A Novel Adversarial Deep Learning Method for Substation Defect Image Generation.

Area of Science:

• Electrical Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Substation equipment defects pose significant risks to power transmission safety.
• Accurate and timely defect detection is critical for maintaining grid reliability.
• Supervised deep learning models for defect detection are hindered by insufficient defect image data, especially with complex backgrounds.

Purpose of the Study:

• To propose a novel adversarial deep learning model for generating realistic substation equipment defect images.
• To address the challenge of limited training data for defect detection models.
• To improve the performance of object detection models in identifying surface defects.

Main Methods:

• Developed the Abnormal Defect Detection Generative Adversarial Network (ADD-GAN).
• ADD-GAN generates defect images by segmenting local areas, avoiding global style distortion.
• Employed a joint discriminator for both overall and defect images to enhance focus on local defect features.

Main Results:

• The ADD-GAN method generated a high-fidelity dataset of substation equipment defects.
• The YOLOV7 object detection model, trained on ADD-GAN generated data, achieved 81.5% mean average precision (mAP).
• The proposed method outperformed existing image data augmentation and generation techniques.

Conclusions:

• The ADD-GAN effectively generates realistic defect images, significantly improving training datasets.
• This approach enhances the accuracy of deep learning-based defect detection models for substation equipment.
• The developed method contributes to safer and more reliable power transmission through improved defect identification.