This paper integrates Large Language Models (LLMs) with deep learning to propose an LLMDL framework for the entire Domain Knowledge Graph (DKG) construction process. The framework includes: data preprocessing drawing on text classification principles; semi-automated domain ontology construction via LLMs; domain-adaptive optimization of the W2NER named entity recognition (NER) model with LLM-based data annotation and result verification; relation extraction considering inter-relation correlations; entity alignment through the integration of SBERT and LLMs; and construction of high-quality DKG with scenario-specific applications. Experimental results show that the proposed method effectively balances text value while controlling text length; improves the F1 score of named entity recognition by 2.24% compared with the original model; achieves a 22.07% F1 score improvement in relation extraction compared with traditional BERT models; and 84.85% of entities achieve standardized expressions in knowledge fusion.