The rapid development of information technology and the deep advancement of Industry 4.0 are driving the widespread application of digital twin technology in the intelligent transformation of manufacturing. However, existing digital twin models in manufacturing still have limitations at the semantic level, specifically manifested as: inadequate multidimensional attribute descriptions of physical assets, insufficient dynamic interaction capabilities, and lack of unified semantic standards and specifications. To address this, this paper proposes a semantic reconstruction method for digital twin models based on large language models (LLMs). This method utilizes a digital twin definition language to build a semantic information asset library and integrates large language models with knowledge graph technology to generate and supplement asset semantic information in real time, thereby achieving dynamic semantic reconstruction of digital twins. Finally, experiments are conducted on a piston processing production line for validation. The results show that compared to traditional manual modeling methods, the semantic reconstruction method based on large language model significantly improves efficiency and stability in handling complex tasks, can quickly respond to dynamically changing production demands, and has advantages in visual presentation.