ScienceDirect Scalable Detection of Defects in Additively Manufactured PLA Components

Abstract

The present work delineates a novel and scalable approach to characterization of defects in additively manufactured components. The approach is based on digital image correlation and involves characterization of surface speeds during rigid body rotation of the workpiece, followed by normalization with respect to rotation speed. Towards this, two different imaging sources were tested, viz. smartphone camera and sophisticated high-resolution/high-speed camera. The proposed approach successfully delineated horizontal and vertical notch defects in a simple FDM fabricated component. Accuracy of this approach was tested with concomitant laser based scanning. Some limitations of this approach were discussed.

Full text

Seediscussions,stats,andauthorprofilesforthispublicationat:https://www.researchgate.net/publication/323150464
ScienceDirectScalableDetectionofDefectsin
AdditivelyManufacturedPLAComponents
Preprint·February2018
DOI:10.13140/RG.2.2.20071.55204
CITATIONS
0
READS
31
8authors,including:
Someoftheauthorsofthispublicationarealsoworkingontheserelatedprojects:
ShapefromMotionViewproject
DirectAdditiveandSubtractiveHybridManufacturing--DASHViewproject
AmolKulkarni
PennsylvaniaStateUniversity
1PUBLICATION0CITATIONS
SEEPROFILE
AmeyNVidvans
PennsylvaniaStateUniversity
3PUBLICATIONS15CITATIONS
SEEPROFILE
GuhaP.Manogharan
PennsylvaniaStateUniversity
24PUBLICATIONS188CITATIONS
SEEPROFILE
SaurabhBasu
PennsylvaniaStateUniversity
37PUBLICATIONS188CITATIONS
SEEPROFILE
AllcontentfollowingthispagewasuploadedbySaurabhBasuon13February2018.
Theuserhasrequestedenhancementofthedownloadedfile.

Available online at www.sciencedirect.com
ScienceDirect
Procedia Manufacturing 00 (2018) 000–000
www.elsevier.com/locate/procedia
2351-9789 © 2018 The Authors. Published by Elsevier B.V.
Peer-review under responsibility of the scientific committee of NAMRI/SME.
46th SME North American Manufacturing Research Conference, NAMRC 46, Texas, USA
Scalable Detection of Defects in Additively Manufactured PLA
Components
Amol Kulkarnia, Amey Vidvansb, Mustafa Rifata, Gregory Bicknellc, Xi Gongc
Guha Manogharanc, Janis Terpennya, Saurabh Basua,*
aHarold and Inge Marcus Deoartment of Industrial and Manufacturing Engineering, Penn State University, University Park, PA, 16801, USA
bis now at George W. Woodruff School of Mechanical Engineering, Georgia Institute of Technology, Atlanta, GA, 30313 USA
cDepartment of Mechanical and Nuclear Engineering, Penn State University, University Park, PA, 16801, USA
* Corresponding author. Tel.: +1 814 863 2447; fax: +1 814 863 4745
E-mail address: sxb514@psu.edu
Abstract
The present work delineates a novel and scalable approach to characterization of defects in additively manufactured
components. The approach is based on digital image correlation and involves characterization of surface speeds during
rigid body rotation of the workpiece, followed by normalization with respect to rotation speed. Towards this, two
different imaging sources were tested, viz. smartphone camera and sophisticated high-resolution/high-speed camera.
The proposed approach successfully delineated horizontal and vertical notch defects in a simple FDM fabricated
component. Accuracy of this approach was tested with concomitant laser based scanning. Some limitations of this
approach were discussed.
© 2018 The Authors. Published by Elsevier B.V.
Peer-review under responsibility of the scientific committee of NAMRI/SME.
Keywords: Additive Manufacturing; Defect Detection; Digital Image Correlation; Optical Imaging
1. Introduction
Additive Manufacturing (AM) involves production of
component geometries in a layer-by-layer, i.e. bottom-
up approach. Recent advancements in AM have
resulted in their widespread use in fabrication of re-
entrant geometries in rapid design-to-manufacture
environments [1]. Negating many requirements of
conventional machine tools, e.g. tool inserts and
fixtures gives AM an advantage in remote settings,
where access to traditional manufacturing machine
tools is not readily available. Although AM has several
advantages over traditional manufacturing processes,
the primary reason preventing its broad acceptance is

