Research Of Indoor Fingerprinting Localization Based On Deep Learning

The remarkable development in mobile Internet and mobile devices has fostered the flourish of indoor location-based services applications,and then strongly drive the development of location determination technologies.WLAN fingerprinting-based indoor localization has become a hot research topic in recent years,due to the low cost deployment,existing infrastructure,ease of implementation and its promising performance.Fingerprinting localization usually consists of two basic phases: the offline phase which is for fingerprint database construction and the online phases in which positioning systems estimate the position of the mobile device.Received signal strength(RSS)and channel state information(CSI)are widely used in fingerprinting positioning systems.Many existing indoor localization systems use RSS as fingerprints due to its simplicity and low hardware requirements.However,RSS is a coarsegrained value,which not only fluctuates over time but also is not unique for a specific location due to rich multipath effects and shadow fading in indoor environments.Unlike RSS,CSI is fine-grained high-dimensional channel information in physical layer.Nevertheless,on one hand,the CSIs of different carriers have correlation and noise;on the other hand,the high-dimension characteristic of CSI improves the computational complexity of positioning algorithm.Motivated by above problems,we leverage a reference point(RP)clustering algorithm based on spectral clustering to exploit the correlation of wireless channels and a CSI dimension reduction and feature extraction method based on deep auto-encoder.Then,RSS and CSI are combined to enrich the fingerprint features of location.Furthermore,a positioning algorithm based on deep learning positioning is proposed to improve positioning accuracy and stability.The main contributions of this thesis are summarized below:(1)In order to improve the localization accuracy of 3D indoor localization,we propose a spectral clustering and weighted backpropagation neural networks(SWBN)method.To be specific,to exploit the correlation of wireless channels,RSS is first partitioned into several groups using a tailored spectral clustering based method.Afterward,the weighted BPNN algorithm is adopted for location estimation.Experimental results demonstrate that SWBN could reduce localization median error by 11.38% and the training time is significantly reduced by 41.48%,compared to neural networks based method.(2)To further improve the positioning accuracy,an indoor positioning algorithm based on deep learning combine with RSS and CSI leverages(DADN).To extract features hidden in CSI data and reduce the computational complexity of localization,a deep auto-encoder based CSI dimension reduction and feature extraction method is designed to encode CSI.Then,the fingerprint database is constructed combined with RSS and CSI code.Moreover,a positioning method based on deep learning is proposed to improve positioning accuracy and stability.The advantages of DADN are strong scalability(the fingerprint database is established based on AP to conveniently apply different AP selection methods to positioning algorithm according to actual scene),and transferability(the trained deep auto-encoder and deep neural network models can be fine-tuned when updating the fingerprint database),and high positioning accuracy(the experimental results have exhibited that the average positioning error of DADN is 0.72 m and the median error is 0.5m).(3)Targeting at the actual indoor positioning scenario in which there are a bundle of mobile clients with high concurrent positioning requests during a short time,we design and develop an indoor positioning system based on Netty,Spring Boot and Rabbit MQ framework.The system includes the server and the computer client,the design idea of which is modularizing the server by function and reducing the coupling of modules by using the system core scheduler based on asynchronous concept,to improve system performance and stability.According to simulate the concurrent indoor positioning scenario by computer clients,the results show that the system can support 1500 users' stable positioning services under the condition of 500 ms or 700 ms positioning request interval.In summary,the designed indoor positioning system can provide fast,efficient and stable positioning services for users.