Robotics and Computer-Integrated Manufacturing 39 (2016) 32–42

Contents lists available at ScienceDirect

Robotics and Computer-Integrated Manufacturing

http://d0736-58

n CorrE-m

journal homepage: www.elsevier.com/locate/rcim

Regular Article

Bead modelling and implementation of adaptive MAT path in wire andarc additive manufacturing

Donghong Ding, Zengxi Pan n, Dominic Cuiuri, Huijun Li, Stephen van Duin, Nathan LarkinSchool of Mechanical, Materials, and Mechatronics Engineering, Faculty of Engineering and Information Sciences, University of Wollongong, Northfield Ave,Wollongong, NSW 2500, Australia

a r t i c l e i n f o

Article history:Received 23 May 2015Received in revised form16 October 2015Accepted 10 December 2015Available online 19 December 2015

Keywords:Bead modelArc weldingAdditive manufacturingMedial axis transformation (MAT)Path planning

x.doi.org/10.1016/j.rcim.2015.12.00445/& 2015 Elsevier Ltd. All rights reserved.

esponding author.ail address: zengxi@uow.edu.au (Z. Pan).

a b s t r a c t

Wire and arc additive manufacturing (WAAM) is a promising alternative to traditional subtractivemethods for fabricating large aerospace metal components that feature high buy-to-fly ratios. This studyfocuses on the development of an automated manufacturing system in order to free the operator fromintervening in the analysis of the CAD model, planning the deposition path, and then manually settingthe welding process parameters. Firstly, the relationship between single bead geometry and weldingprocess parameters is established through an artificial neural network (ANN) model. Then, the adaptivemedial axis transformation (MAT) algorithm for void-free deposition with high geometrical accuracy isintroduced. The adaptive MAT path is implemented by using the single bead ANN model together with apreviously developed multi-bead overlapping model. Finally, the adaptive MAT path planning strategyand the established bead models are tested through experimental deposition of two metal components.The results show that the developed bead model and adaptive MAT-based path are capable of producingdepositions with high quality (void-free) and geometrical accuracy through automated selection ofprocess variables for the WAAM process.

& 2015 Elsevier Ltd. All rights reserved.

1. Introduction

Interest in additive manufacturing (AM) has expanded drama-tically in recent years due to the numerous advantages that thisprocess provides over traditional manufacturing [1]. Although themajority of recent work in AM has been focused on three-di-mensional printing of polymers, AM techniques for fabricatingmetal alloys have been available for more than a decade [2,3]. Wireand arc additive manufacturing (WAAM) using either the gas metalarc welding process (GMAW) or the gas tungsten arc weldingprocess (GTAW) attracts great interest due to its high depositionrate, environmental friendliness, and cost-competitiveness [4–8].In particular, WAAM becomes a promising alternative to conven-tional subtractive methods for fabricating large aerospace alloycomponents that feature high buy-to-fly ratio [9,10].

Generally, process planning for a WAAM system involves CADmodelling, 3D slicing, 2D path planning, weld bead modelling,weld setting, robot code generation, and post-process machining,as shown in Fig. 1. 3D CAD models are firstly sliced into a set of 2Dlayers. Then the path planning module generates deposition pathsfor each of the sliced layers. After the paths are generated, the

desired bead geometries along the path are determined accord-ingly. Bead modelling, on one hand, controls the path planningvariables. On the other hand, it determines the optimum weldsettings corresponding to the desired bead geometry. Subse-quently, the deposition path together with the selected weldingparameters is transferred into an integrated robot code filethrough the robot code generation module. Finally, a near-netshape deposit is produced automatically by the robotic arc weld-ing system. Post-process machining is usually required to producea finished component with the desired dimensional tolerances.Among these steps, proper 2D path planning and weld parameterselection for WAAM are crucial for achieving defect free depositionwith high quality and high geometrical accuracy [11,12]. Due tothe large deposition size of WAAM as compared to powder-basedAM, bead modelling dominates the main variables in 2D pathplanning and weld parameter selection. This requirement distin-guishes WAAM from powder-based AM that has a higher deposi-tion resolution.

As an important step in process planning, 2D Path planninggenerates tool paths to fill the 2D sliced layers with full densitydeposits as required by the component design. Compared topowder-based AM, path planning in WAAM is more complex be-cause it is constrained by both weld bead geometry and layergeometry. As shown in Fig. 2a, the cross-section profile of a singledeposited weld bead is a parabola-like shape. Therefore, multiple

Fig. 1. Automated process planning for robotic WAAM system.

