Scholarlink

<p>As solar energy adoption continues to grow globally, ensuring the optimal performance and longevity of photovoltaic (PV) modules is crucial for maximizing energy efficiency and minimizing operational costs. Faulty PV modules can lead to significant reductions in energy output, making early detection and timely maintenance vital. Reliable and efficient inspection methods are therefore essential to detect defects and prevent energy losses, ensuring the long-term sustainability of solar energy systems. In response to this need, this study implements a hybrid feature extraction and classification framework for PV module defect detection, combining feature fusion strategies with traditional machine learning techniques to improve performance. Pre-trained CNN architectures, including DenseNet201, ResNet50, DarkNet53, and Inceptionv3, are fine-tuned, and rather than using their direct outputs, features from the fully connected and pooling layers are extracted and fed into classifiers such as k-Nearest Neighbor, Random Forest, and Support Vector Machines. The hybrid model’s performance is assessed through extensive experiments on augmented images from a publicly available dataset containing 6 different class. The study examines the impact of using individual and fused features, applying both single machine learning methods as well as an ensemble approach for traditional machine learning models. The results demonstrate that the model achieves high performance metrics, with accuracy reaching to 98.66%, respectively. The feature fusion and ensemble methods significantly enhances the model’s overall accuracy.</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

Improving photovoltaic module inspection with convolutional neural networks and classifier integration

  • 期刊论文

  • 发表日期：

    2025年7月9日

查看全文
  • Hatice Okumus

摘要

As solar energy adoption continues to grow globally, ensuring the optimal performance and longevity of photovoltaic (PV) modules is crucial for maximizing energy efficiency and minimizing operational costs. Faulty PV modules can lead to significant reductions in energy output, making early detection and timely maintenance vital. Reliable and efficient inspection methods are therefore essential to detect defects and prevent energy losses, ensuring the long-term sustainability of solar energy systems. In response to this need, this study implements a hybrid feature extraction and classification framework for PV module defect detection, combining feature fusion strategies with traditional machine learning techniques to improve performance. Pre-trained CNN architectures, including DenseNet201, ResNet50, DarkNet53, and Inceptionv3, are fine-tuned, and rather than using their direct outputs, features from the fully connected and pooling layers are extracted and fed into classifiers such as k-Nearest Neighbor, Random Forest, and Support Vector Machines. The hybrid model’s performance is assessed through extensive experiments on augmented images from a publicly available dataset containing 6 different class. The study examines the impact of using individual and fused features, applying both single machine learning methods as well as an ensemble approach for traditional machine learning models. The results demonstrate that the model achieves high performance metrics, with accuracy reaching to 98.66%, respectively. The feature fusion and ensemble methods significantly enhances the model’s overall accuracy.