Defect detection of photovoltaic modules based on improved VarifocalNet

Area of Science:

• Electrical Engineering
• Computer Science
• Materials Science

Background:

• Defective photovoltaic modules significantly reduce power generation efficiency.
• Current deep learning methods often prioritize either speed or accuracy, limiting practical application.
• A balanced approach is needed for effective defect detection in solar panels.

Purpose of the Study:

• To propose an improved VarifocalNet model for enhanced detection speed and accuracy of defective photovoltaic modules.
• To address the limitations of existing methods that focus on single performance metrics.
• To improve the overall efficiency and reliability of photovoltaic power generation.

Main Methods:

• Designed a novel bottleneck module with standard and dilated convolutions to increase network depth and receptive field.
• Developed a smaller-parameter bottleneck module to improve detection speed.
• Incorporated a feature interactor in the detection head to enhance classification accuracy.
• Utilized an improved intersection over union (IoU) in the loss function for better bounding box prediction.

Main Results:

• The proposed method achieved the highest detection accuracy among compared techniques.
• The improved VarifocalNet demonstrated faster detection speeds compared to most existing methods.
• The modifications successfully enhanced both the speed and accuracy of defective photovoltaic module detection.

Conclusions:

• The enhanced VarifocalNet effectively balances detection speed and accuracy for photovoltaic module defect identification.
• The proposed architectural and loss function improvements offer a significant advancement in the field.
• This method contributes to more reliable and efficient solar energy generation through improved defect management.