Researchers Use Machine Learning to Detect Pitting Corrosion

ARTIFICIAL INTELLIGENCE · ULTRASOUND · CORROSION MONITORING

Could artificial intelligence identify the early signs of pitting corrosion hidden inside an ultrasonic signal? Researchers in Norway are exploring how machine learning could reveal changes that are difficult to distinguish using conventional signal-processing methods.

In their 2024 AMPP paper, Pitting Corrosion Detection by Ultrasound Monitoring, Magnus Wangensteen, Tonni Franke Johansen, Ali Fatemi, and Erlend Magnus Viggen investigated whether time-lapse ultrasonic data could be combined with neural networks to detect and estimate the depth of small defects in steel.

From AI Predictions to Corrosion Research

At the 2024 AMPP Conference + Expo in New Orleans, physicist Michio Kaku discussed a future in which artificial intelligence and machine learning could influence coatings and corrosion prevention. For the Trondheim-based research team, those technologies were already becoming part of the corrosion-monitoring toolbox.

Their work focuses on whether machine learning can recognize patterns in ultrasonic measurements that may be difficult to isolate using traditional processing techniques.

Why Pitting Corrosion Is Difficult to Detect

Pitting corrosion is a localized form of attack that can produce small cavities in a metallic surface. Unlike broadly distributed wall loss, a pit may occupy only a small area while penetrating more deeply into the material.

Ultrasonic inspection is an established nondestructive testing technique, but identifying small pits close to a reflective back-wall surface can be challenging because the pit response can interact with the much stronger reflection from that surface.

Four ultrasound sensors recorded data from the opposite side of the steel plate while holes were drilled incrementally into the specimen.

Where Machine Learning Enters the Picture

Instead of looking only at a single ultrasonic measurement, the researchers captured multiple traces over time. Those measurements could be combined into two-dimensional time-lapse images showing changes in ultrasonic reflectivity.

The resulting images became input data for neural networks trained to recognize whether a pit-like defect was present and to estimate how deep it was.

The diagram illustrates how a scattered wave from the defect can reach the transducer before the primary back-wall pulse.

Building a Controlled Dataset

One challenge in training machine-learning models is knowing the true condition of the material associated with every measurement. For this study, the researchers used a high-precision drilling machine to create holes incrementally in a steel plate. Because the depth of each machined hole was controlled, they could establish a high-confidence reference for the defect depth represented in the ultrasonic data.

Random, time-ordered combinations of pulse-echo measurements were then used to construct time-lapse datasets. A classification neural network was trained to identify whether pit-like defects were present, while a regression network estimated defect depth.

The researchers tested the models on data from a transducer that had not been included in the training process, providing a way to evaluate how the system performed on unseen measurements.

Moving From Simulation Toward Real-World Measurements

The team’s 2024 experiment followed earlier research into ultrasonic guided waves and machine learning. Their 2023 AMPP work relied on ultrasound simulations and synthetic data. The newer project moved the research closer to physical measurements by collecting ultrasonic data from an actual steel specimen as controlled defects were created.

That transition matters because laboratory simulations allow researchers to control parameters that may vary substantially on real equipment. Surface condition, geometry, temperature, material variability, noise, instrumentation, and irregular corrosion morphology can all complicate field measurements.

Why AI Could Be Useful for Corrosion Monitoring

The potential advantage of machine learning is not that it changes the physics of ultrasonic inspection. Instead, it may provide another way to interpret complex measurement patterns.

A conventional analysis may struggle to isolate a useful parameter when several signals overlap. A trained model, given enough representative data with known outcomes, may be able to recognize combinations of features that correlate with specific material conditions.

In a future monitoring system, that capability could potentially help operators prioritize large volumes of sensor data and identify measurements that warrant closer investigation.

Beyond Pitting Detection

The researchers see potential applications beyond identifying individual pits. Time-lapse ultrasonic monitoring combined with machine learning could also be explored for recognizing other changes in material condition.

What Could This Mean for Plant Operators?

Long-term monitoring can produce large quantities of data. If validated for real corrosion conditions, AI-assisted analysis could eventually help screen those measurements, identify subtle trends, and surface signals that deserve additional attention.

Sensorlink, where Wangensteen and Fatemi work, also has ultrasound data from field installations that the team hopes can support additional model development.

The larger opportunity is not necessarily a fully autonomous corrosion diagnosis. It is the possibility of using machine learning to extract useful information from monitoring signals that might otherwise be difficult to interpret efficiently.

About the Researchers

A Promising Step—With More Validation Ahead

This research shows how machine learning may complement established nondestructive testing techniques by extracting information from complex ultrasonic signals. In the controlled drilled-hole experiment, the models demonstrated the ability to classify defects and estimate their depth with promising accuracy.

The next challenge is more difficult—and more important for practical corrosion monitoring: demonstrating that similar approaches can reliably identify the irregular geometry and changing conditions associated with actual corrosion pits.