2 Author name / Procedia Manufacturing 00 (2018) 000–000
part quality, which tends to be variable. This is due to
the natural albeit stochastic generation of performance
compromising surface roughness features and
volumetric detects [2], characterization of which has
proven to be challenging. It has been realized that
process parameters in AM play a primary role in the
resulting part quality[3]. As a result, manufacturers
adjust process parameters on an experiential trial and
error basis. This is unsustainable in platforms where
rapid deployment of parts featuring complex
geometries is necessary. Unfortunately, even such
manufacturing protocols largely rely on subjective
characterization approaches, primarily because on-
line defect characterization is often prohibitively
expensive [4].
Challenges towards accurate and rapid on-line
detection of defects originate from of manufacturing
defects in the presence of manufacturing system
variability that can result from the dynamics of
manufacturing machine tools tied to shop floor
vibrations or from variability that naturally exists in
materials. These defects often feature high aspect
ratios that are defined as their characteristic depths
normalized with respect to their characteristic span in
the other two dimensions. For instance, the aspect ratio
of a defect artificially introduced on the surface of a
perfect cylinder would be ar =
, where h is the
height of the defect and l=min{a,b}is its span. Defects
featuring high aspect ratios ar and low characteristic
spans l can be challenging to characterize in scalable
frameworks due to insensitivity to visible light [4].
Techniques exist that can circumvent these
fundamental limitations, for instance by optical-
interference profilometry or through incidence of
high-energy/low-wavelength electromagnetic waves
as in scanning electron microscopy. However, these
techniques are slow, i.e. involve sample preparation
and therefore cannot be used for on-line monitoring.
In this regard, these techniques cannot be used for
validation of manufactured components as part of the
build routine.
This paper proposes a simple approach to
characterization of manufacturing defects of additive
components. The underlying spirit of the proposed
approach is scalability. In this regard, the approach is
based on optical imaging and can utilize standard
smartphone cameras as well as advanced optical
imaging with sophisticated lens assemblies. In this
regard, this research presents theoretical arguments
that will enable quantification of resolution of the
imaging source used for characterization of AM
defects.
The remaining parts of this paper are organized as
follows. Section 2 describes challenges associated
with on-line defect characterization and some
approaches that have been devised to circumvent the
same. Subsequently, section 3 describes some
standard defects in AM components. Section 4
describes our proposed approach towards optical
imaging and defect detection. Section 5 describes
experiments performed to demonstrate a proof of
concept capability of the proposed approach to defect
characterization. Section 6 discusses some results that
validate a proof of concept of this approach.
Subsequently, section 6 also describes
characterization of defects in a simple 3D printed PLA
part. This leads into section 7 that discusses
shortcomings of this approach and provides
suggestions for future work that can improve the
throughput of this methodology.
2. Literature Review
Optical imaging with subsequent processing has
been extensively used to detect defects in 3D- printing.
On-line implementation of the same includes
platforms such as that in Ref.[5], where three cameras
optically image an AM component synchronously
during its build routine. These images are
subsequently analyzed to provide guides, e.g. if the
higher wavelength component of the part geometry is
being additively manufactured within a specified
threshold error. A closed loop approach given in Ref.
[6] involves characterization of anomalies in optical
images of the geometry being additively fabricated.
These anomalies are delineated with respect to
expected process signatures that can be modeled a-
priori from the input parameters of the build, e.g. part
geometry, material and process parameters. More
recent approaches have looked at refinement of image
analysis with the primary motive of accelerating their
throughput. For instance, Ref.[7] has looked at multi-
fractal analyses towards characterization of
manufacturing defects within families. This approach
is motivated by defect characteristics being strongly
rooted in process physics. Therefore, it can be
anticipated that a few specific defect types will result
if the process physics can be specified. As this is the
case in most manufacturing platforms, including AM,
these defects types can therefore be learned using
machine-learning algorithms [4]. Eventually, this
knowledge can accelerate detection of defects. Finally,

