Feature Optimization Integrated With Hybrid Regression Based Machine Learning Using Received Signal Strength Measurements For Indoor Localization

With the development of the Internet of Things technology and wireless communication technology,people's demand for indoor positioning services is increasing.The indoor positioning service is based on the data provided by the indoor location.The mobile terminal provides indoor positioning services to users by collecting information and matching a spatial fingerprint database composed of geographic location coordinate information and related location feature information.The indoor positioning problem based on the received signal strength indication(RSSI)can be transformed into a machine learning problem solving.Using RSSI as a data feature and position coordinates as training labels for training positioning models have been widely used in indoor positioning research.However,most of the existing indoor positioning solutions do not effectively deal with the fingerprint database.The disadvantages are as follows: directly reducing the data with PCA to reduce the number of invalid RSSIs in the training data;artificially deploying routers to collect RSSIs,not only has a large workload and cost High,there are also significant limitations in actual deployment.At the same time,most existing positioning systems are modeled using a single training model.In a large area,due to the different indoor environments in different areas,it is difficult for a single model to establish a high-precision positioning model for all indoor areas.Therefore,this thesis studies the establishment of the RSSI fingerprint database and the establishment of different regional positioning models.According to the current research status of indoor positioning at home and abroad,the main work of this thesis is:(?)Aiming at the problems of high acquisition cost when using RSSI fingerprint database and doped interference data during dimensionality reduction,this thesis proposes a feature optimization algorithm(RPCA)combining Relief F and PCA.RPCA uses the Relief F algorithm to assign weights to different features based on actual physical coordinates,sets a threshold value through training observations to dynamically screen features that are favorable for classification,and then uses the PCA algorithm to reduce the dimension of the filtered fingerprint database.The RPCA algorithm eliminates the need to deploy a router during the data collection phase,and only needs to collect the existing RSSI to build a fingerprint database,which greatly saves work costs and does not have to worry about the problem of interfering with data during the dimensionality reduction process.We set up an experimental environment on the fourth floor of library,Nanjing University of Posts and Telecommunications to test.The experimental results show that the RPCA algorithm proposed in this paper is applied to the fingerprint positioning algorithm,and the average number of fingerprint database features is reduced by 98%,the average positioning error is 1.50 m.Compared with the fingerprint positioning algorithm without RPCA algorithm,the average positioning error is 2.43 m and the average positioning time is 0.449 s.This method reduces the positioning error by 0.93 m and improves the operating efficiency by 62.14%.(?)In order to further improve the positioning accuracy based on the fingerprint positioning algorithm,this thesis makes three optimizations to the fingerprint positioning algorithm: First,in order to solve the problem of large randomness of the clustering results of the K-Means ++ algorithm during the offline region positioning phase,a pre-divided area and K are proposed.-Means ++ algorithm combined method to achieve unsupervised clustering,which solves the problem of large randomness of clustering results and difficult to distinguish data between different clusters.Then,for the problem of low positioning accuracy of a single training model,this paper proposes an indoor positioning algorithm based on hybrid regression(HRN)in the offline regression learning stage.HRN integrates two commonly used positioning algorithms: support vector machine(SVM)and k-nearest neighbor(k NN)algorithms.Different regions learn with multiple algorithms.By comparing the root mean square error(RMSE)of different algorithms in the region,the smaller the choice As a training model for the region.The HRN algorithm can better train positioning models suitable for different regions,and improve the positioning accuracy of the entire indoor.Finally,in the online test phase,this paper designs a fingerprint scoring system(FSP).The FSP determines whether the previously trained model is valid by calculating the matching degree between the test phase data set and the offline training data set.If it is valid,it locates it,and if it fails,it Data needs to be collected again for training,which improves the stability of the system.Through experimental tests,we reduced the average positioning error to 0.92 meters,and the average online execution time was 0.069 s.Compared with the traditional fingerprint positioning algorithm,the average positioning error is reduced by 1.51 m,and the program execution efficiency is improved by 84.63%.