Gabriel Buendia, Beatriz Pulido, MIguel Torralba, Manuel Reyes, Felix Arevalo, Hector Climent
DOI Number: N/A
Conference number: IFASD-2024-171
Artificial neural networks (ANN) are known for solving complex problems and detecting nonlinear relationships between the variables of a database in a fast and accurate manner.
Moreover, their storage memory optimization makes them an attractive tool for aeronautical applications such as flight control algorithms, health monitoring of structural components and flight simulators [1]. In order to study the feasibility of applying feedforward backpropagation networks to dynamic loads problems, two applications on a military transport aircraft have been analysed:
- Assessment of potential aircraft overloading in hard landings events, where classification neural networks were used.
- Prediction of Fatigue continuous turbulence loads, where regression neural networks were used.
The following procedures have been explored for the proper development of these neural networks:
- Exploration of the database inputs and outputs. This includes an initial assessment of the relevance of the input parameters, the selection of the suitable outputs to be monitored and the identification of the densest regions in the database.
- Neural training and hyperparameter tuning using Keras/TensorFlow 2.0 [2]. Sensitivity studies are performed to select the combination of the parameters that specify the details of the learning process which provide the best results, either for predicting a single or multiple outputs simultaneously. The definition of the cost function and metrics is of special relevance.
- Interpolation and extrapolation capabilities assessment. Evaluations on the ability of the trained networks to predict results from inputs not used in the training, which are inside or outside the limits of the variable space in which they have been trained, were performed.
- Performance evaluation with recorded flight data. Influence of the errors coming from conservative estimations of the neural networks on the fleet operations is evaluated.
The extension of this methodology to other dynamic loads problems such as fatigue discrete gust, taxi loads, and other overloading events will be explored in the future.