A Study Of The Positioning Method Driven By Machine Learning Of Urban Road Test Data

With the large-scale development of the Internet technology,the "Internet plus" action merging the Internet into traditional industries,provides a new opportunity for economic development.At the same time,the demand of high precision and low cost localization is more urgent.Having studied a variety of traditional localization methods,including Wi Fi localization,RF localization,Bluetooth localization,and fingerprint location without additional hardware support,One of the most important issues affecting accuracy is found.Most of the traditional location technologies rely on signal propagation,but the signal is unstable and changes over time,and machine learning just makes up for that.Machine learning has a long history,but the development of it is slow before.Because it has a huge amount of calculation,and computer performance is limited.With the emergence of high-performance computers,it has made it possible to achieve fast implementation,and has also made good applications in artificial intelligence,image processing,face recognition and other fields,making machine learning the most popular research direction at present.The use of machine learning algorithms to achieve positioning is clearly becoming a new direction of wireless localization technology.In this paper,we study Deep Neural Networks(DNN)algorithm,Support Vector Regression(SVR)algorithm and k-Nearest Neighbor(k-NN)algorithm,Combining with the characteristics of urban test data,we propose three different localization methods.The original urban test data was pre-processed,including raw data cleaning,and data normalized.A deep network called Stacked Denoising Autoencoder(SDA)is proposed to determine the initial parameters of DNN algorithm,improving the robustness of the algorithm.The accuracy of DNN algorithm is 54.1m and the confidence is 79.4%.The accuracy of SVR algorithm is 180 m and the confidence is90%.The accuracy of k-NN algorithm is 20 m and the confidence is 89.26%.Compared with the results of the three algorithms,the accuracy of the SVR algorithm is too poor,the k-NN algorithm has certain promotion value.However,although the DNN algorithm has poor result,there is still much room for improvement of the accuracy and confidence,which is the focus of the follow study.