Scholarlink

<p>Rapid expansion of large-scale solar farms has increased the need for accurate and efficient photovoltaic (PV) panel defect detection. The resilient defect localization approach presented in this study combines a powerful synthetic data augmentation strategy based on deep convolutional generative adversarial networks (DCGAN) with a light, custom-made YOLOv12n detection model. Replacing Conv with DWConv + Pointwise Conv in YOLOv12n achieves a computational complexity of 4.5 GFLOPS and a tiny model size of 1.807 million parameters. At the core of correcting severe class imbalance is the DCGAN module, which enhances sparse anomalies by generating high-quality synthetic images for each defect class. Based on experimental findings on the PVEL-AD and PV-MultiDefect datasets, the proposed method achieves real-time inference speeds of 2.3 ms and 4.3 ms per image, respectively, and a mAP@0.5 of 93.8% and 93.6%. Besides surpassing most state-of-the-art detection models in terms of effectiveness and accuracy, the work highlights the significance of DCGAN-based synthetic augmentation toward achieving even more generalization benefits for real-world PV defect detection. The suggested framework is most suitable for deployment on solar farm monitoring industrial edge devices because of its low parameter size and high inference rate.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Synthetic data generation with deep convolutional generative adversarial network for solar photovoltaic fault detection using resource-constrained framework

  • 期刊论文

  • 发表日期：

    2026年7月9日

查看全文
  • Amudala Jashwanth,
  • Nimmagadda Kesav Satya Sai,
  • Rayappa David Amar Raj,
  • Rama Muni Reddy Yanamala,
  • Archana Pallakonda

摘要

Rapid expansion of large-scale solar farms has increased the need for accurate and efficient photovoltaic (PV) panel defect detection. The resilient defect localization approach presented in this study combines a powerful synthetic data augmentation strategy based on deep convolutional generative adversarial networks (DCGAN) with a light, custom-made YOLOv12n detection model. Replacing Conv with DWConv + Pointwise Conv in YOLOv12n achieves a computational complexity of 4.5 GFLOPS and a tiny model size of 1.807 million parameters. At the core of correcting severe class imbalance is the DCGAN module, which enhances sparse anomalies by generating high-quality synthetic images for each defect class. Based on experimental findings on the PVEL-AD and PV-MultiDefect datasets, the proposed method achieves real-time inference speeds of 2.3 ms and 4.3 ms per image, respectively, and a mAP@0.5 of 93.8% and 93.6%. Besides surpassing most state-of-the-art detection models in terms of effectiveness and accuracy, the work highlights the significance of DCGAN-based synthetic augmentation toward achieving even more generalization benefits for real-world PV defect detection. The suggested framework is most suitable for deployment on solar farm monitoring industrial edge devices because of its low parameter size and high inference rate.