Practical guidelines for developing BP neural network models of measurement uncertainty data

The predictability of measurement uncertainty is a critical issue in quality assurance. The performance of such a process cannot be predicted if no proper mathematical model is available. For manufacturing processes where no satisfactory analytical model exists, or where a low-order empirical polynomial model is inappropriate, neural networks offer a good alternative predictive modeling approach. This paper considers the primary decisions and activities that arise during back-propagation (BP) neural network model construction, selection, and validation for this novel application. Computational experiments were designed to cross-examine the two types of hidden layers of networks with different hidden neurons, training tolerances, and testing tolerances based on the v-fold cross-validation technique. Hypothesis testing is used to evaluate different kinds of prediction errors from the models developed. The best models were very accurate in generalizing the measurement uncertainty prediction, and thus are suitable for use in a manufacturing environment. This study reveals no statistical advantages of using a two-hidden-layer net over a one-hidden-layer net in modeling the measurement uncertainty data and offers some general and simple rules for BP model selection and validation.