Wind Turbine Blade Damage Detection using CNN

Wind turbine blades suffer cumulative damage — surface cracks, leading-edge erosion from rain and particles, lightning strikes and subsurface delamination. Manual inspection means rope-access technicians or expensive downtime, and early-stage defects are easy to miss in thousands of drone photographs. This project applies transfer learning with an EfficientNet-B0 backbone to classify blade inspection images into five damage categories, turning a tedious visual review into a fast, repeatable screening step. Students learn a complete applied-vision workflow: dataset curation and annotation, augmentation for outdoor lighting variation, fine-tuning a pre-trained CNN, and honest evaluation with per-class metrics. The interactive demo walks through the inference pipeline on real blade photographs with representative model outputs.

1. Blade inspection photographs are collected from drone and rope-access surveys and annotated into five damage classes.
2. Images are resized to 224×224, normalized, and split 80/20 into stratified train and validation sets.
3. Augmentation (rotation, brightness jitter, horizontal flip) simulates the lighting variation of outdoor inspection.
4. EfficientNet-B0, pre-trained on ImageNet, is fine-tuned with a new classification head and dropout.
5. Training runs with Adam and cosine-decay learning rate, with early stopping on validation loss.
6. The held-out split is evaluated once: accuracy, confusion matrix and per-class precision/recall are generated for the report.
7. In the web demo, a sample photo passes through the same preprocessing and the trained model to produce the damage prediction.

Tech stack:

• Python 3, TensorFlow/Keras
• EfficientNet-B0 (transfer learning)
• NumPy, Matplotlib, scikit-learn
• Jupyter Notebook (training)
• HTML/CSS/JS inference demo
• Pillow (image preprocessing)

Dataset & model details

• Dataset: Custom curated blade-inspection set — approximately 2,400 drone and rope-access wind turbine blade photographs assembled for the build, annotated into 5 classes: healthy, crack, leading-edge erosion, lightning damage, delamination; stratified 80/20 train/validation split.
• Task: 5-class image classification; input = 224×224×3 blade photo, output = probability distribution over the 5 damage classes.
• Model: EfficientNet-B0 backbone (ImageNet pre-trained) → GlobalAveragePooling → Dropout(0.3) → Dense(5, softmax); ~5.3M parameters.
• Metrics: Validation accuracy ~93% (design target), macro F1 ~0.90 (design target), per-class precision/recall, confusion matrix. No accuracy is claimed as measured until the training run is executed for the order.

• Five-class damage classifier Healthy, crack, leading-edge erosion, lightning damage and delamination — a purpose-built label set for blade inspection, with per-class precision and recall reported.
• Curated inspection dataset Around 2,400 drone and rope-access blade photographs, annotated by damage class, with a documented stratified train/validation split.
• Transfer-learning training notebook EfficientNet-B0 fine-tuning with augmentation for outdoor lighting, early stopping, and training-curve plots — fully reproducible.
• Interactive inference demo Single-file web app: pick a sample blade photo, run the analysis pipeline, and see the predicted class with probability bars.
• Per-class evaluation report Confusion matrix and precision/recall/F1 for each damage class, so the viva can discuss exactly which defects confuse the model.
• Inference-ready model export Saved weights plus a preprocessing pipeline, ready to run on new inspection photos without retraining.
• Viva kit Report PDF, presentation PPT and a Q&A document covering CNNs, transfer learning, augmentation and evaluation metrics.

• Curated, annotated blade-image dataset with split files
• Complete training & evaluation Jupyter notebook
• Trained CNN model file with preprocessing code
• Interactive inference web demo with sample photos
• Confusion matrix, per-class metrics and training-curve plots
• Project report PDF (background, methodology, architecture, results)
• PPT presentation for final review
• Viva Q&A preparation document

• Trained on the curated five-class set — rare defect types outside these classes are out of scope.
• Outdoor inspection photos vary in lighting and angle; heavy glare or motion blur degrades predictions.
• 93% is a design target for the training run, stated honestly — the report documents the actual achieved figure.
• The classifier screens images; it does not replace certified structural inspection of blades.
• Subsurface delamination is only detectable where it shows surface symptoms in the photo.