Research On Indoor Positioning Method Based On Machine Learning

In recent years,the development and popularity of the Global Positioning System(GPS)can actually basically meet the needs of the vast majority of people for outdoor Positioning services.However,people spend most of their time indoors,and GPS can hardly meet the needs of indoor positioning.Under this background,more and more scholars begin to study indoor positioning technology.Among them,indoor positioning technology based on Wi Fi is gradually known and recognized by people due to its low cost and accurate positioning,and is gradually put into practical application scenarios.In this paper,the current mainstream indoor positioning algorithm is analyzed from the perspective of ranging and non-ranging,and the positioning method based on location fingerprint is studied at last.This method is based on the establishment of fingerprint database.In order to ensure the authenticity of data,this paper does not use the simulation data,but uses the Wi Fi module ESP8266 as the access point(AP)point.The signal strength from each access point at each designated location is collected as the original fingerprint through the Android application written by myself,and the collected data is preprocessed to deal with the gross error.Finally,the fingerprint database is constructed according to the different algorithms.Secondly,this paper uses the WKNN algorithm with adaptive K value and the traditional SVR algorithm to achieve indoor positioning.The preprocessed data are randomly sampled,and 75% of the data are selected as training data for regression training.The remaining 25%was used as test data to test the model,and the cumulative error was calculated by comparing the test results with the actual results.Experiments show that the positioning accuracy of the adaptive K value WKNN algorithm is higher than that of the SVM algorithm in this particular scene.Thirdly,the convolutional neural network structure designed in this paper is used to train the data and save the model.The input layer is the signal strength value collected at each grid point,and the position label of the point is taken as the classification result,that is,the output.Sixty-four filters are defined in the middle convolution layer for feature extraction.In order to prevent overfitting,the Dropout layer is added to reduce the sensitivity of the network to data.Compared with the WKNN algorithm and SVR algorithm with adaptive K value,the accuracy of the location predicted by this network model is significantly improved.Finally,in order to further improve the positioning accuracy,this paper will through adaptive WKNN algorithm of K value and it is concluded that the positioning of the convolutional neural network coordinates results weighted,it is concluded that the new positioning coordinates,finally from the two aspects of the error mean and variance to separate two kinds of algorithm and a comparison of two algorithms after combination algorithm.Experimental results show that the accuracy of weighted positioning by using the WKNN algorithm with the adaptive K value is 49% higher than that obtained by the WKNN algorithm with the adaptive K value alone,and 14% higher than that obtained by the WKNN algorithm with the convolutional neural network alone.