Small Data Set Learning with Synthetic Samples and Area Membership Values

摘要

A common strategy in manufacturing systems is to execute pilot runs before the mass production. We model the limited data obtained from pilot runs,and try to shorten the lead time required for predict future production in this study. A manufacturing system is usually comprehensive;Artificial Neural Networks are commonly applied to extract management knowledge from acquired data for its non-linear properties. It is the fundamental assumption for Artificial Neural Networks to get as large a number of training data as needed;nevertheless, this is often not achievable for pilot runs because there are few data obtained during trial stages,and theoretically this means that the obtained knowledge is fragile. The purpose of this research is through the proposed procedure to decrease the Artificial Neural Networks prediction error rate in small data set problems. Based on a consideration of dependent data attributes, the proposed procedure is designed to utilize extreme value theory and statistical prediction interval calculations to derive fuzzybased synthetic samples to fill sparse data information gaps. After synthetic samples are generated,area membership values are applied to combine data values,and occurrence possibility for every sample is used as the input of Artificial Neural Networks. The results of this research indicate that the prediction error rate can be significantly decreased by applying the proposed procedure to a very small data set with attribute dependency.