中图分类号：TP391.41 文献标志码：A

Improved YOLOv8 for gas flame state recognition under low-pressure conditions

SAl Qingyi1，ZHAO Jin1，YAN Yonghui1，BI Degui² （1.SchoolofEnergyandPowerEngineering，UniversityofShanghaiforScienceandTechnologyhanghai 2o93，China; 2.SchoolofEnvironmentandArchitecture，UniversityofShanghaiforScienceandTechnologyShanghai 2Oo93，China）

Abstract： Gas flame detection in high-altitude low-pressure environments faces dual challnges of insufficient recognition accuracy and real-time performance degradation. This study proposes an improved lightweight YOLOv8n algorithm with multi-dimensional optimizations for enhanced performance. First， a GhostConv module was constructed in the backbone to reduce computational parameters. Second， the C2f module was restructured into a C2f_RepGhost configuration that maintained feature representation capability while simplifying inference processes. Finaly， the convolutional block attention module （CBAM） attention mechanism was incorporated to strengthen fine-grained flame feature extraction， and the WIoU loss function was adopted to improve localization accuracy. Experimental results based on mechanism testbed data demonstrate that the improved model's parameter count decrease by 12.64% ， computational load decrease by 12.2% ， and precision increase by 21.2% while maintaining frame rate detection. This method provides an effective lightweight solution for flame state recognition in low-pressure environments.

Keywords： low-pressure; flame detection; deep learning; YOLOv8; attention mechanism

以甲烷为主要燃料的燃气炉对燃烧稳定性的要求极高，不稳定燃烧或熄火现象极易触发爆炸事故。高原地区低压、低氧环境会降低甲烷的燃烧效率[1]，导致火焰形态[和稳定性[3]等发生变化，进而影响燃烧过程的安全性。因此，精准识别火焰状态对燃烧过程的实时监控与安全预警至关重要，关系到生命与财产安全。

早期的燃气火焰状态检测技术依靠红外摄像机、温度传感器、压力传感器等硬件设备[4]，易被干扰并且准确率不高。近年来，深度学习技术在图像检测领域取得了很好的应用效果，逐渐成为了主流检测手段。白卫东等[5]采用支持向量机（supportvectormachine，SVM）对锅炉燃烧火焰图像进行分类识别，实现了准确的燃烧状态监测。Ronquillo-Lomeli等l运用概率神经网络（probabilisticneuralnetwork，PNN）算法对锅炉火焰进行分类，准确率达到了 90% 以上。Gangopadhyay等[7]使用3D神经网络和长短期记忆网络，实现了对实验室燃烧火焰的稳定与不稳定状态的识别。Zhao 等[8结合YOLOv8与VQ-VAE，提高了实际检测的准确率。Yar等改进了YOLOv5s模型，减少了模型尺寸和复杂度，并提高了准确率。王冠博等[]提出了改进YOLOv4-tiny 轻量化模型，通过增加感受野并对模型进行剪枝，提升准确率的同时大幅减少了模型参数量以及权重文件。陈露萌等[1]改进了YOLOv5s模型，在特征提取部分添加协同注意力机制，并改进 a -CIoU，提高了模型的收敛速度。

尽管这些研究在平原环境下取得了良好的效果，但在高原环境下，火焰的边缘形状、面积和颜色变化迅速：随着气压降低，火焰边缘形状膨胀变得更加明显；……