Abstract:
Objectives The causal factors behind carrier-based aircraft launch and recovery anomalies are characterized by scarcity and unpredictability. To mitigate the hallucination problem of large language models (LLMs) in generating such anomaly causes, we proposes a Scenario Generation for Carrier Aircraft Driven by Large–Small Model Collaboration (SGCAD) method.
Methods First, a carrier-based aircraft launch and recovery knowledge base is constructed using professional literature and retrieval-augmented generation (RAG) techniques, forming a dataset of normal operation descriptions. Then, normal descriptions are used as templates combined with scenario-specific prompts, enabling the large model to generate potential anomaly causes. A smaller model is subsequently employed to discriminate between reasonable and unreasonable anomaly causes. Finally, the large model is fine-tuned through Direct Preference Optimization (DPO) and iteratively refined to progressively increase the proportion of reasonable anomalies.
Results Experimental results show that after multiple SGCAD iterations, the proportion of reasonable anomaly causes in the generated set reaches 94%, effectively alleviating hallucination issues and significantly improving the rationality and realism of the generated content. Expert evaluation further confirms that the generated anomaly causes cover diverse and complex scenarios, employ standardized terminology, and comply with physical laws.
Conclusions The proposed approach provides a valuable reference for the analysis of carrier-based aircraft launch and recovery anomalies.
下载: