Keywords: Deep learning, multi-modal data, Non destructive analysis Internship Duration: 30/11/-1 - 30/11/-1
Head of the hosting team: vincent Vigneron
Address of the host laboratory: Informatique, BioInformatique, Systèmes Complexes Team Informatique, BioInformatique, Systèmes Complexes 23 Boulevard de France 91037 EVRY CEDEX France
Supervisor 1: vincent vigneronE-mail: vincent.vigneron@univ-evry.fr Phone: 0663568760
Supervisor 2: Hichem MaarefE-mail: hichem.maaref@iut.univ-evry.frPhone: 0663568760
Description Ultrasounds are used for non-destructive testing (NDT) of industrial parts without harming their integrity. This consists of emitting acoustic waves and detecting their interactions with defects in the part. In real-time, the re-emitted waves (echo) are then converted into a digital image of the fault, thus located and characterized. This internship focuses on identifying possible defects in fasteners in a critical system by deep neural networks. The identification of these defects will be based in particular on several ultrasound measurements (multimodal), carried out in situ by the maintenance teams of various partner industrial sites [4]. A promising approach is first to estimate the quality of an acquisition, as many factors can directly lead to poor analysis when it comes to determining the presence or absence of a defect or rendering the acquisition uninformative for this task. Objectives: The (SMART) objectives of this study are as follows: 1 (main): Be able to detect, in an unsupervised way, poor-quality acquisitions automatically 2. Compare results to make them consistent with those of experts. 3. Improve existing models for supervised estimation of acquisition quality.
Methodology The networks envisaged a priori for this research are deep neural architectures, more specifically, Vision Transformer (ViT) for supervised classification with, as inputs, several modalities sent at the same time in different directions, forming a multimodal "image" [2]. Expected results: Deep neural networks will be trained to correct pre-existing annotations by identifying suspect annotations. A statistical study of the behavior of model outputs will be carried out to obtain a confidence index in estimating model outputs. A data augmentation procedure may be developed to reduce the data to be measured and build a new training database.
Références [1] Karakuş etal. Detection of line artifacts in lung ultrasound images of covid-19 patients via nonconvex regularization. IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control, 67(11) :2218–2229, 2020. [2] Lopez etal.. Non-destructive testing application of radiography and ultrasound for wire and arc additive manufacturing. Additive Manufacturing, 21:298–306, 2018. [3] Najafabadi et al.. Deep learning applications and challenges in big data analytics. Journal of Big Data, 2, 12 2015. [4] Shangqin Yuan and Xudong Yu. Ultrasonic non-destructive evaluation of selectively laser-sintered polymeric nanocomposites. Polymer Testing, 90 :106705, 2020.
