Data-Driven Framework for Anomaly Identification and Process Optimization in Continuous Casting

Continuous casting is a critical and complex process in steel manufacturing, in which product quality is highly sensitive to process parameter variations. Early identification of anomalous parameters and timely process optimization are, therefore, essential for reducing quality defects and production losses. However, existing data-driven approaches often face limitations in practical industrial applications, particularly in achieving high accuracy under strict response time requirements. In this study, a data-driven framework is proposed to address the above limitations by identifying key anomalous parameters and inferring feasible parameter adjustments under operational constraints, with interpretability analysis incorporated to enhance consistency with metallurgical knowledge. The proposed framework was trained and evaluated on a real-world continuous casting dataset comprising 191,254 records, where the data were temporally split, and the final 25 pct was reserved as an independent test set. Experimental results demonstrate strong performance in both anomaly identification and process optimization, achieving a Top-10 anomaly identification rate of 96.64 pct and an optimization hit rate within ±5 pct of 91.54 pct, demonstrating its effectiveness and practical applicability.