Author name / Procedia Manufacturing 00 (2018) 000–000 3
delineation of defects has looked at calibrated
characterization of reflective properties of the surface
[4]. For instance, surface roughness features can be
anticipated to result in incoherent scattering, the extent
of which can be characterized and related to the
characteristics of surface roughness, i.e. high
frequency components of its profile.
3. Real defect detection
In the present research, fused deposition modeling
(FDM) will be used for characterization of our defect
detection test bed. Fused deposition modeling is one
of the most widely used polymer 3D printing
technologies [8]. In this process, a thermoplastic
filament is fed into a heated nozzle, whereby the
filament undergoes melting. In this molten state, the
filament is extruded and laid down in a layer by layer
fashion. Process parameters such as nozzle
temperature, printing and extrusion speed and layer
thickness play a prominent role in the quality of the
component produced [3] by FDM. None the less,
several defects arise if the 3D printer isn’t calibrated
to handle complex geometries. Examples of the same
include: elephant foot base bulging, pillowing,
stringing, under-extrusion, over-extrusion, gaps in the
top layer, visible lines at the bottom layer, scars on the
top surface, cave in and dimensional inaccuracies [5].
This present research is concerned with layer
defects, e.g. horizontal and vertical defects. The
horizontal layer defects include misalignment of
layers, skewed or shifting of layers, and missing
layers. In addition, environmental factors, e.g. dust
particles can settle between the AM layers and result
in highly rough features on an anticipated smooth
surface. In contrast, vertical layer defects result from a
common starting point at each layer, where the nozzle
begins to deposit the fused thermoplastic filament.
This results in a notch like vertical bar on the
component as shown in the figure below. Both
horizontal and vertical defects feature high aspect
ratios. For instance, the horizontal as well as vertical
notches shown in Fig. 1 illustrates these defects on a
standard cylindrical part fabricated by FDM. The
present research attempts to characterize these defects
as a proof of concept of our proposed approach to
defect detection. It is anticipated that with this
validation, the proposed methodology can be extended
for characterization of other defect types.
4. Proposed Approach for Defect Detection
Our proposed approach to defect detection is
schematically illustrated in Fig. 2. The approach is
rooted in the rotational kinematics of rigid bodies that
prescribes that the surface speed of a rotating object
can be quantified using the relation: V=rω. Here, r is
the radius of the point of interest with respect to the
axis of rotation and ω is the angular speed. Herein,
points on the work piece further from the axis of
rotation, i.e. those featuring higher radius, e.g. point b
featuring a radius r
b
will feature higher surface speeds
V
b
in contrast with those points on the work piece that
feature a smaller radius, e.g. point c featuring radius
r
c
, during angular rotation of the work piece at speed
ω.
The crux of the
proposed approach
therefore lies in
characterization of
surface speeds
during rigid body
rotation of the work
piece that is being
additively
manufactured. This
is performed using
digital image
correlation (DIC).
This process
involves optical
imaging of the surface of the rotating object [9], [10].
Subsequently, image correlation algorithms are
implemented to characterize rigid body displacement
of features on the region of interest. This rigid body
Fig. 1 Surface of cylinder fabricated with fused
deposition melting shown. The magnified inert on
the right shows horizontal and vertical defects with
black arrows. Fig. 2 Principal behind
proposed approach
towards defect detection.

