NUCLEP, Brazil, Joined Jul 2020, 1
UT characterization of defects using neural networks
Characterization of defects using neural networks
Dear, I have been thinking a lot about the topic of characterization of defects by ultrasound - although many manufacturing standards define some strategies to characterize a defect, in practice, it is not exactly an easy thing to obtain consistent results.
Nowadays we perceive a great effort by the manufacturers of flaw detectors to provide the best image representation in their equipment. But even so, when we are going to characterize defects, the analysis of the A-Scan signal ends up being the most conclusive - in some cases.
As our test method is based on sound, why should we turn sound into an image to try to be more successful in characterization? Why not start from A-SCAN (RF Waveform)?
As we all know, when we think of sound, the tunings of instruments are based on the A-440 Hertz note, however, a well-trained listener can identify whether that note is being played by a piano, a guitar or a violin - and all of these differences will be in the waveforms of these instruments.
In the same way, the signals from cracks, inclusions and lack of fusion will present specific characteristics in their waveform that alone should be sufficient to characterize them. Perhaps teaching the machine to characterize itself based on the characteristic of the signal, using mathematical models such as neural networks and starting from defects of different shapes, patterns and sizes, can give greater reliability to our results. Do you know any research in this area of activity and is anyone else interested in the field? Tell me your impressions.
India, Joined Aug 2015, 17
Re: UT characterization of defects using neural networks In Reply to Vinícius Sardou at 15:03 Jul-07-2020 (Opening).
A lot of work is going on in this field. As you said, defects have different scattering and different effects on the frequency of the received wave etc. There are a lot of research groups currently working on first gathering the information about how the scattering changes dependent on size, location etc and then once that information is established training neural networks. Teh difficulty is characterizing the scattering from defects as a lot of parameters might influence the so called scattering matrix, which depends and is not limited to the type of wave hittinf the defect, the angle of incidence, frequency, material in which defect is embedded, type of defect, geometry of defect, size etc. In the end to assure you, most of the groups in research have this as priority and are already producing results related to simple defects and defect characterization and identification
Re: UT characterization of defects using neural networks In Reply to Chirag Anand at 16:04 Jul-07-2020 .
Really old stuff!!!
Worked more than 20 years ago with a software called ICEPACK invented by Dr Rober Hay in Canada - his company that time was TEKTREND - later bought by TECRAD which is now Olympus NDT - 2001 Robert Hay posted here https://www.ndt.net/forum/thread.php?msgID=4084#4084
Seems that this package is now distributed by TISEC (http://www.tisec.com/products/matlibe/pattern_recognition.htm)
See this old report from Swizerland: https://lib.dr.iastate.edu/cgi/viewcontent.cgi?article=1376&context=qnde
or this newer report: https://www.michigan.gov/documents/mdot/RC1629_Part_II_486744_7.pdf
If you do further search with the term Pattern Recognition, you see that Neuronal Networks have been already used in the mid and late 1990ties.That time the Computer (Intel 80386, 80486 and first Pentium Processor and the old Windows 95 limited of course the performance, nevertheless the Software was able to deal with a-scans - that time full waveform recordings like today were not common. So a preprocessing of A-scans had been used to reduce the amount of data.
Thats the reason I say: "Old Stuff"
If you look for today's deep learning approaches in NDT and in special the methods used for Ultrasound, a lot is based on Neuronal Networks. Much is programmed in Python, which is a good platform for such approaches.
Check the internet for publications, some applications and codes are published in Github.
May be you find also information here: