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

Paper Number: 135274

135274 - A Neural Style Transfer Data Augmentation Strategy as Applied to Total Focusing Method Images: A Classification Task Perspective 

Abstract: 

Ultrasonic testing techniques (UT) are widely applied to assess the integrity of structures and materials in different industrial sectors ranging from the aeronautic, the oil & gas to the nuclear one, for instance. In particular, in the nuclear industry context the UT inspections are made challenging due to specimens sizes (e.g., thick pipes) considered, the needed of small (e.g., millimetric) defects detections constraints and the inspection of coarse-grained materials (e.g., cast austenitic stainless steel (CASS)) inducing high scattering noise into probed signals that possibly hide the presence of defects.

The use of numerical simulations has proved to mitigate the need of UT experimental measurements in order to design and develop inspection procedure and assess their performances. Nevertheless, if challenging inspection problems are addressed as in CASS materials, the noise induced into the signals, due to the interaction of ultrasonic waves with the material, cannot accurately be accounted due to the material structure variability from the microstructure point of view.

In this context, enhancing the realism of simulated data is an important aspect to better assess the inspection performance. Toward this end, this paper proposes a preliminary analysis based on the use of neural style transfer paradigm as applied to UT simulated total focusing method (TFM) images obtained via CIVA software for enhancing the realism of synthetic TFM images computed for the inspection of a CASS pipe.

More specifically, the article presents a data augmentation procedure adopted for creating a large set of synthetic TFM data aiming at embedding characteristic noise patterns that can be observed in experimental images that cannot be reproduced straightforwardly in simulations due to the lack of knowledge of material microstructure characteristics. The paper will analyse the possible improvements of the proposed strategy compared to simulated-only TFM images as applied to machine learning based defects detection in experimental TFM images.

Presenting Author: Roberto Miorelli CEA

Presenting Author Biography: Roberto Miorelli received the Ph.D. degree in physics from the University of Paris-Sud XI, Orsay, France. He is Research Engineer in electromagnetic modeling and machine learning (ML) as applied to the nondestructive testing and evaluation (NDT&E) problems at University Paris-Saclay, CEA List Institute, Palaiseau, France. His research interests include electromagnetics theory and modeling with emphasis on fast forward numerical solutions of electromagnetic problems via dyadic Green function integral-equation-based methods in layered media. His actual research is focused on the study and the development of ML strategies targeting highly efficient solutions of forward and inverse problems as well as statistical studies as applied to NDT&E and structural health monitoring relying to ultrasound and electromagnetic methods.

Authors:

Roberto Miorelli CEA
Leonard Le Jeune CEA
Steve Mahaut CEA
Souad Bannouf IRSN