Study On Wi-Fi Fingerprint Positioning Algorithm Based On Autoencoder

The scenes of modern society are more extensive and complex,and the demand for location services is increasing.However,GNSS positioning cannot receive positioning signals indoors.For modern large and complex indoor environments,other signal sources must be used for positioning.In this paper,the fingerprint localization technology based on Wi-Fi signal is studied in detail focusing on two aspects of fingerprint database,data preprocessing and fingerprint localization method.Based on the machine learning method,the fingerprint positioning is more accurate than the traditional method.The main research contents are as follows:Firstly,by pre-processing the collected Received Signal Strength Indication(RSSI)data and comparing the effects of the two filtering methods for processing RSSI data,the use of Gaussian filtering in twice is proposed,which can significantly reduce the fluctuation of fingerprint signal,the effect of noise and gross error measurements.The original data collected with the twice Gaussian filtering is averaged as the location fingerprint database,then cluster the fingerprint database.Based on the high dimensional characteristics of the RSSI fingerprint vector,a hierarchical clustering method based on adjusted cosine similarity is proposed and is used in fingerprint clustering.And the machine learning classification method is used to classify the online measurement points,and the online measurement points can be first divided into the sub-fingerprint database,and then the subsequent positioning processing is performed.To solve the problem of high dimensionality of location fingerprint database,autoencoder network is used to learn its main features.Through the research and analysis of several kinds of autoencoder,a stacked denoising autoencoder technology which is suitable for processing Wi-Fi fingerprint data is proposed.The optimal deep network structure is found through experiments,which makes the dimensionality of fingerprint database be reduced almost without loss.The softmax layer is used to convert the output of the autoencoder into the probability of the reference point's RSSI.Based on this method,a WKNN method is proposed to obtain the location result of the stacked denoising autoencoder.Finally,two machine learning regression methods,SVR and RF regression,are used to locate the online measurement points.The traditional KNN,WKNN and the proposed Sparse SDAE WKNN method are used to locate the online points.The two positioning modes are studied and compared: the "rough positioning and precise positioning" mode and the direct positioning mode.The experimental results show that the proposed Sparse SDAE WKNN method has the best positioning effect and higher precision.With two steps of "rough positioning and precise positioning",the positioning accuracy is much higher than the direct positioning mode.Two-step positioning mode is used for dynamic experiment positioning.With the experiments of direct Wi-Fi positioning and Kalman filtering positioning,the results show that the trajectory of Kalman filtering positioning is smoother,which can effectively solve the point overlap and jump phenomenon of Wi-Fi positioning trajectory and improve the accuracy of dynamic positioning.