Research On Indoor Localization Algorithm Based On Extreme Learning Machine And Iterative Self-organizing Data Analysis Clustering

With the development of wireless telecommunication technology and mobile internet technology,location based services(LBS)becomes more and more important in many practical applications.However,satellite positioning system which has been widely applied does not perform well in indoor environments,because the signals can be blocked with the buildings.Thus,indoor positioning has become a hot research spot in recent years.Because of the high-density layout of the Wi-Fi node,the coverage of Wi-Fi signal in indoor environment is very large.Moreover,Wi-Fi receivers have been embedded into a large number of mobile terminals.Therefore,it is very easy to receive the Wi-Fi signals.As a new kind of feed-forward neuron network,the hidden node parameters of extreme learning machine(ELM)can be randomly given.As a result,ELM has high learning efficiency and a strong generalization ability.So,in this thesis,we will study the Wi-Fi based indoor positioning algorithms using the ELM technology.The research mainly includes the following aspects:(1)First,the fingerprint positioning model and working principle by the received signal strength indicator(RSSI)measurements have been studied.Then,traditional fingerprint matching methods are introduced.Next,some existing ELM based indoor localization algorithms are analyzed,which gives the foundation for follow-up research.(2)For the large-scale localization requirement,a new locational algorithm using multiple kernel extreme learning machine(MK-ELM)and iterative self-organizing data analysis(ISODATA)is proposed.In the off-line phase,the received RSSI measurements are pre-processed by the ISODATA technique and then the label of each RSSI measurement can be obtained.Thus,the RSSI-label training set and RSSI-position training subsets which are divided by different labels are formed.Then,the MK-ELM technique is used for both the classification learning and the regression learning by the above training data set.The RSSI measurement based classification function and position based regression functions are also obtained.In the on-line phase,after the RSSI measurement classification result of the received RSSI measurement,proper position based regression function is chosen for final position estimation.Because of the unsupervised clustering data pre-processing utilization,the proposed algorithm does not require the prior knowledge of measurement information and the operation is easy.Moreover,since the ELM technique and multiple kernel function are used for localization,the real-time capability and accuracy of the proposed algorithm can be improved.(3)A new indoor localization algorithm with received signal strength indicators fingerprints by extreme learning machine(ELM)technique is proposed,when the large amount of RSSI measurements are unlabeled.In the off-line phase,the ISODATA techniques algorithm method,as an unsupervised clustering analysis,is used to reveal the inherent nature of the RSSI measurement training data and obtain the class of each RSSI measurement.Then,in order to increase the classification capability and robustness,the multi-kernel ELM learning method which consider more than one kernel function,is proposed for the RSSI classification learning and obtain more accurate RSSI measurement classification function.Next,a two-stage RSSI measurement feature extraction algorithm is proposed.For each RSSI training subset.The kernel principal component analysis(KPCA)which can enable the linear operations in a high dimensional feature space is used to obtain coarse Wi-Fi feature at first.And then the deep learning network based ELM is introduced to get high level refined feature of RSSI measurement.At last,each training data subset with the refined RSSI feature is used for semi-supervised regression learning and obtain the position regression functions.In the on-line phase,after the RSSI measurement classification and feature extraction of the received RSSI measurements,the target position can be estimated with the corresponding position regression function.Compared with the single-kernel learning technique,the proposed algorithm has better regression capability and robustness.Moreover,the off-line learning performance of the proposed algorithm can be improved,since more efficient feature of the RSSI measurements can be obtained by the proposed two-stage feature extraction technique.