T

Ab

Xa

b

c

a

ARAA

KABC3P

1

acrsmpiciwre

udi(A

U

1h

Journal of Manufacturing Processes 15 (2013) 432–443

Contents lists available at ScienceDirect

Journal of Manufacturing Processes

j ourna l ho me pa g e: www.elsev ier .com/ locate /manpro

echnical paper

n integrated CNC accumulation system for automaticuilding-around-inserts

uejin Zhaoa,b, Yayue Pana, Chi Zhoua, Yong Chena,∗, Charlie C.L. Wangc

Epstein Department of Industrial and Systems Engineering, University of Southern California, Los Angeles, CA, USASchool of Mechanical Engineering, Shandong University, Ji’nan, Shandong, ChinaDepartment of Mechanical and Automation Engineering, The Chinese University of Hong Kong, China

r t i c l e i n f o

rticle history:eceived 1 May 2013ccepted 31 May 2013vailable online 11 July 2013

eywords:dditive manufacturing

a b s t r a c t

In this paper a non-layer-based additive manufacturing (AM) process named computer numerically con-trolled (CNC) accumulation process is presented for applications such as plastic part repairing andmodification. To facilitate the CNC accumulation process, a novel three-dimensional (3D) laser scanningsystem based on a micro-electro-mechanical system (MEMS) device is developed for in situ scanning ofinserted components. The integration of the scanning system in the CNC accumulation process enablesthe building-around-inserts with little human efforts. A point processing method based on the algebraic

uilding around insertNC accumulationD scanningrocess planning

point set surface (APSS) fitting and layered depth-normal image (LDNI) representation is developed for con-verting the scanning points into triangular meshes. The newly developed 3D scanning system is compactand has sufficient accuracy for the CNC accumulation process. Based on the constructed surface model,data processing operations including multi-axis tool path planning and motion control are also investi-gated. Multiple test cases are performed to illustrate the capability of the integrated CNC accumulationprocess on addressing the requirements of building-around-inserts.

iety o

© 2013 The Soc

. Introduction

Manufacturing processes can be generally classified intodditive-based, subtractive-based, and deformation-based pro-esses [1]. Each process has its own unique capability in addressingequirements such as materials, speed, cost, quantity, accuracy,hape complexity, and flexibility. To take advantage of differentanufacturing processes, it is desired for future manufacturing

rocesses to have the capability of building-around-inserts. Thats, a manufacturing process is capable of incorporating existingomponents that are fabricated by other manufacturing processesn its fabrication process [2,3]. Such capability, if fully realized,

ould enable new product components with heterogeneous mate-ial properties and multi-functionality be developed for futurengineering systems.

Over the past twenty-five years, a collection of additive man-facturing (AM) processes have been developed for the need ofirect digital manufacturing. A unique capability of AM processes

s that they can fabricate parts directly from computer-aided designCAD) models without part-specific tooling or fixtures. CurrentM processes are mostly planar-layer-based, that is, materials

∗ Corresponding author at: McClintock Avenue, GER 201, Los Angeles, CA 90089,SA. Tel.: +1 213 740 7829/3715.

E-mail address: yongchen@usc.edu (Y. Chen).

526-6125/$ – see front matter © 2013 The Society of Manufacturing Engineers. Publishettp://dx.doi.org/10.1016/j.jmapro.2013.05.009

f Manufacturing Engineers. Published by Elsevier Ltd. All rights reserved.

are accumulated based on uniform layer-by-layer deposition ina single building direction. While the planar layers make thealgorithm development, software coding, and hardware imple-mentation simpler, the allowable motions between the tool andthe work piece are limited. In most cases, only translationalmotions in the X, Y, and Z axes are allowed in AM processes, e.g.stereolithography apparatus (SLA) and selective laser sintering (SLS).Consequently, embedding existing components in such processesis usually difficult. For example, Kataria and Rosen [2] identifiedsome major problems related to building-around-inserts in theSLA process including laser shadowing and vat recoating. Both ofthe problems are related to the uniform layer-based fabricationapproach.

Recently a non-layer-based AM process named the ComputerNumerically Controlled (CNC) accumulation process was developed[4]. The tool path of the CNC accumulation process can havemulti-axis motions; consequently, the CNC accumulation processis flexible in incorporating existing components in the buildingprocess.

1.1. Multi-axis CNC accumulation process

Similar to AM processes such as the laser- or DMD-basedSLA processes [5,6], the CNC accumulation process [4] is basedon selectively curing photocurable resin. However, differentfrom the SLA process, the curing tool in the CNC accumulation

d by Elsevier Ltd. All rights reserved.

X. Zhao et al. / Journal of Manufacturing Processes 15 (2013) 432–443 433

ing an

piciopppotaap

mutpsattw

Tcmmp

Fig. 1. A comparison of (a) the CNC machin

rocess is merged in liquid resin. Hence the tool can cure resinn various orientations with unconstrained material feeding. Aritical problem to be addressed in the CNC accumulation processs how to ensure the newly cured resin will attach to the baser previously built part instead of the accumulation tool. In ourrevious work [4], the attaching forces in the CNC accumulationrocess were studied. The feasibility of the CNC accumulationrocess has been demonstrated based on: (1) applying certain typef coating (e.g. Teflon film) on the tip of the curing tool such thathe attaching force between the resin and the tool can be reduced;nd (2) ensuring adequate over curing of liquid resin such that thettaching force between the newly cured resin and the base or thereviously built part will be sufficiently large.

