Performance of Cellular-based Positioning with Machine Learning

Abstract: From 1G to nowadays 5G, the development of cellular systems stirs more demands and expectation for high quality of cellular-based positioning service. Precise positioning (also called localization) is one of the key applications of 5G and beyond. time of arrival (ToA) techniques is one of the main positioning techniques utilized today. However, in the process of positioning, the noise resulted from various factors definitely will affect the accuracy and reliability of positioning service within a cellular system. Machine learning is gradually reported as one of the useful and significant ways to reduce the noise, and there are various machine learning algorithms related to cellular-based positioning. Both Convolutional Neural Network (CNN) and denoising autoencoder (DAE) have been proven that can improve the positioning accuracy, and these two algorithms aims to regenerate the distance between user equipment (UE) and base station (BS) directly after training. However, the distance information are with more feature that may be hard to be regenerated directly. A new method aims to extract the noise data at first and subtract the noise data from the noisy distance data to obtain the distance between UE and BS, which is called noise learning based algorithms. The corresponding noise learning based denoising algorithms of CNN and DAE are called Denoising Convolutional Neural Network (DnCNN) and noise-learning based denoising autoencoder (nlDAE). In this project, we evaluate the performance of these four algorithms in different noise environments in ToA based and cellular based positioning systems. The hyperparameters of each algorithm in different noise environment are searched by keras tuners. The results show that these four algorithms significantly improve the positioning accuracy, and the regular denoising algorithms have the same positioning performance with the noise learning based denoising algorithms with the hyperparameters searched by keras tuner.