A Multilevel Deep Fusion Framework for FeO Content Prediction in Sintering Process

Iron ore sintering is a critical procedure in steel production, and the content of ferrous oxide (FeO) in the finished sinter directly reflects sintering quality. This key indicator is commonly obtained by data-driven soft sensor modeling technology. However, existing methods struggle to fully utilize the multisource and heterogeneous characteristics of the sintering dataset, and this may limit the improvement of FeO prediction accuracy. To address this issue, this paper proposes a multilevel deep fusion framework (MDFF) for detecting FeO content in the sintering process. This novel framework integrates a multilevel fusion of multisource heterogeneous data, i.e., data-level, feature-level, and decision-level. First, shallow features are extracted from images based on expert experience for data-level fusion. Then, multiscale encoding network and shared representation learning network are used for feature extraction and fusion. Finally, the prediction results of shared representations and modality-specific features are combined at the decision level to realize real-time sensing of FeO content. The proposed MDFF fully exploits the intrinsic correlations between multisource information at multiple levels, and experimental results on real sintering datasets further prove its accuracy and stability.