The CNC accumulation process has great similarity to the CNCachining process. As shown in Fig. 1a, the CNC machining process

ses a machining tool to remove the material that is in touch withhe subtracting tool. Hence, for a given work piece (W) and toolath (Si) with tool orientation in each cutter location (Oj), the con-tructed shape (M) will be M = W − U(T)Si+Oj

. In contrast, the CNCccumulation process, as shown in Fig. 1b, uses an accumulationool to add material that is in touch with the curing tool. Hence, forool path (Si) with tool orientation (Oj), the constructed shape (M)ill be M = U(T)Si+Oj

.Two good properties of the CNC accumulation process are: (1)

he allowable motions between a tool and a work piece are signifi-

antly increased; and (2) the feeding of liquid resin to facilitate theaterial accumulation process is straightforward since an accu-ulation tool is merged inside liquid resin during the building

rocess.

Fig. 2. A test case of building-around-inserts

d (b) the CNC accumulation processes [4].

1.2. Automatic building-around-inserts

The feasibility of using the CNC accumulation process in applica-tions such as building-around-inserts has been demonstrated [4].A critical challenge identified in our previous study is how to gen-erate tool paths for an inserted component whose digital modeland relative position are unknown. Similar to the CNC machiningprocess, the collision between a tool and a workpiece needs to beidentified and avoided in the tool path planning of the CNC accu-mulation process. In addition, an appropriate gap between the tooltip and the base surface is required in the fabrication process toensure the newly cured resin will attach to the base surface [4].

Fig. 2 shows a test case based on a gear with a broken tooth.Suppose the plastic gear is fabricated by the injection molding orother AM processes. During its usage, one of the teeth is broken.Instead of throwing the part away, the broken gear can be repairedusing the CNC accumulation process. However, the tool path plan-ning for adding the desired tooth can be challenging since (1) thephysical object to be repaired may be positioned arbitrarily on theplatform; (2) the surface of the broken tooth, which will be used asthe base surface for the missing tooth, can be irregular.

Hence, a critical challenge in using the CNC accumulationprocess in building-around-inserts is to automate the tool pathplanning for any given objects to be built on. If successfullyaddressed, the capability of remanufacturing parts will not only

reduce the material and energy waste, but also expand thefunctionality of product components. Additive manufacturingprocesses such as Tungsten Inert Gas (TIG) welding and LaserEngineered Net Shaping (LENS) have been developed before for

using the CNC accumulation processes.

4 cturing Processes 15 (2013) 432–443

risWa

1

btmp3iaccaf

2

cgspsssb

ntnciptcr

nitoiutnppgjo

lSdsp

e(

34 X. Zhao et al. / Journal of Manufa

epairing and modifying metallic parts; however, automatic repair-ng and modification of plastic parts are still lacking. Such productustainability issues attract increasing attentions in recent years.

e are motivated to develop the CNC accumulation process inddressing such needs.

.3. Contributions

To increase the capability of the CNC accumulation process inuilding-around-inserts, an in situ scanning system that can cap-ure the shape of inserted objects and related geometric processing

ethods have been developed for the multi-axis CNC accumulationrocess. Test results indicate that the accuracy of the developedD scanner is satisfactory. A set of geometric processing methods,

ncluding the APSS point processing, LDNI-based surface retrievingnd offsetting, tool path planning, and 5-axis accumulation motionontrol, have also been developed. Experimental results verify theapability of the integrated CNC accumulation system in building-round-inserts. The process may offer an effective and efficientabrication process for repairing and modifying plastic parts.

. An integrated CNC accumulation system

Various 3D scanning systems have been developed before andommercially available. Although such scanners can be used inenerating the digital model of an inserted object, such an ex situcanning approach will require additional effort on registering theosition and orientation of the object inside the tank. In compari-on, an in situ 3D scanner that is integrated in the CNC accumulationystem can directly capture the geometric information of a baseurface related to accumulation tools. Hence the tool paths foruilding given shapes on the base surface can be quickly planned.

A desired in situ 3D scanner needs to be compact on size. It alsoeeds to be flexible on accommodating the position and orienta-ion of a mounted component in the tank. In addition, the scannereeds to have sufficient accuracy for the CNC accumulation pro-ess. The scanning process is desired to require little or no humannvolvement such as adding markers, or constantly registering theosition of the scanner related to the tank. Note that, in order to usehe CNC accumulation process in building-around-inserts, only theoncerned surface model needs to be captured; there is no need toeconstruct the 3D model of the whole object.

Current 3D scanning systems can be classified into contact andon-contact systems. The contact 3D scanners use a probe to phys-

cally touch the surface of an object. In addition to a firm fixtureo hold the object, a motion system is needed to translate andrientate the touch probe related to the target surface. In compar-son, non-contact 3D scanners emit radiation or light beams, andse detectors to receive the signals reflected from the object [7]. Ariangulation method is one of the most commonly used 3D scan-ing approaches [8]. In the triangulation method, a prepared lightattern is projected onto a surface in order to capture the pointositions on the surface. The light pattern can be a dots, lines, orrids. A camera or light detector is then used to capture the pro-ected light. Consequently, the 3D coordinates of the projected lightn the surface can be calculated by a triangulation model [8].

Fig. 3 shows a typical 3D scanning head based on the triangu-ation method. Both the laser and sensor are fixed inside the head.ome scanners have a rotating hexagon mirror to convert a laserot into a laser line [9]. However, the limited freedom of the lasercanning path requires the use of an external motion system to

osition the scanning head in order to scan a surface [9–12].

