Research on Life-Survivability Assessment Methods for Submarine Damage-Induced Flooding Based on Large Language Models

ObjectiveTo address the lack of domain knowledge and the “hallucination” problem encountered when general-purpose large language models are applied to submarine damage-control scenarios involving damage-induced flooding, this paper proposes a large language model adaptation method for submarine survivability assessment under damage-induced flooding based on Low-Rank Adaptation (LoRA) and Retrieval-Augmented Generation (RAG). The aim is to improve the model’s professional knowledge comprehension and survivability-index calculation capabilities in damage-induced flooding survivability assessment scenarios.MethodsA domain-specific fine-tuning dataset for damage control, containing 10,089 samples, was constructed based on authoritative materials such as professional knowledge in submarine damage-induced flooding damage control, academic monographs, technical reports, and national standards. A damage-control domain knowledge base was also established using lightweight markup language text as the minimum knowledge unit. LoRA was adopted to conduct domain-knowledge alignment training on the baseline model, thereby enhancing the model’s understanding of submarine damage-control knowledge. Meanwhile, RAG was integrated to accurately retrieve calculation procedures for survivability indicators under damage-induced flooding, guiding the model to complete survivability-index calculations.ResultsAfter LoRA fine-tuning, the large model for submarine damaged-flooding survivability assessment outperformed the baseline model in knowledge question-answering tasks in terms of both Bilingual Evaluation Understudy (BLEU) and Recall-Oriented Understudy for Gisting Evaluation (ROUGE) metrics. In the task of calculating submarine damaged-flooding survivability indicators, the accuracy of the model’s calculation results increased from 33.33% for the baseline model to 93.33%, while the compliance rate of the calculation process increased from 13.33% to 96.67%.ConclusionThe submarine damage-induced flooding survivability assessment large language model designed in this paper can effectively improve the knowledge comprehension and computational reasoning capabilities of general-purpose large language models in the field of submarine damage-induced flooding damage control, providing standardized and traceable decision-data support for emergency response in submarine damage-control scenarios.