Date of Award

Spring 2022

Document Type

Thesis - Open Access

Degree Name

Master of Science in Aerospace Engineering

Department

Aerospace Engineering

Committee Chair

Dr. Hever Moncayo

First Committee Member

Dr. Richard Prazenica

Second Committee Member

Dr. Yan Tang

Abstract

A pilot is a highly nonlinear and incredibly complex controller whose responses are difficult to predict. Many accidents have occurred from pilot error before or after failures and almost always after entering areas of the flight envelope considered as Loss-of-Control regimes. If a pilot's inputs to the flight control system can be predicted, then the introduction of dangerous flight conditions can be readily avoided. Avoidance could take the form of a warning indicator or augmentation of the pilot's inputs. The primary difficulty lies in how to actually predict how the pilot will perform in the future.

Methods to solve this problem are focused around the McRuer pilot model which simplifies the pilot response to a four-parameter equation that has been the focus of most recent solutions. Many recent attempts at solving this problem have found promising results in Wavelets, Most Likelihood Estimation, Extended Kalman Filters, and Unscented Kalman Filters. This thesis applies two new methods to the estimation problem and suggests a modification to one.

The three methods investigated in this thesis are a modified form of the Unscented Kalman Filter, Fourier Transform Regression with Time Domain derivatives, and Adaptive Neural Networks. The Unscented Kalman Filter holds merit in many estimation problems for its ability to handle model nonlinearities and noise in the systems and sensors. In this respect, it held the best solution for this work given that it could correctly estimate the parameters. However, the filter had to be finely tuned to reach a solution. The Fourier Transform Regression method could only handle time-invariant pilot model parameters due to its usage of batches of data. Once the parameters began varying with time, the solutions began having singularities. The adaptive neural networks showed promise being that they are stochastic estimators, but the solutions held within show they need more development to become a viable solution to this problem. It is recommended that deep reinforcement learning or combinations of these algorithms be applied to this estimation problem in the future to determine a more robust solution that can estimate the pilot's intent online.

Share

COinS