4 Author name / Procedia Manufacturing 00 (2018) 000–000
displacement naturally results from surficial motion of
the rotating object. The DIC characterized surface
speeds are thereafter normalized with respect to the a-
priori known angular velocity. Repetitive
characterization in a similar fashion of the radii of
points of the surface results in a point cloud that is
representative of the geometry of the work piece.
This characterization is unfortunately skewed by
the naturally occurring foreshortening error in digital
imaging. This phenomenon manifests slower apparent
relative speeds in stationary objects further away,
when viewed with respect to a moving reference than
stationary objects in close vicinity. Towards our
proposed approach to geometry characterization, this
phenomenon is addressed by calibrating the camera
resolution N≡N(d) (pixels/mm) with respect to object
distance d (mm) from the camera lens as described in
Ref [10].
5. Experiments and Methods
Experiments for demonstrating the proof of concept
involved imaging the rigid rotation of a variety of
geometries. This rigid rotation was performed on a
stage fabricated in-house. The rotation stage
comprised a 3D printed PLA platform that was
coupled to a worm gear and actuated using a Power
HD AR-3606HB servo motor. Motion to this stage
was controlled using an Arduino micro-controller that
was programmed in a standard windows computer.
Some imaging was also performed on a manual tool
indexing device.
Imaging was performed using two devices. First: a
smartphone camera featuring a manufacturer
calibrated focal distance of 31 mm and a f/2.2 lens and
second: a PCO high speed camera equipped with an
infinity telescopic microscope lens. Towards imaging,
the lenses were oriented perpendicularly with respect
to respective goniometer axes of rotations. The error
that can result in imperfect vertical alignment on the
resulting point clouds is discussed in the forthcoming
sections. Imaging was performed in full resolution in
both, smartphone and PCO cameras. However, the
smartphone camera distance was adjusted such that the
work piece featured a width smaller than 400 pixels
close to the center of the image. This specification was
derived from lens distortion characteristics present in
our smart phone cameras as described in the next
paragraph.
Lens distortion was characterized by imaging a
checkered pattern featuring alternate black and white
squares of size ~0.71mm that was printed with a HP
printer featuring a resolution ~42μm. Subsequently,
the locations of the vertices on the checkered pattern
registered in the digital image were found using a
Matlab routine. Displacement of these vertices with
respect to their expected locations were then
characterized as a distortion field δ(x,y). Here, x and
y refer to the ideal locations of vertices on the
checkerboard. For quantifying errors in defect
characterization that result from camera distortion, this
field was then differentiated spatially as ∂δ/∂x. With
subsequent analysis as in Ref.[10], it was shown that
the images acquired by the camera lens featured
negligible lens distortion with regards to defect
characterization within a zone that was ~ 400 pixels
wide close to the center of the image. Therefore,
imaging was confined to this zone by adjusting the
object work piece distance.
Preliminary experiments involved imaging of a
variety of objects featuring regular surface curvatures.
In this article, experiments performed with acorn
pericaps are described. These pericaps feature a
diameter ~20 mm and exhibit a variety of surface
features (see forthcoming sections). Nonetheless,
investigations with engineered geometries such as
pristine and crushed soda cans are described elsewhere
[10] to test the validity of this approach towards
characterization of simpler shapes. The shape of
these objects was characterized using the proposed
imaging methodology. Subsequently, accuracy of the
methodology was verified by comparison of some DIC
characterized point clouds with those obtained from
concomitant laser based scanning in a Makerbot laser
scanner. Comparison of point clouds was performed
using the open source software: CloudCompare. This
involved characterization of normal distances between
the DIC laser scanning characterized point clouds as
prescribed in Ref. [10].
Subsequently, a simple cylindrical object was made by
FDM for characterization of manufacturing defects.
The cylindrical component was made of PLA
(polylactic acid) plastic, which along with ABS
(acrylonitrile butadiene styrene) is the most common
plastic used in 3D-printing [8]. PLA belongs to a class
of aliphatic polyester family which is non-toxic, bio-
degradable and has low shrinkage which makes it an
ideal material for 3D printing[11] . However, an
additively manufactured PLA component is far from

