Proceedings of the 1998 IEEE

International Conference on Robotics & Automation

Leuven, Belgium. May 1998

On-Line Detection of Defects in Layered Manufacturing

Tong Fang, I. Bakhadyrov, Mohsen A. Jafari', and G. Alpan

Dept. of Industrial Engineering
96 Frelinghuysen Road
Rutgers University
Piscataway, NJ 08854

standard representation of output CAD data is a

Abstract collection of triangular planar facets, called .STL
Layered manufacturing is an additive manufacturing format. The STL model is then sliced by
process where the parts are built layer by layer intersecting it with a set of horizontal planes. Each
based on the geometry provided by a CAD system. horizontal slicing plane constitutes a 2D piece-wise
There are several diferent technologies for layered linear contour. The generation of the tool-path is the
manufacturing. Our focus will be on &sed final stage before the fabrication. The goal at this
deposition of ceramics. Here layers are formed by stage is to fill the interior of the 2D slicing planes
depositing material through a liquefier-head on with various tool path patterns, e.g., contours and
adjacent roads in the shape of contours and/or rasters. A raster tool path is faster and easier to
rusters. As this technology is being pushed into new generate and with its orientation altering by 90deg
>ontiers to buildfinctionul parts, the quality of the between consecutive layers, it provides a stronger
fabricated parts is becoming a major driving force part but with a more rough surface [Weeren 951. A
for improving the technology. Here we are contour tool path, on the other hand, provides a
presenting our preliminary results on an online smoother surface. To combine the advantages of
monitoring system that can be used to detect defects. both patterns the boundaries of the slice are
The ultimate objective is to use such a system to fabricated using contour tool path and the rest of the
detect and remove the defects on any given layer interior is filled with a raster tool path pattern. The
prior to the next layer. density of a raster pattern is controlled by the air gap
between raster segments. The tool-path information
1. Introduction and the design parameters are then merged into a
Layered Manufacturing (LM) significantly design file and downloaded to the FD machine.
reduces part specific manufacturing lead times. It
has been primarily used in building prototypes for In conventional manufacturing, e.g., metal
design conceptualization, verification and cutting, due to its subtractive nature, some errors can
simulation. There are a number of different be corrected later without adversely affecting the
processing techniques that are used for layered part quality. In layered manufacturing, due to its
manufacturing and rapid prototyping [Yan 19961. additive nature, errors in a layer must be removed
Here we will only focus on Fused Deposition (FD), prior to the deposition of the subsequent layers. This
which is used for the fabrication of polymer, wax indeed makes the on-line process monitoring and
and ceramic parts. The FD process starts by on-line removal of defects a very important step in
designing a CAD model of the part using a CAD the evolution of layered manufacturing from its
package (I-DEAS or Pro-E). Currently the industry current status to its next level of manufacturing high

' All correspondence must be made to this co-author.

This work was partially supported by the Office of Naval research under grant 8 N-0014-96-1-1175 and by a grant from the NJ
commission on Science and Technology.

