Lightweight Transmission Line Fault Detection Method Based on Leaner YOLOv7-Tiny

Qingyan Wang, Zhen Zhang, Qingguo Chen, Junping Zhang, Shouqiang Kang2024 · DOI

TL;DR

This paper introduces "Leaner YOLOv7-Tiny," a significantly lighter and more efficient object detection model for identifying faults in transmission lines from aerial images. By optimizing the YOLOv7-Tiny architecture with depthwise separable convolutions and a novel attention mechanism, the model achieves high detection accuracy, particularly for small targets, making it ideal for deployment on resource-constrained drones.

What Problem Does This Paper Address?

Detecting faults in transmission lines is crucial for maintaining power grid reliability. Traditional methods often rely on complex image processing and machine learning techniques, which can be computationally intensive. The rise of drone-based inspections offers a promising avenue, but existing deep learning models, while accurate, are often too large and resource-hungry for practical drone deployment. This paper addresses the need for a lightweight yet accurate fault detection system that can operate efficiently on drones, overcoming the limitations of parameter complexity and high computational load in current state-of-the-art models.

Methods

The core of this work is the development of "Leaner YOLOv7-Tiny," a streamlined version of the YOLOv7-Tiny object detection model. The key modifications include:

  • Depthwise Separable Convolutions (DepthSepConv-S): The standard convolutional layers in the YOLOv7-Tiny backbone are replaced with depthwise separable convolutions, inspired by the PP-LCNet architecture. This technique decomposes standard convolutions into depthwise and pointwise operations, drastically reducing the number of parameters and computational cost without significant loss in feature extraction capability.
  • SP Multi-Scale Spatial Attention Mechanism: A novel attention mechanism, "SP attention," is introduced. This mechanism uses convolutional kernels of varying sizes (3x3, 5x5, 7x7) to capture features at multiple scales. This is particularly beneficial for detecting small targets, which are often challenging to identify due to their limited visual information.
  • Improved FCIoU Loss Function: The standard CIoU loss is enhanced to FCIoU loss. This modification aims to better balance the contributions of high-quality and low-quality samples during training. By giving more weight to high-quality samples (larger IoU), it accelerates model convergence and improves overall detection accuracy, especially in the presence of varied image quality often encountered in real-world scenarios.

Datasets

The dataset used in this study comprises authentic images of transmission lines collected by the State Grid Harbin Power Supply Company's Inspection Center. This includes aerial photos of overhead transmission lines, frames from aerial videos, and publicly sourced insulator images. The dataset contains various fault types, such as insulator flashover ("Boom"), missing equatorial ring ("Fall"), and bird nests ("Birds"). To enhance the model's generalization and robustness, image transformations like brightness enhancement/reduction, noise addition, color shifting, and exposure adjustments were applied to the original dataset.

Key Insights and Ideas

  • Lightweighting for Practical Deployment: The primary insight is that significant reductions in model size and computational load are achievable by strategically replacing standard convolutional layers with depthwise separable convolutions, making advanced deep learning models feasible for drone applications.
  • Multi-Scale Attention for Small Targets: The SP attention mechanism highlights the importance of capturing features at various scales to effectively detect small and often difficult-to-spot targets like transmission line faults.
  • Loss Function Optimization for Real-World Data: The FCIoU loss demonstrates that tailoring the loss function to better handle imbalanced data quality (common in aerial imagery) can lead to faster convergence and improved accuracy.
  • Trade-offs in Lightweighting: While lightweighting reduces parameters, it can initially decrease accuracy. The paper shows that strategic integration of attention mechanisms and improved loss functions can not only recover but even surpass the original model's performance.

Key Results

  • Model Size Reduction: Leaner YOLOv7-Tiny achieved a 20% reduction in model size compared to the original YOLOv7-Tiny, decreasing from 6.02 MB to 4.67 MB.
  • Detection Accuracy: The proposed model achieved the highest detection accuracy among the compared lightweight models, with a [email protected] of 98.5% and [email protected]:0.95 of 69.5%.
  • Robustness: Experiments showed that Leaner YOLOv7-Tiny consistently and accurately detected all faults with no instances of misses or false detections, outperforming other mainstream lightweight models in robustness.
  • Small Target Detection: Attention heatmaps indicated that Leaner YOLOv7-Tiny exhibited a higher level of attention towards small objects compared to other models, confirming its effectiveness in small target detection.

Limitations and Future Work

  • Detection Time: While significantly improved in accuracy and size, Leaner YOLOv7-Tiny exhibits a slight increase in detection time compared to some other lightweight models due to the introduction of the attention mechanism. However, it still meets real-time requirements.
  • Further Optimization: The paper suggests that future work could explore further optimizations to reduce detection time while maintaining high accuracy, potentially through more advanced architectural designs or hardware-specific optimizations.
  • Dataset Expansion: While data augmentation was used, the authors could explore larger and more diverse datasets to further enhance the model's generalization capabilities across a wider range of environmental conditions and fault types.

Summary updated

BibTeX
@misc{wang2024lightweight,
  author = {Qingyan Wang and Zhen Zhang and Qingguo Chen and Junping Zhang and Shouqiang Kang},
  title = {{Lightweight Transmission Line Fault Detection Method Based on Leaner YOLOv7-Tiny}},
  year = {2024},
  doi = {10.3390/s24020565},
}
Figure