SUN Mingzhu, ZHANG Yiran, ZHOU Jicheng, WU Guangxi, JIN Xuanyue, ZHANG Ji, ZHANG Wenhao, SHI Ruchuan, HAN Tao
The ironmaking-steelmaking interface is a critical production interface linking blast-furnace tapping, hot-metal transportation, hot-metal pretreatment, and basic oxygen furnace steelmaking. Its operation involves interactions among multiple production entities, strong cross-process coupling, stringent temporal constraints, and complex boundary conditions. As production information systems continue to accumulate multi-source heterogeneous data, conventional analytical approaches based primarily on single-field thresholds and expert experience cannot adequately integrate anomaly detection, causal explanation, and evidence tracing. Moreover, direct reasoning by general-purpose large language models may produce conclusions that are inconsistent with observed process data. To address these limitations, this study proposes an intelligent-agent framework for anomaly detection and root-cause attribution in hot-metal ladle circulation by integrating metallurgical process rules with retrieval-augmented generation. Production records collected over seven consecutive months from three steelmaking lines at a large iron and steel enterprise in Jiangsu Province were used to develop and evaluate the proposed method. Through object alignment and temporal normalization, the primary circulation process chain of hot-metal ladles was reconstructed and linked to events involving the hot return of residual molten steel from continuous casting. A deterministic anomaly-detection module based on metallurgical process rules and an anomaly-attribution module incorporating external knowledge retrieval, evidence binding, and conflict resolution were then developed. The intelligent agent performs user-intent recognition, tool-call parameter generation, module selection, and structured result presentation. Functional regression testing showed that the data-processing, tool-invocation, and agent-orchestration pipelines functioned correctly under the predefined test cases. On 80 expert-annotated samples, the full proposed method achieved a Hit@1 of 82.50%, outperforming the direct-generation baseline. Ablation experiments further showed that metallurgical process rules, external knowledge retrieval, and evidence verification made distinct contributions to candidate-cause identification, evidence support, and inspection-item coverage. The proposed method establishes a closed-loop analytical workflow encompassing data association, anomaly detection, knowledge retrieval, causal explanation, and human review, thereby providing interpretable and traceable support for anomaly tracing, process knowledge accumulation, and intelligent operations and maintenance at the ironmaking-steelmaking interface.
