Automatic defect localization and characterization through machine learning based inversion for guided wave imaging in SHM - CEA - Commissariat à l’énergie atomique et aux énergies alternatives Accéder directement au contenu
Article Dans Une Revue AIP Conference Proceedings Année : 2019

Automatic defect localization and characterization through machine learning based inversion for guided wave imaging in SHM

Résumé

Guided Wave (GW) based Structural Health Monitoring (SHM) relies on the permanent integration of sensors on a structure to measure propagated guided waves and deduce information regarding its current structural health. GW imaging takes advantage of the GW measurements to establish a cartography of the health of the inspected structure, leading to the detection and the localization of the possibly present defects. Characterization of the defects from the obtained maps is often no directly possible, as the size and magnitude of the defect indication depends on a substantial number of parameters beyond the defect size, such as the wavelength of interrogation or the reflectivity of the flaw. So far, characterization of defect size through guided wave imaging in SHM context has therefore only been achieved in controlled environments for well-known flaws. This communication presents the use the GW-SHM simulation models developed at CEA-LIST to build a large database of GW imaging results and train a machine learning based inversion algorithm on it. Firstly, a training set accounting for various inspection parameters, e.g., frequency, flaw position and size is generated. Secondly, imaging algorithms are applied to each database sample and these images are used, in the so-called training phase, to build an inverse model with a supervised machine learning algorithm. In practice, this corresponds to fit a regressor (e.g., kernel ridge regressor, support vector regressor, Gaussian process regressor, etc.) on the set of signals/parameters pairs. To assess the performance of the inversion strategy, (i.e., the capability to retrieve flaw positions and/or dimensions), the inverse model is evaluated on new datasets (usually called test sets), which were generated independently from the training phase. The inversion results are analyzed with respect to accuracy and CPU time efficiency
Fichier principal
Vignette du fichier
050005_1_online (2) (1).pdf (1.63 Mo) Télécharger le fichier
Origine : Fichiers produits par l'(les) auteur(s)

Dates et versions

cea-04555919 , version 1 (23-04-2024)

Identifiants

Citer

Roberto Miorelli, Andrii Kulakovskyi, Olivier Mesnil, Oscar D’almeida. Automatic defect localization and characterization through machine learning based inversion for guided wave imaging in SHM. AIP Conference Proceedings, 2019, Volume 2102: 45th Annual Review of Progress in Quantitative Nondestructive Evaluation, 2102, pp.050005. ⟨10.1063/1.5099771⟩. ⟨cea-04555919⟩
0 Consultations
0 Téléchargements

Altmetric

Partager

Gmail Facebook X LinkedIn More