Semi-supervised Learning Approach to Abnormality Detection with Complementary Features

This paper introduces a machine learning based solution to the practical task of identifying the semi-molten abnormal working condition of fused magnesium furnace. The primary challenge faced by the task is the insufficiency of labeled samples for classifier training. This problem is tackled under the semi-supervised learning framework by combining two complementary features i.e. the smelting currents which are unlabeled and the monitoring images which are partially labeled. An entropy regularized form of cost function is designed which brings the distribution pattern of the smelting current to the training of image based classifier, and an efficient optimization algorithm based on cross entropy method is presented. The proposed solution is tested on industrial dataset showing remarkable result in accuracy.