Research On The Performance Problem Of Online Indoor Positioning Service In Large-scale Environment

With the popularization of smartphones and the advent of mobile Internet,online indoor positioning makes significant contributions on diverse fields like smart home,smart mall and public safety emergency response,by providing convenient navigation service and mining valuable business information from the positioning data.However,when it is used in high-speed railway stations,exhibition centers or other large environment,copious wireless access points and smartphone carriers will bring massive computing and storage pressure,such as large amount of high-dimension training samples and strong-concurrency data throughput,which make orthodox positioning algorithms and stand-alone server architecture demanding to sustain high quality service.Under these circumstances,the main work in this thesis aims at improving the algorithm performance for indoor positioning and the 10 performance for data storage model,which is arranged as follows:1)A subspace partition algorithm based on twice clustering is proposed.By aggregating the samples with highly analogial fingerprints to separate subspaces,the searching scope of samples in the process of seeking the nearest neighbor for kNN classifier can be decreased.In order to get high cohesion clusters,a center initialization method is put forward on the basis of the characteristics of indoor positioning data.Experiment results demonstrate that the total cohesion of the clusters generated by the improved algorithm is 18.7%higher compared to the original kmeans algorithm with randomly initalled cluster centers.2)An improved indoor positioning algorithm based on dimensionality reduction is proposed.Dimensionality reduction can be accomplished by removing the weak relativity items from the RSSI feature vector,considering the logarithmic normal distribution model established from the scanning frequency of access points,which makes kNN classifier calculate the positioning coordinates in low dimensional vector space consequently.Experiment results demonstrate that in this means the average dimension of the eigenvector is only 13%of the original sample.Furthermore,the improved positioning algorithm manifests a significant performance advantage and competitive positioning accuracy compared to the existed algorithm proposed in other papers.3)A Redis-MySQL mixed storage model is proposed.The model maintains more efficient data access by caching the hotspot positioning data in Redis,a high performance database based on memory,compared to traditional relational database.As for non-hot data,an asynchronous persistence store mechanism based on distributed message queue RabbitMQ is designed according to the producer consumer model,which decouples the positioning data persistence task from positioning server.Experiment results demonstrate that the speedup ratio of service response time on location query is 1.48.Simultaneously,the positioning server can save 90%blocking time and MySQL server can be protected from crash in case of data burst in a short time with the asynchronous mechanism.4)A Redis cluster mechanism on account of horizontal fragmentation strategy is designed.In this way different positioning data produced by different mobile terminals or environments will be routed to different Redis nodes by means of hash mapping,which achieves data query parallelly effectively.Experiment results demonstrate that the Redis cluster mechanism perfroms a good speedup ratio of service response time for front-end request in high-concurrency scenario.