Welding in shipbuilding remains highly dependent on skilled labour, while expert know-how is difficult to formalise and existing automation solutions are often fragmented at the system level. This paper presents a welder-centred intelligent welding framework that links IoT-enabled data collection, a welding data and knowledge repository and an LLM-based welding copilot. At the core of the copilot, we develop WeldGPT, a welding-domain large-model module built on Qwen-7B with LoRA-based fine-tuning and retrieval-augmented prompting over structured welding manuals and experience data. Welding tasks are represented by a unified Task and Condition Description, which, together with entries retrieved from Welding Manual and Welding Knowledge and Experience, is transformed by WeldGPT into Robot-Executable Welding Procedures in terms of parameter settings and robot-oriented operation plans. A prototype Weld Copilot system is instantiated using historical data and procedures from a robotic fillet-weld production line at a medium-sized shipyard located in Zhuhai, China. Preliminary offline results indicate that WeldGPT can reproduce the main parameter-setting patterns encoded in existing procedures and provide practically useful starting points for configuring and refining robotic welding programs, supported by a web-based interface for interactive inspection and adjustment.
Cao et al. (2026) studied this question.