Application au contrôle non destructif ultrasonore d'un algorithme de déconvolution utilisant des réseaux de neurones convolutifs

Abstract

Compared to other non-destructive testing methods, such as radiography, ultrasound is often said to lack resolution and penetration into the volume. The purpose of this study is to propose a method of processing A-Scans ultrasonic signals, imitating deconvolution but using machine learning techniques. Deconvolution produces very good results in theory but loses a lot of effectiveness in practice. The objective is twofold : to allow inspection at a lower frequency so that the signal is less attenuated while maintaining an equivalent or even better resolution. The training of the machine learning algorithm had to be done with the minimum number of experimental cases, hence simulations were used. Therefore, a convolution layer neural network architecture was developed. The learning was perfomed with Pogo FEA, a software using graphics card calculation. These simulations were processed to add typical experimental noise. Experimental measurements were carried out to test the resolution of the algorithm. Reflectors at a distance of half a wavelength at 2.25 MHz in aluminium could be distinguished. An experimental noise amplified to 20 and then 5 dB was added. The maximum resolution decreases but the number of false detections increases with noise. Two examples of the use of the CNN developed here are presented. First, the axial resolution of an image produced by a total focusing method (TFM) was significantly improved by pre-processing the data. Two interfaces spaced 0.96 mm apart in aluminum are very easily distinguishable while they are not on a conventional TFM image. An attempt to identify a reflector by drawing the outline of an interface is also presented.

Date	20 Jan 2020
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Pierre Bélanger (Supervisor)