SOLVING MULTI-RESPONSE PROBLEMS THROUGH NEURAL NETWORKS AND PRINCIPAL COMPONENT ANALYSIS

Abstract

ABSTRACT In this study, we propose a four-phased procedure based on neural networks and principal component analysis to resolve the parameter design problems with multiple responses. The quality characteristics of a product are first evaluated through Taguchi's quality loss function. A backpropagation neural model is then trained to map out the functional relationship between control factors and responses' quality loss. The functional relationship is then fed into the principal component analysis procedure to transfer a set of responses into a set of uncorrelated principal components. A feasible combination of control factors can be obtained through the recalling function of a backpropagation neural model. Therefore, the conflict for determining the optimal combination of control factors in a multi-response problem can be greatly reduced. The proposed procedure is illustrated with a case study of a fused optical fiber coupling process. The confirmation experiment also yields a satisfactory result. The proposed procedure is relatively simple and could be implemented easily by using ready-made neural and statistical software.