Autonomous Monitoring of Contaminants in Fluids

The litigation and mitigation of maritime incidents suffer from a lack of information, first at the incident location, then throughout the evolution of contaminants such as spilled oil through the surrounding environment. Prior work addresses this through ocean and oil models, model directed sensor guidance and other observation methods such as satellites. However, each of these approaches and research fields have short-comings when viewed in the context of fast-response to an incident, and of constructing an all-in-one framework for monitoring contaminants using autonomous mobile sensors. In summary, models often lack consideration of data-assimilation or sensor guidance requirements, sensor guidance is specific to source locating, oil mapping, or fluid measuring and not all three, and data assimilation methods can have stringent requirements on model structure or computation time that may not be feasible. This thesis presents a model-based adaptive monitoring framework for the estimation of oil spills using mobile sensors. In the first of a four-stage process, simulation of a combined ocean, wind and oil model provides a state trajectory over a finite time horizon, used in the second stage to solve an adjoint optimisation problem for sensing locations. In the third stage, a reduced-order model is identified from the state trajectory, utilised alongside measurements to produce smoothed state estimates in the fourth stage, which update and re-initialise the first-stage simulation. In the second stage, sensors are directed to optimal sensing locations via the solution of a Partial Differential Equation (PDE) constrained optimisation problem. This problem formulation represents a key contributory idea, utilising the definition of spill uncertainty as a scalar PDE to be minimised subject to sensor, ocean, wind and oil constraints. Spill uncertainty is a function of uncertainty in (i) the bespoke model of the ocean, wind and oil spill, (ii) the reduced order model identified from sensor data, and (iii) the data assimilation method employed to estimate the states of the environment and spill. The uncertainty minimisation is spatio-temporally weighted by a function of spill probability and information utility, prioritising critical measurements. In the penultimate chapter, numerical case-studies spanning a 2500 km2 coastal area are presented. Here the monitoring framework is compared to an industry standard method in three scenarios: A spill monitoring and prediction problem, a retrodiction and monitoring problem and a source locating problem.