Transmission line foreign object segmentation based on RB-UNet algorithm

Area of Science:

• Electrical Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Identifying foreign objects on transmission lines is critical for operational integrity.
• Challenges include scattered distribution and elevated heights, posing risks and operational difficulties.

Purpose of the Study:

• To develop an advanced semantic segmentation network for accurate foreign object detection on transmission lines.
• To improve upon existing methods for enhanced power line monitoring and safety.

Main Methods:

• Introduced RB-UNet, a novel semantic segmentation network.
• Enhanced U-Net by replacing the encoder with ResNet50 for feature extraction.
• Incorporated batch normalization in the decoder and a combined Focal and Dice loss function to address class imbalance and boost accuracy.

Main Results:

• Achieved a mean Intersection over Union (mIoU) of 88.43% for foreign object detection.
• Demonstrated significant superiority over other semantic segmentation techniques.
• RB-UNet proved proficient in detecting and segmenting diverse foreign objects like nests, kites, balloons, and rubbish.

Conclusions:

• The proposed RB-UNet algorithm offers a highly effective solution for foreign object detection on transmission lines.
• The enhanced architecture and loss function contribute to superior performance in challenging monitoring scenarios.
• This technology enhances the safety and reliability of power transmission infrastructure.