Novel trajectory prediction and flight dynamics modelling and control based on robust artificial intelligence algorithms for the UAS-S4

Abstract

Aerial Traffic Management (ATM) and Aerial Collision Avoidance (ACA) are the most important issues in aviation transportation. Accurate Aircraft Trajectory Prediction (ATP), precise Flight Dynamics Model (FDM), and efficient Flight Dynamics Control (FDC) are the main fundamental requirements, for trajectory-based operations, such as ATM and ACA. The aim of this thesis is to design, and further develop these above-mentioned three fundamental requirements for critical trajectory-based operations. For each fundamental requirement, a thorough research study was conducted to meet its related objectives. The first study focused on accurate Aircraft Trajectory Prediction (ATP). This study began with formulating an ATP as a time series regression problem. Next, six data-driven Neural Network models were designed and fine-tuned to produce accurate ATPs. Their architectures were based on Logistic Regression (LR), Support Vector Regression (SVR), Deep Neural Network (DNN), Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), and Long-Short Term Memory (LSTM). The six ATP models were evaluated in terms of prediction accuracy, and their super-efficiency was confirmed. Despite their excellent performance, we could generate adversarial samples to mislead them. This issue presents a security concern regarding Neural Network-based ATP models. Therefore, a defense algorithm was designed based on an adversarial retraining strategy. The new, robust learning-based ATP models showed excellent performance against adversarial attacks while still performing their ATP tasks accurately. The second study addressed the design of an efficient Flight Dynamics Controller (FDC). Relying on the available UAS-S4 Local Linear Scheduled Flight Dynamics Model (LLS-FDM) at the Laboratory of Research in Active Controls, Avionics and Aeroservoelasticity (LARCASE), a nonlinear FDM was designed based on the Takagi-Sugeno (TS) Fuzzy Logic approach. Simultaneously, the desired reference model was determined, and then stabilized using a Linear Quadratic Regulator (LQR). Regarding the reference FDM, a “model-based” FDC was designed for the UAS-S4 FDM, which worked very well based on tracking errors. Next, a Fuzzy Logic Controller (FLC) containing robust adaptive gains was designed in order to consider the nonlinearities and uncertainties due to fuzzification and external disturbances. The results confirmed that the robust adaptive fuzzy logic controller could stabilize the flight dynamics and accurately track the reference model state variables. The third study was conducted for the design of an accurate Flight Dynamics Modelling (FDM) method. Accurate FDMs allow engineers to design highly efficient model-based FDCs. Flight tests were conducted on the UAS-S4 Ehecatl (at the LARCASE), and 216 local FDMs were obtained by using the in-house Local Linear Scheduling Flight Dynamics Model (LLS-FDM) designed to handle 216 flight conditions. The LLS-FDM was further augmented using the knearest neighbor interpolation and extrapolation methodologies. Relying on this augmented data, Support Vector Regression (SVR) was used as a benchmarking algorithm for the LLSFDM regression. The trained SVR could predict the UAS-S4 FDM for the entire flight envelope. A Root Locus diagram was utilized to validate the UAS-S4 SVR-FDM by evaluating the predicted eigenvalues’ closeness to their original values. The SVR prediction accuracy was studied for different flight conditions, number of neighbors, and a range of kernel functions. Despite the excellent performance of the trained SVR, the FDM was vulnerable to adversarial samples. Hence, an Adversarial Retraining Defense (ARD) was developed by relying on adversarial FDMs, that were created via the Adapted Fast Gradient Sign Method (AFGSM) to design a Robust-SVR FDM. The Robust-SVR FDM worked very well under adversarial attacks while providing better time domain properties for state variable stabilization than the LLS-FDM.

Date	7 Mar 2022
Original language	American English
Awarding Institution	École de technologie supérieure
Supervisor	Ruxandra Botez (Supervisor)