0-7803-4300-X-5/98$10.00 0 1998 IEEE 254

quality functional parts. With the current practice, type of applications that we are dealing with
unless a major fault happens during a fabrication requires us to detect faults as small as a few
process, the whole part is built and only after that microns.
the part is tested for some of its quality or functional
characteristics. For a relatively complex part Very few studies have been reported in the ]literature
geometry, the built process alone could take more on the surface inspection of ceramics material. The
than ten hours. It is therefore, very undesirable and closest work that we came across to was i l project
economically infeasible to find out only after so carried out by the Coors Ceramic Company and the
many hours that the part is defective and the whole Oak Ridge National Laboratory where a variety of
process must be repeated. An on-line process surface defects of rectangular ceramic substrates
monitoring system similar to what we are were studied [Patek 961. These surface defects could
developing will eliminate the processing faults after range from surface contamination, blemish,es, open
each layer is built. Thus after all the layers are built, cracks, edge chips, burrs, pits, dents, ridges, blisters
it is much less likely to have a defective finished to hairline cracks. Another recent study on ceramic
part. surfaces can be found in [Ritter 961. In this study,
photo-thermal techniques are used to detect faults
Our approach to (on-lineprocess monitoring is: (1) to including cracks, delamination, and micro-crystal
take an optical irnage of each slice or layer after the changes of structure. In this case, the production
layer is built, (2) analyze the image to identify method is inot layered manufacturing and such
various defects, (3) calculate different geometrical defects can be caused during the machining
properties of defects together with their position processes where some mechanical force need to be
within a pre-defined coordinate system, and (4) applied to the sample (e.g., drilling a hole m.ay cause
feed-back this inlformation to a closed loop control cracks in the sample where force is applied). It is
system in order to iremove the defects on-line. This assumed that the changes in mechanical properties
is very much like any typical application of machine due to these defects are correlated to the chianges in
vision in an automated environment with a number the thermal properties. Based on this assumption,
of exceptions: (1) The texture analysis of each built some form o f thermal waves are transmitted onto the
slice is extremely difficult due to the way the slices sample's surface and a thermal response ffrom the
are built and due to type of material that is used. (2) surface is captured and imaged by a sensor. The
Since the defects, must be corrected or removed on- thermal response of the defected region is shown to
line possibly with no manual intervention, the be different than that of the non-disturbed region
tolerance that is required in terms of computing the [Schmitz 961. By changing the frequency of the
various geometrical properties of the defects thermal wave, it is possible to inspect the sample
including their location are extremely tight. (3) Due surface as well as beneath the surface within a range
to material binding that must take place between of several microns up to a few millimeters. As a
adjacent roads, there is only limited time window to continuation of the study, the laser and thermal
accomplish all these tasks. (4) Defects are mainly of treatments are used for an attempt to correct these
two types in this application - under-fills or over- micron-level faults that are discovered during
fills, with less tolerance toward under-fills. To inspection.
remove an over-iill or to fill an under-fill, we must
know, respectively, the height or the depth of the In terms of complexity and resolution, the
defect. Such a three-dimensional image analysis type of image processing problem that we deal with
requires stereo-type camera setup. With our required very much resembles to those in semiconductor
accuracy and resolution stereo-typing adds to other industry. A technique called Spatial Signature
challenges that we are facing. (5) Lighting is another Analysis (SSA) was proposed [Gleason 961 [Tobin
source of complexity in this environment as the light 971 to analyze the distribution of defects over
reflectivity of some of the ceramic materials is very wafers. The distribution of defects was then
sensitive to the lighting source and setup. At the classified using fuzzy classifiers arid was
present time, we are developing such as an on-line categorized according to a problem in the
system. Our focus is on ceramic materials and the manufacturing process. This was not an on-line

