Research On Indoor Localization Algorithm Of RSS Based On Deep Learning

Localization-based Internet of Things(IoT)applications have been developed and deployed,the indoor localization methods based on Wi-Fi(Wireless Fidelity)are utilized due to Wi-Fi’s wide distribution.Compared with indoor localization methods based on other Wi-Fi measurement methods,the indoor localization methods based on RSS can realize the location of low-cost equipment.There are many challenges in RSS-based indoor localization coming from multipath effects and noise,environment dynamics,device heterogeneity,and security.This thesis focuses on the deep learning-based RSS fingerprint methods.The main research contents and innovations are summarized as follows:1.In order to solve the problem that RSS time series is susceptible to multipath effect and sensitive to time-varying environment,an algorithm combining Kalman filter and deep neural network is proposed.Taking advantage of the time dependence of RSS sequence,this thesis presents a time-varying RSS filtering algorithm based on Kalman filter and refined post-processing,which can effectively suppress time-varying noise.Feature extraction based on Deep Neural Network(DNN)was carried out for one dimensional RSS data.On the BUPT dataset,compared with the traditional DNN algorithm,Kalman-DNN algorithm improves the localization accuracy at least 25%.When the average positioning time is 0.02 ms,the localization accuracy of Kalman-DNN algorithm is at least 10%higher than that of Kalman-CNN(Convolutional Neural Network)algorithm.2.To solve the problem that RSS data is susceptible to multipath interference and environmental dynamics,a CapsLoc algorithm based on capsule network was proposed.Among them,capsule network can effectively extract hierarchical structure from RSS data for indoor localization.Experimental results show that,On the BUPT dataset,CapsLoc can achieve indoor localization with an average localization error of 0.68 m,compared with traditional machine learning methods such as KNN and SVM and deep learning methods such as CNN and SAE-CNN(Stacked Autoencoder,SAE),its localization accuracy is improved by at least 37%.3.An EdgeLoc algorithm based on capsule network and edge computing is proposed to solve the problem that RSS data is easily affected by device heterogeneity.Capsule neural network can effectively extract incremental features from RSS data.Firstly,a multi-step data stream is designed to convert RSS fingerprints into image-like data,and the feature matrix is used to reduce the absolute sensing error caused by device heterogeneity.Secondly,an edgeIoT framework is designed to train the deep learning model on the edge server to realize the real-time localization of heterogeneous IoT devices.The experimental results show that on the BUPT dataset,the average localization error of the EdgeLoc algorithm is 0.68 meters,and the average positioning time is 2.05 milliseconds;on the UJIIndoorLoc dataset,the localization accuracy of EdgeLoc is improved by 14.4%compared with the SAE-CNN algorithm.4.Aiming at the AP(Access Point)attack problem,an indoor localization SE-Loc algorithm of RSS based on semi-supervised deep learning is proposed.AP selection method based on Pearson correlation coefficient is designed.And a semi-supervised deep learning network based on denoising autoencoder and convolutional neural network is proposed for robust feature learning and location matching.Experiments show that in the face of up to 100 AP malicious attacks,SE-Loc can still achieve an error fluctuation of 1.7m and an average localization accuracy of 8.9m on the UJIIndoorLoc dataset.