Document Type

Article

Source Publication Title

Proceedings of Society for the Advancement of Material and Process Engineering (SAMPE) Conference 2018

Abstract

The state of art non-destructive inspection techniques for composite materials detect the presence of defects in the composite material, but they do not identify what type of defect it is, and hence, further visual inspection of the details are needed. This visual classification is a costly and time-consuming process, and it’s difficult to distinguish all of the defects effectively. Broadband Dielectric Spectroscopy (BbDS), has been an established tool for dielectric material characterization in polymer industries for a long time. Dielectric spectra of heterogeneous materials are altered by constituent interfaces, with changes in morphological heterogeneity, electrical and structural interactions between particles, and shape and orientation of the constituent phases of the material system. Machine learning and Artificial Neural Networks (ANN) are computing systems that behave like our brains, storing and learning from previous data (training data) fed into it. In this terminology, classification is identifying the data according to the subset it belongs to. In this paper, we propose a Non-destructive inspection technique by combining the concepts of the Broadband Dielectric Spectroscopy with Machine Learning Algorithms and Neural Network Computing systems. This technique not only detects the presence of the defects, but can also accurately predict and classify the various defects based on their dielectric properties, as the presence of the various defects varies with the spectra of the interfaces. An experimental procedure for obtaining the dielectric properties of the composite materials with various defects and the classification of the defects by Random Forest Classifier Algorithm and Neural Networks are discussed in this research work.

Disciplines

Engineering | Materials Science and Engineering

Publication Date

5-1-2018

Language

English

Comments

We would like to acknowledge the Institute of Predictive Performance Methodologies at the University of Texas at Arlington, for funding and supporting this research work, its interns for assisting with the data collection process for their tremendous help in this research work.

Available for download on Wednesday, January 01, 3000

Share

COinS