Micro-electro-mechanical systems (MEMS) is a rapidly growing,merging technology. Products such as direct digital micromirrorDMD) devices have been used in AM processes [6,13]. Recently a

Fig. 3. An illustration of the triangulation method and a scanning head with a fixedlaser and sensor.

dual-axis analog MEMS pointing mirror device (TALP1000B), hasbeen developed by Texas Instruments (Dallas, TX) [14]. The goldcoated mirror (∼3.2 mm × 3.6 mm) of the MEMS device is con-structed of single crystal silicon. It can be driven with precisepointing resolution. In addition, each rotation axis of the deviceis individually and independently controlled.

Such a compact dual-axis rotation mirror enables a laser dot toscan in a certain range. Hence it is possible to cover a surface byusing the controlled movement of such a micromirror instead oftranslating the scanning head that is much larger and more bulky.In this paper, a 3D scanner based on the MEMS technology is devel-oped. An in situ scanning system based on the MEMS device iscompact, cheap and easy to be integrated. A base surface may bescanned with a single setup to reduce the requirements of motionplanning or data registering between multiple setups.

In the following sections, the development of a 3D laser scan-ning system based on the dual-axis MEMS pointing mirror will bediscussed. The schematic design, prototype development and scan-ning point retrieval algorithms are presented. The integration ofthe developed 3D scanner in a CNC accumulation system is alsodiscussed. Based on the dual-axis mirror, the developed scannerreturns scanning data of a surface portion where material needs tobe added. For the scanning result that is usually sparse, the alge-braic point set surface (APSS) algorithm is adopted in processingthe scanning points for constructing related surface models. Auto-matic tool path generation based on the scanning results is alsopresented. Fig. 4 shows the data flow pipeline of an integrated CNCaccumulation system for building-around-inserts.

The remainder of the paper is organized as follows. Section 3presents the in situ laser scanner based on the MEMS device. Section4 describes the point processing algorithms for constructing thedigital model of a scanning surface. Section 5 discusses the 5-axistool path planning operations based on the scanning data. Two testcases are presented and analyzed in Section 6. Finally, conclusionsare drawn in Section 7.

3. Development of an in situ 3D scanning system

A dual-axis analog pointing mirror device (TALP1000B fromTexas Instruments) has been identified as a potential solution for

an in situ 3D scanner for the CNC accumulation process. With theuse of a dual-axis mirror, a laser dot can scan along both the X and Ydirections by tilting the mirror along the two axes (NS and WE). Dif-ferent angle combinations of the mirror can reflect the laser dot to

X. Zhao et al. / Journal of Manufacturing Processes 15 (2013) 432–443 435

Fig. 4. A pipeline of an integrated CNC accumulation system for automatic building-a

ad(m

3

dtotts(dd

round-inserts.

ny position within a pre-defined scanning area. The moving free-om extended from a 1-dimensional (1D) line to a 2-dimensional2D) area enables the scanning with no or little external

ovement.

.1. Design of an integrated 3D scanning unit

Fig. 5 shows the schematic of a 3D scanning unit based on theual-axis rotation mirror. A dot laser shoots a laser beam towardhe mirror. The laser dot is then reflected by the mirror and shootsn the surface of a scanning object. By rotating the mirror alonghe NS and WE axes, the laser dot can scan any position withinhe scanning area. The scanning area that can be covered in one

etup is dependent on the maximum rotation angles of the mirror±5◦). An appropriate scanning depth is determined by a calibrationatabase. A camera is installed with the mirror to detect the laserot reflected by the part surface. Based on the planned laser line

Fig. 6. The setup of the 3D scanning unit with the dual-a

Fig. 5. A schematic illustration of the dual-axis mirror based scanning unit.

and the captured image data, the 3D position of the intersectionpoint can be calculated based on the aforementioned triangulationmethod.

Fig. 6 shows a prototype laser scanner that is integrated intoa 5-axis CNC accumulation system. In the 3D laser scanning unit,(1) a red dot laser by LaserGlow Technologies (Toronto, Canada)is selected based on the wavelength range of the coated mirror(700 nm to 10 �m). Its brightness and focus range can be adjusted.(2) The TALP1000B is a high-performance MEMS device designedfor light steering applications. The rotation range of the two-axismirror is ±5 degrees and the mirror size is about 9 mm2. The switch-

ing time of the mirror is less than 5 ms. The device provides positionfeedback with 13-bit resolution. (3) A small web camera with animage resolution of 640 × 480 is used as the light sensor.

xis mirror in the 5-axis CNC accumulation system.

4 cturing Processes 15 (2013) 432–443

eoirtTc

fisttatt

idpsX

obipa2ctr

3

ebctFwori

iwptw

ktccecaa

(

Scab

36 X. Zhao et al. / Journal of Manufa

The dot laser and the dual-axis mirror are installed on the twonds of an aluminum bar. The laser dot reflected by the mirror fallsn the platform, which is covered by a chessboard paper. Cameras installed side away from the mirror to capture the image of theeflected laser dot. Other than the laser scanning unit, there arehree translation stages X, Y and Z, and two rotation stages A and B.hese five stages are designed for the 5-axis accumulation tool. Aontrol board is used to control the 5-axis movements.

Based on such an integrated system, a given workpiece is firstxed in a resin tank that is placed on the platform. The lasercanning unit is then used to collect the scanning data relatedo the deposition surface. With the scanning data, tool paths forhe accumulation process are computed using the point processinglgorithm and the tool path generation operations. The accumula-ion tool, driven by the 5-axis motion stage, then adds material onhe inserted workpiece.

