Prompt Engineering Approach Study for Supervised Fine-Tuned (SFT) Large Language Models (LLMs) in Spacecraft Fault Diagnosis

Traditional correlation and causality analyses based on telemetry data often lack depth in analyzing the multi-level coupling of faults and their emergent effects in spacecraft. Large Language Models (LLMs) offer a viable solution to this issue, necessitating meticulous research and design in prompt engineering. Against this backdrop, we have conducted a series of studies on using LLMs for diagnosing coupled faults in spacecraft. Initially, we systematically integrated a comprehensive fine-tuning dataset that includes prior expert knowledge, historical fault records, maintenance reports, and operation logs, to fine-tune the ERNIE-Bot-turbo model on the Baidu AI Cloud Qianfan platform. Subsequently, in the context of practical spacecraft fault diagnosis, we quantitatively explored the impact of different prompt engineering approaches on the fine-tuned LLMs using a specially designed multi-label spacecraft fault diagnosis test dataset. Finally, we proposed an Advanced Min Max Ant System (AMMAS) algorithm for autonomously iterating to explore the optimal prompts. AMMAS leverages the strong language generation capability of LLMs to optimize prompt text generation, using F1-score as the fitness function, aiming to autonomously iterate and optimize the prompt generation process, uncovering the most effective prompts for spacecraft fault diagnosis.