Calibration curve of a multi-component balance using MLP artificial neural network with learning endowed with loading uncertainty

Abstract

One of the traditional approaches of curve fitting to the calibration data set is the polynomial fitting by the least squares method. As an alternative to the polynomial approach, one can use Artificial Neural Networks to interpolate the data points, and this is the subject of the present work. The system to be calibrated consists of the external aerodynamic balance of the subsonic wind tunnel no. 2, the TA-2, of the Brazilian Aerospace Institute (IAE). The Multilayer Perceptrons (MLPs) are the class of neural networks chosen in this study because the mathematical modelling of the external balance calibration is multivariate. Studies regarding the convergence of functions were carried out taking into consideration different architectures of this network class, in order to obtain adequate models for different calibration sets. The results of the least squares regression, fitted to the polynomial nowadays employed at TA-2, are chosen as reference. Measurement uncertainties were considered through weighting the neural network learning algorithm by the angle reading uncertainty and uncertainties of the loads applied in the calibration process of the external balance. A comparison between the common practice of disregarding the uncertainties and regarding them in the MLP learning process is highlighted.