Dynamic Flame Image Generation for Ammonia Combustion Monitoring Using Trend-Encoded Diffusion Models

Abstract

Ammonia is a promising green energy carrier, but ammonia combustion, often accompanied by high NOx emissions, poses a great challenge for real-time monitoring. Although flame image-based soft sensors offer a solution for online NOx estimation, the training of this type of data-driven models is constrained by the scarcity of high-quality labeled samples in practical scenarios. To achieve high-quality generation of complex dynamic combustion data for soft sensor modeling, this study proposes a temporally conditioned diffusion generation framework integrated with trend encoding. This framework constructs a trend-aware encoder based on sparse attention and motion bias, designed to extract knowledge of dynamic flame evolution patterns and incorporate it into the diffusion model training process to generate flame sequences that conform to combustion dynamics. Furthermore, a state consistency regularizer is introduced to mitigate label drift during generation and promote alignment between the textures of generated images and their corresponding pollutant labels. Experiments conducted on an ammonia combustion data set demonstrate that the proposed method outperforms conventional approaches in image generation and significantly improves the prediction accuracy of the established soft sensor model. Compared with the unaugmented baseline, the proposed method reduces the mean prediction error for NOx and O2 by 40.63% and 39.16%, respectively.