Abstract:

Meteorological data, as a typical spatiotemporal big data, faces severe data security challenges while empowering economic and social development. Addressing current issues in meteorological data security monitoring, such as insufficient semantic understanding, low accuracy in data feature recognition, and poor generalization capability, this study proposed an intelligent semantic recognition framework for meteorological data based on large language models. By constructing high-quality training datasets and domain knowledge bases, integrating Retrieval-Augmented Generation (RAG) with LoRA lightweight model technology, applying Chain-of-Thought (CoT) fine-tuning, and selecting PPO as the reinforcement learning algorithms to continuously optimize the recognition performance of the meteorological data security model. Experimental results demonstrate that this method effectively improves the accuracy of meteorological data feature recognition.

Key words: large language models, data security, semantic intelligent recognition, RAG, CoT