Research On Multi-floor Location Algorithm Based On Extreme Learning Machine Theory

Since its better performance in training speed,prediction accuracy and generalization ability,extreme learning machine(ELM)based on machine learning algorithms have been proposed for indoor localization.However,most of the existing research works are based on 2 dimension(2D)indoor environments.In this paper,we will extend the application scenarios and study the ELM based localization techniques in indoor multi-floor environments.The main contributions are as follows:1.The problem of localization in the multi-floor environment is formulated as a machine learning problem which is solved with the ELM based learning technique.The proposed algorithm contains two main phase:(1)coarse phase: Floor localization and(2)refined phase: position estimation.First,the floor localization and the position estimation are transformed as a classification problem and the regression problem,respectively.Then,the ELM based techniques are used to solve the above two machine learning problem and obtain the final target coordinates.2.For the floor localization,in the off-line phase,the principal component analysis(PCA)is proposed training data pre-processing at first.It can extract a set of linearly uncorrelated variables from the correlated RSSI fingerprints using orthogonal transformations.Since the dimension of training data is reduced,the computational loads of off-line learning are decreased.Then,the ensemble ELM learning by combining multiple individual ELM learning is proposed for classification learning and obtaining the floor classification function.Because it can achieve better generalization performance than single ELM learning,the learning performance can be improved.In the on-line phase,after the PCA pre-processing of the received RSSI measurement,the floor level can be estimated with the classification function.3.For position estimation,in the off-line phase,in order to obtain better learning performance,the training data set is classified into some training data subsets at first,according to the geographical location information.Moreover,the center of each training data subset is also obtained by the K-means clustering algorithm.Then,the PCA approach,as a mathematical tool for dimensionality reduction,is used for RSSI training data subset pre-processing which can reduce both the noise and redundancy.At last,each training data subset is used for position regression learning in turn and obtains the position regression functions.In the on-line phase,through the distance comparison between the received RSSI measurement and the center of each training data subset,the regression function corresponding to the training data subset with the smallest distance is chosen for final position estimation.4.An RSSI fingerprint collecting software based on the Android system was developed to describe the performance of the proposed algorithm.The field test shown that for floor localization,the proposed algorithm can reach or even exceed the performance of the traditional high-cost approaches,such as KNN method,K-means method.Because of the PCA technique and training data set partition pre-processing,the performance of position estimation of the proposed algorithm improves dramatically than traditional ELM based techniques.