Insulator Defect Detection Method Based on Mixture of Experts Learning

Insulator defect detection is a critical technology for ensuring the safe operation of power systems. However, factors such as diverse defect types, small target scales, and complex backgrounds severely limit the performance of traditional detection methods. This paper proposes a multi-modal insulator defect detection method based on a mixture of experts learning. The method uses Grounding DINO as the base framework and introduces multiple lightweight LoRAbased expert modules in the shallow layers of the image encoder, enabling the model to learn specific features of different defect types. During training, a random expert sampling strategy is adopted to enhance generalization capability, and a two-stage training scheme is used to optimize overall performance; during inference, all expert outputs are fused to fully utilize complementary features. Experimental validation on a self-built dataset containing 2000 real-scene images shows that the proposed method significantly outperforms existing mainstream methods in detection accuracy, particularly excelling in detecting small-target defects while maintaining high inference efficiency. This research provides an effective technical solution for intelligent insulator defect detection in complex scenarios.