A multimodal feature fusion and large language model approach for the combustion stability diagnosis of 660 MWth coal-fired boilers

Under low-load flexible operation conditions, where coal-fired boilers often need to respond quickly to deep load changes, the combustion state of boiler burners is vital for furnace efficiency and safety. Although image-based flame detection offers higher monitoring accuracy than traditional methods, it cannot provide quantitative evaluations and clear diagnoses of flame stability. To address these issues, this study introduces a large-language-based approach utilizing a pre-trained Llama 3.2-3B model with lightweight fine-tuning for flame stability diagnosis. Temporal features from operational data and spatial features from flame videos are extracted using a Long Short-Term Memory network and a Vision Transformers advanced by exploring intrinsic inductive bias model. A custom multimodal fusion network then blends these complementary features to create a unified representation of combustion characteristics. Expert annotations are integrated during fine-tuning to improve understanding and diagnostic reasoning specific to combustion. Additionally, a multimodal feature database of stable conditions is built to enable quantitative evaluation of flame stability. Tests conducted on a 660MW opposed-fired boiler demonstrate the effectiveness of the proposed model. The results show the model's ability to distinguish between stable and unstable flame states across different operational loads. It accurately detects transitions between safe and unsafe combustion states and offers interpretable recommendations for adjustments. This provides a practical pathway for the safe, efficient, and intelligent operation of coal-fired boilers under variable load conditions.