Developing Ship Electronic Lookout Using LoRA Fine-Tuned Large Language Model
摘要
To alleviate the pressure on crew members in the lookout and warning tasks during ship navigation, this research developed a ship electronic lookout (SEL) using the Low-Rank Adaption (LoRA) fine-tuned Large Language Model (LLM). The SEL can synthesize heterogeneous information from multiple sensors, such as CCTV and AIS, to depict the current navigation environment in real-time and provide lookout alerts. Specifically, real-world navigation data is first collected, and the sensor information is processed into a dataset (LookOut). Subsequently, Qwen2-7B-Instruct was selected as the baseline to balance model responsiveness and safety. Finally, LoRA was used for supervised fine-tuning (SFT) to develop the SEL. Experimental demonstrate that LLM fine-tuned for specific application scenarios can mimic the language norms of the ship lookout and be combined with perception systems to achieve electronic lookout.