Abnormal furnace condition perception of blast furnace based on cross-modal knowledge collaboration under low resource samples

To address the challenges of furnace condition perception in blast furnace production caused by data scarcity and complex anomalies, this article proposes a cross-modal knowledge collaboration-based framework for abnormal furnace condition perception. This framework integrates data-driven and knowledge-driven features to achieve high-precision identification of multiple categories of abnormal states under low-resource conditions. Principal component analysis is applied to high-dimensional production data (including key process parameters such as blast pressure, blast volume, and permeability index) for dimensionality reduction and key feature extraction, aiming to eliminate noise while retaining primary variation information. A domain graph integrating equipment, process, and anomaly knowledge is constructed, and through knowledge embedding and cross-modal feature fusion, an enhanced input is generated. Finally, the Light Gradient Boosting Machine model is employed for efficient identification. Experiments demonstrate that this method achieves an accuracy of 97.2% in identifying five categories of furnace condition anomalies under weak supervision, significantly outperforming random forest (78.3%) and XGBoost (96.2%).