Scholarlink

<p>In this study, faults on PV cells were detected with high accuracy using images obtained from a thermal camera. The dataset used in this study consists of 2401 original UAV-based thermal images collected from PV power plants located in ten provinces of Turkey with different climatic conditions. Powerful CNN-based structures with different architectural families, basic deep learning models, transformer-based and hybrid approaches, lightweight focus structures, and classical machine learning models were grouped in a multi-class manner. Within these groups, 11 different AI models were compared under the same experimental conditions. Four different classes were evaluated for classification, “healthy”, “hotspot anomaly”, “bypass diode fault”, and “string fault”. Among the architectures evaluated, DenseNet121 achieved the highest numerical accuracy at 98.7%. However, performance differences were relatively small, particularly among the top-performing models such as DenseNet121, CNN+Transformer, and ResNet50.</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

Multi-class performance evaluation of artificial intelligence based methods for PV fault detection from thermal images

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月28日

查看全文
  • Kübra Kaysal

摘要

In this study, faults on PV cells were detected with high accuracy using images obtained from a thermal camera. The dataset used in this study consists of 2401 original UAV-based thermal images collected from PV power plants located in ten provinces of Turkey with different climatic conditions. Powerful CNN-based structures with different architectural families, basic deep learning models, transformer-based and hybrid approaches, lightweight focus structures, and classical machine learning models were grouped in a multi-class manner. Within these groups, 11 different AI models were compared under the same experimental conditions. Four different classes were evaluated for classification, “healthy”, “hotspot anomaly”, “bypass diode fault”, and “string fault”. Among the architectures evaluated, DenseNet121 achieved the highest numerical accuracy at 98.7%. However, performance differences were relatively small, particularly among the top-performing models such as DenseNet121, CNN+Transformer, and ResNet50.