Enhancing Wind Turbine Nacelle Fire Detection via Stable Diffusion

Wind turbine nacelle fire detection plays a critical role in ensuring production safety. Deep learning-based monitoring systems offer distinct advantages in efficiency, cost-effectiveness, and real-time performance. However, the lack of high-quality fire-specific training data for wind turbine nacelles remains a significant challenge. In this work, we propose a novel framework to synthesize realistic wind turbine nacelle fire monitoring datasets by leveraging the strong prior knowledge of Stable Diffusion (SD) and Large Language Models (LLMs). Specifically, we use the Low-rank Adaption(LoRA) and noise inversion techniques, as well as a vision-language model to generate annotated wind turbine nacelle fire data, we build a training dataset with 5,000 images for the application scenario. Experimental results demonstrate that our method significantly improves the accuracy and robustness of fire detection systems compared to baseline approaches.