A Robust Indoor Localization Algorithm Based On Extreme Learning Machine Theory

Since the large-scale deployment of Wi-Fi networks in indoor environment and most of mobiles have the embedded Wi-Fi receivers,Wi-Fi indoor localization became a hot research topic in both academia and industry.With the development of machine learning,the problem of Wi-Fi localization can be formulated as a machine learning problem.As a new machine learning technology,extreme learning machine(ELM)has the advantages of fast learning speed,low computational complexity and good generalization performance.However,when the ELM theory is used for indoor localization,it will exist the over-fitting problem in off-line learning phase,the robust of the localization result becomes weak,especially,in abnormal training data condition.In order to overcome the above shortcomings of the ELM based localization algorithm,in this paper,a robust indoor localization technique is studied by the ELM theory.The main research contents include:(1)At first,the range-based Wi-Fi localization algorithms are summarized.Then the system model and operating principle of Wi-Fi signal based received signal strength indicator(RSSI)fingerprint localization are described.Next,the traditional fingerprint matching algorithm,such as K-nearest neighbor(KNN)and support vector machine(SVM)are described for both the off-line phase and the on-line phase.They will give a theory support for the following research work.(2)An ELM based localization algorithm using ridge regression technique is proposed.In the off-line phase,all the RSSI measurement of Wi-Fi signals are collected at different locations and then form the RSSI fingerprint-position training data set.In order to solve the problem of measurement noise,a robust localization algorithm using the ELM theory and ridge regression technique is proposed to obtain more stable prediction results and better generalization ability.The ridge parameter is obtained from the variance of the training error.In the on-line phase,the estimated position can be calculated with the obtained position regression function straightly.Because the deviation of the off-line training phase reduces,a more stable and accurate position estimation can be obtained by the proposed algorithm.(3)An ELM based localization algorithm using decision level fusion is proposed.In the off-line phase,the fingerprint data set preprocessing based on base station feature extraction is performed at first using K-means clustering method.Then several training subset with RSSI fingerprints from different base station class are formed.Then,several off-line regression learnings are performed with the training data subset by ELM method.By comparing the learning errors of each training subset,several position regression functions are obtained.In the on-line phase,after extracting the base station feature of the received RSSI measurements,the inter-mediate position estimations are calculated by the position regression functions.The final position estimation can be obtained by the decision level data fusion technique.Due to the different weights of each position regression function,the position estimation of the proposed algorithm is more accurate.Moreover,multiple off-line learning can alleviate the impact of abnormal training data set and improve the generalization ability of off-line learning.