by João Gante, Gabriel Falcão (University of Coimbra) and Leonel Sousa (INESC-ID)
Despite being available to civilians since the 1980s, the Global Positioning System is still the standard method for positioning. While unquestionably precise enough for most uses, it requires a dedicated antenna and a significant amount of energy from mobile devices. Using 5G’s millimetre wave networks and machine learning, our work shows that we can obtain similar accuracies without these drawbacks.
The advent of 5G is expected to bring new wireless communication capabilities, but it will be accompanied by other challenges. One of 5G's highlights is the introduction of millimetre wave (mmWave) communications, defined by the use of wavelengths between one and ten millimetres, unlocking a significant block of untapped bandwidth. However, with mmWave transmissions, the propagation properties change dramatically: the resulting radiation not only has severe path loss properties, but also reflects on most visible obstacles.
In wireless communication systems, the physical size of the antennae is proportional to the selected wavelength. Therefore, for mmWave systems, whose wavelength is much smaller than most system’s, mobile devices can opt between spending less volume or having more antennas. If they follow the latter option, they will have access to beamforming (BF), a signal processing technique that can counteract the aforementioned mmWave drawbacks through a steerable and directive radiation pattern. In fact, when line-of-sight (LOS) communications are unattainable, the focused beam can be aimed towards obstacles, such that its reflection reaches the desired target. As a consequence, multiple signals can co-exist in the same frequency band with reduced interference levels, with careful beamforming execution. No wonder mmWaves are so desirable, at least in theory – they belong to an underutilised part of the spectrum where high levels of spatial multiplexing are possible.
The recent focus in mmWave communications also led to the proposal of new positioning systems. The accuracy achievable in certain conditions is remarkable, having sub-metre precision in indoor and ultra-dense LOS outdoor scenarios. Nevertheless, to be broadly applicable to outdoor localisation, a positioning system must also be able to accurately locate with devices in non-line-of-sight (NLOS) locations, using a limited number of base stations (BS). These requirements, allied to multiple, often overlapping non-linear propagation phenomena such as reflections and diffractions, pose serious challenges to the traditional geometry-based positioning methods. In fact, recent mmWave experimental work conducted at New York University demonstrates that geometry-based methods cannot be directly applied to accurately locate NLOS targets, and thus new solutions are needed.
Our team at Instituto Superior Técnico (Lisbon) and Universidade de Coimbra (Coimbra) started addressing this problem in 2018 with the following question: if the BF process can deliver such spatial selectivity, can we use that selectivity to gather spatial information? In mmWave transmissions obstacle interactions are deterministic, resulting in some attenuation and in a change of direction, which impacts the distance travelled by the signal. Given that the time between transmission and reception depends on the distance traversed by the signal, and that 5G is designed for relatively short ranges, it becomes very difficult to sample the signal so as to identify individual interactions, resulting in a significant amount of lost spatial information. However, when BF is available, we can focus the signal in a particular direction, and isolate the interactions from that spatially selected transmission. Repeating this process so as to cover all possible transmission directions, the receiver is able to sample a set of information-rich signals, which we’ve termed “beamformed fingerprint” (BFF) [1]. With the availability of fingerprint data, machine learning methods and hierarchy techniques were proposed to infer accurate position estimates. Using a single BS, our method achieved Global Positioning System (GPS)-level results for single-point estimates (3.3 m), in a scenario containing mostly NLOS positions, providing positioning capabilities whenever there is mmWave coverage.
The goal of a positioning system is to estimate the position of a target, which is a direct consequence of its movement. The movement of a user, in turn, is limited by physical restrictions, such as velocity and acceleration, as well as human-made constraints, such as traffic rules. As a consequence, it is possible to leverage additional sources of information if sequences of positions are considered, as opposed to single-point estimates. In [2], we employed temporal convolutional networks (TCNs) when sequences of BFF are available to the system, effectively enabling the system to track a mobile device. Our work with TCNs [L1] achieved the state of the art for NLOS mmWave outdoor positioning, having an average estimation error as low as 1.78 m, with a root mean squared error one order of magnitude smaller than the second-best work for the same problem and, more impressively, more accurate than low-power GPS implementations.
If a positioning method is to displace the GPS as the default positioning method, it must boast similar accuracy levels and lighter hardware and energy requirements. In fact, being a 1980s technology that requires coordination with satellites, the GPS receivers are locked to specific frequencies, which require dedicated antennae, and have power-hungry signal processing requirements. Our most recent work [3] was aimed at answering these practical questions, regarding the BFF positioning system. In essence being a mmWave positioning system, it shares its hardware requirements with mmWave communications, having no additional requirements in mmWave-enabled devices. Finally, the proposed BFF positioning system is also 47 times and 85 times more energy efficient per position fix (for continuous and sporadic fixes, respectively) than low-power GPS implementations, as shown in Figure 1.
Figure 1: Average error vs average energy required per position fix for the positioning technologies discussed in this article. The proposed BFF positioning system has an accuracy comparable to low-power GPS implementations, while achieving energy efficiency gains exceeding 47× per position fix.
Having shown that 5G and machine learning methods can reach GPS-level accuracy levels with lower hardware requirements and much higher energy efficiency, our goal is to advance the industry towards a new paradigm – one that empowers smaller positioning-enabled devices and does not result in space debris.
Links:
[L1]: https://github.com/gante/mmWave-localization-learning
References:
[1] J. Gante, G. Falcao and L. Sousa: “Beamformed Fingerprint Learning for Accurate Millimeter Wave Positioning”, in IEEE VTC2018-Fall.
[2] J. Gante, G. Falcao and L. Sousa: “Deep Learning Architectures for Accurate Millimeter Wave Positioning in 5G”, in Neural Processing Letters.
[3] J. Gante, L. Sousa and G. Falcao: “Dethroning GPS: Low-Power Accurate 5G Positioning Systems using Machine Learning”, in IEEE JETCAS.
Please contact:
Gabriel Falcão, University of Coimbra, Portugal
Leonel Sousa, INESC-ID and Universidade de Lisboa, Portugal
