Distance Evaluation Between Two Embedded BLE Systems Using Machine Learning Methods

This article presents a new method for smartphone distance estimation using the Bluetooth Low Energy (BLE) standard (2.402-2.480 GHz) with machine learning method for a low-resource access controller from the brand AKIDAIA [1]. Using two types of antennas (Patch or Yagi), extensive Received Signal Strength Indicator (RSSI) collection was conducted. This data collection allowed us to generate two datasets with different characteristics, demonstrating the importance of dataset construction on overall accuracy with preprocessing. Subsequently, we generate two distance estimation learning methods and compared them to increase the global accuracy. This work aims to identify the most efficient method according to the dataset composition and according to the machine learning defined on a low-resource embedded device with limited computational and memory capacity. The best accuracy was obtained with the Random Forest learning machine, with an accuracy of 67.77% over a range of up to 10 meters using the IQR preprocessing algorithm and with one of the two datasets presented in this article.