Application of artificial intelligence to tunnel fire monitoring and early warning systems

In tunnel fire safety prevention and control, artificial intelligence (AI) has gradually become a key means of accurate monitoring and intelligent warning of tunnel fires because of its ability to solve problems of traditional monitoring, such as long response delays, long prediction times, and high false alarm rates. Prediction models based on empirical formulas take minutes to calculate, failing to meet the needs of early intervention. Therefore, using AI to accurately predict fire development trends (such as smoke spread and temperature distribution) is crucial for formulating emergency plans and ensuring safety. In particular, fire monitoring and intelligent early warning methods based on AI have become an important research direction in tunnel fire safety research.

Applications of AI to tunnel fires include the development of few-shot and self-supervised learning methods to enhance model generalization ability. They also involve promoting system integration and standardization to realize platform-based collaborative management. In multisource data collection, multi-sensor fusion adopts an improved hierarchical architecture based on D–S evidence theory. It integrates temperature, smoke, and gas data, thereby improving fire identification reliability by 45% in complex environments. Video monitoring relies on CNN (convolutional neural network) and YOLOv8 algorithms, combined with tunnel CCTV (closed-circuit television) systems, to analyze flame and smoke characteristics. It achieves 96% recognition accuracy and reduces the false alarm rate by 30%. Edge computing has achieved up to 96% accuracy and supports real-time alarms. At the platform level, AI-based disaster prevention and response systems (e.g., Shanghai's intelligent system) enable real-time visualization of fire locations and temperatures. They automatically trigger coordinated control of ventilation and sprinkler systems, reducing response delays by more than 50% compared with manual operation. In terms of intelligent early warning, generative AI, such as GANs (generative adversarial networks) and Transformers, can generate fire spread simulations within 5 s. LSTM–TCNN (long short-term memory-temporal convolutional neural network) reduces temperature field prediction from minute-level to second-level (with 90% accuracy), and digital twins construct 1∶1 virtual tunnels to generate synthetic data, thereby reducing the demand for training data by 50%.

AI can effectively improve detection accuracy and response efficiency in tunnel fire monitoring and early warning. However, several challenges remain, including the scarcity of real-world samples (applying highway models to railways reduces accuracy by 15%–20%), the limited ability of traditional algorithms to capture global features, a lack of standardization in system integration, and high deployment costs. Future research will focus on using generative diffusion models to generate high-fidelity data and alleviate the sample scarcity issue, while reinforcement learning will be employed to optimize the collaborative control of equipment. In addition, a three-dimensional visualization platform based on BIM (building information modeling) and digital twins will be developed to enable VR/AR-based simulations. Further improvements in multimodal fusion are expected to enhance data reliability and cross-scenario adaptability, thereby advancing the intelligence of tunnel fire prevention and control. This research will contribute to improve the intelligence level of tunnel fire early warning and emergency response and promote the practical application of AI in tunnel fire engineering.