| dc.description.abstract | The use of autonomous systems such as Unmanned Aerial Vehicles (UAVs) in civilian and military operations has drastically increased over the last 10 years, requiring significant advances in health management technologies to maintain safe and reliable operations. Most of these technologies employ some form of fault-tolerant control, which seeks to alter the parameters of the controller or switch controllers to accommodate faults when they occur. These methods are focused on low-level control and are primarily used to maintain a safe trajectory. What they do not account for, however, is the health of the individual components and the subsequent degradation in the health of the overall system after the fault occurrence. This recognition and response to the degradation has been termed health-aware decision making, where component and system state of health information is used to generate safer and more reliable plans when faulty conditions occur during a flight. Operating UAVs in environments which are not fully observable poses unique challenges, especially considering the overall system health, environmental conditions, and unforeseen faults, all of which adversely affect system operations. In this context, we assume that the system is operating in a degraded but functional state capable of satisfying system-level safety and performance constraints, even after an abrupt fault occurs in the system. The combination of the fault and the preexisting degraded state means the UAV is no longer able to finish its current mission and safely return to base, but could potentially finish part of its existing mission or perhaps an alternate mission.
This thesis centers around the development of an end-to-end framework for health-aware replanning of UAVs with varying levels of use-based degradation during operations in environments where conditions may change from time to time. Accurate state of health estimates are derived through system-level prognostics, where the remaining useful life is estimated using model based or data-driven methods. Furthermore, it is desired for these models to be executed online in resource constrained settings. Therefore, the key contributions are a systematic study of model compression methods specific to Long-Short Term Memory (LSTM) networks for regression tasks, which account for the temporal characteristics as in the prognostics problem. Second, is a framework for health-aware replanning, in which an online replanning agent must reason about both faults and use-based degradation. A comprehensive study using multiple UAVs of varying ages and degradation levels, operational profiles, and different faults demonstrates the effectiveness of our proposed framework in a simulation environment, as well as on flight-certified hardware tested at NASA Ames Research Center. | |
