Foreign object detection in power transmission lines using SESYOLO.

Area of Science:

• Electrical Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Foreign object attachment to power transmission lines is a frequent cause of electrical faults.
• Existing object detection methods struggle with accurate identification and sample scarcity for foreign objects on power lines.

Purpose of the Study:

• To address the scarcity of foreign object samples on power lines.
• To develop a high-precision, low-latency foreign object detection algorithm for power transmission lines.

Main Methods:

• Utilized Artificial Intelligence Generated Content (AIGC) for abundant, high-quality training data.
• Introduced Spatial and Channel Reconstruction Convolution (SCConv) for efficient feature learning.
• Implemented Efficient Reparameterized Generalized-FPN (Efficient RepGFPN) for effective information exchange.
• Developed Squeeze and Excitation Detect (SE-Detect) for richer feature extraction with fewer parameters.
• Employed WIoU loss function and a distillation schema for performance enhancement.

Main Results:

• Achieved a 9% increase in mAP@.5 and a 9.1% improvement in recall rate compared to YOLOv8.
• Demonstrated a mAP@.5 of 93.9% specifically for bird's nest detection.
• The algorithm effectively detects small and irregular foreign objects in cluttered visual environments.

Conclusions:

• The developed algorithm significantly enhances the detection of foreign objects on power transmission lines.
• The combination of AIGC, SCConv, Efficient RepGFPN, and SE-Detect improves precision and reduces latency.
• The model shows particular effectiveness in aerial inspection scenarios with small targets and complex backgrounds.