Session: 06-02: Machine Learning and Statistical Methods in NDE II

Paper Number: 145561

145561 - Generating Synthetic Ut Data With Deep Learning 

Abstract: 

EPRI has been leveraging NDE to assess cost savings on power plant equipment and infrastructure. EPRI has collected extensive ultrasonic wave data sets in the NDE Program. These ultrasonic testing (UT) data sets have the potential for helping create another method for synthesizing UT data to supplement or replace flaw response simulation approaches. This project aims to investigate and understand the feasibility and performance of machine learning techniques capable of synthesizing UT data. Such models and techniques are considered highly valuable, as the industrial NDE has the potential to serve various applications, including personnel training and, ultimately, demonstration.

NDE techniques play an important role in evaluating both structural components and systems in a safe, reliable, and cost-effective manner. NDE methodology is particularly important in providing the detection of potential flaws and corrosion-induced defects that may otherwise cause a failure.

A study was performed on machine learning techniques capable of synthesizing ultrasonic testing (UT) data sets. Synthesizing UT data sets can assist in the development of NDE techniques across many industries. Using five data sets of 1700 ultrasonic amplitude scan (A-scan) waveforms, each with differing levels of corrosion, synthesizing techniques were studied. These techniques yielded both quantitative and qualitative results of deep-learning-based generative models.

The results show that realistic synthetic A-scans can be generated with a non-significant amplitude difference when compared to real A-scans. Further improvements and stabilization can likely be implemented through additional synthesizing techniques and research.

Presenting Author: Caleb Watson EPRI

Presenting Author Biography: Caleb Watson is a Data Scientist at the Electric Power Research Institute (EPRI). He conducts key data analysis and machine learning model development within the NDE transformative technologies group. The NDE transformative technologies group creates cutting edge technology, research, and innovative strategies for utilities and ultrasonic sensor parties.

Watson joined EPRI as a student employee in the NDE Program. Before joining EPRI in 2022, Watson completed his undergraduate degree in data science at the University of North Carolina at Charlotte in which he prominently contributed insights to a wide variety of big data-based problems. During his time at the University of North Carolina at Charlotte he conducted research relating to both sensor-based data to provide solutions to technologically oriented problems as well as sensitive personal data to provide solutions to social issues. His research at the University of North Carolina at Charlotte led to his initial responsibilities given at EPRI.

Watson’s research at EPRI has focused on developing applications and insights relating to ultrasonic testing instruments and artificial intelligence. He served as one of the primary data scientists involved in many other data analysis efforts in the transformative technologies group. Most recently he has generated synthetic ultrasonic data using deep learning algorithms. He has also conducted studies of experienced ultrasonic data technicians using biometric tracking software to gather the best data analysis methods for knowledge transfer and ultimately digital twin development.

Authors:

Caleb Watson EPRI
Terrill Massey EPRI
Mark Dennis EPRI