Large Scale Corrosion Detection for Overhead Lines

Over the past year and a half we’ve been hard at work to create an automated, high-precision system for detecting corrosion on transmission towers. Our approach blends a combination of computer vision techniques and state-of-the-art models to deliver fast, accurate and objective condition assessments. Our work has been recognised by the Institute of Asset Management, where we won the Eason Award for Digital Innovation last December.

Try it out below!

We have a longstanding relationship with National Grid, where we’re deploying the solution for everyday use – that’s over 22,000 transmission towers, analysed automatically with each survey.

Multistage corrosion detection pipeline

Current ways to review the state of the network are purely manual. Towers are surveyed by either a drone or helicopter and in some cases even require a climbing survey, which is expensive and dangerous. Images from these surveys are then assessed by a skilled analyst who is able to discern the level and spread of corrosion on the tower. That’s anywhere from 10 to 30 images, per tower. 

Corrosion reduces a tower’s lifespan by 50%

In tandem with our forecasting model, we’ve been developing a deep neural network solution for automatic corrosion detection, aptly named Deepsteel. Deepsteel scales to thousands of images, accurately and rapidly. It can detect the presence and extent of corrosion with pixel-level granularity. Where an analyst can only go over a handful of assets per day, Deepsteel can process thousands.

Our 3-stage pipeline takes images from helicopter or drone surveys, processes the data and outputs an estimate of the tower’s corrosion levels and spread.

To achieve this, we begin with a simple question: what in the image is “steel” and what can be discarded? Our first model separates the steelwork from everything else (background, conductors, insulators, etc). We then give this as input to our second model, which works in a similar way but has instead been trained on hundreds of examples of different types of rust, on a pixel level.

Finally, our localiser model is able to identify different parts of the tower present in each image. This allows us to calculate the corrosion metrics on a more granular level; we are able to determine the state of, say, the “top left crossarm”. 

We aggregate the results from each image to arrive at an overall condition score for the tower’s parts, which we present to the analysts for approval or correction.

Consistent and objective results across the network

Deepsteel offers more than just speed and accuracy. Early on in our development, we discovered that our models are more objective than humans; when two analysts were presented with the same tower, they reached an agreement level of only ~70%. That is to say that not everyone grades corrosion the same way, whereas Deepsteel is consistent in its output.

Reliably forecasting corrosion state into the future

A well maintained tower, with regular painting, can play its part in the network for over 100 years. On the other hand, a poorly maintained tower lasts less than half of this time. National Grid knows this, which is why their Condition Monitoring team uses our KAI Platform every day to assess each tower’s condition. 

“We are able to take real time data and use it to predict when assets on our network need attention.”

– Mark Simmons, Condition Monitoring Manager at National Grid

These assessments, paired with historic data and atmospheric data, allowed us to model the progression of corrosion. Atmospheric conditions such as humidity and rainfall are key external drivers but what’s more important is the current state of the tower when forecasting future state. In other words, a well maintained or protected tower will last longer than one which isn’t. In this analysis, “age” is a proxy for the type of steel used and the protection applied when the tower was built.  The UK network was built in waves and each wave used different standards of steel or protection.

Using this model we are able to much more accurately forecast the future state of the network. 

Business as usual

We are taking careful steps to ensure that our AI enhances the work analysts do. Deepsteel’s accelerated processing can spot dangerously corroded towers as soon as they’ve been surveyed and alert the Condition Monitoring Team to these anomalies.

The benefits of Deepsteel are:

      1. Save millions across the network: improved asset lifetime and maintenance lead to huge savings
      2. Time saving: assessing a tower takes <1 minute, compared to 25 minutes now
      3. Save on surveys: our forecasting helps you target which towers to survey, resulting in 8% less surveys per year
      4. Greater objectivity: assessments are consistent between towers and time
      5. Faster actionable insights: get the state of assets as soon as the survey is uploaded.

    Interested in integrating Deepsteel into your daily operations?

    About Petar Gyurov

    Petar is passionate about climate change and conservation, and wants to apply novel machine learning methods to revolutionise these fields.

    More from

    Petar Gyurov

    Sign Up for More!

    Sign up for news, info, and to stay ahead of AI industry trends.