The developed scanning unit can be mounted on a separate flex-ble arm. The unit can also be fixed on the motion system that isesigned for the CNC accumulation tool. Consequently, the work-iece can be moved up and down using the Z stage to enlarge thecanning volume in the Z direction, and moved sideways using the

and Y stages to enlarge the lateral scanning area.For a fixed setup, as shown in Fig. 5, the scanning area depends

n the rotation angle of the dual-axis mirror, and the distanceetween mirror and the scanning plane. In addition, the camera

mage range needs to be larger than the laser scanning area. In therototype system as shown in Fig. 6, the defined scanning range has

Z height of 25.4 mm. On the top plane where the Z value is equal to5.4 mm, the scanning area is set at 25 mm × 25 mm to ensure suffi-ient accuracy. When a surface is outside the scanning range, eitherhe scanner or the workpiece needs to be moved. Accordingly, theelated transforming matrix can be applied to the scanning result.

.2. Calibration of 2D camera

Camera calibration is required for identifying the intrinsic andxtrinsic parameters of a CCD camera. Accordingly, the relationetween the captured 2D image and the 3D position of a point Pan be established. The camera calibration is a time-consumingask despite the availability of some camera calibration toolbox.or example, Zhang [15] introduced a camera calibration method,hich is used in the OpenCV image processing library and many

ther toolboxes. However, the calibrated camera parameters andelated computation models cannot fully capture the nonlinearityn a working volume.

To capture such nonlinearity for improved accuracy, the work-ng volume (refer to Fig. 5) is subdivided into 11 layers in our study

ith a distance of 2.54 mm between two neighboring layers. Arinted chessboard is fixed on the platform, which can be moved inhe Z axis. Accordingly the 11 layers from z = 0 mm to z = 25.4 mmith �z = 2.54 mm can be calibrated separately.

For a layer K at height zw, all the corners on the chessboard havenown coordinates (xw, yw, zw). An image of the chessboard is cap-ured. Based on it, the image coordinates (Xi, Yi) of all the cornersan be identified. Accordingly the relationship between an imageoordinate of each layer and the related world coordinate can bestablished as: (Xi, Yi)Layer K = P(xw, yw, zw). By repeating the pro-ess, a database between the 3D points within the working volumend the image points that are captured by the camera can be builts:

Xi, Yi)Layer K = PK (xw, yw, K × �z), 0 ≤ K ≤ 10. (1)

uch relations combine both intrinsic and extrinsic matrixes of theamera. Since the layer distance is small, a bilinear interpolationpproach is used to approximate the world coordinates (x, y, z)ased on two neighboring calibrated planes PK(x, y, K × �z) and

Fig. 7. Laser calibration using the top plane (layer 10) and the bottom plane (layer0).

PK+1(x, y, (K + 1) × �z), when z ∈ [K × �z, (K + 1) × �z], and the fourcorners of a checker box.

3.3. Calibration of laser scanning lines

The rotation angles of the mirror are controlled by the out-put values (NS and WE) provided by two Digital-Analog-Converters(DAC) in a control board. To eliminate the DAC output noises andother hardware construction errors, the laser line positions for dif-ferent signal values also need to be calibrated.

Different from the camera calibration, the reflected laser line isstraight from the top to bottom planes. Consequently, a signal value(NS, WE) can be calibrated by the laser dots on the top and bottomplanes that define a laser line. As shown in Fig. 7, suppose (xi 10,yi 10) and (xi 0, yi 0) are the image position of a laser dot on the topand bottom planes. The world positions of the laser points (Xw 10,Yw 10, Zw 10) and (Xw 0, Yw 0, Zw 0) can uniquely define the laser lineinside the scanning volume.

A database consists of the positions of all the laser lines and therelated DAC signal values can be built as: f (NS, WE) = (Xw 0, Xw 10,Yw 0, Yw 10, 0, 10 × �z), where NS and WE are the DAC numbersof the two axes of the mirror. (Xw 0, Yw 0) and (Xw 10, Yw 10) arethe related world positions of the laser line on the bottom and topplanes, respectively. Since a related laser line is straight, an inter-section point at a level Z can be computed by the linear interpolationbetween two related points L0, and L10.

3.4. Computing 3D points

In the scanning process, the dual-axis mirror is rotated basedon a given DAC number (NS, WE). The laser is first turn off tocapture an image of the scanned object (without the laser dot).The laser is then turned on for capturing an image with the laserdot. Consequently, an image with the laser dot can be obtainedby the subtraction between the two images. Based on the calibra-tion methods the 3D computation is modified from the standardtriangulation process.

Let (Xi, Yi) denotes the image position of the laser dot. Accord-ingly, a set of world coordinate points PK(xw, yw, zw) related to (Xi,Yi) for each layer K × �z can be identified from the camera cali-bration database (refer to Eq. (1)). Ten camera line segments PiPi+1

X. Zhao et al. / Journal of Manufacturin

Fac

ciitstlbpwsb

3

a

Fa

TS

ig. 8. Point retrieval by calculating the closest point between a laser line (L0–L10)nd camera line segments (PiPi+1). Laser points L0 and L10 are calculated by (ns, we);amera points P0–P10 are calculated by the image position of the laser dot (Xi , Yi).

an be computed by linear interpolation. Similarly, the correspond-ng laser line related to the current DAC number (NS, WE) can bedentified through the laser point calibration database. Accordingly,he laser points LiLi+1 can be calculated by linear interpolation. Ashown in Fig. 8, the eleven layers divide the working volume intoen regions. Each region contains a camera line segment and a laserine segment. The closest point between the two line segments cane calculated in each region. For example, X0 and X10 are the closestoint in regions layer 0–1 and layer 9–10, respectively. The pointith the smallest distance is selected as the object point in the 3D

pace. For example, as shown in Fig. 8, X4 in region layer 4–5 woulde the computed 3D position.

