Title: Comparing prediction procedures for functional data in aeronautic
Authors: Feriel Boulfani - Mathematics institute of Toulouse (France) [presenting]
Anne Ruiz-Gazen - Toulouse School of Economics (France)
Xavier Gendre - ISAE-SUPAERO (France)
Martina Salvignol - Airbus operations SAS (France)
Abstract: The oil temperature of an aircraft generator indicates the well functioning of the generator. By predicting the changes of the oil temperature, an abnormal behavior can be detected. For this reason, predicting the oil temperature behavior at a given time $t$ is of particular interest. We propose to use as explanatory variables the observations of some auxiliary functions measured by the aircraft during a period preceding time $t$. The purpose is to consider a functional data analysis framework and compare it with the neural network, random forest and linear regression statistical procedures. To avoid overfitting, we apply the dropout technique to these procedures. This technique is used in neural network and consists of dropping randomly units in the network, which can be seen as a way of adding noise to the model. The dropout technique for linear regression model drops randomly some dimensions of the input while for random forest, it drops randomly some decision trees. The methods are compared using some real data set from the aeronautic sector.