Sonstegard, Jack2023-11-282023-11-282023-04https://hdl.handle.net/11299/258612University of Minnesota M.S.M.E. thesis. April 2023. Major: Mechanical Engineering. Advisor: Perry Li. 1 computer file (PDF); xii, 82 pages.An adaptive feedforward control strategy that enables real-time online learning of engineinputs and outputs for a unmanned arial vehicle (UAV) drone was improved and expanded upon in this work. Through the use of recursive least squares for online learning, a variable-direction forgetting factor was used to enhance tracking while ensuring forgetting is only applied to the information-rich subspace. Compared to uniform forgetting methods, the variable-direction method was shown to improve tracking and to bound the singular values of the covariance matrix when a signal was not persistently exciting. Additionally, a detection method for large environmental disturbances was used for covariance resetting where this resetting significantly helped the system adapt to disturbances. In order to use the variable-direction forgetting, a gradient flow approach was used to compute the singular value decomposition of the covariance matrix efficiently. In simulation, this iterative method was shown to converge and to track the singular values with low error. Furthermore, improvements to the algorithm included the use of multiresolution techniques where a fine and coarse mesh were combined using a mixing coefficient to smooth out the response of the engine. This allowed for a simpler representation of the system to contribute the majority of the input when the system was in an unexplored region of the learned map. The improved algorithm was then used to learn a higher dimensional model. To evaluate the safety of the system, a one dimensional drone model was used to monitor the position of the drone based upon the inputs provided by the engine. In order to search this higher dimensional engine map, a new search algorithm was devised that considered constraints of combustion phasing and indicated mean effective pressure as well as a minimization of fuel mass input. In simulation, the position of the drone was shown to track the desired trajectory well. With this work, a foundation for future research in active learning of the engine map is motivated.enAdaptive FeedforwardRecursive Least SquaresSingular Value DecompositionVariable-Direction ForgettingThesis or Dissertation