Caractérisation non-destructive de joints collés utilisant des réseaux de neurones convolutives sur des ondes guidées Lamb

Abstract

Bonded joints are currently being developed in a wide range of industries. These joints offer numerous advantages such as a very good mechanical resistance, cost reduction and they have an aesthetic asset. However, currently there are no nondestructive method to guarantee and quantify the quality of a bonded joint. Indeed, the current solutions are mainly destructive solutions, such as the use of overlap joints. Thus, the aim of this project is to develop a method which allows identifying the adhesion level of a bonded joint. Ultrasonic acquisition were used, the sample were excited with ultrasonic guided waves in order to obtain the different modes. Then, the analysis of the dispersion curves allows the identification of the mechanical parameters. In this project an algorithm capable of identifying the mechanical characteristics based on dispersion curves was developed. For this work, a neural network was implemented, it was trained on finite element simulations of wave propagation in bonded joints. Then, it was tested on experimental data to determine the real characteristic of the joint. The main hypothesis in this work is that the simulated Young’s modulus of the adhesive can be used to estimate the quality of the joint. The results obtained in simulations were promising: the estimation of the modulus of the adhesive of a joint was achieved with an error under 0.05 % on the whole test dataset. Then, more promising results were obtained when monitoring a bonded joint during curing. Finally, when inverting acquisitions on five real joints, the estimates allowed to differentiate and classify the samples according to their bonding quality. At first, the network failed to stabilize on correct estimates, then the implementation of a double network allowed to identify the quality of the joints. These joints were made by CNRC which has expertise in the manufacture of bonded joints, and also, the can manufacture joints voluntarily weak. Although those results were promising, it is important to point out the weak aspects of the method. Firstly, only five joints with the same geometry were used in this study. Secondly, when comparing the experimental dispersion curves with the dispersion curves predicted by the inversion method some discrepancies could be observed. Finally, this study shows that it is possible to perform an inversion of bonded joints with convolution neural network based on ultrasonic guided waves to determine their mechanical properties such as the Young’s modulus. This study also showed that the method was also able to evaluate more abstract aspect such as overall joint adhesion.

Date	4 Dec 2022
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Pierre Bélanger (Supervisor) & Matthew Toews (Co-supervisor)