.5. 3D scanning verification

The aforementioned calibration databases for both the camerand the laser have been built. Scanning tests based on them are

ig. 9. Scanning results of the top surface of a mechanical part. (a) Distribution of scannill the scanning points on the Z axis.

able 1tatistical analysis of scanning tests.

Test Median (mm) AVG (mm)

1. Platform (z = 0) 0.202 0.203

2. Platform (z = 2.54) 2.814 2.816

3. A part (z = 3.9) 4.148 4.146

4. Platform (z = 5.08) 5.080 5.072

g Processes 15 (2013) 432–443 437

performed to verify the accuracy of the developed scanning system.In the first test, the platform was raised by the z-stage to differentZ heights (0.0 mm, 2.54 mm, and 5.08 mm). The planar surface isscanned and reconstructed based on the calibration databases. Inthe second test, a mechanical part with a height of 3.8 mm waspositioned at a height of 3.9 mm. Its top surface is a planar planepainted in black. The scanning results of its top surface are recorded.Fig. 9a is the computed scanning points related to the average plane.Fig. 9b shows the deviation of all the points in the Z axis.

Table 1 presents the statistics of the scanning results, the com-puted point Z heights, in the verification study. The average heightdeviations of the computed 3D points in the four test cases are lessthan 0.3 mm. The scanning errors may come from the sources suchas the resolution of the webcam, the control resolutions of the dual-axis mirror, and the errors in the line segment approximation andthe laser dot computation. They are discussed in more details asfollows.

(1) A low-cost webcam is used in our prototype system. The cam-era has a limited resolution (640 × 480 pixels). In our setup, it isabout 0.18 mm/pixel. In addition, the brightest image positionof the laser dot is used to represent the laser dot. Both cam-era resolution and the image point identification may introduceerrors in the final scanning result.

(2) Two digital analog convertors (DAC) are used in controlling thevoltage and current of the coils that drive the dual-axis mirror.The resolution of the DAC in our control board limits the rota-tion resolution of the mirror. In addition, there may be vibrationand other mechanical noises in the hardware setup.

(3) The laser dot used in our system has a certain focus distance.The spot size slightly changes from the top to bottom planes.In addition, the linear interpolation between two neighboringcamera points may also introduce errors.

Approaches, such as adopting a better camera, using more

layers in building the calibration database, and computing 3Dpoints based on non-linear interpolation, can be developed tofurther reduce the scanning error. Nevertheless, the experimentalresults show that the scanning error of the developed 3D scanner

ng points–all the points under the average Z plane are hidden and (b) deviation of

Max (mm) Min (mm) VAR (mm)

0.487 0.0 0.00753.319 2.515 0.0144.393 3.844 0.0075.080 4.934 0.0005

438 X. Zhao et al. / Journal of Manufacturing Processes 15 (2013) 432–443

F at poio s the p( d, the

ifg

4

satlnsdasat

srscpnmdipppfd

4

mne

p

wp

o

In degenerate cases, V corresponds to the coefficients of a planeequation with v0 representing the plane’s distance from the originand [v1, . . ., vd]T being its normal. With normal constraints, V(x) canbe computed by the following equations [16]:

ig. 10. Overview of the point processing framework. (a) Estimation of the normalsf a LDNI point x onto the APSS: blue point denotes query point x, red point denoteshown in red line). (For interpretation of the references to color in this figure legen

s generally less than 0.5 mm in the Z axis. Such accuracy is satis-actory for the CNC accumulation process. More scanning tests areiven in Section 6.

. Processing of scanning points

The scanning results obtained by the point-by-point based 3Dcanner have the following properties. (1) The sampling pointsre relatively sparse compared to other scanning systems such ashe structured light based systems. Instead of generating samp-ing points that are dense with various noises and outliers, a smallumber of key points on the surface are sampled for surface recon-truction. (2) The point clouds of a surface may have unevenensities in some areas. (3) The scanning result can be a portion of

surface. (4) The scanning results may be distorted due to mea-urement noises. (5) A base surface to be scanned for the CNCccumulation process is usually continuous since the accumulationool path is usually a set of continuous curves.

Hence a 3D surface reconstruction method that can processparse sampling points of a continuous surface is required in theelated point processing software system. The algebraic point seturfaces (APSS) algorithm proposed by Gross et al. [16] is a goodhoice in handling high curvatures and sparse point cloud. In thisaper, we extended the APSS algorithm and used the layered depth-ormal image (LDNI) representation [17] in generating triangulareshes of the scanned surface. Fig. 10 shows the framework of the

eveloped point processing method. The scanning data comput-ng process is divided into the following steps: (1) preprocessingoints including denoising, normal estimation, and orientationropagation; (2) constructing an implicit surface based on the APSSrojection; and (3) generating sampling points of the implicit sur-ace for constructing triangular meshes. Each step is described inetails as follows.

.1. Preprocessing scanning points

Normal estimation and orientation play an important role in theesh surface reconstruction. For a given set of sampling points,

ormal is chosen as the eigen-vector corresponding to the smallestigen-value. It can be computed by the co-variant matrix of points:

∑i ∈ N′(p)

(pi − P̄)(pi − P̄)T. (2)

here N′(p) is the point set of the k nearest neighbors of a queryoint p.

