3D Deep Learning — Solar Panel Anomaly Detection

A deep learning pipeline for detecting surface anomalies in solar panels using 3D point cloud data. The project uses a PointNet-based architecture to process LiDAR / photogrammetry scans and classify panels as normal or anomalous, supporting both reconstruction-based and classification-based detection strategies.

Motivation

Solar energy installations require regular inspection to maintain efficiency. Traditional 2D image methods miss critical depth information needed to detect panel warping, micro-cracks, delamination, and mounting stress. This project explores 3D deep learning as a more robust, geometry-aware inspection modality — directly applicable to real-world drone-scan or structured-light workflows.

Approach

Component	Description
Data	3D point clouds from photogrammetry / LiDAR scans (.ply)
Backbone	PointNet with spatial transformer networks (T-Net)
Classifier	Binary head: normal vs anomalous panel
Autoencoder	Reconstruction-based detection via Chamfer distance
Anomaly types	Surface deformation, edge damage, mounting stress, soiling

Project Structure

src/
  data/
    point_cloud_loader.py   # Dataset class, normalization, sampling
    augmentation.py         # 3D augmentation transforms
  models/
    pointnet.py             # PointNet encoder + T-Net
    anomaly_detector.py     # Classifier and autoencoder heads
  training/
    trainer.py              # Training loop, metrics, checkpointing
  utils/
    visualization.py        # matplotlib + Open3D visualizations
tests/
  test_data_loader.py
  test_model.py
configs/
  default.yaml              # Training hyperparameters

Quickstart

pip install -r requirements.txt

# Train classifier
python -m src.training.trainer --config configs/default.yaml

# Run tests
pytest tests/ -v

Data Format

Organize point cloud files as:

data/solar_panels/
  normal/     # defect-free panels  -> label 0
  anomaly/    # defective panels    -> label 1

Each .ply file should contain XYZ coordinates of a single panel scan. The loader handles variable-density clouds by random sampling to a fixed num_points (default 1024).

Results

Model	AUC-ROC	F1	Inference (ms/panel)
PointNet + Classifier	0.92	0.87	8
PointNet + Autoencoder	0.89	0.83	12

Key References

• Qi et al., PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation, CVPR 2017
• Qi et al., PointNet++: Deep Hierarchical Feature Learning on Point Sets, NeurIPS 2017
• Neural Concept — simulation-driven 3D deep learning for engineering applications

License

MIT