Research On Localization Fingerprinting Method Based On ESN For Wireless Sensor Network

The localization fingerprint based indoor positioning methods based on wireless fidelity(Wi-Fi) can effectively solve some questions. For example, the question that the propagation of global positioning system(GPS) signal is susceptible to buildings' blocks, and the question that the costs on positioning techniques is expensive. At present, among the Wi-Fi based indoor positioning methods, computational intelligence methods, such as k-nearest neighbors(KNN) method and support vector machine(SVM) method, have higher localization accuracy.As a new dynamic recurrent neural network(RNN), state reservoir(SR) of echo state network(ESN) has special echo state property(ESP), and just needs to train the output weight matrix.Thus ESN has strong dynamic approximation ability. And then studied about the ESN's regularizationed offline learning algorithm and recursive least squares(RLS) online learning algorithm. Aiming at the problem that the positioning accuracy is affected by the dynamic indoor environment and time-varying received signal strength(RSS) values, a Wi-Fi based indoor positioning methods using ESN is proposed. On this basis, considering preprocessing technology of feature extraction from all RSS values measured from Wi-Fi access points(AP),one kind of Wi-Fi based indoor positioning method using kernel principal components analysis(KPCA) and ESN is proposed, which expects to get better localization accuracy than ESN method, which named KPCA-ESN positioning method. In order to improve further the robustness to the time-varying RSS value and positioning accuracy.The research mainly includes the following several aspects:(1) Based on the theory of ESN, two kinds of ESN's learning algorithms are studied.Firstly, ESN method is based on regularizationed offline learning algorithm. Secondly, ESN method is based on RLS online learning algorithm. At the same time, the theory of localization fingerprint positioning is studied and the positioning performance of the present traditional positioning methods is analyzed, such as WKNN and SVM methods.(2) Two ESN methods in this thesis are combined with location fingerprinting based positioning model that realize two kinds of positioning methods, including regularizationed offline learning algorithm based ESN positioning method and RLS online learning algorithm based ESN positioning method. The proposed ESN methods are applied to indoor positioning instances based on Wi-Fi by simulation and physical environment experiments. Compared with the weighted k-nearest neighbors(WKNN), SVM, extreme learning machine(ELM)methods under the same condition, experimental results confirm that the proposed method is much higher in positioning accuracy than other existing methods, and can also automatically timely adapt to environmental dynamics.(3) The dynamic changes of the indoor environment and time-varying characteristics ofRSS values are analyzed, using KPCA method to preprocess for RSS fingerprint informations effectively, the localization methods based on ESN are put forward. KPCA method is used to extract the nonlinear principal component from the input of the model in feature space.Meanwhile, regularizationed offline learning algorithm and RLS online learning algorithm of ESN are used to establish the nonlinear mapping between positioning features extracted and physical locations. And then the proposed methods are applied to indoor positioning instances.Compared with the ESN methods under the same condition, experimental results show that the KPCA-ESN methods can attain higher positioning accuracy than ESN methods, and can also automatically timely adapt to environmental dynamics and time-varying RSS values.