The effect of using different feature sets and flight data envelopes on the fidelity of deep learning based system identification of a fighter aircraft

Abstract

Aircraft models that accurately reflect flight dynamics are critically important for aircraft design, development, and certification. To achieve high-fidelity results, flight dynamics models are developed through System Identification using flight data. Recently, Deep Learning-based approaches have gained prominence. However, collecting flight data is highly expensive, and accurately estimating the nonlinear region have significant challenges. This study aims to improve flight dynamics models in the nonlinear region by using different feature sets and optimize the flight test campaign by making flight envelope analyses with using deep learning-based system identification methodology. First, a generic F-16 flight model was modified to obtain flight data. Short Period, Bank to Bank, and Dutch Roll maneuvers were executed in this model at various speeds and altitudes to be used for aerodynamic database predictions. Parameters required for System Identification were measured and obtained during these flights tests. Next, data preprocessing was conducted. This involved preparing an infrastructure to calculate the total force and moment coefficients of the aircraft using mass-inertia, thrust values, and the flight data. Subsequently, an algorithm was developed to prepare deep learning-based models for the prediction algorithm. The models were trained using different feature sets and flight envelopes. The purpose of using different feature sets was to enhance the accuracy of the flight model in the nonlinear region. Instead of using input sets directly for training, different exponents of the inputs and their coupled parameters were added to the feature set using knowledge of flight dynamics and aerodynamics. This provided the model with various input combinations during training the models. Initially, models were trained using simplified feature sets. Then, another feature sets with added exponential and coupled parameters were used for training with the same flight data. The models were compared using validation test maneuvers. The study also focused on the subject of costs of the flight tests, a critical aspect of System Identification studies. Optimizing flight test maneuvers and inputs is essential for cost efficiency. Typically, performance and stability analyses are conducted in isolated regions thus the flight data does not lead to have sufficient distribution to reflect all flight envelope. In other cases, flight envelope is scanned at very high frequency which leads to very high costs. Second part of this study examined focused on model accuracy changes with the amount of data by examining system identification with different flight test maneuvers and envelopes. Models were trained using data from only the nonlinear region, only the linear region, and the full flight envelope, each with different data densities. In total, six different cases were tested and compared. In conclusion, comparing both feature sets for all force and moment coefficients revealed that they produce similar results in the linear regime. However, as nonlinearity increases near the extreme points of the flight envelope, the accuracy of the complex feature sets is significantly higher than that of the simpler feature sets. Flight test comparisons indicate that adding complex parameters to feature sets is necessary to improve accuracy in regions with increasing nonlinearity. Examining the effect of flight test envelopes used during training shows that using data only from nonlinear or linear regions is insufficient to represent the entire flight envelope. Even with increased data density in isolated regions, it is necessary to collect data from both linear and nonlinear regions for an accurate model across the entire flight envelope. Properly executed maneuvers demonstrate that increasing data density does not enhance accuracy cost-effectively. For future applications, optimizing feature sets and flight test campaigns in System Identification studies can yield a model that performs accurately across the entire flight envelope. Additionally, significant budget savings can be achieved by conducting a cost-effective flight test campaign.