Location fingerprinting for IoT systems using machine learning

Student thesis: Doctoral Thesis

Abstract

The Internet of Things (IoT) has evolved rapidly as the number of connected nodes continues to grow, projected to be in excess of Trillions worldwide by 2025. IoT enables a number of application and services that enhance the quality of life of citizens and business practice. The demand for IoT-like connectivity is set to continue; for example, the advent of LPWAN providing a combination of advantageous features such as long-range, low power connectivity gates the deployments of a range of hitherto costly implementations over extended areas of coverage.A spectrum of valuable real-world IoT applications such as tracking, are predicated on location information. However, the provision of a low power, cost effective engineered solution to provisioning location still remains a major challenge, especially within resource constrained IoT deployments. GPS-enabled solutions are power hungry and potentially prohibitively expensive within extensive IoT architectures. Furthermore, ranging-based network-centric methods lack accuracy because of the long distances subject to dynamically varying path characteristics and the ultra-narrow bandwidth. The prevailing state-of-the-art motivates investigations into low-complexity, energy-efficient technique for IoT node localisation.The Thesis presents an empirical investigation into the use of fingerprinting for IoT node localisation within a suburban region in Saudi Arabia subject to varying environmental conditions, ranging from clear sky to sandstorms. The approach is based on the use of Received Signal Strength Indicator (RSSI) within a LoRaWAN network setting.The performance of LoRa transmission as a function of varying coding parameters is determined. The RSSI data gathered during the characterisation phase is exploited to estimate locations of IoT nodes using location fingerprinting. More specifically, k-Nearest Neighbour (KNN) algorithms are used to develop a baseline location model.The accuracy of the LoRaWAN based baseline node localisation is enhanced through the use of Machine Learning (ML). RSSI ratios between pairs of Gateways in conjunction with kernel-based ML techniques - Support Vector Regression (SVR) and Gaussian Process Regression (GPR) – is proven to improve the node localisation models. Moreover, the impact of the kernel function on model performance is evaluated. Further, RSSI measurements at different spreading factors are combined to form more robust location features; two machine learning ensemble techniques - Gradient Boosting and Random Forest - are then employed to determine the impact on the accuracy of node localisation models using combined location features. Results indicate that ensemble-derived models improve accuracy compared to single regression tree methods. In addition, feature transformation is proven to be effective in improving localisation performance.Results confirm the feasibility of IoT network-derived localisation in sandstorm environments. Furthermore, it is demonstrated that the LoRaWAN spreading factor is central to optimising performance.
Date of Award4 Jun 2020
Original languageEnglish
Awarding Institution
  • University Of Strathclyde
SupervisorIvan Andonovic (Supervisor) & Robert Atkinson (Supervisor)

Cite this

'