Related Experiment Video
Updated: Jul 16, 2026
09:19
In Situ Monitoring of the Accelerated Performance Degradation of Solar Cells and Modules: A Case Study for Cu(In,Ga)Se2 Solar Cells
Published on: October 3, 2018
Benchmarking Modern Deep Learning Models for Electroluminescence-Based Solar Cell Defect Detection.
Gökhan Şahin1,2, Ali Cengiz Rüstemli3, Ahmed Yaseen Bishree Al-Ani4
1Copernicus Institute of Sustainable Development, Utrecht University, Princetonlaan 8A, 3584 CB Utrecht, The Netherlands.
Sensors (Basel, Switzerland)
|July 15, 2026
Summary
Deep learning models accurately classify solar cells using electroluminescence (EL) images, achieving over 98% accuracy. EfficientNet-B2 demonstrated top performance, aiding in automated defect detection for photovoltaic systems.
Area of Science:
- Materials Science
- Electrical Engineering
- Computer Science
Background:
- Automated defect detection is crucial for photovoltaic (PV) solar cell quality control.
- Electroluminescence (EL) imaging is a standard technique for identifying defects in solar cells.
- Existing methods may lack the accuracy and efficiency required for large-scale PV manufacturing and maintenance.
Purpose of the Study:
- To develop and evaluate a deep learning framework for automated binary classification of photovoltaic solar cells as defective or normal using EL images.
- To compare the performance of four state-of-the-art deep learning architectures (EfficientNet-B2, ConvNeXt-Tiny, MaxViT-T, ResNet-50) on this task.
- To assess the interpretability of the models using explainability techniques.
Main Methods:
- A balanced dataset of 20,400 EL images (10,200 defective, 10,200 normal) was curated, combining cracked and broken cells into a single defective category.
- Dataset partitioning was strictly performed before preprocessing and augmentation to prevent data leakage.
- Four deep learning models were trained and evaluated under identical conditions, with performance metrics including accuracy, F1-score, ROC-AUC, and confusion matrices.
- Explainability heat maps were generated to visualize model attention on defect regions.
Main Results:
- All evaluated deep learning models achieved classification accuracies exceeding 98%.
- EfficientNet-B2 achieved the highest performance with 99.31% accuracy, 0.9931 F1-score, and 0.9987 ROC-AUC.
- MaxViT-T showed rapid convergence and strong performance, while ConvNeXt-Tiny and ResNet-50 also yielded reliable results.
- Explainability heat maps confirmed that EfficientNet-B2 and MaxViT-T accurately focused on defect areas like cracks and fractures.
Conclusions:
- Deep learning architectures provide a highly accurate and reliable method for detecting photovoltaic cell defects from EL images.
- The unified binary classification framework is applicable to both monocrystalline and polycrystalline solar cells.
- Explainability techniques enhance model transparency, supporting the practical deployment of intelligent inspection systems in PV manufacturing and maintenance.
