Transformer oil leakage has become an increasingly prominent fault in power systems

posing a significant risk to equipment reliability and operational safety. To address the current challenges of limited training data and low image quality in transformer oil leakage detection

this paper proposes a method for generating transformer oil leakage image datasets based on a liquid spreading model and physical modeling. The method simulates the diffusion

evaporation

and absorption of transformer oil on ground surfaces and numerically solves the process using grid partitioning and the finite difference method

thereby obtaining the density distribution of oil stains on the ground. In the process of simulating the spreading of oil stains on the ground

anisotropic diffusion coefficients and spatial modulation of surface material properties are introduced to enhance the interpretability and diversity of oil stain morphology. By integrating simulations of multi-scene and multi-view within 3D virtual environments

the method expands data coverage and provides sufficient training samples for detection models. Experiments are carried out on image similarity and transformer oil leakage detection based on synthetic datasets. The generated data have high visual and structural similarity with real images

and can effectively serve transformer oil leakage detection tasks.