Process operation performance assessment of electro‐fused magnesium furnace based on deep auto‐encoder transfer generative adversarial network

Abstract

AbstractThe process operation performance assessment (POPA) of the smelting process of electro‐fused magnesium furnaces (EFMFs) plays a very important role for improving production quality and pursuing the highest economic benefit. Most of the existing methods are based on the assumption of abundant labelled training samples. However, the lack of samples for the POPA is a challenging problem in the smelting process of EFMFs. Traditional methods for POPA make it difficult to obtain satisfactory results under small samples. A new deep auto‐encoder transfer generative adversarial network (DAETGAN) based on source domain data is proposed to improve the performance of POPA with small samples in EFMF smelting process. Firstly, a general transfer framework is proposed, in which the data of source domain is used as input to generate adversarial network (GAN), and a large number of images, sound, and current samples are generated with the same distribution as the data of target domain so as to improve the transfer learning effect. Secondly, a POPA model is proposed with the multi‐source heterogeneous information generated by DAETGAN and the original multi‐source heterogeneous information in target domain. Finally, it is verified by experiments that the DAETGAN model can generate data with the same distribution as the original data, and the accuracy of the assessment of operational performance is effectively improved.