Interpretable recognition of sintering conditions via multi-source fusion and mechanism consistency

Stable and fluctuating operating conditions in iron ore sintering arise from the coupled evolution of bed permeability and thermal state. Because state changes are coupled and time-lagged, single-signal monitoring is often insufficient. This study develops a multi-source fusion framework, termed MGFNet, for recognizing fluctuating sintering conditions by integrating process variables and tail-end cross-section features from variable-length sequences. A mechanism consistency constraint is introduced to align the two sources and reduce non-physical correlations. Experiments show that MGFNet achieves an accuracy of 0.96, recall of 0.94, and F1-score of 0.96, outperforming random forest, support vector machine, long short-term memory, and Transformer baselines. The constraint reduces missed detections from 11% to 6% with limited false alarms. SHapley Additive exPlanations analysis further provides interpretable feature insights.