Enhance pilot's decision : Determination of balanced field length using neural network

Abstract: The data reliability is crucial in aeronautics because the least miscalculation can lead to crash. Among these data, the balanced field length (BFL) is defined as the shortest field length at which both the take-off and the acceleration-stop can be performed. As the BFL is a critical data, it is subject to certifications that add constraints for all its stages of calculation. An in-house software to calculate the BFL is developed at Dassault Aviation but it can not be embedded and its use requires an expert. Due to the nonlinear dependencies and because an available data set is available, neural networks are proposed to predict the BFL with a maximum relative error less than 2%. The data set of simulations has been set up from Falcon 7X in different configurations of take-off. However, different modelings were used for test- ing purposes and so, the data base is contaminated with points which do not respond to our issue. First of all, it was necessary to find a way to identify these points and a data cleaning algorithm is first implemented. A bagging consensus of neural network is then added to it to detect and filter the other mislabeled data points. Different neural networks are finally trained on the data set and are aggregated to propose the best model. The final model is tested on different database and in this instance, on a Falcon 8X one which has very closed characteristic. It turns out that the algorithm worked for both airplanes and reaches the expected results.