However, the eigen-vector analysis cannot guarantee a correctrientation. Re-orienting the normal vectors is necessary based on a

nt pi; (b) propagation of a consistent normal orientation from pi to pj; (c) projectionrojection of x, and the circle center is c; (d) an illustration of the projection surface

reader is referred to the web version of the article.)

direct flipping strategy. The orientation of pi is flipped by the closestpoint:

ni = −ni, if ncpi · ni < 0 (3)

4.2. Algebraic point set surface (APSS) algorithm

Algebraic point set surfaces algorithm is a newly developedpoint set surface definition based on moving least squares (MLS)fitting of algebraic spheres. The surface representation can beexpressed by either a projection procedure or in implicit form.Compared to existing planar MLS, the key advantages of the APSSapproach include significantly improved stability of the projectionunder low sampling rates and in the presence of high curvature. Themethod can approximate or interpolate the input point set and nat-urally handle planar point clouds. The following weighting functionis used in the fitting process [16]:

wi(x) = f( ||pi − x||

hi(x)

), hi(x) = K · D̄,

f (x) ={

(1 − x2)4; x < 1

0; otherwise.

where pi is a point in the input point set and x is a query point thatis near the surface; wi(x) represents the weight of the point pi forthe query point x; hi(x) is defined by a constant factor K timing theaverage distance D̄ of pi’s k nearest neighbors. The weighting func-tion f is a smooth and decreasing function. The weighting functionproposed in [16] is used in our study. Fig. 11 shows the LDNI-basedAPSS projection algorithm.

An algebraic sphere is defined as the 0-isosurface of the scalarfield Sv(x) = [1, xT , xT x] · V , where V = [v0, . . . , vd+1]T is the vec-tor of scalar coefficients to describe the sphere. For vd+1 /= 0, thecorresponding center O and radius R can easily be computed [16]:

O = − 12vd+1

[v1, . . . , vd]T and R =√

OT O − v0

vd+1; (4)

with d being the dimension.

V(x) = C−1(x)b̂(x), (5)

C(x) = DT W(x)D; b̂(x) = DT W(x)b. (6)

X. Zhao et al. / Journal of Manufacturin

Fig. 11. Framework of the APSS fitting using the LDNI array points.

wtc

ˇ

ioic

x

x

W(x) =

⎡⎢⎢⎢⎢⎢⎢⎣

wi · · · 0

0 ˇwi

...

.... . .

...

0 0 ˇwi

⎤⎥⎥⎥⎥⎥⎥⎦

;

D =

⎛⎜⎜⎜⎜⎝

1 pTi

pTipi

0 eT0 2eT

0pi

......

...

0 eTd−1 2eT

d−1pi

⎞⎟⎟⎟⎟⎠ ; b =

⎡⎢⎢⎢⎢⎣

0

eT0ni

...

eTd−1ni

⎤⎥⎥⎥⎥⎦ ;

here {ek} represent the unit basis vectors of the coordinate sys-em. Scalar ̌ should be carefully chosen to weight the normalonstraints. In our study, it is calculated based on:

= 106 ×∑

iwi(x)hi(x)∑iwi(x)

The LDNI representation [17] is selected for converting themplicitly defined APSS surface into triangular meshes. An “almost”rthogonal projection [18] is used in computing LDNI points. Thats, the coordinate of a query point used in the LDNI projection pro-edure is calculated as follows:

′0 = min Ext(0) + Ix × �x; (7)

′1 = min Ext(1) + Iy × �y; (8)

g Processes 15 (2013) 432–443 439

x′2 = min Ext(2) + i × �z;

�z = ceil(max Ext(2) − min Ext(2))s − 1

(9)

�n = �x(−�O)r

(10)

where �x and �y denote resolutions in the X and Y directions,respectively. min Ext(0), min Ext(1), and min Ext(2) represent theminimum value of the X, Y, and Z coordinates of the input point set.max Ext(2) represents the maximum value of the Z coordinate of theinput point set. Ix, Iy are the index of the X and Y axis, respectively.S is the number of points per ray, and i is an integer between 0 andS.

5. CNC accumulation tool path planning

Tool path planning is one of the most critical steps in the machin-ing processes such as CNC milling. Extensive research has beenperformed in the area with various tool path planning algorithmsdeveloped. Many commercial software systems such as Master-CAM, Unigraphics, GibbsCAM, and Powermill, are available for theCNC machining tool path planning. In comparison, less study onmulti-axis tool path planning for additive manufacturing processesis found. Previous efforts (e.g. [19–21]) are mainly developed for themetal-based AM processes such as the Laser-aided Metal Deposi-tion and Laser Engineered Net Shaping processes.

For the multi-axis CNC accumulation process, the tool path plan-ning is equally important. Similarity between the CNC machiningand accumulation processes exists on the tool path planning. (1)Both processes require one or a set of tools to manipulate materi-als. The subtractive process removes materials using cutters whilethe accumulation process adds materials using curing tools. (2) Thetools in both processes move along a predefined trajectory (i.e. toolpath) for forming desired geometry. (3) The incremental distance ofthe tool path is decided by the tool size (i.e. the diameter of the tooltip for the subtractive process, and the LED spot size for the accu-mulation process). (4) Both processes have a fabrication depth. Thecutting depth of the subtractive process is determined by factorssuch as the cutter shape, material properties, surface quality andbuilding speed. In comparison, the curing depth of the accumula-tion process is determined by factors such as the resin properties,the UV light source, and the bonding force between the tool and thebase surface. (5) The planned tool paths in both processes shouldavoid collisions with the workpiece.

