Research On Indoor Positioning Algorithm Based On Extreme Learning Machine In Incomplete Data Sets

With the continuous development of machine learning,the problem of fingerprint localization can be formulated as machine learning problem.Since the extreme learning machine(ELM)algorithm has the advantages of fast learning speed,low computational complexity and good generalization performance,the ELM based localization algorithm has been widely studied.However,in indoor localization scenarios,the received signal strength(RSS)measurement in the fingerprint training data is vulnerable to attacks by the malicious nodes.Moreover,due to the large costs of off-line database collection,the number of training data is often less.Therefore,in this paper,the ELM based localization algorithms which can give the solutions of the above two problems are proposed.The research contents are as follows:(1)First,the system model and working principle of fingerprint location technology are introduced.Then,the traditional fingerprint matching algorithms are described.Next,the related theories of the extreme learning machine are studied and the knowledge of activation function is also described in detail which lays a theoretical foundation for the following research work.(2)Under the malicious node attack condition,a secure location algorithm based on online sequential extreme learning machine(OS-ELM)and hierarchical clustering technique is proposed.In the off-line phase,the attacked training data samples can be identified by the hierarchical clustering technique.Then,the OS-ELM method is used to train the relationship between the received signal strength measurement and the position information by the unattacked training data.Finally,the position regression function is obtained.In the on-line phase,when the received RSS measurement is obtained,the final position can be estimated by the position regression function straightly.Since the hierarchical clustering technology is used to identify the attacked training data samples,the off-line learning can be improved.Moreover,by the online learning ability of the OS-ELM technique,the accuracy of the on-line position estimation can be increased.(3)Under the small number of training data condition,multiple activation function based ELM localization algorithm is proposed.In the off-line phase,in order to increase the nonlinearity and flexibility,a multiple activation function that combines multiple types of activation functions is proposed for training.Meanwhile,the cross validation technique is introduced to obtain the optimized weight parameters of each single activation function and can improve the learning capacity and generalization ability.In the on-line phase,after the feature mapping into the obtained multiple activation function,the final position can be estimated with the received RSS measurements.Moreover,the training error of the proposed algorithm is theoretically analyzed.The theoretical derivation gives the upper and lower limits of the positioning error of the proposed algorithm.So,the superiority of the localization performance of the proposed algorithm is proved.