Abstract: The application of artificial intelligence technology in power inspection is an important technological innovation, especially the object detection technology based on deep neural network. However, the research shows that deep neural networks are vulnerable to adversarial example attacks, which leads to the wrong judgment of object detection model and increases the security risk of power system. This paper studies the adversarial examples defense problem of object detection algorithm based on deep learning, dedicates itself to improving the robustness of the model while maintaining the natural accuracy of clean examples, proposes an adversarial training defense strategy based on contrast learning with multiple disturbance, and uses SimCLR contrast learning framework for adversarial training of the model. The model learns the consistency of adversarial examples and clean examples from the feature level, and uses the reuse disturbance strategy to generate adversarial examples to improve the efficiency of adversarial training. Experiments on China's power insulator data set prove that the proposed method successfully defuses a variety of countermeasures and maintains the accuracy of clean examples, effectively improves the robustness of the object detection model, and is suitable for maintaining accurate output in the face of complex scenarios in the field of power inspection, effectively improving the intelligence level of power inspection.