A GMM-Assisted Learning Approach for Ranging with IEEE802.15.4 devices

IEEE 802.15.4 is the core of multiple industrial standards such as WirelessHART, ISA100.11a, and 6TiSCH. It provides ultra-low-power operation and high reliability for battery-powered devices requiring a 5-10 year lifetime. However, due to its narrow frequency bandwidth, it is rarely used as a wireless ranging technique. Its low resolution in the time domain makes it difficult to precisely determine the arrival time of the signal, which is essential for Time-of-Flight ranging. Emerging approaches, such as WiFi fingerprinting, show promising localization performance by using learning methods to find the pattern mapping between signal characteristics and positions.This paper explores the potential of using learning approaches over IEEE 802.15.4 radios for ranging, based on RSSI and Round Trip Timing (RTT) samples. As a preprocessing step, groups of RTT samples are fed into a Gaussian Mixture Model (GMM), extracting two sets of features: the mean, variance, and weight of two distributions (line-of-sight and non-line-of-sight). Together with the mean and variance of the Received Signal Strength Indicator (RSSI) and RTT samples, 10 features are used as the input to a multilayer perceptron network to predict the distance between the requester and reflector.The approach has been evaluated using IEEE 802.15.4 devices in three different scenarios: an empty football field, an underground parking lot, and an office area. The results show that with dedicated models trained using datasets collected from each scenario, the ranging prediction errors of 0.25 meters, 0.42 meters, and 0.58 meters at 80% probability are achieved in these three scenarios, respectively. With a single model trained using a combined dataset from all three scenarios, prediction errors of 0.41 meters, 0.60 meters, and 0.97 meters at 80% probability are achieved.