Ranking input importance in neural network modeling of engineering problems

This paper addresses the issue of identifying important input parameters in building a multilayer, backpropagation network (BPN) for engineering applications. Since the identification of important and/or redundant input parameters of a BPN leads to reduced size, shortened training time, and possibly more accurate results of the network, it is an issue of great practical as well as theoretical interests. We compare three different methods for ranking input importance-sensitivity analysis, fuzzy curves, and change of MSE-and analyze their effectiveness on BPNs trained to model simple nonlinear functions as well as a real, production use network that has been built to model the cement bonding quality in petroleum engineering. Based on the analysis and our experience in building the BPN for predicting cement bonding quality, we also propose a general methodology for building BPNs in typical engineering applications.