Session: 16-02-01: Poster Session: NSF Research Experience for Undergraduates (REU), NSF Posters

Paper Number: 99954

99954 - Void Detection in a Pml-Truncated Semi-Infinite Solid Material Using a Convolutional Neural Network 

The failure point of any structure is at the location of the weakest point. Within solid materials of uniform shape these weak points are the result of the absence of material that is expected to be filled. Known as voids these defects cause internal stress concentrations which result in unexpected points of failure that can result in costly and catastrophic damage if undetected. These unexpected points of failure can be detrimental to the longevity of civil engineering designs and are a large reason structural health monitoring is necessary. Identifying void locations in materials before use in complex machinery or large structures can help prevent the use of inferior components which would cause premature failure.

Detection of voids have previously been modeled with methods like full waveform inversion and finite element analysis. Where each data acquisition requires lengthy analysis to produce useful results. This research topic involves the use of a Convolutional Neural Network (CNN) architecture and a wave-based data input to produce an output showing where voids are present in a PML-truncated semi-infinite solid material. As the waves propagate through a solid material with unknown void locations the presence of a void will affect the propagation of the wave. As stated above traditional methods of analysis to find the location of these voids are computationally complex and take a substantial amount of time to perform. By using previously calculated void locations, a CNN is trained to recognize what wave abnormalities are likely to be caused by a void in the material. CNN effectively learns from a large input data set seeking patterns such as those produced when a wave is interrupted by a void. The advantage in using a CNN is that the intense computation is done only once in the training of the CNN. This allows analysis of new data to produce results quickly. Once fully trained and presented with input data, the CNN produces a two-dimensional plot showing where voids have been detected within a solid object with a degree of probability. A fully darkened pixel on the plot denotes a definite void in the material, where a fully white pixel denotes an area where there is certain to be no void. This work builds off on the authors’ previous work showing that CNNs are powerful tools that can predict voids efficiently and accurately in multiple-dimensional domain. The CNN results are within a reasonable range of accuracy and are produced in a fraction of the time compared to several minutes or hours which traditional methods take to perform and predict the same result.

Presenting Author: Jacob Thomas Central Michigan University

Presenting Author Biography: Jacob A. Thomas is an undergraduate student at Central Michigan University pursuing a Bachelor of Science in Mechanical Engineering.

Authors:

Jacob Thomas Central Michigan University
Bruno Guidio Central Michigan University
Fazle Pranto Central Michigan University
Shashwat Maharjan Central Michigan University
Chanseok Jeong Central Michigan University