Session: 06-01: Machine Learning and Statistical Methods in NDE I

Paper Number: 134803

134803 - Deep Autoencoder Framework Applied to the Generation of Realistic Guided Wave Signals in Structural Health Monitoring Under Varying Environmental and Operational Conditions 

Abstract: 

Ultrasonic guided waves (GWs) are widely studied in structural health monitoring (SHM) for assessing the integrity of structures as they enable the inspection of larger areas with fewer sensors than conventional ultrasonic testing techniques due to the large propagation distances of such waves and their high sensitivity. Therefore, GWs have great potentials for embedded continuous assessment of the integrity of civil and industrial infrastructures, as well as in the transport sector, throughout the lifetime of the structure (continuous monitoring).

One of the most challenging issues in analyzing and exploiting GW signals in realistic scenarios is the impact of environmental and operational conditions (EOC). Over the last decade, signal processing techniques have been designed to overcome the effect of some parameters, mainly focusing on the temperature. However, depending on the application, such methods may have a limited application range. More recently, machine learning (ML) methods have shown promising results when applied to ultrasonic GWs signals to enhance the analysis capabilities of established techniques.

In this context, this paper proposes a ML framework relying on the use of a tailored autoencoder architecture enabling the fast generation of ultrasonic GWs signals close to experimental ones. The core idea is based on the joint learning of a structured latent space from experimental and numerical data. The addition of EOC parameters, such as temperature, in the model enables to take them into account in the resulting data [RA1] even if the initial simulation was not. This data generation can then be used for statistical studies or to train new inversion models (e.g., classification, localization, sizing of defects) in a very efficient manner.

In this talk, the autoencoder architecture will be described and its performances analyzed based on the use of simulated data obtained from CIVA software and experimental data representative of application.

Presenting Author: Clément Fisher Université Paris-Saclay, CEA LIST

Presenting Author Biography: Clément Fisher was recruited at the CEA in 2019 to develop signal processing and analysis methods in the context of integrated health monitoring (Structural Health Monitoring) using guided ultrasound waves. In particular, these methods are based on machine learning and deep learning methods to automate signal analysis to perform not only detection, but also localization and characterization of defects

Authors:

Roberto Miorelli CEA
Vivek Nerlikar CEA
Arnaud Recoquillay CEA
Oscar D'Almeida SAFRAN Tech