Fig. 2. (a) Schematic diagram of single bead geometry. (b) Top view of the de-position process of single bead. (c) Multiple bead overlapped deposition. w re-presents the width of single bead; d represents the step-over distance in multiplebeads deposition; n represents the bead number. The lines with arrows pointing tothe right represent the deposition paths and the travel directions.

D. Ding et al. / Robotics and Computer-Integrated Manufacturing 39 (2016) 32–42 33

paths deposition should have a certain amount of overlap betweenthe neighbouring paths so that the overall deposit has a relativelyflat upper surface without deep ridges that are difficult to fill insubsequently deposited layers. As shown in Fig. 2c, the step-overdistance, d, which is defined as the distance between the twoneighbouring paths, must be properly designed to produce partswith good surface smoothness. Moreover, to maintain the geo-metrical accuracy, the number of paths (n) needs to be carefullyselected to match the dimension of the layer at the boundaries.The effective deposition width of n paths, in this study, is nd withthe outside materials for post-process machining as shown inFig. 2c. The deposition process of a single bead can be consideredas a constant-radius disk with a diameter of w (bead width) beingswept along the planned path (green lines) as shown in Fig. 2b.

Weld parameter selection is another important step in processplanning for WAAM, since the cross-sectional geometry of a singlebead is determined by the process parameters, predominantlywire-feed rate and travel speed. Generally, the welding parametersare selected based on experience or from a procedural manual. Anoversized geometry is normally deposited and post-process ma-chining is used to achieve the desired geometrical dimension. Inthis study, however, we focus on achieving near-net deposition tominimise the required machining and waste material. Therefore, itis essential to develop a bead model that relates process variablesto the bead geometry.

Several articles related to welding based additive manufactur-ing can be found in the literature; for example solid cubic mildsteel parts have been built by [13]; mild steel and titanium wallshave been built by [14]; large high-value structures have beenproduced by [15,16], and control of process works have been re-ported by [17,18]. However, most of the research provides onlypreliminary developments and few have considered the practicaldevelopment of highly automated process planning systems toproduce shapes with reasonable geometrical accuracy. This paper,based on previous path planning work [19,20] and the overlappingbead model [21], makes a further step of adaptive MAT pathplanning. To our best knowledge, the introduced adaptive MATpath is the first attempt to build components using varied weldingparameters during the deposition, which is preferred for WAAM toproduce parts with high quality and high geometrical accuracy atthe layer boundaries. In addition, several other aspects of processplanning are also investigated in some detail, including single beadmodelling and weld parameter selection. Moreover, the practicalimplementation of an automated deposition system is reportedthrough case studies.

The rest of paper are organised as following. Section 2 presentsartificial neural network (ANN) models for single bead and ana-lytical overlapping model for multiple beads, to cater for produ-cing thin-walled and thick-walled structures, respectively. Section3 describes the adaptive path planning strategy, followed by im-plementation of the adaptive MAT path planning process in

D. Ding et al. / Robotics and Computer-Integrated Manufacturing 39 (2016) 32–4234

Section 4. Experimental results and discussions of the automatedproduction system are provided in Section 5.

Fig. 3. Weld bead geometry.

Fig. 4. Relative error of bead cross-sectional area prediction using parabola model.

Fig. 5. Experimentally measured bead widths and bead heights of all tests.

2. Bead modelling

Modelling both single weld bead geometry and multi-beadoverlapping model are important for producing metal parts withgood surface smoothness and geometrical accuracy. For the WAAMprocess, moreover, the single bead model determines the controlvariables in 2D path planning.

In this Section, a neural network model for single bead geo-metry is established based on experiments using the equipmentthat will be detailed in Section 5. The model is used to predict thecombination of optimal welding process parameters, namely wire-feed rate and travel speed, which will produce the desired beadgeometry. A multi-bead overlapping model that has been in-troduced in a previous publication will also be briefly reviewedand then incorporated into the methodology, to determine theoptimum spacing between adjacent weld beads.

2.1. Single bead profile

Experiments were carried out for a total of 81 combinations of9 different wire-feed rates (F) and 9 different travel speeds (S), assummarised in Table.1. The details of welding process, wire dia-meter, and material used will be introduced in Section 5. Fig. 3schematically shows a deposited bead and the cross-section of thebead geometry with a width, w and a height, h.

All 81 bead profiles are fitted with a parabola function usingcurve fitting method described in Ref. [21]. The relative error ofarea prediction, E, is defined as the percentage of the area differ-ence between the predicted and the actual bead area over theactual bead area,

=−

×( )

EA A

A100%

1p a

a

where Ap is the prediction of the bead area by the parabola model.The actual area, Aa of a weld bead cross-section, namely metaldeposition rate per unit length, can be calculated as

π πλ= = ( )A

FDS

D4 4 2a

w w2 2

where, Dw is the diameter of the wire electrode, λ is the ratio ofwire-feed rate to travel speed.

As shown in Fig. 4, the relative errors of area prediction for the81 tests are within 2%, indicating that parabola function matchesthe bead geometry with reasonably high accuracy. The weld beadheight and width of all the tests are shown in Fig. 5. It can be seenthat the obtained bead widths range from 2 to 12 mm and beadheights range from 1 to 3 mm. With the changing of bead height,the range of bead width at the same height also varies.

2.2. Artificial neural network (ANN) model

Single bead geometry has a significant bearing on weldingprocess parameters (wire-feed rate and travel speed). Therefore, a

Table1Welding parameters for weld bead model.

Parameters Value

Wire-feed rate, F, m/min 2.0, 2.8, 3.6, 4.4, 5.0, 5.6, 6.4, 7.2, 8.0Travel speed, S, m/min 0.4, 0.47, 0.54, 0.61, 0.68, 0.75, 0.82, 0.91, 1.0

Fig. 6. Neural network architecture for predicting bead geometry.

model relating process parameters and the single bead geometryis necessary to set optimum process parameters in WAAM. In thisSection, a neural network model is established, as shown in Fig. 6.

Fig. 7. Measured bead height and width vs. predicted bead height and width usingANN model. Circles represent the bead heights of the independent 8 testing data.Squares represent the bead widths of the independent 8 testing data. Units for bothaxes are millimetres.

Fig. 8. Surface plots of the bead height and bead width as the function of wire-feedrate and welding travel speed.

D. Ding et al. / Robotics and Computer-Integrated Manufacturing 39 (2016) 32–42 35

Wire-feed rate and travel speed are chosen as the two input pro-cess variables, while bead width and bead height are the two re-sultant outputs.

The relationships of 81 combinations between process para-meters and the weld bead geometry were randomly divided intotwo datasets, of which 65 were used for training and 8 for vali-dation. The remaining 8 combinations were reserved for in-dependent testing. The experimental data were linearly normal-ised for ANN training, validation and testing. The best architectureof 2-13-2 network with the mean square error of 0.0019 wastrained. The predicted bead geometry in the training of the net-work was confirmed by correlation with the independent 8 testingexperimental data as shown in Fig. 7. It can be seen that there ishigh correlation between the predicted and experimental values ofthe bead geometry, indicating the ANN has the ability to accuratelypredict the bead geometry from welding parameters input.

In order to optimise the welding parameters based on the de-sired bead geometry, a database of 3721 combinations (61 by 61input matrix) of welding parameters has been built, with the wire-feed rate ranging from 2 to 8 m/min in the steps of 0.1 m/min andthe travel speed ranging from 0.4 to 1 m/min in the steps of0.01 m/min. An output matrix (61 by 61) of the predicted beadheight and width is then generated through the trained neuralnetwork as shown in Fig. 8. This database is used to select theoptimal welding parameters that will produce a given beadgeometry.

2.3. Multi-bead overlapping model

The step-over distance between paths plays an important rolein achieving high surface quality and geometrical accuracy of thesliced layer, as shown in Fig. 9. Some preliminary investigations onmulti-bead overlapping models have been made in recent years[22, 23]. Referring to our previous work [21], the tangent over-lapping model (TOM) is described as follows.

