Wind Turbine Rotor Blade Generalised Ice Detection with Generative AI

Supplementary material for the application paper "Domain-Invariant Icing Detection on Wind Turbine Rotor Blades with Generative AI for Deep Transfer Learning" (in submission).

A repository with code for predicting blade icing on images of turbine rotor blades using supervised (neural style transfer) and unsupervised (CycleGAN) techniques.

Download and use of the repository:

To download this repository and its submodules use

git clone --recurse-submodules https://github.com/malvela/WindTurbine-IceDetection_GenerativeAI.git

Individual files and functionality:

This repository contains Python files for generalised icing prediction (domain-invariant - independent of the wind park the AI model has been trained on) on wind turbine rotor blades using a tiny computer.

• CycleGAN/cyclegan_generativeai_icing.py : Used to train the CycleGAN model from scratch (or leverage the pre-trained Summer2Winter Yosemite model).
• Fast_Style_Transfer/Overlay_Images.ipynb : Used to overlay the styled image to the rotorblade using labelled masks.
• Fast_Style_Transfer/StyleTransfer_Notebook_BladeImages.ipynb: Used to modify the content images with the pretrained style transfer model.
• StyleTransfer-TrainFromScratch/NST_TrainingFromScratch.py: Used to modify content images by training neural style transfer algorithm (based on VGG-19) from scratch.

Cite as:

If you are using this repository in your research, please cite it as:

Chatterjee J., Alvela Nieto M.T., Gelbhardt H., Dethlefs N., Ohlendorf J.-H., Greulich A., Thoben K.-D., "Domain-Invariant Icing Detection on Wind Turbine Rotor Blades with Generative AI for Deep Transfer Learning" (in submission)

References:

1. Reference for CycleGAN original model (and original pre-trained models used - such as summer2winter): (https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix).
2. Reference for Fast Style Transfer technique: https://www.tensorflow.org/tutorials/generative/style_transfer
3. Reference for Neural Style Transfer Algorithm: https://www.tensorflow.org/tutorials/generative/style_transfer

License:

This repo is based on the MIT License, which allows free use of the provided resources, subject to the original sources being credit/acknowledged appropriately. The software/resources under MIT license is provided as is, without any liability or warranty at the end of the authors.