Fault diagnosis of blast furnace based on incomplete multi-source domain adaptation with feature fusion

Abstract

Aiming at the model mismatch caused by changes in data distribution, transfer learning (TL) has been introduced to fault diagnosis of the blast furnace (BF) ironmaking process. However, most existing TL methods require that the category space of each source and target domain be identical, and ignore the semantic information of multi-source data under domain adaptation. To address these issues, we propose a novel method based on incomplete multi-source domain adaptation with feature fusion for fault diagnosis of BF. Firstly, a multi-scale convolutional network is set to effectively extract diverse features while enabling information interaction through point-wise convolution. Secondly, Transfer Vision Transformer is constructed for each source domain to fuse global and local features, and extract domain-specific knowledge with more semantic information. Finally, the model weights each source classifier based on the inter-domain similarity to obtain the result. Experiments on actual BF data validate the effectiveness of the proposed method.