Solar Panel Defect Detection from EL Images using CNN

Photovoltaic modules lose power silently: micro-cracks from transport and hail, hot spots from cracked interconnections, and snail trails from silver nanoparticle migration all show up long before the naked eye can see them. Electroluminescence (EL) imaging makes these defects visible — a forward-biased cell glows under an EL camera, and any dark region marks an inactive area. Today this inspection is done by humans paging through hundreds of cell images, which is slow and inconsistent. This project replaces the human eyeball with a trained convolutional neural network: the CNN takes an EL image of a single solar cell and classifies it into one of four states — healthy, micro-crack, hot spot or snail trail — with confidence scores and a Grad-CAM heatmap showing exactly which region drove the decision. The model is trained on the public ELPV dataset of 2,624 EL cell images with expert defect labels, using transfer learning so it reaches strong performance without a GPU farm. A working web demo accepts any EL capture and returns the verdict in under 100 ms on a CPU. The complete training pipeline, evaluation methodology and architecture rationale are documented for a confident viva.

1. EL images are loaded from the ELPV dataset and reorganized into the four working classes: healthy, micro-crack, hot spot and snail trail.
2. Each image is preprocessed with contrast-limited adaptive histogram equalization (CLAHE), resized to 224×224 and normalized, with augmentation (rotations, flips, brightness jitter, noise) simulating EL camera variance.
3. An EfficientNetV2-S backbone pre-trained on ImageNet is attached to a custom head (global average pooling, dropout, dense layer) with the top blocks fine-tuned on EL data.
4. Training uses categorical cross-entropy with class weights to handle the rarer hot spot and snail trail classes, plus early stopping on validation F1.
5. The held-out test split is evaluated once: confusion matrix and per-class precision/recall/F1 are generated and documented in the report.
6. Grad-CAM is computed from the last convolutional block to produce the defect-localization heatmap shown in the demo.
7. In the web demo, an uploaded EL image goes through the identical preprocessing pipeline and returns the predicted class, confidence, per-class probabilities and a recommended maintenance action.

Tech stack:

• Python 3, TensorFlow/Keras
• EfficientNetV2-S (transfer learning backbone)
• NumPy, OpenCV (CLAHE preprocessing), Matplotlib, scikit-learn (metrics)
• Jupyter Notebook (training & evaluation)
• HTML5 + JavaScript (EL inspection web demo)

Dataset & model details

• Dataset: ELPV (Electroluminescence PV) dataset by Deitsch et al. — 2,624 EL images of mono- and polycrystalline silicon solar cells with expert defect annotations, published as a public benchmark. The build re-organizes the expert labels into the four working classes: healthy, micro-crack, hot spot, snail trail.
• Task: 4-class image classification; input = 224×224×3 preprocessed EL image, output = probability distribution over the four defect classes.
• Model: EfficientNetV2-S backbone (blocks 1–4 frozen, 5–6 fine-tuned) + global average pooling → Dropout(0.35) → Dense(256, Swish) → Dense(4, softmax); approximately 1.2M trainable parameters (design target).
• Metrics: Macro-F1 92.4% (design target for the built-to-order training run), per-class precision/recall, confusion matrix. No metric is claimed as measured until the training run is executed for the order; the delivered report documents the actually achieved figures.

• [4-class EL defect classifier] Healthy, micro-crack, hot spot and snail trail classification from a single EL cell image, built on an EfficientNetV2-S backbone fine-tuned on real electroluminescence imagery.
• [Grad-CAM defect localization] Every prediction ships with a heatmap overlay highlighting the cell region the network focused on, so the verdict is explainable rather than a black box.
• [Working web demo] Upload any EL capture (or use the built-in sample gallery) and get the classification with full class probabilities, defect findings and a recommended action in seconds.
• [Real public dataset] Trained and evaluated on the ELPV dataset (Deitsch et al.) — 2,624 EL images of mono- and polycrystalline cells with expert labels, not synthetic data.
• [Full training notebook] Preprocessing (CLAHE, resize, normalization), augmentation for EL camera variance, training loop, early stopping and evaluation in one reproducible notebook.
• [Honest evaluation report] Confusion matrix, per-class precision/recall/F1 and training curves — measured on a held-out test split, with the actual figures documented from the order's training run.
• [Batch QC mode] A review queue that triages many EL captures at once and flags defective cells for manual review, matching how a production-line inspection station works.

• Complete training & evaluation Jupyter notebook
• Trained CNN model weights (.h5) with the preprocessing pipeline
• EL inspection web demo (upload, classify, Grad-CAM overlay, batch QC queue)
• Confusion matrix, per-class metrics and training-curve plots from the order's run
• Project report PDF (background on EL imaging, CNN architecture rationale, methodology, results)
• PPT presentation for final review
• Viva Q&A preparation document (transfer learning, CLAHE, Grad-CAM, class imbalance, EL physics)

• 92.4% macro-F1 is a design target, stated honestly — the report documents the actual achieved figure after the order's training run.
• Trained on single-cell EL crops — full-module EL images must be cropped to cells first, included as a preprocessing utility but not fully automatic.
• ELPV labels carry known label noise (some cells have multiple defects); the report discusses how this affects measured metrics.
• Very faint early-stage snail trails may be missed; the evaluation quantifies the per-class recall honestly.
• The model is an educational QC build, not a certified inspection system — it does not replace professional module certification.