主管：中华人民共和国应急管理部
主办：应急管理部天津消防研究所
ISSN 1009-0029  CN 12-1311/TU
消防科学与技术

消防科学与技术 ›› 2026, Vol. 45 ›› Issue (6): 79-89.DOI: 10.20168/j.1009-0029.2026.06.0079.11

• • 上一篇    下一篇

基于深度学习的隧道火灾温度场实时反演模型：构建与部署

章忠华1, 刘江2, 刘磊3, 吴珂3, 张天航3   

  1. （1.杭州市建设工程质量安全监督总站，浙江 杭州 310000； 2.中铁隧道股份有限公司，河南 郑州 450007； 3.浙江大学 城市火灾安全工程研究中心，浙江 杭州 310058）
  • 收稿日期:2025-07-02 修回日期:2025-12-03 出版日期:2026-06-15 发布日期:2026-06-15
  • 作者简介:章忠华，杭州市建设工程质量安全监督总站，工程师，主要从事市政隧道建设运营与消防管理工作，浙江省杭州市拱墅区莫干山路100号，310000。
  • 基金资助:
    国家自然科学基金项目（52478422）；物联网技术应用交通运输行业研发中心开放课题（202401T）

A deep learning framework for real-time reconstruction of tunnel fire temperature fields： model design and deployment

Zhang Zhonghua1, Liu Jiang2, Liu Lei3, Wu Ke3, Zhang Tianhang3   

  1. (1. Hangzhou City Construction Engineering Quality and Safety Supervision Station, Hangzhou Zhejiang 310000, China; 2. China Railway Tunnel Group Co., Ltd., Zhengzhou Henan 450007, China; 3. Research Center for Urban Fire Safety Engineering, Zhejiang University, Hangzhou Zhejiang 310058, China)
  • Received:2025-07-02 Revised:2025-12-03 Online:2026-06-15 Published:2026-06-15

摘要： 隧道火灾具有突发性强、烟气蔓延快、能见度低等特点，严重威胁交通基础设施和通行人员的安全。为应对这一挑战，本文建立了基于转置卷积神经网络（TCNN）的隧道火灾温度场实时反演模型，并提出“层级感知火场策略”，实现了将局部隧道训练的AI模型应用于全局隧道，突破了传统测温手段在空间分布、实时性和成本方面的限制。TCNN模型通过转置卷积层高效捕捉火灾温度场的空间特征，能够精准反演隧道火灾中的温度分布。通过模块化FDS仿真数据库开展了模型设计、训练与性能评估，在此基础上设计了部署方案，结合Kafka消息队列与WebSocket实时推送机制，实现模型推理结果的异步传输与前端可视化展示，并通过双线程机制完成系统状态的智能切换与控制。该方法为隧道智慧消防系统的设计与工程部署提供了参考路径与技术支持。

关键词: 隧道火灾, 转置卷积神经网络, 温度场反演, 深度学习, 部署策略

Abstract: Tunnel fires are characterized by their sudden occurrence, rapid smoke propagation, and low visibility, posing severe threats to transportation infrastructure and the safety of tunnel users. To address these challenges, this study proposes a real-time temperature field reconstruction model for tunnel fires based on a Transposed Convolutional Neural Network (TCNN). Furthermore, a “hierarchical perception fire scene strategy” is introduced, which enables AI models trained on local tunnel segments to be deployed in full-scale tunnel environments, thereby overcoming the limitations of traditional temperature measurement methods in terms of spatial coverage, real-time capability, and cost-effectiveness. The TCNN model efficiently captures the spatial features of the fire temperature field through transposed convolution layers, allowing precise inversion of the temperature distribution in tunnel fires. The model is developed, trained, and evaluated using a modular Fire Dynamics Simulator (FDS) simulation database. On this basis, a deployment framework is designed that integrates Kafka message queues and WebSocket protocols to achieve asynchronous transmission of inference results and real-time front-end visualization. A dual-thread control mechanism is also implemented to enable intelligent state switching and system control. The proposed method provides a viable technical pathway and engineering reference for the design and implementation of intelligent fire protection systems in tunnel environments.

Key words: tunnel fire, TCNN, temperature field inversion, deep learning, deployment strategy