Desenvolvimento de um sistema de posicionamento interno por radiofrequência utilizando aprendizado de máquina

Abstract

Parallel to the rapid development of wireless communication systems in recent times, an increase in the demand for positioning systems is also observed. To meet this demand, several solutions have been introduced to the market, such as the Global Positioning System and Beidou, which are capable of locating devices on most outdoor sites. However, these satellite-based solutions are not sufficiently accurate, making it necessary to develop alternate techniques for this purpose. Among these, one promising technique is the radiofrequency-based location system, which can use the existing transmission infrastructure of signals that permeate indoor areas, like Wi-Fi and Bluetooth, to detect the position of a mobile device using different metrics, like channel state information. Nonetheless, radio channels are naturally volatile and hard to model analytically, because of the several effects that influence these signals. One possible way to tackle this problem is the use of machine learning, that aims to find a statistical model to relate the mobile device position with empirical channel observations. Therefore, this work’s goal is to implement an indoor positioning system using channel state information and regression models based on Support Vector Machine, k-Nearest Neighbor and Random Forest algorithms. For each of these algorithms, 192 different models were trained using numerically simulated data, varying the number of channels used and the central frequency valued, in order to find the best combination among these quantities. The best simulated model, based in kNN, had a root mean squared error of 0.781 m and a variance of 0.298 m2, spending, on average, 0.260 ms on each prediction. The best physical model, on the other hand, got a root mean squared error of 0.340 m, a variance of 0.0615 m2 and an average time per prediction of 0.418 ms.