Construction, evaluation, and application of an intelligent agent for diagnosis in sintering process

As a key procedure of iron and steel metallurgy, the sintering process has long depended on expert experience and is confronted with challenges including strong multivariable coupling and dynamic response lag. The existing data-driven methods have deficiencies in the full utilization of textual knowledge and interpretability. Oriented to the intelligent transformation of the sintering process, this study aims to construct an intelligent question-answering system that deeply integrates domain knowledge and data models. The system can realize accurate reliable and interpretable automatic answers to professional questions and provide support for process optimization and personnel training. This study proposes a solution for large language models that integrates efficient low-rank adaptation(LoRA) fine-tuning and retrieval-augmented generation(RAG). The system takes the DeepSeek-R1-Distill-Qwen-7B model as the base framework. A sintering domain dataset containing 35 019 high-quality question and answer pairs was adopted for LoRA fine-tuning. The method significantly improved the model's semantic understanding of professional terminologies and process logic at a low cost. An RAG module was integrated simultaneously. The module provides accurate external knowledge for model generation through real-time retrieval of a local process knowledge base, which restricts model output and alleviates model hallucinations. Experimental results show that the joint training framework greatly improves the model's performance in answering professional sintering questions. The logic of model answers increases from 1.2 to 2.8 with a full score of 3.0. The correctness of model answers rises substantially from 0.7 to 6.5 with a full score of 7.0. Factual errors of model output are effectively reduced. This study verifies that the joint strategy combining domain fine-tuning and retrieval augmentation effectively improves the practicability and reliability of large language models in vertical domains. The research provides a feasible technical path and practical paradigm for the intelligentization of complex industrial processes based on large language models. Further research can be conducted on multimodal information fusion and real-time decision-making in the future.