A Novel 24 h × 7 Days Broken Wire Detection and Segmentation Framework Based on Dynamic Multi-Window Attention and...

Area of Science:

• Electrical Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Substation wire damage detection is difficult due to variable lighting and insufficient data.
• Existing methods struggle with accuracy and robustness in real-world conditions.

Purpose of the Study:

• To develop a robust 24/7 framework for detecting and segmenting damaged substation wires.
• To address challenges of low-light conditions and limited annotated data.

Main Methods:

• Proposed a novel framework integrating low-light image enhancement, a YOLOv11n-based detection/segmentation network with dynamic multi-scale window attention (DMWA), and meta-transfer learning.
• Constructed a dataset of 3760 RGB images.
• Evaluated performance across lighting conditions from 10 to 200,000 lux.

Main Results:

• The proposed framework significantly enhanced broken wire detection and segmentation performance.
• Demonstrated improved robustness across a wide range of lighting conditions.
• Meta-transfer learning effectively supported small-sample training and mitigated negative transfer.

Conclusions:

• The novel framework offers a substantial improvement for automated substation wire inspection.
• The dynamic multi-scale window attention and meta-transfer learning strategies are key to robust performance.
• This approach enhances safety and efficiency in substation maintenance.