A Semi-Supervised Learning-based Dynamic Prediction Method for Semi-molten Condition of Fused Magnesium Furnace

Abstract

Fused magnesium furnace (FMF) is an important equipment for producing magnesium oxide, which is prone to occurring the semi-molten abnormal condition during the production. If the abnormal condition is not predicted in time, the furnace shell will be burned through, endangering the personal safety of the staff on site. Therefore, it is necessary to predict the semi-molten abnormal condition in time and accurately. Existing machine learning-based methods adopt static models for recognizing and predicting anomaly. However, the model accuracy will degrade as data features shifting over time and melting processes. To address the above problems, this paper proposes a dynamic prediction method for semi-molten abnormal condition of multiple FMFs based on semi-supervised learning. We introduce a consistent regularization strategy and dynamically update the model weights by learning multiple FMF smelting process video data with a sparse set of condition labels. The algorithm is able to dynamically adapt to the shifted data features for accurate anomaly prediction. The proposed algorithm can predict the semi-molten abnormal condition in real time and accurately under the condition of only 1% label, enabling the safe and reliable operation of FMF.