Research on Deep Learning-Based Multi-Level Cross-Domain Foreign Object Detection in Power Transmission Lines

Qingxue Liu, Xia Wang, Yun Su, Wei Jiang, Zheng Zhang, F. Shen, Lingxi ZhuSensors 2025 · DOI

TL;DR

This paper introduces CO-YOLO, a novel deep learning model for detecting foreign objects on power transmission lines. By integrating the YOLOv11 architecture with the ConvNeXt network and employing Bayesian optimization, CO-YOLO significantly improves detection accuracy, especially for small and occluded objects, while maintaining high real-time performance.

What Problem Does This Paper Address?

Power transmission lines are critical infrastructure, and foreign objects like bird nests, kites, and debris pose significant safety hazards, potentially leading to short circuits, fires, and widespread power outages. Traditional manual inspection methods are inefficient and lack real-time capabilities. While deep learning has advanced object detection, existing models often struggle with detecting small or occluded objects and can be too computationally intensive for edge deployment or real-time monitoring. This paper aims to address these limitations by developing a more robust, accurate, and efficient foreign object detection system for power transmission lines.

Methods

The core of this research is the enhancement of the YOLOv11 object detection model. The authors propose a "CO-YOLO" (ConvNeXt-You Only Look Once) model that integrates the ConvNeXt network into the YOLOv11 architecture. This integration aims to leverage ConvNeXt's ability to capture richer appearance features and spatial dependencies of foreign objects.

Key modifications and techniques include:

  • YOLOv11 Backbone Enhancements: Several modules were integrated into the YOLOv11 backbone to refine feature extraction:
  • C3k2-Dual Module: Replaces traditional convolutions with Dual Conv to improve computational efficiency and parameter reduction while enhancing feature extraction.
  • NonLocalBlockND Module: Captures long-range dependencies across feature maps to better understand global context.
  • C2PSA-DHSA Module: Replaces the conventional C2PSA module's attention mechanism with DHSA to improve image restoration and focus on critical regions without overfitting.
  • DySample Module: An advanced upsampling module that uses a dynamic point-sampling strategy to better preserve detailed features and semantic content in low-resolution feature maps.
  • ConvNeXt Integration: The ConvNeXt-B network is incorporated to capture feature relationships, leading to richer appearance features and spatial dependencies of foreign objects.
  • PIoU Loss Function: Replaces the SIoU loss function to improve localization precision and overall detection performance by accounting for angular discrepancies between predicted and ground-truth bounding boxes.
  • Bayesian Optimization: Used for automated hyperparameter tuning, accelerating model convergence and improving generalization capability.

Datasets

The researchers compiled a custom dataset of 979 images for foreign object detection on transmission lines. This dataset includes four categories: bird nests, balloons, branches, and plastic debris. To augment the dataset and simulate adverse conditions, they applied various data augmentation techniques, including random cropping, flipping, rotation, grayscale transformations, and adjustments to hue, saturation, exposure, and brightness, increasing the dataset size to 2739 images.

They also conducted generalized experiments on two public datasets:

  • FOTL (China Insulator Dataset): Includes insulator data from railway, communication, and power sectors.
  • CPLID (Southeast Asian Power Industry Dataset): Includes bird nests, branches, flying objects, and regional flora and fauna.

Key Insights and Ideas

  • Multi-level Cross-Domain Fusion: The core idea is to fuse features from different levels and domains (visual features from YOLOv11 and spatial appearance cues from ConvNeXt) to create a more comprehensive understanding of foreign objects.
  • Attention for Global Context: The NonLocalBlockND module is crucial for capturing long-range dependencies, which is vital for understanding the context of foreign objects on vast transmission line structures.
  • Dynamic Upsampling for Detail Preservation: The DySample module addresses a key challenge in object detection: preserving fine details in feature maps, especially for smaller objects.
  • Bayesian Optimization for Efficiency: Automating hyperparameter tuning with Bayesian optimization significantly speeds up the process and leads to better model performance.
  • PIoU Loss for Better Localization: Moving beyond SIoU to PIoU loss is a subtle but important improvement for precise bounding box regression, especially for objects with unusual aspect ratios.

Key Results

The proposed CO-YOLO model demonstrated significant improvements:

  • Accuracy: Achieved an [email protected] of 98.4% and [email protected]:0.95 of 66.1%.
  • Efficiency: Reached an inference speed of 303 FPS.
  • Comparison to YOLOv11: CO-YOLO improved [email protected] by 1.9% and [email protected]:0.95 by 2.2% compared to the original YOLOv11 model.
  • Comparison to ETLSH-YOLO: Achieved improvements of 3.7% and 8.6% in mAP metrics.
  • Comparison to TFD-YOLOv8: Showed remarkable gains of 33.0% in [email protected] and 80.4% in [email protected]:0.95.
  • Generalization: Demonstrated strong performance on public datasets (FOTL and CPLID), with gains up to 6.4% in [email protected]:0.95 compared to baseline models.

Limitations and Future Work

  • Limitations: The paper acknowledges that the model's performance might degrade under extreme environmental conditions (e.g., fog, snow) or for rare object types not present in the training set.
  • Future Work: The authors suggest future research directions including:
  • Developing lightweight models and acceleration techniques.
  • Exploring multimodal perception (e.g., combining visual data with thermal or radar).
  • Achieving task-level generalization for broader applicability.

Summary updated

Abstract

With the rapid advancement of deep learning technology, deep learning-based methods have become the mainstream approach for detecting potential safety hazards in transmission lines, playing a crucial role in power grid safety monitoring. However, existing models are often overly complex and struggle with detecting small or occluded targets, limiting their effectiveness in edge-device deployment and real-time detection scenarios enhanced the YOLOv11 model by integrating it with the ConvNeXt network, a multi-level cross-domain analysis detection model (ConvNeXt-You Only Look Once) is proposed. Additionally, Bayesian optimization was employed to fine-tune the model's hyperparameters and accelerate convergence. Experimental results demonstrate that CO-YOLO [email protected] reached 98.4%, [email protected]:0.95 reached 66.1%, and FPS was 303, outperforming YOLOv11 and ETLSH-YOLO, in both accuracy and efficiency. Compared with the original model, CO-YOLO model improved by 1.9% in [email protected] and 2.2% in [email protected]:0.95.
BibTeX
@inproceedings{liu2025research,
  author = {Qingxue Liu and Xia Wang and Yun Su and Wei Jiang and Zheng Zhang and F. Shen and Lingxi Zhu},
  title = {{Research on Deep Learning-Based Multi-Level Cross-Domain Foreign Object Detection in Power Transmission Lines}},
  booktitle = {Sensors},
  year = {2025},
  doi = {10.3390/s25165141},
}
Figure