A new method based on YOLOv5 and multiscale data augmentation for visual inspection in substation

Area of Science:

• Electrical Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Artificial intelligence (AI) shows promise for substation visual inspection and defect detection.
• Practical applications face challenges due to dynamic shooting environments and limited datasets, leading to suboptimal accuracy and stability.
• Existing methods struggle to consistently identify defects in complex substation settings.

Purpose of the Study:

• To enhance the accuracy and stability of AI-based defect detection in substations.
• To address limitations posed by dynamic shooting environments and insufficient data.
• To propose a novel approach combining advanced data augmentation and model optimization.

Main Methods:

• An enhanced multiscale data augmentation technique was developed to improve recognition accuracy in time-varying conditions.
• YOLOv5 (You Only Look Once version 5) was utilized for defect detection using multi-scale image data.
• A novel model pruning method was implemented to strategically reduce YOLOv5 parameters, enhancing model stability and accuracy.

Main Results:

• The proposed enhanced multiscale data augmentation effectively mitigated environmental variability.
• Hybrid pruning of YOLOv5 significantly improved defect identification accuracy and model stability.
• Experimental results on substation defect images validated the methodology's effectiveness.

Conclusions:

• The hybrid pruning YOLOv5 and multiscale data augmentation approach offers a robust solution for substation defect detection.
• This method enhances AI performance in challenging, real-world substation environments.
• The findings contribute to more reliable and accurate automated visual inspection systems for critical infrastructure.