An Interpretable Knowledge-Enhanced Soft Sensor Framework for Basic Oxygen Furnace Endpoint Prediction Via LLM-driven Feature Engineering and SHAP-based Threshold Analysis

Basic oxygen furnace (BOF) steelmaking is a high-temperature chemical reaction process where endpoint carbon content and temperature are critical variables determining molten steel quality. Traditional data-driven models rely on manual feature engineering and iterative hyperparameter tuning, with predictions difficult to translate into actionable guidelines. This paper proposes an interpretable soft sensor framework integrating large language model (LLM)-based knowledge feature engineering, Tabular Prior-data Fitted Networks (TabPFN), and SHapley Additive exPlanations (SHAP)-based threshold analysis. The framework generates physicochemical knowledge-enhanced features via multiple LLMs, combines correlation screening with recursive feature elimination with cross-validation (RFECV) to select the optimal feature subset, applies TabPFN for rapid endpoint prediction, and extracts sensitive intervals and operational thresholds via SHAP dependence fitting. Using 814 heats from a 220-ton BOF, the model achieved test-set coefficient of determination (R2) of 0.5318 and 0.8315, hit rates (HR) of 79.75% (±0.01%) and 92.64% (±10 °C), and training times of 0.3805 s and 0.3944 s for carbon content and temperature, respectively. Against state-of-the-art models, R2 increased by 5.75% and 1.66%, HR by 3.06 and 1.23 percentage points, and training speeds by factors of 402 and 453, respectively. SHAP analysis identified a high-sensitivity interval of [1682, 1792] kg for light-fired dolomite on carbon content, and an efficient temperature regulation region within 700 kg of ore consumption. Over 52 industrial heats, HR reached 94.23% and 100% for carbon content and temperature. These results demonstrate that the framework enables rapid BOF endpoint prediction and operational threshold extraction, providing a quantitative basis for endpoint control.