Based on the scanned surface and the CAD model of to-be-addedgeometry, a general tool path planning method for the 5-axis CNCaccumulation processes has been developed. The method consistsof the following major steps. (1) Thicken the scanned surface byusing an offsetting operation with an offset distance determined bythe curing depth of the tool. A general offsetting method for an inputpolygonal model and an arbitrary offset distance is presented in[22] and used in our study. (2) Subtract the target geometry and theoffset model to compute the shape of a thin layer. Robust Booleanoperations based on the layered depth-normal image (LDNI) aredescribed in [17] and used in our study. (3) Generate a set of con-tours to fill the computed layer shape with an interval distancedefined by the curing spot size. (4) Use an off-the-shelf CNC machin-ing software system to generate numeric control G-code based onthe planned tool path. (5) Use a CNC accumulation controller to loadthe generated G-code, and accordingly send motion commands to

a controller to drive the 5-axis motion. At the same time, a micro-controller is used to control the curing state (i.e. turning the LEDON/OFF). Fig. 12 illustrates the aforementioned steps. Two moreexamples are presented in Section 6.

440 X. Zhao et al. / Journal of Manufacturing Processes 15 (2013) 432–443

eratio

6

tTi

6

Titatr

ou

F

Fig. 12. Tool path gen

. Experimental results

Based on the planned tool paths, accumulation tools can selec-ively cure liquid resin into desired shape on the scanned surface.wo test cases were designed to verify the effectiveness of thentegrated CNC accumulation system in building-around-inserts.

.1. Experimental setup

The hardware setup of our prototype system is shown in Fig. 6.he 5-axis motion configuration of an accumulation tool is shownn Fig. 13. A light guide is used to provide UV light from a LED tohe tip of the fiber optics. A Teflon film is applied on the tool tips the coating media [4]. A 5-axis motion system including X, Y, Zranslations and A, B rotations is designed to allow the tool to cure

esin in various positions and orientations.

During the building process, an inserted component is mountedn a resin tank that is placed on the platform. The surface to besed as the base is first scanned by the developed 3D scanner. The

ig. 13. Multi-axis motion configuration of the CNC accumulation system. (a) A schemati

n and CNC controller.

tank is then filled with UV curable resin. The building process canstart after the tool paths have been planned. After the accumulationprocess finishes, the built object is taken out for cleaning.

6.2. Repairing a broken gear

A test case on repairing an existing plastic part is shown inFigs. 14 and 15. In the test, a gear was built using a commercialSLA machine. We manually break one tooth of the gear (refer toFig. 14a). Note that the broken surface on which material will bedeposited is irregular. With the integrated CNC accumulation sys-tem, it is possible to repair the broken tooth.

First, the working surface, on which a new tooth will be built,is scanned. As shown in Fig. 14b, the scanning point set is sparsewith noises. The scanning points are processed using the point

processing algorithms as described in Section 4. The computed LDNIpoint set after the point processing step is shown in Fig. 14c. Finallya triangular mesh as shown in Fig. 14d is constructed from the LDNImodel.

c drawing and (b) an illustration of an accumulation tool used in the system.

X. Zhao et al. / Journal of Manufacturing Processes 15 (2013) 432–443 441

F art; (bA the L

CpAmatpFpwtC

Fo

ig. 14. Surface reconstruction for the CNC accumulation process. (a) A broken pPSS-LDNI fitting of the scanning points; (d) the base surface model generated from

Assuming the scanned surface has been aligned with the targetAD model (refer to Fig. 15a). A set of offsetting operations can beerformed based on the scanned surface and a given curing depth.ccordingly multiple layers can be computed based on the offsetodels and the target tooth model. As shown in Fig. 15b, there

re 15 layers to be built in order to repair the missed tooth. Note allhe layers have a uniform thickness. For each layer, continuous toolaths can be planned based on the layer model as shown in Fig. 14d.ig. 15c shows the tool path of one layer in the CNC accumulation

rocess. Note that the planned tool path is not on a planar plane,hich requires 5-axis motions. Based on the planned tool paths,

he repairing work of the broken gear can be performed by theNC accumulation process. Fig. 15d shows the repaired gear. For

ig. 15. Fabrication process based on a scanned surface. (a) Alignment of the surface withf a layer; (d) the repaired gear.

) scanning points of the broken surface; (c) generated LDNI points based on theDNI points.

demonstration purpose, a different resin is used in the CNC accu-mulation process. If the same type of resin was used, the repairedtooth would be identical to other teeth.

6.3. Modifying a curved surface

As discussed in Section 1, each manufacturing process has itsown advantages and disadvantages. For example, complex geome-tries such as surface textures and characters are difficult to be

fabricated by forming processes such as injection molding andextrusion. Such complex features will significantly increase thetooling cost in the forming processes. Assume a teapot with no sur-face textures or characters is fabricated by the injection molding

the gear model; (b) offsetting results for 15 accumulation layers; (c) the tool path

442 X. Zhao et al. / Journal of Manufacturing Processes 15 (2013) 432–443

n exis

pbepa3aup

e

F

Fig. 16. Scanning and post processing of the scanned surface. (a) A

rocess (refer to Fig. 16a). In the test, the model was still fabricatedy a SLA machine due to the lack of injection molding tools. How-ver, such simple shape can easily be fabricated by the formingrocess in a much shorter time and at a much lower cost. Furtherssume the final design needs to have complex features such asD textures, which are difficult to be injection molded. In order todd such features, the integrated CNC accumulation system can be

sed in modifying an existing part that is fabricated by the formingrocesses.

Suppose some characters such as “USC” are required on thexterior surface of the teapot. Similar to the previous test case,

ig. 17. The tool path and fabricated part. (a) The designed model; (b) a sliced layer; (c) th

ting teapot; (b) scanned points; (c) LDNI model; (d) mesh surface.

the portion of the surface where the new feature will be addedis scanned. The scanning result is shown in Fig. 16b. The scanningpoints are then processed by the point processing algorithms asdescribed in Section 4. A much smoother and denser LDNI point setis generated as shown in Fig. 16c. A mesh surface can be constructedfrom the LDNI points (refer to Fig. 16d).

