Assessment of Artificial Neural Networks' performance for the synthetic model-free Angle of Attack and Sideslip Estimator

Redundancy in aviation, especially in Air Data Systems, is crucial for flight safety and reliability, yet achieving the necessary levels - ranging from dual or triple in civil aircraft and UAVs to quadruple in high-performance jets - is challenging due to space, complexity, and weight constraints. Since 1949, USAF studies have led to the development of synthetic sensors capable of estimating aerodynamic angles without physical sensors, offering benefits such as enhanced reliability, maintainability, reduced weight, and lower power consumption and emissions. Although synthetic flow angle estimators solve equations that model data fusion from onboard avionics, their performance is often compromised by uncertainties in operational data. Neural networks trained on corrupted data have been explored to mitigate these errors, but this data-driven approach remains aircraft-specific. To develop a more adaptable, model-free solution independent of specific aircraft data, this paper examines various neural network architectures - including Multi-Layer Perceptrons (MLPs), regularization networks such as the Generalized Radial Basis Function Networks (GRBF-NNs), Recurrent Neural Networks (RNNs), and the innovative Kolmogorov-Arnold Network (KAN) - in solving a model-free algorithm, named "Angle of Attack and Sideslip Estimator" (ASSE). Results demonstrate that RNNs, and particularly LSTMs, are ideally suited for online deployment, owing to their exceptional capacity for processing sequential data and capturing temporal dependencies, as well as their remarkable robustness to noise inherent in sensor-derived flight data.