255
process monitoring system. In this work we will use parameters (e.g., offsets, tool path, etc.) set by the
a similar idea. We will define process signature that, operator, but also depend on the dynamic of the
in some unique way, relates to the built process and positioning and deposition systems.
to the material used for fabrication. The process
signature is then used to filter the original image and Detection of Defects on A Slice
from the filtered image various defects were In this article, we will focus only on the
identified. detection of defects and we will not discuss any
issues that relate to the three-dimensional
2. A Framework for the Online Process complexity of the problem. In particular, we will
Monitoring focus on two-dimensional extraction of defects. We
As we discussed earlier, each slice is a will do this by studying the texture of a slice. We
collection of contours and rasters with each contour define what we call “Ideal Process Signature”
or raster being a collection of line segments. The which, theoretically speaking, should be attainable if
orientation of the rasters often changes from one various parameters, including material properties,
slice to the next in 90deg in order to improve the road offsets, thermal properties, etc. were known.
strength of the fabricated part. Each line of raster is Given that one knows the Ideal Process Signature
a road that is deposited as the nozzle on the for a given case, one should be able to simulate and
liquefier-head moves according to a pre-defined virtually build each slice by applying this signature
path. To meet functional properties, it is important to the ideal or reference tool path. Comparing the
that no unintentional gaps are present between the simulated model of an image to the actual image of
roads. Depending on the viscosity of the material each built slice, one could then make deductions
that is being used to build the part, the material about possible defects in that slice. We note that
could flow away from the roads often creating since we use the ideal tool path to generate the
undesirable defects. The overlapping of the roads virtual slice, any defects obtained by the comparison
can be controlled, to some degree of accuracy, by a reflects errors in both positioning and deposition
proper setting of an offset value. Positive setting of systems. One could also use the Ideal Process
the offset value would mean that roads do not Signature to filter the original slice image by
overlap but should only be laid down next to each moving the signature along paths that are defined on
other according to a pre-defined geometry. Negative the image. At the present time we are working with
offsetting, on the other hand, means that adjacent the other groups in the same project to define Ideal
roads overlap. The temperature of the chamber or Process Signatures for various cases. We intent to
the work envelope also plays a major role in shaping build a knowledge base which could be used for this
the roads. If out-of-control, a change in temperature purpose. As of now such a knowledge base does not
could cause unwanted flow of material, unwanted exist - We can only form a process signature that is
overlapping between adjacent roads, or unnecessary partially deducted from the image of a slice.
gaps between the adjacent roads. We illustrate the concept of using the
process signature on a sample. Figure 1 gives the
What we explained above basically defines original image of an interior region of the sample.
the natural texture of each slice provided that the For this sample, the offset between roads is set to
built process is under control to a high degree of some negative value and also the material property
accuracy. What we mean by this is that the tracking is such that material laid down on a road flows to
or contour errors caused by the positioning system some extent, to the empty side. As this also happens
are negligible. Furthermore, the deposition system is in the next adjacent road, the flow from the two
also under control all the way from extruding the adjacent roads overlap and form hill as indicated by
filament from the top of the liquefier-head through a fading line in the image. The boundary of a road is
the very small opening of the nozzle at the bottom recognized by two hills on the two sides. Figure 2
of the head. In practice, we know that this is often gives what we call “Positive Process Signature”.
not the case - There will be position and deposition This signature in a way defines the scaling factor
errors almost in every case. In other words, the that will be used to filter the original image by
actual texture of a slice will not only depend on the masking the image pixels that are outside the hills.
Figure 3 illustrates the positively filtered image. As two dimensional basis. This idea works quite nicely
it is obvious from this image, only the hills defining on the interior regions provided that the roads are
the road boundary and the hills formed by flow of linear. The process signature changes drastically as
