Date of Award

Spring 2026

Access Type

Thesis - Open Access

Degree Name

Master of Science in Engineering Physics

Department

Physical Sciences

Committee Chair

William MacKunis

Committee Chair Email

mackuniw@erau.edu

College Dean

Jayathi Raghavan

Abstract

This thesis investigates deep neural network (DNN)-based adaptive control strategies for unmanned aerial vehicles (UAVs) operating under aerodynamic uncertainty and complex actuator dynamics.

The first contribution presents a control strategy employing a concurrent learning (CL)-based DNN training algorithm, which combines online adaptive DNN weight adaptation with offline batch-like training updates using a recorded data stack. The analysis focuses on the closed-loop performance improvements resulting from the use of optimum CL data-selection algorithms, which ensure that the recorded data stack maintains sufficient data diversity to provide a statistically meaningful representation of the operating conditions using a reduced data set. Specifically, this result includes a detailed error system development formally incorporating realistic measurement errors in the recorded data stack; a Lyapunov-based stability analysis deriving a sufficient condition on the diversity of the recorded data stack under which the tracking error and DNN weight estimation error are proven to be uniformly ultimately bounded; and numerical simulation results demonstrating that the use of the data selection algorithm yields up to an 83% reduction in DNN training time and a 15% reduction in mean-squared tracking error compared to a CL DNN system with no data selection algorithm.

The second contribution presents a robust and adaptive DNN-based control method that formally compensates for unmodeled system dynamics in addition to actuator model uncertainty in a class of nonlinear, non-affine, overactuated dynamic systems. This result is inspired by recent developments in biaxial-tilt actuation unit (BTAU)-based multirotor aerial vehicle technology, which renders multirotors overactuated through two controllable tilt angles in each propeller rotor. The proposed method features an innovative error system development that re-casts the non-affine dynamic model in a control-affine form, a singularity-robust pseudoinverse-based control structure, and a bank of dynamic filters enabling robust compensation using only velocity measurements for feedback. Numerical simulation results demonstrate that the feedforward adaptive DNN element reduces the mean-squared tracking error by up to 31.5% while reducing the mean-squared commanded control magnitudes by 1.5% compared to a standard robust feedback control law.

The third contribution presents a hardware implementation and experimental validation of a concurrent learning (CL)-based adaptive control system for online parameter estimation on a nonlinear pendulum testbed. Classical adaptive control methods rely on persistence of excitation (PE) to guarantee parameter convergence, a condition that is difficult to satisfy in practice. CL relaxes this requirement by leveraging a recorded history stack of input-output data concurrently with instantaneous measurements, with convergence rate governed by the minimum singular value of the stored data matrix. A Lyapunov-based stability analysis is presented, which establishes uniform ultimate boundedness (UUB) of the tracking error, auxiliary error, and parameter estimation error, and formally incorporates realistic acceleration measurement noise in the recorded data stack. The proposed CL adaptive law is experimentally validated across three distinct inertia configurations and benchmarked against a standard gradient-based adaptive controller. Results demonstrate that the CL approach reduces steady-state parameter estimation error by up to two orders of magnitude relative to the standard adaptive controller, achieving errors as low as 1.27% compared to 77.05% for the baseline, while simultaneously reducing mean-squared control effort by up to 24.6% in the highest inertia case. These results confirm that CL-based adaptive laws are practically implementable on embedded hardware and that their theoretical convergence guarantees translate to significant and measurable improvements in parameter estimation accuracy on physical systems.

Download

Share

COinS