A probabilistic framework for the assessment of robust autonomy in unmanned aircraft systems

In the near future, Uninhabited Aircraft Systems (UAS) will bring significant benefits to a rapidly growing number of operations beyond defence. When it comes to national security, UAS could be used for border patrol, policing, detecting trafficking and smuggling, rapid emergency response, search and rescue, bushfire and flooding monitoring, and early detection of threats to national assets that could be targeted by terrorism. In civilian applications, UAS have the potential to be used for airborne mineral explorations, crop-yield monitoring, traffic monitoring in transportation systems, digital mapping, inspection of power-grid lines, and data collection for environmental and climate research. Due to this potential, nations worldwide are seeking to develop further their UAS capabilities. Over the past ten years, research activities in industry and academia have been growing, and companies are designing and developing the next generation of UAS. A global market with a worth close to 50 billion $USD is projected within the next decade. There is, however, a fundamental challenge that needs to be addressed before we can realise the full potential of routine UAS operations. This challenge is the integration of UAS into the National Airspace System (NAS). This integration requires UAS to demonstrate levels of reliability and safety equivalent to current aviation standards. As the level of autonomy in UAS operations increases, UAS must be able to either continue operating in the presence of faults and anomalous conditions, or shut down safely, whilst not beaching a minimum level of safety - this characteristic is described as robust autonomy. Most of the recent research effort in UAS has been focused on improving robust autonomy as well as aircraft ability to conduct different missions. This thesis takes a different approach and focuses on how robust autonomy in UAS can be quantified so National Airworthiness Authorities (NAA) can certify UAS for operations in the NAS. A novel framework is proposed whereby characteristics related to missions, aircraft, and operational and weather conditions are used in a Bayesian probabilistic assessment of how the autonomy handles anomalous conditions. The end result is a set of predicted probabilities of satisfactory performance called Measures of Robust Autonomy. To demonstrate the proposed framework, a general reconnaissance mission of a UAS is considered and measures of robust autonomy in relation to actuator faults and autonomous decision-making about control re-configuration are computed. To achieve this, the design of a low-computational complexity autopilot with capability for actuator reconfiguration is considered. We propose a design for joint control allocation and motion control scheme whereby the actuator configuration is used to map actuator constraints into the space of the aircraft generalised forces. This allows one to constrain the output of the flight controller and consider an unconstrained allocation problem since the controller always outputs feasible commands. We develop tuning rules and demonstrate robust stability to parametric uncertainty. The controller is then combined with a diagnosis system, which informs the controller about actuator faults so it can adjust the constraints and trigger the re-configuration of the allocation problem. The fault diagnosis includes fault detection and isolation. The detection is achieved by injecting a signal in the null space of the actuator configuration matrix and testing residuals between aircraft pitch and roll rate and those predicted by low-order models. With such signal, the motion of the UA is undisturbed unless there is a fault. Once a fault is detected a dedicated test signal is used to isolate the fault. The NAA certification problem can be cast as a problem of decision under uncertainty. The solution of this problem requires the probabilities that capture uncertainty in the outcomes of missions and utilities that quantify the consequences of certification decisions. The results in this thesis provide a novel flexible framework for evaluating the probabilities required to solve this decision problem.