The nature of fire is amorphous and its characteristics vary based on the space, environment, and materials involved. Particularly, early fire detection is a crucial task in preventing large-scale accidents. However, there is a significant lack of learnable early fire datasets for machine learning approaches. This study presents an early fire detection system tailored to specific spaces, achieved through a digital twin-based automatic fire learning data generation model. The proposed method starts by automatically generating realistic particle simulations to create synthetic fire data in RGB-D images. These images are matched to the view angle of monitoring cameras to replicate the digital twin environment closely resembling the actual space. In essence, our approach produces synthetic fire data that captures diverse fire scenarios unique to each specific location. Subsequently, these datasets are employed for transfer learning, enhancing the capabilities of state-of-the-art detection models. The improved models are then deployed on AIoT devices within the real space. This spatially optimized synthetic fire data generation process enhances the accuracy and reduces false detection rates in comparison to existing fire detection models that lack adaptability to specific spaces.