Research on Assistant Decision-making System for Reversible Quality Deviation Recovery in Long-process Steelmaking based on DBSCAN-LightGBM and IPORF

Abstract

To address the challenges of undetectable quality deviations and the lack of quantitative evaluation of downstream compensation capability in long-process steelmaking, this paper proposes an active collaborative control method based on DBSCAN-LightGBM and an Improved Parrot Optimization algorithm combined with Ordinal Regression Random Forest (IPORF). First, multi-source heterogeneous data from L2, MES, and LIMS are integrated, and anomaly diagnosis and restoration are performed using the Isolation Forest algorithm under metallurgical mechanism constraints, resulting in a high-dimensional structured dataset comprising 12,458 heats. Second, the DBSCAN algorithm is employed to automatically mine and label deviation patterns without prior labels, and a classification model is constructed based on LightGBM, achieving a weighted F1-score of 0.945 for deviation identification. Furthermore, the IPORF model is proposed to quantify the compensation capability of downstream processes for upstream deviations. Hyperparameters of the Ordinal Regression Random Forest are optimized using the Improved Parrot Optimization algorithm, yielding a model accuracy of 94.5% and a mean absolute error of 0.08. Experimental results demonstrate that the proposed method effectively enables closed-loop management from deviation perception to strategy generation, providing theoretical support and engineering demonstration for intelligent quality governance in steel manufacturing processes.