For the low-concentration gas direct combustion system in coal mines, traditional sensors such as temperature probes and ionization rods are insufficient to accurately reflect the subtle fluctuations in the combustion process. To address this problem, a collaborative monitoring strategy for multi-window low-concentration gas combustion states is proposed. Firstly, a low-concentration gas combustion state dataset is constructed using high-speed cameras (HSC) and high-definition cameras (HDC) with different viewing windows. Data enhancement analysis, such as Gaussian blur, is performed on the dataset to expand its diversity to improve the adaptability of the model. Further, a low-concentration gas combustion state monitoring model is constructed, and backbone, neck, and head models are established, respectively. The backbone model improves the ability to extract local features; the neck model enhances the sensitivity to the dynamic change trend; the head model optimizes the efficiency of gas combustion state feature extraction and reduces the extraction of redundant features. Finally, on the premise that the model parameter quantity meets the deployment requirements, the low-concentration gas combustion state monitoring model is deployed to the embedded platform to monitor the complex combustion state changes of the gas in real-time. The experimental results show that the precision of the low-concentration gas combustion state monitoring model is 96.11%, the accuracy is 95.80%, the monitoring rate is 87.5 fps, the number of model parameters is 40.1×106, the average error is no more than 0.14, the average precision of the model deployed on the embedded platform is 96.10%, and the average monitoring speed is 30.10 fps. It meets the precision, real-time, and adaptability requirements of the gas combustion state monitoring of the low-concentration gas direct combustion system, and has certain industrial application value.