Single Domain Generalization in Vibration Damper Detection Using Frequency Domain Attention and Scale-Invariant Learning

Improving the generalization capability of vibration damper detection model for transmission lines is essential for its application in complex power scenarios. Variations in vibration damper types, and unmanned aerial vehicle(UAV) shooting distances lead to domain shift, which impacts detection performance. This paper proposes a single domain generalization (SDG) detection method for vibration dampers that integrates structural knowledge to guide the model in extracting domaininvariant features (DIF), combining frequency domain attention (FDA) and scale-invariant learning (SIL). Different types of vibration dampers exhibit combinatory structural invariance, while dampers of varying scales demonstrate aspect ratio invariance. First, a learnable weight matrix is applied to the amplitude spectrum after Fourier transformation, preserving low-frequency information while attenuating high-frequency components. This approach retains structural features and reduces morphological differences. Next, SIL is introduced to emphasize aspect ratios during bounding box regression, improving the model's capacity to handle scale variations and enhancing feature generalization. Experimental results demonstrate that without significantly increasing the number of parameters and model size, the proposed method achieves AP50 of 60.8% and 77.1% in the two target domains, respectively, representing increases of 3.3% and 4.0% over the original model, thus offering strong support for the practical application of intelligent detection models