SAMPE Journal - May/June 2020 - 40

FEATURE / AFP INSPECTION

RESULTS
Figure 6 displays a number of testing images and
their respective predictions. It can be observed that
the software is capable of identifying the locations
and types of defects with a fair accuracy. The green
color on the prediction maps signify the existence
of gaps while orange signifies the existence of overlaps. The UI with functioning inspection capabilities was assessed by an inspection system operator
and feedback was given to improve the system. We
can make a few notes as to the effectiveness of the
defect detection algorithms. Firstly, it should be
noted that any change of material has to potential
to affect the end accuracy of our system.
This means that while there is a potential to do proper identification of dry fiber or thermoplastic material,
it will often skew results in the detection of smaller defects in the case of thermoplastics and the identification of wrinkles and puckers in rough dry fiber.
It can also be noted that while there are a number of
defects that our system can detect with a fair degree of
accuracy, there a number of classes that are observed a
handful of times and thus have limited opportunities
to train with. In these cases, our detection algorithms
are capable of indicating that a defect is present, but
misclassification is a distinct possibility.
CONCLUSION
An ML algorithm based on Fully Convolutional
Neural Networks (FCN) that allows for pixel-by-pixel classification is used. This permits the software
to identify the exact size and shape of a defect in
addition to its location on the part. The utilization
of a UI gives the system operator precise control
over the inspection process and allows for the possibility to correct inaccurate predictions from the
ML algorithms. The ability to continually train the
FCN from these corrections implies continually
improving accuracy in the inspection process. A
Raspberry Pi 3 B+ is used to host a USC AFP Defect Database as a MySQL server that can be easily
accessed from other software tools. The system's
information availability is ideal for integration with
rapid analysis tools or machine parameter correlation, moving the system to an Industry 4.0 concept.
In addition, the storage of data means that a better digital twin can be instanced, aiding in product
lifecycle management (PLM). The ML network was
constructed with the use of Glorot initialization
for 2D convolutions and batch normalization. The
network architecture is a FCN variant of the ResNet
network incorporating skip functions every three
layers. In total the network is 45 layers deep. An
Nvidia Titan Xp graphics card was used to accelerate training and improve prediction speed.
The execution of an AFP inspection system re-

SAMPE JOURNAL

quires more than simply proper identification of
defects. Rather, the system outline separates itself
through both identification and presentation. Defect information can be utilized in a constructive
manner, and the hesitancy of using automated inspection, particularly those consisting of machine
learning algorithms, can be mitigated. Machine
learning can be applied in the context of a tool rather than a proof of concept. Integrating an inspection
system with other manufacturing analysis tools can
spur greater efficiency, innovation, and quality.
There are a number of important notes when
creating ML-based systems. It is an absolute prerequisite to have a suitable amount of data for input
vectors that have a large number of features. The
data must also have adequate distribution over all
of the classes that are to be identified. For AFP defects, this implies that the preponderance of gaps
and overlaps can pose a potential problem for data
collection efforts. This can be mitigated through
the application of data augmentation algorithms.
In the system presented in this publication, a sine
wave distortion was introduced to certain collections of data that were evaluated to be underrepresented in the dataset.
The identification of multiple defects beyond
the gap and overlap focus of many of the automated inspection systems is a principle priority of
the ongoing development of this AFP inspection
platform. In certain cases of rare defect types, our
algorithms are often able to identify that a defect
is present, but will tend to misclassify the defect
type. This is manageable through the UI, however it points to the need for a potential standard
AFP defect training set in a manner similar to the
ILSVRC or other image recognition and classification competitions. This will alleviate the potential
miss-identifications due to the early learning stages, when the system is in 'setup' and/or initial trials
are being conducted, i.e. different material setup or
such. Concerted work to yield above 1000 training
images may be the necessary tipping point to push
the current system into being accurate enough for
industrial integration with minimal operator intervention. Developing this additional training data
will be a focus for the continuing improvement of
inspection system. A number of additional tasks for
future work include:

M AY/J U N E 2 0 2 0

* Expanding data management tools
* Exploration and optimization of
profilometery settings
* Developing network retraining
capabilities from operator input

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

http://www.sampe.org

SAMPE Journal - May/June 2020

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