Transmission line defect detection via an integrated improved YOLOv8 and deep neural random forest framework

Power transmission lines often suffer from various defects that can impact their functionality and safety. Detecting these defects is crucial for timely maintenance and prevention of potential failures. This study introduces a novel framework that combines an enhanced version of the YOLOv8 algorithm with a deep neural random forest (DNRF) model to improve defect detection in transmission lines using UAV imagery.

The researchers developed a spatially deformable convolution (SDC) algorithm to improve feature extraction from the images captured by UAVs. This technique allows the model to better capture the spatial relationships between different features in the images. Additionally, they introduced a more comprehensive hybrid loss function that enables the model to recognize defects of varying scales with greater accuracy.

After processing the images, the enhanced model feeds the candidate regions into a deep neural decision forest (DNRF) for fine-grained classification. The DNRF model categorizes the defects into precise types, such as normal, stains, cracks, scratched surfaces, and surface peeling.

Extensive experiments demonstrated the effectiveness of the proposed algorithm. Compared to traditional methods like Faster R-CNN, the new framework achieved a recognition accuracy of over 92% for the five defect types mentioned earlier. It also demonstrated notable improvements in terms of speed and efficiency. Specifically, the algorithm increased the detection speed by 36 frames per second (FPS) while reducing the number of parameters by 54.57% compared to Faster R-CNN.

Furthermore, when compared to YOLOv7-M, YOLOv9-C, and YOLOv11-M, the algorithm achieved a slight 2.6% improvement in mean average precision (mAP) for YOLOv9-C, a 2.38% gain for YOLOv7-M, and a 1.9% increase for YOLOv11-M. Although there was a slight reduction in inference speed, the overall performance gains in accuracy and parameter efficiency made the proposed approach highly effective for identifying multi-scale defects in power transmission lines.

This research was partially supported by the State Grid Hebei Electric Power Company Science and Technology Guide Project (kj2023-004). The corresponding authors, Liangshuai Liu, Lingming Meng, Anchang Li, Peng Yan, and Yuntao Zhao, are affiliated with the State Grid Hebei Electric Power Research Institute in Hebei, China. The study is published under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, allowing non-commercial use and sharing of the research findings while crediting the original authors and sources.