Author name / Procedia Manufacturing 00 (2018) 000–000 5
perfect and warps around the edges. Towards
characterization of geometric horizontal and vertical
errors in the cylinder fabricated by FDM, two images
were taken one degree apart from each other. This was
done using the high-speed camera that was equipped
with a high-resolution (1.3 μm/pixel) telescopic
microscope. Further, imaging was limited to the
vertical defect as identified in Fig. 1. For comparison,
the FDM manufactured cylinder was also scanned
with a Keyence LJ-V7200 laser scanner.
6. Results
Figure 3 shows the results of shape characterization
using the proposed methodology. Figure 3a shows a
subset of the sequence of images of the acorn pericarp
featuring a step angle Δθ=10
o
over an angular range
θ∈[10
o
,90
o
]. The magnified view of the acorn
pericarp at an angular orientation θ=50
o
is shown in
Fig. 3b and the reconstructed and rendered point cloud
is shown in Fig. 3c. It is evident from this figure that
the DIC characterization can reproduce the dominant
features of a 3D geometry in a point cloud. For
instance, a growth defect in the acorn pericarp was
readily delineated by DIC based reconstruction. This
growth defect is demarcated within a dashed line in the
optical and reconstructed/rendered point clouds.
In order to test the accuracy of characterization, the
geometry of another acorn pericarp was characterized.
The point cloud obtained from the same was compared
with that obtained from laser based scanning. Figure
4a shows the optical image of the acorn pericarp, this
exhibiting similar geometric features as its DIC
characterized counterpart, a rendered point cloud for
which is shown in Fig. 4b. Figure 4c shows the DIC
characterized point cloud (gray scale) overlaid on that
obtained from laser scanning. Figure 4d shows the
difference in point clouds, exhibiting small mean 0.16
mm and standard deviation errors 0.12 mm.
The aforementioned results clearly validate some
capabilities of the proposed methodology in
characterization of geometries featuring gradual
surface curvatures. Towards characterization of
surface features exhibiting sharp curvatures (e.g. high
aspect ratios), the methodology was deployed in
delineating defects in our FDM manufactured
cylinder. Imaging for the same involved orienting the
cylinder over a step angular span of Δθ=1
o
and
acquiring two digital images. These images are
elucidated in Figs. 5a and feature a high resolution
~1.3μm/pixel. Characterization of these defects was
subsequently performed by DIC on a zone demarcated
within the dashed line in Fig. 5a.
Figure 5b shows the DIC characterized
displacement field on the zone demarcated in Fig. 5a.
Herein, the pair of vertical lines (cf. arrows in Fig. 5a)
associated to the vertical defect exhibit high gradients
in surface displacement with respect to their
neighboring zones. Figure 5c shows the mean
horizontal displacements obtained from this exercise,
this exhibiting two dominant minima in the vicinity of
the vertical defects. The DIC distance between these
minima was found to be 0.5016mm, lying in vicinity
of its real counterpart at 0.4902 mm. The
displacement field also show zones of horizontally
Figure 3. Shape characterization of acorn pericarp
with growth defects. The dashed line shows a
growth defect in the acorn pericarp.
Figure 4. Delineating accuracy of characterization
methodology.

6 Author name / Procedia Manufacturing 00 (2018) 000–000
oriented minima, pointed using horizontal black
arrows in Fig. 5b. These zones correspond to
horizontal defects shown in Figs. 1 and 5a.
Figure 6 shows the laser scanned counterpart of
AM defects on the fabricated components. In
comparison to results obtained from DIC, the laser
scanned surface features exhibits a defect width of 2.5
mm for the vertical defect, thereby overestimating by
~ 2 mm. The laser scan feature exhibits periodically
occurring undulations, i.e. horizontal defects on the
surface of our FDM components, in contrast with the
DIC characterized defect that exhibits abrupt surface
undulations. Reasons for these differences between
laser and DIC based characterization lie in their
respective imaging conditions. Some of these factors
will be discussed in the discussion section.
7. Discussion
It is clear from results described in section 6 that
DIC based reconstruction is a viable method for
detection of defects in AM components. Nonetheless,
discrepancies were found between defects This section
identifies some limitations of the process.
7.1 Characterization errors arising from
intrinsic factors: surface curvature in fabricated
component
In order to avoid effects that naturally arise from
foreshortening, characterization errors resulting from
surface curvature are identified here using simulated
reconstruction. Towards this, a highly curvilinear
ob ject as sh own in F ig. 7a w as gen era ted in MATLAB.
Rotation of this object about its central axis was
simulated, projections from which were extracted. The
characterization of its shape was thereafter performed
using the algorithm described in section 4. Figure 7b
shows this reconstructed shape. In comparison with its
original shape, the reconstructed geometry exhibits
dull features where highly curved surface features are
anticipated. This has been pointed using arrows shown
in Figs. 7a and 7b.
Reasons for this discrepancy lie in limited spatial
resolution during imaging of highly curved features.
Figure 5. Characterization of defects in part
fabricated by FDM. The inset on bottom left shows
orientation of rotation.
Figure 6. Laser based characterization of AM
defect.
Fig. 7 Errors in shape characterization
resulting from high curvature features.

