Experimental Analysis of Neural Approaches for Synthetic Angle-of-Attack Estimation

Synthetic sensors enable flight data estimation without devoted physical sensors. Within modern digital avionics, synthetic sensors can be implemented and used for several purposes such as analytical redundancy or monitoring functions. The angle of attack, measured at air data system level, can be estimated using synthetic sensors exploiting several solutions, e.g., model-based, data-driven, and model-free state observers. In the class of data-driven observers, multilayer perceptron neural networks are widely used to approximate the input-output mapping angle-of-attack function. Dealing with experimental flight test data, the multilayer perceptron can provide reliable estimation even though some issues can arise from noisy, sparse, and unbalanced training domain. An alternative is offered by regularization networks, such as radial basis function, to cope with training domain based on real flight data. The present work's objective is to evaluate performances of a single-layer feed-forward generalized radial basis function network for AoA estimation trained with a sequential algorithm. The proposed analysis is performed comparing results obtained using a multilayer perceptron network adopting the same training and validation data.