Photovoltaic (PV) energy plays a key role in reducing global carbon emissions. As the number of PV systems installed continues to grow, ensuring the long-term reliability of PV modules becomes increasingly important. Failures can significantly reduce energy yield, yet many degradation mechanisms remain difficult to quantify using conventional inspection methods. Multispectral imaging enables the detection of such failures by capturing complementary information from different spectral domains.
The objective of this project was to develop a quantitative methodology for acquiring and automatically analysing multispectral images of PV modules. The proposed framework combines ultra-high-resolution multispectral imaging with artificial intelligence to detect defects visible at the solar cell level and correlate them with electrical performance losses. A single ultra-high-resolution multispectral camera was used to acquire visible (VI), electroluminescence (EL), and ultraviolet fluorescence (UVf) images.
During the project, a dataset of 241 PV modules corresponding to 11076 solar cells was collected and processed. A dedicated segmentation pipeline automatically extracted individual cell images from module images. More than 10% of the dataset was annotated by experts, providing a high-quality training and validation dataset for machine learning models. The core component of the methodology is a multispectral defect classification model based on the Channel Vision Transformer architecture. The model was trained using a progressive transfer-learning strategy: pre-training on ImageNet, fine-tuning on large publicly available EL datasets, and final adaptation to the project’s multispectral images. Despite the limited size of the proprietary dataset, this strategy enabled robust defect detection across different module technologies. The model achieved >90% accuracy on labelled project data and approximately 94% accuracy on an independent external dataset (EL only). For unlabeled images, the system achieved >70% correct failure identification, meeting the project’s performance targets.
The developed approach enables the detection of several PV failure modes, including cell cracks, inactive regions, corrosion, delamination, and discolouration. A key improvement over conventional methods is the ability to detect multiple simultaneous defects, reflecting the complex degradation patterns typically observed in aged modules. Combining multiple spectral channels also allows identification of defects not visible in EL imaging alone.
In addition to defect classification, the project demonstrated a data-driven method to estimate module power loss directly from multispectral images. A deep learning regression model based on ResNet-18 was trained to predict the loss in maximum power using VI, EL, and UVf images. This establishes a quantitative link between degradation patterns in images and losses in electrical performance.
Some limitations remain. The methodology performs best for defects with clear optical signatures. Rare defects and complex combinations remain challenging due to limited training data. Furthermore, degradation mechanisms without visible optical signatures cannot be detected. The analysis also focused on cell-level visible defects, while module-level visible degradation patterns, such as potential-induced or ultraviolet-induced degradation-related EL patterns, were not included.
Overall, the project demonstrates that AI-based multispectral image analysis enables scalable, quantitative, and automated detection of PV module defects and links visual degradation patterns to loss of electrical performance. The results highlight the strong potential of AI-enhanced multispectral imaging as a next-generation tool for PV reliability assessment in research laboratories, industrial production, and operation and maintenance of PV systems.
Main Findings
- A multispectral artificial intelligence (AI) model based on the Channel Vision Transformer (ChannelViT) architecture was developed to automatically detect and classify defects in photovoltaic (PV) modules three complementary imaging modalities: visible (VI), electroluminescence (EL), and ultraviolet fluorescence (UVf) images
- When trained and tested using EL images of crystalline silicon PV modules from publicly available datasets, the model achieved 94.4% classification accuracy on previously unseen images for common EL-visible defect classes such as good, cracks, cross cracks, dark regions, and corrosion. This demonstrates strong performance for standard EL-based defect detection.
- The inclusion of visible (VI) and ultraviolet fluorescence (UVf) images enable the detection of additional defect classes that are typically not visible in EL images alone, such as discolouration and delamination. However, larger labelled datasets are required to fully capture these defects across different PV module technologies.
- When combining all three image types (VI, EL, and UVf), the developed model achieved >90% accuracy on expert-labelled multispectral data and ≥70% correct defect identification on unlabeled data, demonstrating reliable performance for multispectral defect detection.
- A complementary AI-based method for estimating PV module performance directly from multispectral images was also developed. Using VI, EL, and UVf images together reduced the mean absolute error (MAE) in predicted power loss from 5.5% (EL-only) to 4.8%, demonstrating that combining multiple spectral imaging modalities improves the accuracy of performance loss estimation.