MDFF: A Multilevel Deep Fusion Framework of Multisource Heterogeneous Data for FeO Content Prediction in Sintering Process

As the key quality indicator in the iron ore sintering process, the ferrous oxide (FeO) content in the finished sinter directly reflects its thermal state and denotes sintering quality. Currently, data-driven soft sensor modeling technology has become the mainstream method for predicting this quality indicator. However, existing approaches either fail to leverage the multimodal properties of the sintering dataset or simply neglect to explore the interrelationships between multisource heterogeneous data, thus limiting the improvement of prediction accuracy. To address this issue, we propose a multilevel deep fusion framework (MDFF) for real-time FeO content detection, which combines data fusion, feature fusion, and decision fusion methods together to form an end-to-end framework for soft sensor modeling. First, shallow features are extracted from cross section images based on expert knowledge for data-level fusion. Then, two multiscale encoding networks and a shared representation learning network are employed to realize feature-level fusion. Subsequently, the final predicted FeO content is obtained by fusing the prediction results of shared representations and modality-specific features at the decision level. By integrating multisource data and designing multilevel information fusion, the proposed MDFF method can make full exploitation of the intrinsic interactions among multisource information at various levels. The experimental results of the actual sintering process show that MDFF is superior to the existing models in terms of prediction accuracy and stability.