Enhancing Scrap Steel Yield Identification Precision by Community Division of Knowledge Graph

Accurately identifying scrap steel yield rates remains challenging due to the diverse types, mixed sources of scrap, and complex furnace working conditions. This paper proposes a mechanism and data joint-driven identification method, and identification precision is enhanced by community division of a knowledge graph. Firstly, a knowledge graph for scrap charging is constructed, and the label propagation algorithm (LPA) is used to divide communities with similar charging patterns. Then, a physics-informed neural network is designed for each community to identify scrap steel yield rates. Finally, the shapley additive explanations approach is employed to assess and quantify the influence of these factors on scrap steel yield rates. Experimental results indicate the following: (1) The proposed model for scrap steel yield rate based on knowledge graph community division achieves the highest identification precision, with a Root Mean Square Error (RMSE) of 3.20 tons and a Mean Absolute Error (MAE) of 2.62 tons. (2) Compared with the baseline joint-driven model without community division, the proposed method reduces the MAE by 26.6% (from 3.57 t to 2.62 t) and significantly improves the hit rate within ±5 tons by 13.97 percentage points (reaching 87.55%). These improvements validate the effectiveness of the community-based divide-and-conquer strategy in handling complex charging patterns.