Eventos Anais de eventos
COBEM 2017
24th ABCM International Congress of Mechanical Engineering
Analysis of MLP Artificial Neural Network architecture for E423 airfoil
Submission Author:
Bruno Diniz , BA , Brazil
Co-Authors:
Thalita Medeiros, Leonardo Lopes Nascimento, Bruno Diniz
Presenter: Leonardo Lopes Nascimento
doi://10.26678/ABCM.COBEM2017.COB17-0924
Abstract
This work analyzes if a Artificial Neural Network with multilayer perceptron architecture is able to develop the Lift coefficient X α curve, without the need for numerous experimental tests. The airfoil used for the architecture analysis was the Eppler 423. The coefficients of lift (C1), coefficients of drag (Cd), different angles of attack (α) and different Reynolds numbers (Re) of this airfoil were used. For the network architecture two input data were used, and the output is the Cl. A total of 1000 training epochs were used, which means that 1000 complete presentations of the training set were made during the learning process, this number of times acts to the synaptic weights and the bias levels stabilize to obtain the Lower RMS. The cross-validation process is used in the network because it acts as an aid criterion to evaluate the robustness (data repeatability). The main analysis of the network is given by the robustness, which is found by graphically comparing the dispersion between the RMS test and RMS training. The results obtained, allows us to reach the conclusion that there is a good repeatability and the network is satisfactory, which was obtained a graph behavior quite satisfactory for some Reynolds numbers.
Keywords
Eppler 423 airfoil, Artificial neural networks, architecture, RMS
