SAMPE Journal - May/June 2020 - 35

INTRODUCTION
Purpose
The advent and popularization of AFP manufacturing in the aerospace industry has led to large
jumps in the size and throughput of various composite structures that are realistically possible in a
traditional manufacturing setting. This increase in
the speed of material deposition, however, has a
number of notable drawbacks. Principally among
them, automation with robotic components leads
inevitably to a lack of control in the quality of what
is produced on a system. The result is potential for
the rapid production of defects in manufacturing
and thus a need to detect, identify, and characterize
defects. A collection of common AFP defects can
be found in1. Traditionally, this has been accomplished through the use of human inspectors, and
thus the inspection time and quality are subject to
human variations. The need for human inspectors
also placed an additional restraint on the availability of experts needed to accomplish the task. In
recent years, the concept of automated inspection
systems coupled with automated manufacturing
systems has generated a great deal of interest in
the both the academic and professional communities. Combined with the advancements in machine
learning techniques automated inspection has the
potential to reduce inspection time and cost and
increase consistency while having high accuracy.
Reliable rapid inspection also allows for an additional data source that can be accessed for further
study. This enables advances in product lifecycle
management (PLM), the development of a digital
twin, and the post-manufacturing analysis of a particular structure2.

et al.6 use machine learning techniques to incorporate multiple sensor inputs for the identification of
defects on composite aerospace structure.
Benítez et al.7 outlines the creation of a thermographic inspection system to identify defects in
composites structures, with defect identification
being accomplished through the use of both the
support vector machine (SVM) described in8 and
radial basis function networks. Brüning et al.9 couples an infrared inspection system mounted on the
AFP machine head with several process parameters
used in manufacturing to utilize machine learning
for the optimization of the process parameters for
AFP. The use of SVM for the classification of porosity, inclusion, and delamination in composite
structures through ultrasonics is demonstrated
in10. The use of ML techniques in the evaluation of
eddy current data for the classification of defects
in composite structures was studied in11. The authors utilized a number of algorithms including a
U-BRAIN approach that showed promise.
The first iteration of our approach was demonstrated in12. In the ensuing period, automated AFP
defect detection has remained an area of active
research and development. Thus, the continued
improvement of our system, with particular em-

Proposed Solution
In previous work we extensively discuss the architecture of the network used for defect recognition,
and the process by which the defects are classified1.
In this publication, we intend to outline the most
recent additions to the system which now incorporates a UI for the USC developed Machine Learning
tool, as well as an AFP defects MySQL database that
can be shared with other projects.
Literature Review
Machine learning in visual inspection tasks has
made a steady gain in the literature since the popularization of the convolutional neural network
(CNN) in3 for image classification tasks. Further developments outlined in4 pushed some image recognition tasks above human performance. Meng
et al.5 utilize CNNs to classify defects in composite
materials from ultrasonic scanning methods. Kuhl

w w w. s a m p e . o r g

M AY/J U N E 2 0 2 0

SAMPE JOURNAL

http://www.mccleananderson.com/ http://www.sampe.org

SAMPE Journal - May/June 2020

Table of Contents for the Digital Edition of SAMPE Journal - May/June 2020