Wind Turbine Maintenance Can Become A Lot Easier With The Use of Image Classification

While using drones to inspect wind turbines is great, there is still room for automation in the process. Image classification can help.

We can all tell if a car has not been well maintained — rust, peeling paint, and broken bumpers are all giveaways that a vehicle could use some TLC. It may still be able to make morning commutes, but its faults are telltale signs that it may be older and more prone to other issues.

One can make inferences about anything using sight, from leftovers in the fridge to cars and even wind turbines. The latter is especially important, as wind turbines are one of the most important forms of clean energy in the world. Maintaining them is crucial; when they are working at their best they are able to produce as much electricity as possible. There’s one problem with this, however — how can we see something and make sure that it is well maintained if it’s over 300 hundred feet in the air?

The answer is drones, and they have been used for years to capture pictures of turbines that can then be inspected. While very useful, the process still has some room for automation. A human is necessary to identify the difference between damaged and not damaged, and this can make it very time-consuming to check every turbine. Using image classification and AI, it can be made significantly faster and more efficient.

What is Image Classification?

Image classification is the use of computer vision, a subset of AI that replicates human vision, to categorize and classify images. In training, it is given pictures that it breaks down into pixels. These images have labels attached so the model can learn what different patterns between pixels mean, and eventually categorize new images that are fed through.

My Image Classification Model

I built an image classifier to identify the difference between damaged and non-damaged wind turbines from an image. With this, damage can be identified faster, leading to faster repairs and more overall energy being produced.

I worked in PyCharm and used scikit-learn, a popular Python library for machine learning, to construct the body. I was able to achieve an accuracy rate of almost 70%, which is relatively high for a machine-learning model.

While building this, one of the many challenges I faced was trying to find the right workspace to write my code in. I have only ever coded artificial intelligence in Google Colab, so I had to pivot when I realized I could not load my file of data onto the platform. I ultimately worked in PyCharm, where I was able to upload the images.

I had never built an image classifier before, so it was a learning experience modifying code to adapt to images instead of written data. After reading lots of articles and watching a lot of tutorials (This one in particular was super helpful), I was able to build a working model and gain skills along the way.

The Data

I trained my classifier using images similar to the ones above, which I sourced from Kaggle. For the purpose of creating the dataset, I used pictures of broken parts and rust as “damaged” turbines, and ones without as “not damaged”. While not perfect, I chose them because I think they give a good representation of what surface damage looks like on a wind turbine.

Why Should We Care About Image Classification in Wind Turbine Maintenance?

When wind turbines are damaged, the time until they are repaired is an amount of time when they are not producing as much electricity as possible. This is not good for increasing the usage of clean energy, as we can only use what is produced, and have to rely on fossil fuels for the rest.

By using AI and image classification to identify damage, repairs can be made at a much quicker rate. More electricity can be produced because of this, and as a society, we can rely on wind energy more to replace fossil fuels. Additionally, an increase in wind energy can lower the price for consumers, making clean energy more accessible and affordable.

Future Projects

For my next project, I want to use generative AI to display what different aspects of our climate will look like in the upcoming years. I have never built with generative AI before, so it may be a bit difficult to figure out what code is necessary for it to function. I am up for the challenge, however, and am very excited about what I will create next.