Data-driven controller design for aircraft maneuver control with bayesian optimization

Abstract

Aircraft maneuver control has been an important component in diverse areas ranging from traditional aircraft designs to drone development and computer simulations. For instance, high-resolution aerial engagement simulation is a key to reliable battle experiments, and it requires a detailed aircraft maneuver model to reflect the aircraft and the air-dynamics in the real world. High-resolution aerial engagement simulation requires a maneuver control model that satisfies various maneuver goals, such as waypoint navigations, attitude control, etc. To meet the maneuver goals, a control model can be implemented using a conventional controller such as Proportional Integral Derivatives (PID) control method. This traditional approach suffers from two aspects. First, tuning parameters of PID is a difficult task, which requires a deep understanding of the target domain including the dynamics model of the controlled aircraft

and the tuning tasks can be time-consuming, especially when the complexity of the control system and the numbers of parameters to be tuned are high. Second, the PID controller requires many assumptions and predefined air-dynamics characteristics. These data are garnered from extensive tests, and the measured data are applied to a general dynamics model with various assumptions. This dissertation presents a new method of designing a flight control model suitable for the aircraft dynamics model used in the engagement simulation. The proposed method is a data-driven control method and utilizes the black-box model of the controlled process. In my study, the Gaussian process (GP) is adopted to obtain a black-box response model of underlying nonlinear systems. GP is a representative nonparametric method and flexible enough to describe complex nonlinear systems. Experimentally collected data from flight tests are used to obtain black-box models of the controlled aircraft. These black-box models do not require detailed knowledge of the plant dynamics. And I apply the Bayesian optimization method to find the global optimal value with only a small number of experiments. After identifying the black-box model of the plant using the Gaussian process and the Bayesian optimization method, I try to find optimal control with unknown controller structure using an optimal piecewise constant control. In reality, a pilot generates a micro-adaptive control, but we consider an experienced pilot might keep a constant control if the pilot can expect the given control will deliver the desired state. If we have information on the structure of the controller, we can find the optimal control based on it. I suggest a method to automatically tune controller parameters without requiring the underlying dynamics model. I consider proportional-integral-derivative controllers (PID controllers) In the first study, I propose a new method of designing a data-driven controller for an aircraft maneuver. Assuming that we do not have knowledge on the controller and the controlled aircraft, I propose a controller design with explorations on the control inputs and their responses from the aircraft. Specifically, I utilize Bayesian optimizations with Gaussian process (GP) regression for black-box modeling on the aircraft responses from the explored controls, which are selected as samples to experiment in Bayesian optimization. I test the proposed controller with a rigid six degree-of-freedom (DoF) nonlinear aircraft model by varying kernel structures of GP regressions. My proposed method shows shorter flight time and smaller deviations to navigate fixed waypoints compared to the manually tuned PID controller. The proposed controller can be an alternative to the PID control, particularly when both the controller structure and controlled process model information are unknown. The second study focuses on the development of an automatic controller tuning framework combining Bayesian optimization with PID tuning for aircraft maneuvering. If we have information on the structure of the controller, we can find the optimal control based on it. I suggest a method to automatically tune controller parameters without requiring the underlying dynamics model. I consider proportional-integral-derivative controllers (PID controllers) as a gray-box controller model. This method can produce micro-level inputs in continuous time. The proposed studies are data-driven control approaches to design a controller directly using I/O data without any information on the mathematical model of the controlled process. Using the proposed methods, we can find out the optimal design with only a few experiments on the target plant and find globally optimal control value, not the local optimal value. The proposed methods will be useful for developing the aircraft maneuver control model which is essential in the engagement simulation.