基于CvT模型的高原环境下燃气燃烧火焰图像识别方法研究
Research on Flame Image Recognition Method for Gas Combustion in Plateau Environment Based on CvT Model

针对高原低压环境下燃气锅炉燃烧状态监测难题，本文提出基于CvT模型的火焰图像智能识别方法。在青海西宁的一个1.4 MW燃气锅炉试验台上进行实验，采集典型工况燃气火焰图像，提取、分析其火焰图像特征，在此基础上，提出CvT-13混合模型，融合CNN局部特征提取与Transformer全局时序建模优势对火焰状态进行识别，实现模型特征与图像特征参数的协同分析。结果表明，CvT-13模型可实现高原低压环境下火焰状态的多维度表征，在背景、稳定及不稳定燃烧状态的分类测试中达到99.51%平均准确率，其中熄火不稳定状态召回率达99.67%，背景识别精确度100%。本文为高原低氧环境下的燃烧稳定性实时诊断提供了高精度、强鲁棒性的解决方案。
To address the challenges of combustion state monitoring in gas-fired boilers under plateau low-pressure environments, this study proposes an intelligent flame image recognition method based on a Convolutional vision Transformer (CvT) model. Experimental investigations were conducted on a 1.4 MW gas-fired boiler testbed in Xining, Qinghai Province, where characteristic flame images under typical operating conditions were collected and analyzed. Building upon extracted flame image features, we developed a CvT-13 hybrid model that integrates the advantages of CNN-based local feature extraction and Transformer-based global temporal modeling for flame state identification, enabling synergistic analysis between model characteristics and image feature parameters. Experimental results demonstrate that the CvT-13 model achieves multi-dimensional characterization of flame states in low-pressure plateau environments, attaining 99.51% average accuracy in classifying background, stable combustion, and unstable combustion states. Particularly notable performance includes 99.67% recall rate for flameout-unstable states and 100% precision in background recognition. This research provides a high-precision and robust solution for real-time diagnosis of combustion stability in hypoxic plateau environments.