TMF-Det: a scenario-based flame and smoke detection method for early warning of externally caused mine fires

Abstract

To address the difficulty of detecting small early-stage flames, weak-boundary smoke, low illumination, and strong interference from dust and water mist in externally caused mine fires, this paper proposes a scenario-based flame and smoke detection method named TMF-Det. Based on YOLOv11s, the method constructs the MineBG v3 hard-background dataset and introduces real underground images without fire or smoke to constrain false positives from tunnel lighting, miner-lamp reflections, metal highlights, dust, and water mist. A dual-path augmentation strategy combining general fire augmentation and mine-domain augmentation is used to simulate underground imaging degradation, while a Transformer attention enhancement module and a bidirectional feature pyramid network BiFPN multi-scale feature fusion structure are introduced to improve contextual discrimination and cross-scale feature representation. Evaluation on the fixed test set of the cleaned Fire and Smoke Detection Dataset showed that TMF-Det achieved a Precision of 91.60%, a Recall of 88.65%, an mAP@0.5 of 93.01%, and an mAP@0.5:0.95 of 77.21%. Compared with the YOLOv11s baseline, TMF-Det improved mAP@0.5 and mAP@0.5:0.95 by 3.16 and 3.25 percentage points, respectively. On an independent set of 3,171 fire- and smoke-free underground images collected from six mines, the number of false-positive boxes decreased from 185 for the YOLOv11s baseline to 37 for TMF-Det, corresponding to an 80.00% reduction. These results demonstrate the effectiveness of TMF-Det in improving flame and smoke detection and suppressing background-induced false positives, while further validation using real underground fire and smoke data remains necessary.