YOLO-WAD for Small-Defect Detection Boost in Photovoltaic Modules

Area of Science:

• Materials Science
• Electrical Engineering
• Computer Vision

Background:

• Photovoltaic module performance is critical for solar cell lifetime.
• Accurate detection of small defects, like hot spots, remains a significant challenge.
• Existing methods struggle with the precise identification of micro-defects.

Purpose of the Study:

• To develop an advanced deep learning model for accurate detection of small defects in photovoltaic modules.
• To enhance the feature extraction and fusion capabilities for identifying subtle anomalies.
• To improve the overall performance and reliability of solar cell inspection systems.

Main Methods:

• Proposed YOLO-WAD model based on YOLOv10n architecture.
• Incorporated C2f-WTConv (wavelet transform convolution) in the backbone for enlarged receptive fields.
• Introduced an ASF structure in the neck for multi-level feature fusion.
• Added an extra detection layer with C2f-EMA (efficient multi-scale attention) for feature redistribution.
• Implemented a DyHead (dynamic head) for comprehensive scale, spatial, and channel awareness.

Main Results:

• YOLO-WAD achieved an overall accuracy of 95.6% on the test dataset.
• Small-target defect detection accuracy reached 86.3%.
• Demonstrated a 4.1% and 9.5% improvement over YOLOv10n and other mainstream models, respectively.

Conclusions:

• The proposed YOLO-WAD model effectively detects small-target defects in photovoltaic modules.
• The integration of wavelet transform convolution, attention mechanisms, and dynamic heads enhances detection accuracy.
• YOLO-WAD offers a feasible and improved solution for solar cell performance monitoring and defect analysis.