Multi-component balance load estimation using neural networks

A 1.5 inch diameter, multi-piece, six-component (fiveforce / one-moment) balance was calibrated at three balance calibration facilities. This paper evaluates the use of neural networks to estimate the balance calibration loading al each facility. For the current work, multiple-input, single-output neural networks are trained using a Levenberg-Marquardt algorithm for each of the balance load gages. The load estimates obtained from these neural network computations are compared to load estimates computed from the more traditional regression-based polynomial math model for the balance response. Cross-validation of the neural network and regression models is performed to evaluate calibration repeatability within a facility and between facilities.