Articles MIN

Leveraging Machine Learning For Nondestructive Testing

What is non-destructive testing?

Nondestructive testing is used in many industries to evaluate the properties of components without altering or causing damage to the component being tested. This includes examining its integrity or condition to see if anything might indicate deviations from normal or optimal characteristics. It is a highly valuable tool to help prevent potential losses and hazards arising from the failure of a component while saving time and money by not compromising its future usage.

Common non-destructive testing methods

There are various methods of conducting nondestructive testing (Exhibit 1), including vibration, guided acoustic waves (Exhibit 2), and electromagnetic induction like eddy currents. After inspection, the collected data has to be analyzed to determine whether a flaw or fault exists. Traditionally, this analysis has been done manually using fixed rules, which can be slow, tedious, and requires explicit knowledge of characteristics that indicate a flaw.

Technique Description Usage Data Type
Electromagnetic testing A test coil is placed around the material being tested to probe possible defects in conductive material that show up as disruptions in the magnetic field. Aerospace and petrochemical industries Images (e.g., 2d or 3d imaging from direct current, eddy current)
Guided wave Ultrasonic waves are introduced into piping using a transducer ring and the propagation information is collected Pipeline inspections in petrochemical industry Time series (e.g., A-scans, C-scans)
Vibration analysis Vibration data are collected from rotating machinery to see how its signature might reflect machinery conditions. Condition monitoring on rotating machinery, and structural health monitoring (e.g., bridges, pipes, turbine blades) Time series (e.g., velocity, acceleration data)
Visual testing Visual data is collected about a material or object to inspect it for flaws. Aerospace, automotive, maritime, mining, petrochemical, power generation, and mining industries Images (e.g., photos, video recordings, drone imagery)


Exhibit 1. Common nondestructive testing methods

 

 

Exhibit 2. Example of piping inspections using guided waves. A transducer ring is placed outside of the piping being tested, and its resulting propagation is collected.
Image source: Sprialboy / CC BY-SA (https://creativecommons.org/licenses/by-sa/3.0)

 

Machine learning optimizes the non-destructive testing process

Machine learning (ML) is a useful tool that can help automate the testing process by detecting the presence of flaws/faults from the collected data in a faster and reliable manner with minimal labor. Further, it does not require prior knowledge of the explicit rules for detecting the presence of flaws/faults. This type of testing will result in significant cost savings from the reduction in downtime and measurement effort.

What is needed to use ML in nondestructive testing

To build an ML algorithm for nondestructive testing, both the collected data and associated ground truth labels about whether a fault exists are needed. This labeled dataset should be reasonably sized so that the model can learn generalizable information that can be applied to new/unseen data. 

In case of no sizable real labeled dataset, a synthetic dataset can be built, using a priori knowledge with respect to how flaws/faults would manifest themselves, to enable automated flaw detection. The caveat is that the extension of the built model to real-world data depends on how successfully the synthetic data captures the characteristics needed for differentiating fault presence in real-world data.

Potential benefits of ML in non-destructive testing

The vibration analysis use case for a risk score model is just one example of how ML can automatically and efficiently quantify faults and provide new insights. Model building is not limited to flaw/fault detection (presence or nonpresence), but can also be flexibly extended to output any measurable quantities related to the fault (e.g., severity, damage size) for which ground truth data is available. Further, model building is not limited to time-series data from vibration testing or guided wave testing but can be adapted to image data used in electromagnetic or visual testing (Exhibit 1).

Unlike manual analyses which require prior knowledge on what information to look out for, ML models can learn directly from the data to elucidate which characteristics are more important for prediction. This quantification of feature importance for inputs to the model is extremely useful for corroborating and extending the current understanding of fault characteristics with respect to knowledge from expert evaluators.

The vibration analysis and guided wave use cases demonstrate how real-life data size limitations can be overcome by using a synthetic dataset that closely represents the real-world data or through the use of bootstrapping techniques. Further, as more real data is collected during model deployment, the model can be improved by fine-tuning it to improve prediction capability for the more recent real data.

Conclusion

ML is a useful tool for automating the nondestructive testing process by providing a quick and low effort way to evaluate the properties of components; ML streamlines and improves the nondestructive testing process by automating post-data-collection analyses, providing new insights, and potentially improving detection performance in a quick and low effort manner with great cost savings.

 

Work with Us

Start a Project

We'd like to help creating your next solution, whether modernizing legacy platforms or developing new AI solutions. Technology moves fast, let's build sustainable solutions.
Get Started