Reprinted from the June 2023 issue of Materials Evaluation, Volume 81, Issue 6.
.
.
by Iikka Virkkunen, Tuomas Koskinen, Topias Tyystjärvi, and Oskar Siljama
.
NDE Is About Making Sense of the Data
Modern equipment used for nondestructive evaluation (NDE) can acquire more data faster than ever before. The benefits from these richer datasets are profound. We can detect damage earlier, find those hard-to-see things that were previously missed, and better characterize what we find to allow better decisions for further action. This increases the value of NDE and extends its use.
At the same time, the problem created by this move to richer data is equally profound. There are now vast amounts of data for the inspector to analyze, and it’s getting more complex. While the number of channels has multiplied, the number of available inspectors’ eyes have decreased. More inspectors are retiring than coming into the industry, everywhere in the world. So, we’ll need to do more data analysis with fewer inspectors.
Deep Learning Takes Automated Data Analysis to New Level
Automated data analysis is not new, but most of the data is still evaluated with human eyes. This is particularly true for the more challenging inspections with richer datasets. That’s because traditional automation is not good at evaluating noisy data and can’t fully exploit the big modern datasets.
All this has changed in the last four years. Modern machine learning techniques based on deep learning have proven capable of analyzing the most difficult of NDE datasets with high accuracy. These vast datasets can be analyzed in seconds and areas of interest can be highlighted for the inspector for final evaluation. AI can automate the most tedious and error-prone parts of the inspectors’ work. Most importantly, these systems are widely available and can handle data for all commonly used methods, including ultrasonic testing (UT), radiographic testing (RT), electromagnetic testing (ET), and visual testing (VT).
Future developments are also becoming clear. First, AI was highlighting areas of interest and leaving the inspector to look at the 5% of the data that mattered. Now, it’s already capable of separating indications, making measurements, and characterizing the findings. It can also track data quality, evaluate coverage, and gather all the information needed to automatically generate full reporting so the inspectors can focus on the final evaluation. In a few years, AI will generate readable explanations of the findings. It will compare different inspections and extract trends and predictions. It will work across disciplines and talk to, say, structural integrity models to make better suggestions.
AI Is Going to Change the Game
The future is already here, but it’s not evenly distributed. Modern AI data evaluation for train axle UT inspection has been in field use since 20191. Nuclear pressure vessel head penetration inspection data was successfully evaluated with AI in a field trial in 20222. Inspections that detect minuscule individual pores to track porosity in composite production have been used for production development since 20211. Real-time VT of pressure vessel inspections saw field trials in 2022.
It is clear that modern AI techniques can be used to analyze even the most complex NDE datasets with high accuracy and reliability, and this is applicable to all the major NDE methods we use today. While the share of AI-evaluated data is still small, it is growing rapidly. In a few years, most data will first be evaluated automatically. Even low-volume special inspections where end-to-end trained solutions are not viable will benefit from supporting models that automatically highlight and quantify salient features of the data, take hints from inspector actions, and learn to support them better as the inspection progresses.
We’ll also see AI-first inspections made commercially available—that is, inspection procedures where the whole data acquisition is designed from the ground up to be evaluated using AI. This brings two additional benefits: first, we can fully unleash modern data acquisition capabilities. Since the data is evaluated by AI in seconds, we can acquire even more channels and use complimentary techniques to detect and characterize defects without imposing an undue burden on inspectors. Second, this allows increasing use of more sophisticated methods where the complexity in data evaluation has hindered widespread use. This includes, for example, modern techniques like the ultrasonic total focusing method, but also age-old things like mode-conversion techniques. With AI, the complexity of data analysis can be automated, making the use of complicated techniques as easy as conventional methods. In a few years, this will allow complex inspections to become commonplace. It’s no longer feasible to work through the data manually. There will be just too much of it.
Even with the 20× efficiency gains of today’s AI, people are sometimes struggling to see the business case for it. They can’t seem to fit AI into their operations. Paradoxically, the even higher gains of the future will not make this easier. To understand, try a thought experiment: imagine there was a new regulation that overnight forced you to evaluate all data 20 times. What would that do to your business? How would you need to change your operation to accommodate? This is how the future will see your current operation. Adopting AI is easy technically, but major efficiency improvements will also affect workflows, planning, and resourcing. Start planning for automated data evaluation now.
What AI brings to NDE is nothing short of a revolution. First, a revolution in productivity and reliability as the current, proven AI techniques are adopted across the industry over the next few years. Second, a revolution in capability as AI-first NDE procedures are developed to a wider range of inspections. To NDE inspectors, this will bring an unprecedented ease of use and freedom to focus on the essence of their craft.
References
1. Koskinen, T., T. Tyystjärvi, O. Siljama, and I. Virkkunen, 2022, ”AI for NDE 4.0 – Recent use cases,” Journal of Non Destructive Testing and Evaluation, 19 (4): 19–23
2. EPRI, 2022, AI tool developed by EPRI significantly cuts analysis time in US nuclear plant field trial, Electric Power Research Institute, https://www.epri.com/research/products/000000003002025510
_______
Authors
Iikka Virkkunen (iikka.virkkunen@trueflaw.com), Tuomas Koskinen, Topias Tyystjärvi, and Oskar Siljama: Trueflaw Oy, Tillinmäentie 3, tila A113, FIN-02330 Espoo, Finland
_______
The June 2023 issue of Materials Evaluation is available to members online at https://source.asnt.org/me-archive/
.
For more information on artificial intelligence and machine learning, read the July 2023 Technical Focus Issue of Materials Evaluation. This issue is open access.:
- Benefits and Concerns of Using Emerging Artificial Intelligence Chatbots with Work in NDT
- Validated and Deployable AI/ML for NDT Data Diagnostics
- Tips for Effective Machine Learning in NDT/E
- Machine Learning Techniques for Acoustic Data Processing in Additive Manufacturing In Situ Process Monitoring – A Review
- Real-Time AI-Driven Interpretation of Ultrasonic Data from Resistance Spot Weld Process Monitoring for Adaptive Welding
- Acoustic Emission Source Localization Using Deep Transfer Learning and Finite Element Modeling–Based Knowledge Transfer
.
Check out the blog post Machine Learning Is Here Now to Make the Inspector’s Life Easier, from March 2021, also written by Iikka Virkkunen.
.