Multi-Level Fusion Transformer Coupled With Mechanistic Model and Multimodal Data for Soft-Sensor Modeling in Sintering Process

Timely and reliable estimation of Ferrous Oxide (FeO) in sintered ore is increasingly critical for blast furnace control under tightening energy and emission constraints. Offline chemical assays incur substantial delays and high costs, depriving operators of the rapid feedback needed for timely set-point correction. Meanwhile, the sintering process is strongly nonlinear and heterogeneous across sensing modalities, making single-source soft sensors brittle in production. To address this, we present a multimodal fusion Transformer for online FeO soft sensing that couples mechanistic models with data-driven learning: thermodynamic state variables from a temperature-field model are fused with process time series at the data level to inject physics-informed context, and a dual-stream encoder performs deep fusion of time series and image sequences via cross-attention to achieve fine-grained temporal–spatial alignment. Experimental results on a real-world sintering dataset demonstrate that the proposed model significantly outperforms baseline methods in terms of RMSE and R2. Furthermore, the fusion strategy generalizes to stronger Transformer backbones, indicating architectural portability and robustness. The results highlight that incorporating domain knowledge with attention-based multimodal fusion is an effective route to industrial soft sensing under real-world constraints.