As shown in Fig. 10a, for two identical beads with individualwidth w and height h, Bead 1 is first deposited on the substrate,and Bead 2 is deposited next to Bead 1 with a step-over distance ofd. When the area of the critical valley is equal to the overlapping

area, the critical step-over d* is obtained. In this case, the uppersurface of the deposit between the bead centrelines (depositionpaths) will be flattest. For the step-over distance d4d*, in-sufficient overlapping results in a surface depression, as shown inFig. 10b. Severe surface depressions are difficult to fill in sub-sequent layer paths, increasing the likelihood of voids within thedeposit. For the step-over distance dod*, excessive overlap pro-duces an increased deposit height as shown in Fig. 10c. This leadsto accumulating height errors over multiple layers. Therefore, theoptimum step-over is the critical distance d*, which depends onlyon the single-bead width [21], and is d*¼0.738 w.

3. Adaptive MAT path planning

3.1. Adaptive MAT path towards void-free deposition with highgeometrical accuracy

When the wire-arc welding process is used for additive man-ufacturing, to obtain optimum results the calculated path shouldfirstly cover the entire geometry without crossover to guaranteeeven and void-free deposition. Secondly, the continuity of thedeposition path should be maintained to avoid frequent start/stopof the wire-feed system. Lastly, the geometrical accuracy at thecomponent boundary should be considered in order to minimiseor eliminate post-process machining.

Many types of path patterns have been developed, such asraster, zigzag, contour or spiral [20]. Raster path patterns contain aset of discontinuous scan lines, which require the depositionprocess to start and stop regularly and therefore not suitable forWAAM. Zigzag path patterns may involve highly convoluted paths,which result in the accumulation of heat in certain regions, andfrequent alternations of tool path travel directions are also not

Fig. 9. Overlapped surfaces at different step-over distances. (a) No overlapping when dZw, where w is single-bead width. (b) Less overlapping when d*rdrw, where d* isthe critical step-over distance. (c) Ideal overlapping when d¼d*. (d) Excessive overlapping when dod*.

Fig. 10. Schematic diagrams of the tangent overlapping model (TOM) [21].

D. Ding et al. / Robotics and Computer-Integrated Manufacturing 39 (2016) 32–4236

suitable for metal deposition using wire-feed AM. Contour pathpatterns, which offset parallel to the boundary of the given geo-metry are often preferred over raster and zigzag path patterns.However, contour path patterns will generally incur voids in themiddle of the component because the paths are offset startingfrom the boundary and then move towards the centre at a con-stant offset, as shown in Fig. 11a. These internal gaps (Fig. 11b) arenot guaranteed to be filled by subsequent deposits because of thephysical difficulty for material to fully flow and fuse into theconfined corners.

To avoid such internal voids, the MAT (Medial Axis Transfor-mation) path was introduced [24] and its extension for complexgeometries was developed recently [19]. MAT paths are generatedby offsetting the medial axis of the geometry from the centre to-wards the boundary. Fig. 11c shows an example of MAT path withthe deposition sequence indicated by numbers. Although void-free

deposition is obtained using MAT paths, this is achieved at the costof creating discontinuity of the path (such as path 3, 4, and 5 inFig. 11c) and extra deposition at the boundary as described inFig. 11d. Post-process machining must be used to remove the extramaterials and improve the accuracy at the cost of material andenergy wastage.

Step-over distance is always constant for both contour pathpatterns (refer to Fig. 11a) and MAT path patterns (refer to Fig. 11c).For certain geometries, it is not possible to achieve both high ac-curacy (refer to Fig. 11b) and void-free (refer to Fig. 11d) compo-nents using paths with constant step-over distance. However, theWAAM process is capable of producing deposits with differentwidths within a layer through varying travel speed and wire-feedrate, while maintaining a constant deposit height. Therefore, wepropose an adaptive path planning strategy that uses continuouslyvarying step-over distances by adjusting the process parameters todeposit beads of varying width within any given path. The de-veloped adaptive MAT path planning algorithm is able to auto-matically generate path patterns with varying step-over distances(refer to Fig. 11e) through analysing geometry information toachieve better part quality (void-free deposition) and geometricalaccuracy, as shown in Fig. 11f.

3.2. Adaptive MAT path strategy

In order to implement adaptive MAT path on any sliced geo-metry, several modules are required, including medial axis trans-formation, domain decomposition, and path generation on eachdomain, etc. The detailed adaptive MAT path planning methodol-ogy will be provided in another paper. Here, the path planningstrategy is introduced through an example. The followings are themain steps for generating the adaptive path:

(1) The cross-section of a part with two internal holes is shown inFig. 12a. The medial axis of the geometry is computed andrepresented using red lines.

