Multi-Physics Modelling of Autonomous Plume Tracking in Urban Environments

Abstract

This project examines the problem of tracking the sources of airborne pollutants (plumes) in urban environments, using aerial robots equipped with pollution sensors. This is seen as a physical process whose outcome depends on the underlying relationships between the aerial robots hardware, the environment and the plume tracking algorithms. Our primary aim is therefore to understand these relationships and how they may be harnessed to improve plume tracking performance. This work emphasises the role of the aerial robot's energy consumption as a constraint on the plume tracking process. We develop a state-of-the-art plume tracking simulation environment, which employs existing wind and pollution models to generate fast, realistic representations of polluted urban environments. By utilising this multi-physics simulation environment, the reduced experimental costs enable tackling research questions that would be too costly using physical models. Firstly, this work deals with the effects of obstacle representation on plume tracking performance. We identify significant effects related to the disturbed pollution distribution and forbidden zones, with changes in success rates of up to 0.71 and 0.46, respectively. Secondly, we explore enhancing the performance of plume tracking aerial robots through wind-aware energy-minimising path planning. We find that wind-aware aerial robots may offer travel cost savings of the order of 2% even in low environmental wind speed conditions. However, in the context of a plume tracking algorithm based on Particle Swarm Optimisation, wind-aware agents do not lead to noticeably higher success rates compared to wind-unaware agents, due to inadvertent increases in the swarm's wait time. Thirdly, we describe plume tracking as a scalable process, identifying dimensionless parameters that predict and explain its scalable behaviour. This dimensional analysis offers practical benefits by reducing experimental time and costs, achieving dimensionality reduction from O(n^5) to O(n^2).

Date of Award	2 Dec 2024
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Ben Parslew (Supervisor) & Peter Hollingsworth (Supervisor)