Free AI Video Generator for Early Fire Detection in Subway Tunnels

Fire detection is crucial in subway tunnels, where confined spaces and ventilation systems can complicate safety monitoring. Existing approaches often rely on classical machine learning methods that detect fires based on multivariate correlations but struggle with contextual reasoning.

However, recent studies suggest that Large Language Models (LLMs) may help address this challenge by translating structured sensor summaries into concise semantic descriptions. To examine the role of LLMs in fire detection, researchers have proposed a hybrid framework called HyFiD, which employs an LLM as a semantic feature extractor to augment classical classifiers.

HyFiD converts momentary multi-sensor readings (temperature, smoke, O2, CO, and CO2) into textual assessments of environmental states, generating semantic vectors that are fused with numerical features for supervised classifier training. Experiments on Fire Dynamics Simulator (FDS)-based subway scenarios show that the effect of LLM-derived semantic features is strongly dependent on the downstream classifier.

The results indicate a clear trade-off between detection sensitivity and premature-alarm suppression. Lower thresholds generally preserve higher F1 scores but substantially increase Pre-Alarm Rate, indicating more aggressive alarm behavior. For instance, at τ = 0.25 and k = 1.0 s, SVM + LLM and GBM + LLM achieve high F1 scores of 87.45% and 90.85%, respectively, but their PAR values rise to 99.25% and 51.32%. In contrast, higher thresholds suppress premature alarms more strongly but often degrade F1.

The temporal persistence window further modulates this trade-off. Increasing the persistence window from 1.0 to 2.0 s or 3.0 s generally reduces PAR, but excessive smoothing can substantially reduce F1, especially for SVM + LLM and RF + LLM. GBM + LLM is comparatively more robust under moderate smoothing: at τ = 0.50 and k = 2.0 s, it maintains an F1 score of 88.33% while reducing PAR from 47.55% under the default setting to 28.68%. However, further increasing the window to k = 3.0 s reduces F1 to 83.46%, showing that overly conservative alarm persistence can also suppress valid fire detections.

To justify the architectural design of using the LLM as a semantic feature extractor rather than as an end-to-end decision maker, researchers evaluated direct LLM classification under Zero-shot, One-shot, and Few-shot Chain-of-Thought (CoT) prompting strategies. The results show that direct prompting is highly sensitive to both the prompting setting and the LLM backbone when applied to raw multivariate numerical sensor snapshots.

The final alarm behavior is highly sensitive to the selected threshold and persistence window. Researchers retain τ = 0.50 and k = 1.0 s as the default setting for the main experiments because it provides a consistent operating point for comparing classifiers, while the sensitivity analysis clarifies how alternative alarm rules shift the balance between F1, delay, and pre-ignition false alarms.

Under the Zero-shot CoT setting, several evaluated backbones—including Llama-3-8B41, Qwen-2.5-3B58, Qwen-2.5-7B, and Ministral-8B59—produce near-zero Recall and F1, indicating difficulty in directly mapping raw numerical sensor values to stable fire/no-fire decisions. However, Few-shot CoT improves the strongest prompt-only result, with Llama-3-8B reaching an Accuracy of 65.67% and an F1 score of 70.32%. Nevertheless, this remains below the supervised hybrid approach: GBM + Llama-3-8B achieves an F1 score of 90.77% when the LLM is used as a semantic feature generator rather than as a standalone classifier.

These comparisons suggest that, in this setting, LLMs are more reliable when used to produce intermediate semantic representations for supervised downstream classifiers than when used as direct numerical classifiers. This finding has significant implications for machine learning jobs and AI-generated image applications, where accurate decision-making is critical.