(2) Since the geometry with N holes needs to be decomposed intoNþ1 domains, this part is decomposed into 3 domains asshown in Fig. 12b. Each domain is enveloped by a branch loop(red line loop with direction) and either an external boundaryloop (black line loop around orange region) or an internalboundary loop (black line loop within green and blue regions).

(3) Accordingly, adaptive paths with varying step-over distance toaccommodate the width variation of each domain are gener-ated separately. Taking domain 1 as an example, the variationof width along the branch loop is shown in Fig. 13a, whereblue lines represent widths. The domain is expanded along thebranch loop as shown in Fig. 13b, where the variation ofwidths is transferred to the radius (R) as the function of lengthalong the branch loop. To generate the paths for each domain,a key task is to determine the path planning variables nj,which represents the number of path passes in the domain j. Ifthe n1 is chosen to be 3, adaptive paths could be generated

Fig. 11. Illustrations of different deposition paths. External lines represent the boundary of the geometry; internal lines represent the deposition paths with the numbersrepresenting the order of the deposition paths; grey regions are deposited area by the relevant paths. (a) Contour path patterns; (b) The predicted high accuracy depositionbut with internal gaps; (c) MAT path patterns [19]; (d) The predicted void-free deposition but with extra material deposited along the boundary; (e) Adaptive MAT pathpatterns with varying step-over distance; (f) The predicted void-free deposition with high accuracy at the boundary through using adaptive MAT path.

D. Ding et al. / Robotics and Computer-Integrated Manufacturing 39 (2016) 32–42 37

through offsetting the branch loop towards the boundary asshown in Fig. 12c. The range of values for nj is limited by rangeof bead widths that can be produced by the wire-arc weldingprocess at a given bead (or layer) height.

(4) Finally, the adaptive path for the whole geometry is completedas shown in Fig. 12d.

The adaptive MAT path planning algorithm automaticallygenerates a set of continuous deposition paths with varying step-over distance, which is preferred for WAAM to produce parts withhigh quality and high geometrical accuracy at the layerboundaries.

4. Implementation of 2D path planning algorithm

4.1. Determination of 2D path planning variables

To implement the adaptive MAT path algorithm, one importanttask is to determine appropriate values of nj for domain j. Thevalues of nj has significant effects on the desired bead size, withsmall n corresponding to larger weld deposits. If the desired weldbead height is chosen to be h, then a number of beads with thesame height could be chosen by combinations of different wire-feed rate and travel speed. Using the weld bead database matrixdeveloped in Section 2, the range of weld bead width as a functionof weld bead height are shown in Fig. 14 as the shaded area within

the black curve. For certain a height of h, the width of these weldbeads are within a range [w(h)min, w(h)max].

In adaptive MAT path planning, the deposition process can beconsidered as a disk with varying diameter being swept along thecomputed path. As shown in Fig. 15, the domain is expanded alongthe branch loop, with the internal or external boundary re-presented by the black solid line. The maximum step-over distancedmax is obtained at the maximum radius (Rj)max and the minimumstep-over distance dmin is obtained at the minimum radius (Rj)min.

The following relationships can be obtained if nj paths aregenerated

* ≤ ( = ) ≤ *

* ≤ ( = ) ≤ *( )

⎧⎨⎪⎪

⎩⎪⎪

d dR

nd

d dR

nd

3

min minmin

max

min maxmax

max

where *dmin¼0.738 w(h)min and *dmax¼0.738 w(h)max by means ofthe multi-bead overlapping model. Eq. (3) reveals the path plan-ning for WAAM is highly geometry-related which is affected byboth the complexity of the geometry (variations of R) and theshape of deposit (w or d*). By solving n, one can obtain

* ≤ ≤ *

* ≤ ≤ * ( )

⎧⎨⎪⎪

⎩⎪⎪

Rd

nRd

Rd

nRd 4

min

max

min

min

max

max

max

min

Fig. 12. Illustration of adaptive MAT path generation. (a) A geometry and computedmedial axis or skeleton. (b) Domain decomposition. (c) Path generation for eachdomain. (d) Final adaptive paths for the geometry.

Fig. 13. Domain and radius function. (a) Variation of width of a domain.(b) Expanded domain along the length of the branch loop. (c) Path generationthrough offsetting.

