Lightweight Detection Methods for Insulator Self-Explosion Defects.

Area of Science:

• Electrical Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Accurate detection of defective insulators is critical for power grid safety in intelligent inspection systems.
• Traditional object detection models suffer from large parameter sizes, low accuracy, and slow speeds.

Purpose of the Study:

• To develop a lightweight and efficient insulator defect detection model.
• To improve accuracy and speed for real-time power grid inspections.

Main Methods:

• Proposed a lightweight Faster Region-based Convolutional Network (Faster R-CNN-tiny) model.
• Replaced ResNet with EfficientNet in the backbone, utilized feature pyramid networks, and employed depth-wise separable convolutions.
• Implemented transfer learning with a freezing/unfreezing training strategy for small defect detection.

Main Results:

• Faster R-CNN-tiny demonstrated significantly improved mean average precision (mAP) and frames per second (FPS) compared to Faster R-CNN (ResNet).
• The model achieved a substantial reduction in the number of parameters.
• Effectively improved detection of small insulator defects.

Conclusions:

• Faster R-CNN-tiny offers a superior solution for insulator defect detection, balancing accuracy, speed, and model size.
• The proposed model enhances the safety and efficiency of intelligent electrical system inspections.