System Identification-based Fault Detection and Model-Based Control of an Uncrewed Aerial System

Abstract

As the wide-scale use of uncrewed aerial systems, or UAS, proceeds towards civilian airspace, guarantees of operational safety over the entire flight envelope become more critical. However, public opinion is mixed on the general safety of autonomous systems, especially within close proximity to densely populated spaces. Because the opinion of the general public is of paramount importance to the general adoption of this technology, focus must be placed on the development of dependable systems before operations can even be considered.With the stringent payload capacity and size constraints of UAS, advanced control algorithms can fill the gap caused by the reduced redundancy expected of UAS. One such branch of control algorithms is active fault tolerant control, which uses knowledge of failures to update the control loop and maintain performance. However, most works on active fault tolerant control present in literature trivialize fault detection and diagnosis to hard to achieve assumptions of perfect knowledge of failures or the presence of actuator feedback systems, which are currently not widely used. These weight and cost penalties increases the appeal of a software-based system for fault detection, which solely uses measurements from sensors already necessary for autonomous flight, completely eliminating the disadvantages of hardware-based systems.As system dynamics change during failures, system identification is an intuitive way to measure changes in dynamics to estimate actuation failures, especially flight-critical actuators which have a large impact on system performance. In this work, a time domain system identification method based on the eigensystem realization algorithm (ERA) and observer Kalman filter identification method (OKID) is used. These methods used in tandem can generate a model capturing the maximum input-output information from any arbitrary set of data, providing a more generalizable approach than is possible with most system identification techniques. Taking advantage of available flight tests data which contained aileron and rudder failure cases, this combination is applied to flight test data, and the models generated are used to investigate the change in dynamics present during failures, with changes in model parameters correlated to failures.Knowledge of failures can be leveraged by the control loop by an adaptive flight controller if the dynamic model of the aircraft is available after failure. This research presents applications of two different control algorithms capable of adapting to changing dynamic models, model predictive control (MPC) and dynamic inversion. This work adds a non-traditional control increment term, Δu, to the MPC cost function to directly punish large control rates to prevent oscillations and out of phase behavior. The presented dynamic inversion flight controller allows for a more direct approach to system performance design. Its theory implies the ability to maintain performance in the case of faults as long as adequate updates to the model are provided and enough overall control authority is available to make the commanded maneuvers.Typical approaches to simulation validation of fault tolerant control methods are based heavily in assumptions on how failures impact the dynamic model. In direct contrast to this, validation models for this work are based on collected failure flight test data and generated using a Monte Carlo-based technique. This method varies model parameters based on a given normal distribution to find a dynamic model which best fits the measured aircraft states. Nominal simulations results show a similar level of performance of both control algorithms, while failure simulations highlight the satisfactory performance of both controllers as well as the different benefits offered by each algorithm.This work builds off of a previously published conference paper which presents preliminary results of the fault detection method with one data set and only performs failure validation simulation for the LQR and dynamic inversion controllers.