Abstract: The semantic segmentation of the image of power equipment in the transformer substation can provide the capability of scene interpretation for inspection robots, which is the key and the basis to realize the intelligent inspection and management.However, due to the variety of power equipments and the complex background environment, the accuracy of existing semantic segmentation methods that only relies on single-modal images is limited.To solve the above mentioned problem, a semantic segmentation network based on differential fusion of heterogeneozus features is proposed.According to the characteristics of different levels and modalities feature maps, a differentiated fusion strategy is applied to extract the complementary advantages of spatial details and semantic information with infrared and visible images so as to utilize fusion information to guide the decoding process and to realize stable semantic segmentation of heterogeneous images.In order to verify the performance of the proposed method, a large number of heterogeneous images were collected by unmanned aerial vehicles and robot platforms.In this way, semantic segmentation dataset for power scene is manually annotated and constructed.The test shows that the proposed method could accurately identify the various types of power equipment, which has practical value for ensuring the safe and stable operation of power systems.