<p>The assembly process of a coal mining machine generates complex unstructured textual data, which makes it difficult for workers to quickly understand and reason with during the assembly stage, thereby reducing operational efficiency. Traditional BERT-based Named Entity Recognition methods exhibit poor performance when faced with scarce and highly specialized datasets. To address this issue, this paper proposes a domain-specific large language model-based method for Named Entity Recognition. Firstly, foundational large language models are comparatively analyzed to select the one that best adapts to the coal mining machine domain data, enabling effective handling of the complex assembly data. Secondly, by applying the QLoRA fine-tuning method, the training parameters of the large language model are adjusted, which reduces the computational resource demands for building a domain-specific model while maintaining performance. Finally, the model is fine-tuned and evaluated on a real-world coal mining machine assembly dataset. Experimental results demonstrate that the QLoRA fine-tuning leads to a significant performance improvement: the BLEU-4 score increases from 6.1225 to 65.8013, and the F1-Score for the NER task reaches 0.893. The proposed method improves the accuracy for the NER task in coal mining machine assembly and is of significant importance for assisting workers in understanding assembly information.</p>