Online Estimation of Combustion Oxygen Content with an Image-Augmented Soft Sensor Using Imbalanced Flame Images

Area of Science:

• Combustion Engineering
• Image Processing
• Machine Learning

Background:

• Accurate oxygen content measurement is crucial for optimizing combustion efficiency and economic performance.
• Soft measurement techniques using flame images offer a promising approach for oxygen content analysis.
• Challenges remain in acquiring image features across varying oxygen levels and generating images under imbalanced conditions.

Purpose of the Study:

• To develop a novel generative-based regression model for accurate oxygen content estimation from flame images.
• To address challenges in image feature acquisition and generation under unbalanced data conditions.
• To enhance training datasets and improve the accuracy of oxygen content measurement.

Main Methods:

• A generative-based regression model was developed to learn potential vectors and capture flame features.
• The model generates virtually high-quality, labeled flame images, augmenting training datasets.
• A convolutional-based regression model estimates oxygen content directly from the augmented flame images.

Main Results:

• The generative model successfully captured flame features and generated informative synthetic flame images.
• Data augmentation through generated images reduced the need for extensive data collection experiments.
• The proposed method achieved more accurate oxygen content estimation compared to several common techniques.

Conclusions:

• The developed generative-based regression model effectively enhances oxygen content estimation from flame images.
• Synthetic image generation offers a viable solution for data augmentation in combustion analysis.
• This approach improves both the accuracy of oxygen measurement and the efficiency of the process.