Assume the CAD model of “USC” pattern and its position on

the scanned surface are known (refer to Fig. 17a). Accordinglyrelated layers of the “USC” pattern can be computed by offset-ting the pattern surface (refer to Fig. 17b). The corresponding toolpaths for each layer can be generated as shown in Fig. 17c. The

e tool path of a layer; (d) the fabricated object by the two manufacturing processes.

cturin

tseamts

7

fC3trfabmsitbafe

cishp

R

[

[

[

[

[

[

[

[

[

[

[

X. Zhao et al. / Journal of Manufa

eapot modified by the integrated CNC accumulation system ishown in Fig. 17d. Note that, in such a component, two differ-nt manufacturing processes have been used in its fabrication. Inddition, each process utilizes its unique properties: the injectionolding processes is fast and low-cost; while the CNC accumula-

ion process is capable of fabricating complex shapes on arbitraryurfaces.

. Conclusion and future work

Building-around-inserts is an important research issue foruture direct digital manufacturing development. An integratedNC accumulation process has been developed based on a novelD laser scanning system that uses a dual-axis mirror. Comparedo conventional laser scanning systems, the use of the MEMS deviceeduces the size and increases the scanning freedom of the systemrom one-dimensional line segment to a two-dimensional surfacerea. Related point processing algorithms have also been developedased on the APSS fitting and the LDNI representation. Experi-ents verify the accuracy and compactness of the 3D scanning

ystem. Accordingly, a 5-axis CNC accumulation system with then situ 3D scanner has been developed. Experimental results illus-rate the capability of the integrated CNC accumulation process onuilding-around-inserts. Its applications on plastic part repairingnd modification have been demonstrated, which may provide aabrication option for future sustainable manufacturing and het-rogeneous component design.

Considerable work remains to mature the developed AM pro-ess and the related prototype system. Some current work we arenvestigating include exploring new applications of the compact 3Dcanning system, developing a set of CNC accumulation tools withigher resolutions, and extending the multi-axis CNC accumulationrocess into micro-scale fabrication.

eferences

[1] Kalpakjian S, Schmid SR. Manufacturing engineering and technologies. 6th ed.Upper Saddle River, New Jersey: Pearson Prentice Hall; 2010.

[

[

g Processes 15 (2013) 432–443 443

[2] Kataria A, Rosen DW. Building around inserts: methods for fabricating complexdevices in stereolithography. Rapid Prototyping Journal 2010;7:253–62.

[3] Campbell I, Bourell D, Gibson I. Additive manufacturing: rapid prototypingcomes of age. Rapid Prototyping Journal 2012;18:255–8.

[4] Chen Y, Zhou C, Lao J. A layerless additive manufacturing process based on CNCaccumulation. Rapid Prototyping Journal 2011;17:218–27.

[5] Jocobs FP. Rapid prototyping and manufacturing: fundamentals of stereolitho-graphy. Dearborn, MI: Society of Manufacturing Engineers; 1992.

[6] Pan Y, Zhou C, Chen Y. A fast mask projection stereolithography process forfabricating digital models in minutes. ASME Journal of Manufacturing Scienceand Engineering 2012;134:051011.

[7] Chen F, Brown G, Song M. Overview of three-dimensional shape measurementusing optical methods. Optical Engineering 2000;39:10.

[8] Forrest AK. Laser scanners. In: Batchelor B, editor. Machine vision handbook.London, United Kingdom: Springer; 2012. p. 329–54.

[9] Golnabi H. Design and operation of a laser scanning system. Optics and LaserTechnology 2000:267–72.

10] Abuazza A, Brabazon D, El-Baradie MA. Analysis of surface defects using a noveldeveloped fiber-optics laser scanning system. Journal of Materials ProcessingTechnology 2003;143:875–9.

11] Li Y. Beam deflection and scanning by two-mirror and two-axis systemsof different architectures: a unified approach. Applied Optics 2008;47:5976–85.

12] Golnabi H. Image evaluation for the synchronised laser scanning systems.Optics and Laser Technology 1999;31:225–32.

13] Zhou C, Chen Y. Additive manufacturing based on optimized mask video pro-jection for improved accuracy and resolution. SME Journal of ManufacturingProcesses 2012;14:107–18.

14] Texas Instruments Inc. http://www.ti.com/product/talp1000b (retrieved15.11.12).

15] Zhang Z. Flexible camera calibration by viewing a plane from unknown ori-entations. In: International conference on computer vision (ICCV’99). 1999. p.666–73.

16] Gross M, Guennebaud G, Zurich E. Algebraic point set surfaces. ACM Transac-tions on Graphics 2007.

17] Wang CL, Leung Y, Chen Y. Solid modeling by layered depth-normal images onthe GPU. Computer-Aided Design 2010;42:535–44.

18] Alexa M, Adamson A. On normals and projection operators for surfaces definedby point sets. In: In Proceedings of the Eurographics Symposium on Point-BasedGraphics. 2004. p. 149–56.

19] Sundaram R, Choi J. A slicing procedure for 5-axis laser aided DMD process.ASME Journal of Manufacturing Science and Engineering 2004;126:632.

20] Zhang J. Adaptive slicing for a multi-axis laser aided manufacturing process.ASME Journal of Mechanical Design 2004;126:254.

21] Ruan J, Eiamsa-ard K, Liou FW. Automatic process planning and tool pathgeneration of a multi-axis hybrid manufacturing system. ASME Journal of Man-ufacturing Processes 2005;7:57–68.

22] Chen Y, Wang CL. Uniform offsetting of polygonal model based on layereddepth-normal images. Computer-Aided Design 2011;43:31–46.