🔆 Solar Module Health Monitoring via EL Image Segmentation

A deep learning-based pipeline for automated fault detection and classification in solar PV modules using electroluminescence (EL) images. The system performs semantic segmentation to identify 14 defect and 14 feature classes across multiple datasets with varying formats and gray-coded masks.

---

📁 Dataset Structure

dataset_YYYYMMDD/
├── el_images_train/
├── el_images_val/
├── el_images_test/
├── el_masks_train/
├── el_masks_val/
├── el_masks_test/
└── ListOfClassesAndColorCodes.csv

• Grayscale masks have class-specific pixel values defined in the CSV file.
• Datasets include pre-augmented images (flip, rotate, mirror).

---

⚙️ YAML-Driven Configuration

config.yaml:

data:
  root_dir: ./datasets
  version: dataset_20221008
  img_size: 512
  batch_size: 4
  use_rgb_masks: false

train:
  epochs: 50
  lr: 0.0001
  use_class_weights: true
  class_weights: [1.0, 3.5, 5.0, ...]

experiment:
  name: model_run
  output_dir: results/

---

🧠 Model

Supports:

• ✅ DeepLabV3+ (default)
• 🧪 UNet++

Outputs logits over all defect classes per pixel.

---

🚀 Running the Pipeline

🔹 Train

python main.py --config config.yaml

🔹 Test

python test.py --config config.yaml

Results are saved in:

results/
├── images/   # Input EL images
├── masks/    # Predicted RGB masks

🔹 Inference on Single Image

python run.py --image_path path/to/image.png --config config.yaml

---

📊 Performance Metrics

Metric	Value (example)
Train Loss	↓ Decreases steadily
mIoU	~42.7% (varies by dataset)
Per-class IoU	Visualized per epoch
Best Model	Saved automatically

---

🖼️ Sample Output

EL Image	Predicted Mask
results/images/image_001.png	results/masks/mask_001.png

---

🔧 Requirements

pip install -r requirements.txt

• Python 3.8+
• PyTorch, torchvision
• numpy, pandas, Pillow, tqdm, PyYAML

---

📚 References

• Dataset: TheMakiran/BenchmarkELimages
• Paper: ScienceDirect, 2023

---