Intelligent Tunnel Fire Detection Technology Based on the Large Language Model Multi-Agent Collaboration

To address the issues of manual operation dependency and low efficiency in tunnel fire research combining computational fluid dynamics (CFD)with deep learning, this paper proposes a multi-agent collaborative framework based on large language models to automate the entire process of inverting fire source characteristics. The framework decomposes the traditional workflow into four specialized agents, namely physical modeling, data governance, model training, and evaluation analysis, which collaboratively execute end-to-end tasks from CFD scenario generation to model deployment. The results demonstrate that the CNN-LSTM model performs optimally. Under a 6 second observation window and 10 meter sensor spacing, the average R² reaches 0.942, representing a 2% improvement over the baseline LSTM model, while the RMSE is reduced by 28.8%. Under sparse deployment with 30 meter spacing, the average R² remains as high as 0.917, validating the effectiveness of integrating spatial feature extraction with temporal modeling. This work provides an efficient technological pathway for intelligent tunnel fire identification and advances the research paradigm from manual optimization to multi-agent system optimization.