material from the adjacent roads are highlighted. By one moves toward end of road regions. The
applying a standa.rd edge detection algorithm to this boundary between the two regions is quite fuzzy and
image, one can highlight the possible defects present further investigation, possibly using fuzzy logic, is
on the hills or road boundaries, as shown in Figure necessary to identify that boundary. More work is
4. also needed on refining the edges obtained from the
filtered models. The linearity of the tool-path
It is also possible to define a “Negative provided by CAD could be very instrumental here.
Process Image” iin order to mask the hills and to
highlight the roads and flat regions between the References
hills. Figure 5 illustrates this signature for the
interior regions of the sample. By applying this filter [I] Gleason, S . S. , K. W. Tobin, T. P. Karnowski,
to the original image of Figure 1 one obtains the “Spatial Signature Analysis of Semiconductor ]Defects,“
image in figure 6. The edges corresponding to this Solid State Technolow PP. 127-1369 July 1996.
image are shown in figure 7 where the spots on the [2] Patek, D. It., K. W. Tobin and L. Jachter, “Machine
image are possiblle source of defects. Vision Inspection of Technical Ceramics,” Machine
VisionApplications in Industrial Inspection IV, SPIE, San
The way nozzle or the liquefier-head moves Jose, California, SPIE Proceedings vol. 2665, pp. 253-
from one road to another road is very much 257, 1996.
dependent on the positioning system. In FD
technology, the feed is provided by the XY table and [3] Ritter, R., M. Reick, B. Schmitz and Gi. Goch,
the liquifier-head is stationary. As the table reaches “Nondestructive and Contactless Evaluation of Surface
to the end of a road, it slows down or it may even Coatings and Adhesion Defects by Photo thermal
stop for a very short amount of time in order to Radiometry,”Optical Inspection and Micromeasurements
make the turn from one road to another. Conference,SpIE Proceedings Series, Vol. 2782, PP. 662-
673, Besancon, France, 10-14 June, 1996.
Theoretically speaking, this must be accompanied
with a slow down in the flow rate, possibly with [4] Schmitz, B., J. Geerkens and G. Goch, “Photothermal
some very minor delay. To what level of accuracy Inspection of Material Changes in Ceramic Surfaces
this coordination is handled very much depends on Induced be Mechanical Load or Laser Treatment,”
the machine control system and on the operator’s Optical Inspection and Micromeasurements Conference,
experience. What we have seen at these boundary SPIE Proceea’ings Series, vol. 2782, PP. 109-1 19,
regions is too coimplex to show with some simple Besancon, France, 10-14 June, 1996.
signature. The boundary where an interior region
ends and an end of road region starts is fuzzy, but [5] Tobin, K. W., S. S. Gleason and T. P. Karnowski,
there seems to be some explanation for the texture at “Automatic Classification of Spatial Signatures on
Semiconductor Wafermaps,“ s p ~ ~22nd k Annual
the end of road regions. Our preliminary analysis
International ,Yymposium on Microlithographp Santa
shows that there could be a well explained pattern Clara, CA, March 9-14, 1997.
change in process signature from the intermediate
regions to the lend of road regions. Figure 9 [6] Weeren, R. V., M. Agarwala, V. R. Jamalabad, A.
illustrates the pattern change in process signature Bandyophyay, R. Vaidyanathan, N. Langrana, A. Safari,
from the interior regions (a) to the extreme end of P. Whalen, S.C. Danforth, C. Ballard, “ Quality of Parts
road regions (0. We are now working on how to Processed by Fused Position”, Proceedings of the Solid
define the fuzzy boundary between the interior and Freeform Fabrication Symposium, August, 1995.
end of road regions.
[7] Yan, Xue and P. Gu, “ A Review of Rapid
Prototyping Technologiesand Systems,” Computer-
3. Conclusion Aided Design, ’Vol. 28, No. 4, pp. 307-3 18, 1996.
We presented our preliminary results on
how to use process signatures to detect defects on

257
 12 I

Figure 1 Original image of interior regions of the sample Figure 5 Negative signature
(Negative offset between two adjacent roads) X axis: Normalized Signature width
Sample Geometry 19 lmm x 24 5mm Y axis: Gray level Scaling factor
Image Geometry 5 29 mm x 4 79 mm

''7

Figure 2 Positive Signature of interior regions Figure 6 Image filtered by negative signature
X axis: Normalized Signature width. Boundary .IRoad
Y axis: Gray level Scaling factor

Figure 3 Image filtered by positive signature Figure 7 Edge detection of image filtered by negative signature
m Boundary =Road

Figure 8 Original image of boundary region of the sample

Figure 4 Edge detection of image filtered
by positive signature Image Geometry: 5.29 mm x 4.79 mm

258
 -1 -08 -06 0 4 -02 0 02 04 06 08 1

14 1.4

12 1.2

1 1

08 0.8

06 0.6

04 0.4

02 0.2

-0 '
-1 -08 -06 -04 -02 0 02 04 06 08

CO)
14

12

08

06

04

02

Figure 9 Signature of boundary regions(changing from interior to boundary)

259