Entity Recognition for Coal Mine Hydraulic Support Installation Process Driven by LLM LoRA Fine-Tuning

Hydraulic supports, being the pivotal equipment in coal mining face operations, exhibit complex installation procedure knowledge that impedes efficient knowledge extraction and utilization, thereby hindering the provision of scientifically grounded installation guidance and ultimately affecting equipment installation efficiency. This study proposes the development of a domain-specific large-scale model utilizing named entity recognition (NER) for knowledge extraction to enhance the efficiency of hydraulic shield installation. Initially, a few-shot data augmentation method is introduced to enrich hydraulic shield assembly process data, thereby providing a robust dataset for fine-tuning the large language model (LLM). Subsequently, Low-Rank Adaptation (LoRA) fine-tuning techniques are leveraged to optimize large-scale model adaptation. Comparative analysis of the model’s performance post-fine-tuning was conducted using multiple evaluation metrics, revealing that the fine-tuned Deepseek-R1-7b-Distill model exhibited the most superior performance indicators. Ultimately, the fine-tuned Deepseek-R1-7b-Distill model was selected as the domain-specific LLM for NER in hydraulic support installation processes. The experimental results demonstrate that the entity recognition F1 score across all entity types reached 0.8887, validating the efficacy of the methodology. This provides technical support for enhancing the installation efficiency of hydraulic supports.