Research On WiFi Fingerprint Localization Method Based On Deep Learning

Location Based Service(LBS)is widely used in navigation,smart home,service push and Internet of Things,and the demand for LBS is increasing.Although outdoor positioning technologies are relatively mature,indoor positioning still needs further research due to the complexity and variability of indoor environments.WiFi fingerprint-based indoor positioning technology is one of the mainstream positioning methods,but there are still shortcomings.In this thesis,relevant research is conducted to address the existing problems,and the main work of the thesis is as follows:(1)To address the problem that too many WiFi Access Points within a large indoor positioning environment lead to excessively dimensions of fingerprint data,and the fluctuation of WiFi signal affects the accuracy of indoor positioning.This thesis proposes a robust indoor positioning algorithm based on an improved Stacked Denoising Autoencoder(SDAE).The method uses data augmentation and simulates signal fluctuations by adding reasonable noise to the SDAE so that the fingerprint data covers the range of signal fluctuations at the reference point and facilitates the neural network to learn the mapping relationship between fluctuating WiFi signal and coordinates.Finally,the dimensionality reduction of fingerprint data and the extraction of robust signal features are realized.Meanwhile,the average fingerprint data of the reference point is used as the supervision of SDAE,so that the features extracted by the encoder are mapped to a similar low-dimensional space.Experimental results on public datasets show that the robust indoor positioning algorithm approach based on the improved SDAE proposed in this thesis is effective.(2)For the problem of covariance shift caused by the large interval setting of reference points in a large indoor localization range,this thesis proposes a reference point classification method based on fingerprint sequences.The effective signals are sorted according to the signal strength to form a fingerprint sequence,which converts indoor positioning into a reference point classification problem.The sorted formed fingerprint sequence has low dimensionality and is naturally robust to signal fluctuations,which can mitigate the negative impact of the covariance shift problem.In this thesis,a Long Short-Term Memory neural network with attention mechanism and a sequence classification model based on multi-headed attention mechanism are constructed to process fingerprint sequences from both one-dimensional sequence classification and Natural Language Processing perspectives.The experimental results on public datasets are compared to demonstrate the effectiveness of the method.(3)A multi-floor positioning system is designed by combining the above methods.The system can realize floor positioning and coordinate positioning,and the system has a high response speed and can realize real-time positioning.Aiming at the problem of high maintenance overhead of the fingerprint database,the adaptive fingerprint collection function has been added.