Partially Supervised Classification for Industrial System using Deep Neural Network

Abstract

A general classification setting requires prior knowledge (i.e. labeled samples) to cover all classes. However, in many industrial problems, prior knowledge usually does not describe all the classes, and the generation of a complete training set that cover all classes often is a time-consuming, expensive and difficult (if not impossible) task. Our target of this work is, given labeled samples from only a subset of classes, how to assign label to any sample that potentially come from either known or unknown classes in real time data. We test our algorithm on industrial failure classification and experiments show that our method outperforms existing popular baselines.