Attention-enhanced DenseNet for robust solar cell defect detection in electroluminescence imaging
The reliability and efficiency of photovoltaic (PV) modules heavily depend on accurate and scalable defect detection. Electroluminescence (EL) imaging provides high-resolution, non-invasive insights into internal faults such as microcracks, finger interruptions, and inactive regions. However, existi...
Saved in:
| Main Author: | |
|---|---|
| Other Authors: | , , |
| Format: | article |
| Published: |
2025
|
| Online Access: | http://hdl.handle.net/11675/14502 https: |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | The reliability and efficiency of photovoltaic (PV) modules heavily depend on accurate and scalable defect detection. Electroluminescence (EL) imaging provides high-resolution, non-invasive insights into internal faults such as microcracks, finger interruptions, and inactive regions. However, existing inspection methods — including rule-based approaches, classical machine learning, and standard convolutional neural networks (CNNs) — often struggle with generalization, low-contrast anomalies, and variations in imaging conditions. Furthermore, conventional CNNs lack the capacity to model long-range spatial dependencies crucial for detecting dispersed or subtle defects in EL images. This paper proposes a hybrid deep learning framework that integrates a pre-trained DenseNet-169 backbone with a lightweight self-attention mechanism. This architecture is designed to jointly capture fine-grained local patterns and global contextual cues. The model is trained on a unified dataset of over 39,000 EL images from the ELPV and PVEL-AD datasets, reformulated into a binary classification task. Through comprehensive data augmentation, regularization techniques, and attention calibration, the model achieves an AUC-ROC of 0.983 and a macro F1-score of 0.95, outperforming several state-of-the-art baselines. Ablation studies confirm the effectiveness of the attention module and training optimizations. The resulting model demonstrates a strong balance of interpretability, accuracy, and computational efficiency—offering a scalable and robust solution suitable for real-world, high-throughput inspection pipelines in photovoltaic manufacturing. This study directly addresses limitations in current EL-based methods and advances the field of automated solar cell diagnostics. |
|---|