Author name / Procedia Manufacturing 00 (2018) 000–000 7
This limitation prohibits accurate shape
characterization when the radii associated with
features of interest changes rapidly on the surface.
Origins of this can be traced to an averaging effect,
which arises during characterization of radii by DIC.
This effect understates the radii of highly curved
features. Circumvention of this shortcoming can be
performed by improving camera optics, whereby
higher resolution imaging becomes possible.
7.2 Characterization errors arising from
extrinsic factors: limited camera resolution
A dominant source of error in the characterization of
high aspect ratio surface features by DIC lies in the
decaying resolution of a camera with respect to
distance from the object of interest. This is a natural
consequence that exists in optical imaging. To
delineate this effect, let us assume that the resolution
of a camera varies with distance as N≡N(d), where d
is the distance from work piece. If the axis of rotation
of the object is at distance da from the camera, a
surface feature radius r and height Δr would be at
effective distance∆. This implies that for
an angular rotation Δθ, the surficial speed is rΔθ.
Changing the reference axis, we find that that this
digital displacement that will be characterized by DIC
is given in pixels as: ∆∆
∆
For this to be characterizable and error free, minimum
value of parameter Δsp > Δs
err. Here, the parameter
Δserr is the minimum surficial displacement that can be
characterized using DIC [10]. Implementation of this
exercise through our smartphone camera results in Δr
> ~20μm [10].
8. Conclusion
The present work shows that DIC based
characterization can be used for: (i) reconstructing
complex curvilinear geometries and (ii) identifying
defects on components fabricated by additive
manufacturing. The approach was validated with
concomitant laser based scanning. This enabled the
comparison of point clouds obtained from the two
routes to surface characterization using the open
source software CloudCompare. It was seen that
sufficient refinement of camera optics can enable high
fidelity and highly accurate characterization of surface
features, including surface defects on AM
components. Some limitations of DIC based
characterization were discussed.
9. References
[1] Y. Kok, X. Tan, S. B. Tor, and C. K. Chua,
“Fabrication and microstructural
characterisation of additive manufactured Ti-
6Al-4V parts by electron beam melting: This
paper reports that the microstructure and
micro-hardness of an EMB part is thickness
dependent,” Virtual Phys. Prototyp., vol. 10,
no. 1, pp. 13–21, 2015.
[2] A. Townsend, N. Senin, L. Blunt, R. K.
Leach, and J. S. Taylor, “Surface texture
metrology for metal additive manufacturing:
a review,” Precis. Eng., vol. 46, pp. 34–47,
2016.
[3] G. Manogharan, B. Yelamanchi, R. Aman,
and Z. Mahbooba, “Experimental Study of
Disruption of Columnar Grains During Rapid
Solidification in Additive Manufacturing,”
Jom, vol. 68, no. 3, pp. 842–849, 2016.
[4] M. S. Tootooni et al., “Online non-contact
surface finish measurement in machining
using graph theory-based image analysis,” J.
Manuf. Syst., vol. 41, pp. 266–276, 2016.
[5] J. Straub, “Initial Work on the
Characterization of Additive Manufacturing
(3D Printing) Using Software Image
Analysis,” Machines, vol. 3, no. 2, pp. 55–71,
2015.
[6] T. Fang, M. A. Jafari, I. Bakhadyrov, A.
Safari, S. Danforth, and N. Langrana,
“Online defect detection in layered
manufacturing using process signature,” Syst.
Man, Cybern. 1998. 1998 IEEE Int. Conf.,
vol. 5, pp. 4373–4378, 1998.
[7] B. Yao, F. Imani, A. S. Sakpal, E. (Ted)
Reutzel, and H. Yang, “Multifractal Analysis
of Image Profiles for the Characterization and
Detection of Defects in Additive
Manufacturing,” J. Manuf. Sci. Eng., no. c,
pp. 1–21, 2017.

8 Author name / Procedia Manufacturing 00 (2018) 000–000
[8] A. Rengevič, M. Fúra, and N. Čuboňová,
“Analysis of Printing Parameters for
Production of Components With 3D Easy
Maker Printer,” Adv. Sci. Technol. Res. J.,
vol. 10, no. 32, pp. 1–8, 2016.
[9] S. Basu and M. R. Shankar, “Spatial
confinement-induced switchover in
microstructure evolution during severe plastic
deformation at micrometer length scales,”
Acta Mater., vol. 79, pp. 146–158, 2014.
[10] A. Vidvans and S. Basu, “Smartphone based
scalable reverse engineering by digital image
correlation,” Optics and Lasers in
Engineering,. vol 102, pp. 126–135, 2018.
[11] J. Krotkỳ, J. Honzíková, and P. Moc,
“Deformation of print PLA material
depending on the temperature of reheating
printing pad,” Manuf. Technol., vol. 16, no.
1, pp. 136–140, 2016.
View publication statsView publication stats