Session: 16-01-01: Poster Session: NSF-Funded Research (Grad & Undergrad)

Paper Number: 100159

100159 - Convolutional Neural Network for Classification of Elements to Detect Multiple Voids in a 2d Plane Strain Solid Using Elastic Waves. 

Defects of various sizes and shapes, including voids, inclusions, and line cracks, can compromise the structural integrity of civil and mechanical structures. Several approaches have been studied, such as ultrasonic testing, electrical tomography, and so on. Characterizing such defects using wave-based empirical NDT procedures without any systematic numerical method is time-consuming, requires a trained technician, and is only applicable to simple problems (e.g., a problem where only the location of a single line crack of an assumed orientation is detected). Thus, systematic inverse modeling is needed to detect complex cracks subjected to dynamic excitation. Several optimization-based methods are used to tackle inverse modeling, and we can see several limitations to these optimization-based methods. The most prominent drawback is the longer optimization time required for these methods. A number of recent papers have shown that machine learning (ML) would overcome the limitations of optimization-based methods. Machine learning has been studied for the inverse-scattering problems because of its short computing time once training is done and its robustness in terms of accuracy. This paper presents a new method to detect a random number of voids by using a CNN considering elastodynamic waves in a 2D plain strain solid. Conventional FEM requires re-meshing to solve forward problems for a domain, including cracks, which is time-consuming. The extended finite element method (XFEM) or the level-set method can be adopted because they can successfully model cracks or voids in the heterogeneous medium during an iterative process (e.g., optimization) without remeshing.
We consider that an elastic wave source excites the solid, containing a random number of voids, and wave responses are measured by sensors placed around the solid. We use a time-domain analysis in our forward analyses for generating training data. We use the level set method to avoid time-consuming re-meshing for various configurations of voids, which are iteratively updated while we generate training data. We introduce the details of a convolutional neural network (CNN) for tackling the inverse problem as an element-wise classification problem. The network is trained to classify element types (i.e., a regular or void element) from measured wave signals. To this end, we generate training data consisting of input-layer features (i.e., measured signals) and output-layer features (i.e., element types of all elements). When the training data are generated, we utilize the level-set method to avoid an expensive re-meshing process, which is otherwise needed for each configuration of voids. Thus, using the CNN, we identify the location, size, and shape of a random number of voids in the 2D plain strain domain.

Presenting Author: Fazle Mahdi Pranto Central Michigan University

Presenting Author Biography: Graduate Research Assistant, <br/>Solids, Waves, Intelligence, and Mechanics (SWIM) Laboratory,<br/>School of Engineering &amp; Technology,<br/>Central Michigan University.

Authors:

Fazle Mahdi Pranto Central Michigan University
Chanseok Jeong Central Michigan University