Flame analysis and combustion estimation using large language and vision assistant and reinforcement learning

Abstrak

In this study, we present an advanced approach for flame analysis and combustion quality estimation in carbonization furnaces utilizing large language and vision assistant (LLaVA) and reinforcement learning from human feedback (RLHF). The traditional methods of estimating combustion quality in carbonization processes rely heavily on visual inspection and manual control, which can be subjective and imprecise. Our proposed methodology leverages multimodal AI techniques to enhance the accuracy and reliability of flame similarity measures. By integrating LLaVA’s high-resolution image processing capabilities with RLHF, we create a robust system that iteratively improves its predictive accuracy through human feedback. The system analyzes real-time video frames of the flame, employing sophisticated similarity metrics and reinforcement learning algorithms to optimize combustion parameters dynamically. Experimental results demonstrate significant improvements in estimating oxygen levels and overall combustion quality compared to conventional methods. This approach not only automates and refines the combustion monitoring process but also provides a scalable solution for various industrial applications. The findings underscore the potential of AI-driven techniques in advancing the precision and efficiency of combustion systems.