A Human–Robot Team Knowledge-Enhanced Large Language Model for Fault Analysis in Lunar Surface Exploration

Abstract

Human–robot collaboration for lunar surface exploration requires high safety standards and tedious operational procedures. This process generates extensive task-related data, including various types of faults and influencing factors. However, these data are characteristic of multi-dimensional, time series, and intertwined. Also, prolonged tasks and multi-factor data coupling pose significant challenges for astronauts in achieving safe and efficient fault localization and resolution. In this paper, we propose a method to enhance the base large language models (LLMs) by embedding knowledge graphs (KGs) of lunar surface exploration, thereby assisting astronauts in reasoning about faults during the exploration process. A multi-round dialog dataset is constructed through the knowledge subgraph embedded in the request analysis process. The LLM is fine-tuned using the p-tuning method to develop a specialized LLM suitable for lunar surface exploration. With reference to the situational awareness (SA) theory, multi-level prompts are designed to facilitate multi-round dialogues and aid decision-making. A case study shows that our proposed model exhibits greater expertise and reliability in responding to lunar surface exploration tasks than classical commercial models, such as ChatGPT and GPT-4. The results indicate that our method provides a reliable and efficient aid for astronauts in fault analysis during lunar surface exploration.