Detection of Symmetry-Breaking Insulator Defects via Asymmetric Direction-Aware YOLOv11

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

In power transmission systems, insulator strings typically exhibit a distinct periodic geometric symmetry. However, defects such as breakages and flashovers disrupt this symmetry, producing subtle, orientation-sensitive asymmetric patterns that traditional isotropic convolution kernels often struggle to capture. To tackle the challenge of identifying these symmetry-breaking defects, we propose the Direction-aware You Only Look Once version 11 (DA-YOLOv11), a lightweight detection framework designed for asymmetric, direction-aware feature learning. By integrating the PaddlePaddle Lightweight Convolutional Network (PP-LCNet) into the backbone, the model preserves essential structural features while minimizing computational redundancy. We also introduce the Insulator-oriented Simple Parameter-Free Attention Module (INS-SimAM), a direction-aware attention mechanism that employs an asymmetric spatial statistical strategy to adaptively enhance defect-related signals within non-uniform regions. Additionally, the Adaptive Numerically Stable Normalized Wasserstein Distance (ANS-NWD) loss models tiny defects as 2D Gaussian distributions, leveraging their mathematical symmetry to ensure stable regression. Evaluations on a custom dataset and three public benchmarks (IDID, CPLID, UPID) show that the final pruned DA-YOLOv11 achieves an 88.95% mean Average Precision (mAP50) on the core dataset, outperforming the baseline while reducing the parameter count by 34.36%. This work provides a robust solution for detecting geometric symmetry-breaking in complex power inspection scenarios.