Researchers Use Machine Learning to Detect Pitting Corrosion

This high-precision drilling machine has enabled the researchers to obtain measurements that allow high confidence in the ground truth of the depth of the machined holes on the steel plate specimen.

When renowned physicist Michio Kaku delivered the keynote address in March at the 2024 AMPP Conference + Expo in New Orleans, Louisiana, USA, he envisioned a future when artificial intelligence and machine learning will guide the coatings and corrosion prevention industry in exciting new ways.

For four researchers in Trondheim, Norway, the future is now.

In their 2024 AMPP paper “Pitting Corrosion Detection by Ultrasound Monitoring,” Magnus Wangensteen, Tonni Franke Johansen, Ali Fatemi, and Erlend Magnus Viggen propose a technique for early-stage pitting detection through machine learning. Wangensteen presented the paper at the AMPP conference alongside other authors in the machine learning realm.

“I definitely think artificial intelligence is relevant for growth monitoring, and that's what we have leveraged on,” Wangensteen says. “Tiny changes in signals are difficult to distinguish using traditional methods. Machine learning and AI can help us to discriminate what is happening within so that we will be able to do more advanced diagnostics.”

The authors explain that pitting corrosion is a localized form of corrosion that can cause cavities and structural failure in metallic materials, adding that it is more challenging to detect and prevent compared to uniform corrosion.

They continue by stating that ultrasound inspection is an established non-destructive testing technique used to detect internal flaws in materials. 

Four ultrasound sensors (Sensorlink Swarm) are placed on the opposite side of the steel plate recording ultrasound data as the holes are drilled stepwise.

Challenges of Pitting Corrosion

Detecting pitting corrosion near the reflective inner surface, they write, is difficult due to interference effects.

General corrosion is easier to detect using standard methods because it is more uniform, Wangensteen says.

“Pitting is trickier, because it's often a very small pinhole that penetrates the metal,” he says. “The reflected signal from that pinhole is hard to detect using traditional methods, but machine learning could actually detect this.”

By capturing multiple ultrasonic traces over time, two-dimensional time-lapse images of ultrasonic reflectivity can be generated and used as input for a neural network trained for pitting diagnostics, they write.

“I’ve been surprised by how well machine-learning methods have been able to extract information about corrosion data,” Johansen says.

The authors explained the challenges in accurately identifying pitting corrosion using ultrasonic inspection as: 

• Reflection difficulties: Because pitting corrosion occurs near the reflective inner interface, it is difficult to separate the reflection of the pit from the inner wall reflection. 
• Interference effects: When a small, localized pit scatters back an incoming wave, it can interfere with the main back wall plane wave reflection, resulting in minor errors in wall thickness estimation.
• Undulating reflections: In the early phase of pit development, minor distortions can be imposed on the back wall reflection, giving an undulating impression as the pit is emerging. This could give the misleading impression that the wall thickness is increasing, for example. 

Real-World Implications

The same group authored “Ultrasonic Guided Waves and Machine Learning for Corrosion Monitoring in Steel Pipes,” an AMPP 2023 paper based on ultrasound simulations and synthetic data. The 2024 work brings the ultrasound experiment into the real world. 

“The challenge with the synthetic experiment has been transferring it to the real world, because you don't control all the parameters in the real world like you're doing in the synthetic world. The paper from this year is much closer to being applicable on real data.”

In the most recent work, they conducted a drilling experiment to obtain a substantial training dataset for the machine learning model. 

“The data was combined in a way to produce an almost endless amount of training data,” Wangensteen says. “Individual measurements were combined into a time lapse measurement. You can train a machine learning model and do all the predictions on unseen data.”

While they tried to create pitting through corrosion in a controlled manner, it did not form.

“The next step would be to produce real growth pits and see if the trained neural network could detect them,” Wangensteen says. “That’s the natural next step.”

They hope that in the long term, plant operators can use AI to automate diagnostics and filter out important data.

“AI is able to extract information that is not easily extractable,” Viggen adds. “AI is a very good signal processing tool, which gives plant operators information that they would otherwise not have because that information is hard to get at using other means.”