Incomplete Multisource Domain Adaptation for Fault Diagnosis of Blast Furnace

The fault diagnosis of blast furnace (BF) is challenging due to the few fault samples and significant distribution drift. Thus, transfer learning has been introduced into BF fault diagnosis. However, most existing methods focus on knowledge transfer from a single source domain and require that the source and target domains share the same category space. This implies that BF data must encompass all possible fault categories to serve as a source domain. Such constraints are excessively strict in actual BF ironmaking processes. To address this issue, we propose a novel fault diagnosis method called minmax entropy-based adversarial network (MMEAN) for multisource domain adaptation with category shift. First, MMEAN establishes a convolutional neural network (CNN) with shared parameters to extract common features from multiple source and target domains. Based on this, the model utilizes a vision transformer (ViT) for each domain to achieve the fusion of local and global features, extracting domain-specific features with more discriminative information. To achieve knowledge transfer, MMEAN conducts adversarial training by minimizing and maximizing the entropy of unlabeled samples with the feature extractor and classifier, respectively. Finally, MMEAN determines the reliability of each source classifier according to entropy and performs weighted combination to avoid negative transfer. The experimental results on actual BF datasets demonstrate the effectiveness of the proposed method.