Scholarlink

In this paper, a deep learning model based on Vision Transformers (ViT) has been introduced to accurately identify a wide range of photovoltaic (PV) module defects from infrared thermal images. The developed model has been evaluated on a comprehensive dataset encompassing 12 distinct anomaly classes, including cell defects, cracking, diode issues, hot spots, soiling, and vegetation interference, among others. This model performed best in detecting diode anomalies (97% accuracy), identifying normal operating conditions (96% accuracy), and identifying cracking defects (67% accuracy). The model also performed well in distinguishing single-cell and multi-cell faults, a traditionally difficult task in photovoltaic anomaly detection. For binary classification (anomaly or non-anomaly), the model achieved an overall accuracy of 88%. In the multiclass classification task, model performance varied across different anomaly types, with an average accuracy of 81% across the 12 anomaly classes.

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

Anomaly Detection and Classification of Solar Photovoltaic Modules Using Vision Transformers (ViT)

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Yacine Boutana,
  • Sofiane Haddad,
  • Ammar Soukkou

摘要

In this paper, a deep learning model based on Vision Transformers (ViT) has been introduced to accurately identify a wide range of photovoltaic (PV) module defects from infrared thermal images. The developed model has been evaluated on a comprehensive dataset encompassing 12 distinct anomaly classes, including cell defects, cracking, diode issues, hot spots, soiling, and vegetation interference, among others. This model performed best in detecting diode anomalies (97% accuracy), identifying normal operating conditions (96% accuracy), and identifying cracking defects (67% accuracy). The model also performed well in distinguishing single-cell and multi-cell faults, a traditionally difficult task in photovoltaic anomaly detection. For binary classification (anomaly or non-anomaly), the model achieved an overall accuracy of 88%. In the multiclass classification task, model performance varied across different anomaly types, with an average accuracy of 81% across the 12 anomaly classes.