Abstract
High-accuracy oxygen content measurement and control is one key to improving combustion efficiency and economic efficiency. The soft measurement technique of the oxygen content based on flame images is promising. However, image feature acquisition at different oxygen contents and image generation under unbalanced conditions are still challenging. To relieve this dilemma, a new generative-based regression model is developed. It not only learns the potential vectors but also captures flame features well to generate virtually high-quality-labeled flame images. The training data sets can be augmented, thus saving a lot of data collection experiments. Subsequently, a convolutional-based regression model is constructed to estimate the oxygen content using the augmented flame images directly. The designed method generates informative flame images and obtains more accurate oxygen content estimation results than several common methods.