Transformer Discharge Carbon-Trace Detection Based on Improved MSRCR Image-Enhancement Algorithm and YOLOv8 Model

Area of Science:

• Electrical Engineering
• Robotics
• Computer Vision

Background:

• Internal defects in large transformers are challenging to detect visually due to metal enclosures.
• Existing inspection methods lack efficiency and accuracy for complex internal environments.

Purpose of the Study:

• To develop an advanced image-enhancement and deep-learning-based method for internal transformer defect detection using a micro-robot.
• To improve the accuracy and speed of identifying carbon-trace defects in challenging conditions.

Main Methods:

• An improved MSRCR algorithm combined with white-balance was used for image enhancement.
• The YOLOv8 deep-learning model was enhanced with the SimAM mechanism and DyHead framework for feature extraction and detection.
• Model pruning was applied to YOLOv8 to optimize detection speed and efficiency.

Main Results:

• The proposed MSH-YOLOv8 algorithm achieved 91.80% accuracy, a 3.4% improvement over the original YOLOv8.
• The algorithm demonstrated a Frames Per Second (FPS) of 99.2, indicating reduced computational complexity.
• The method significantly outperformed other mainstream target-detection algorithms.

Conclusions:

• The MSH-YOLOv8 algorithm effectively enhances internal transformer defect detection accuracy and speed.
• The developed method is suitable for practical engineering applications with transformer internal-inspection robots.
• This approach offers a robust solution for non-destructive testing in electrical infrastructure.