the double optimization of computational complexity and network parameters is achieved.(4)In order to improve the generalization ability of the fingerprint positioning model in different positioning scenarios,and reduce the workload required to update the fingerprint database,including the collection of fingerprint data and the training of the fingerprint model,using transfer learning combined with Dense Net network to expand and reuse the fingerprint model was proposed.The fingerprint database was reconstructed by using a well-trained fingerprint model,freezing part of the network layer,and fine-tuning another part of the network layer with a small amount of fingerprint data in a new scene.Finally,high positioning accuracy can be achieved while reducing the amount of fingerprint data and network training time. |
PDF Full Text Request