研究目的
This study addresses the growing need to automate the process of classifying PV module defects. In large-scale plants housing millions of operational PV modules, manual inspection of PV modules is simply not viable. Current techniques for fault analysis can be broadly categorised into two groups: techniques that use electrical measurements to detect faults and alternative methods that do not, such as infrared thermography.
研究成果
The proposed approach for defect analysis can be used to assist plant operators to quickly identify as well as rectify defects with minimal effort and cost. In addition, the need for manual inspection will be reduced, further reducing the overall costs of PV module maintenance, making solar energy an even more viable and attractive form of energy production. This study adds to the body of knowledge in PV defect analysis as this approach has not been applied in this field before.
研究不足
The study does not have the luxury of large datasets, which is a limitation for deep learning techniques. Additionally, the soiling class was not used in analysis due to a lack of samples and an unacceptably high within-class variation.
1：Experimental Design and Method Selection:
The study examines a deep learning and feature-based approach for detecting and classifying defective photovoltaic modules using thermal infrared images. The VGG-16 and MobileNet models are used for classification, and the SIFT descriptor combined with a random forest classifier is used to identify defective photovoltaic modules.
2：Sample Selection and Data Sources:
Thermal image data of 398 singular defective and 400 singular non-defective PV modules from three different PV plants were obtained. All three locations under study use crystalline silicon PV modules.
3：List of Experimental Equipment and Materials:
A FLIR Tau 2 640 thermal imaging camera was used to capture the images.
4：Experimental Procedures and Operational Workflow:
Once captured, individual modules were cropped out of the thermal images for analysis. The operating conditions in which the thermal infrared images were captured complied with the recommended operating conditions as stipulated in the International Energy Agency report on infrared thermography field applications.
5：Data Analysis Methods:
The study uses Python 3.6.5 64-bit software to implement the various CNN models. The bag of visual words model is implemented on a Python 3.6.5 64-bit interpreter using the opencv-contrib-python-headless package, whereas the spatial pyramid matching methodology is implemented using Github source code provided by Li. All classification and machine learning algorithms—except for the deep learning models—are implemented in R 3.5.0 using the random forest and caret packages.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