Fig. 14. Ranges of bead width and bead height through varying wire-feed rate andtravel speed.

Fig. 15. Illustration of the deposition process in an expanded domain.

NMM

ji

N

NMM

ji

N

hwhwhw

hwhw

SFSFSF

SFSF

,,,

,,

,,,

,,

1

111

1

111

Fig. 16. The database of weld single bead model.M¼61, and N¼61 as mentioned inSection 2.

D. Ding et al. / Robotics and Computer-Integrated Manufacturing 39 (2016) 32–4238

Eq. (4) can be further simplified as

* ≤ ≤ * ( )Rd

nRd 5

max

max

min

min

Since n must be an integer, the final solution for n is given as

( * ) ≤ ≤ ( * )( )

Rd

nRd

ceil floor ,6

max

max

min

min

where, function ceil rounds the elements within it to the nearestintegers towards infinity and function floor rounds the elementswithin it to the nearest integers towards minus infinity.

In Eq. (6), Rmax and Rmin are sliced layer geometry parameterswhich are obtained from medial axis transformation. For a givenbead height (layer thickness) and the established bead model asshown in Fig. 14, the range of achievable bead width can be

determined, as well as *dmin and *dmax by means of the multi-beadoverlapping model. Therefore, the optimum n for a domain can beeasily solved. If there is no integer solution then the adaptive MATpath algorithm is not suitable for the domain. This means that thecapability of the deposition system to vary the bead width whilemaintaining a constant bead height is not sufficient to accom-modate the width variation of the given domain. In this case an-other path planning method such as non-adaptive MAT is appliedto the domain, at the cost of increased deposition outside of thecomponent boundary as shown in Fig. 11d.

4.2. Determination of welding process parameters

When the desired weld bead geometries have been determinedaccording to the generated adaptive paths, the next step is todetermine appropriate welding parameters to produce the desiredweld beads. To determine the optimal process parameters with thedesired bead geometry, the following database with wire-feed rateand travel speed as the inputs and bead width and bead height asthe outputs is established as shown in Fig. 16.

Given the outputs, the error(i,j) corresponding to all inputcombinations are computed as the relevant error between thepredicted bead geometry and the desired bead geometry. One canobtain the following formula

D. Ding et al. / Robotics and Computer-Integrated Manufacturing 39 (2016) 32–42 39

( ) =−

+−

( = = ) ( )

⎛⎝⎜

⎞⎠⎟

⎛⎝⎜

⎞⎠⎟error i j

w ww

h h

h

i M j N

, ,

1, 2, ... ; 1, 2, ... 7

i d

d

j d

d

2 2

where wd and hd are the desired bead width and bead height,respectively. Through solving for the minimum value of error(i,j),the optimal wire-feed rate and travel speed Fi and Sj are found.

Fig. 18. Case 1, a thin-walled structure with a single internal hole and varied wallthickness.

Table 2Information for Case 1 geometry and the resulting range of bead width. All unitsare millimetres except for n1 and n2 which are dimensionless variables.

h (R1) min (R1) max (R2) min (R2) max wmin wmax *dmin *dmax n1 n2

2 3.45 5.66 3.45 5.66 4.42 7.76 3.26 5.73 1 1

5. Experimental results and discussion

Experimental tests were conducted using a robotic weldingsystem at the University of Wollongong. The robotic WAAM sys-tem and 3D laser scanning system have been integrated into awelding cell to conduct the experiments, as described in Fig. 17.For a detailed description of experimental system, refer to Ref.[21]. Synergic GMAW welding programme is used, in which thewelding current and voltage are automatically controlled by thewelding machine to ensure uniform penetration and weld beadprofile. The wire electrode was copper coated steel wire with thediameter of 1.2 mm. The stick-out length was set to 18 mm tominimise weld spatter for this particular process. The wire-feedrate was set from 2.0 to 8.0 m/min and the welding travel speedwas varied from 0.4 to 1.0 m/min. A 3D laser profile scanner with aresolution of 0.02 mm was used to accurately measure the cross-sectional profile of the weld beads at different locations along thewelding direction.

5.1. Case study 1

Fig. 18 shows a thin-walled structure with a single internal holeand wall thickness ranging from 7 mm to 11 mm. Table 2 providesthe main parameters of the geometry and path planning variables.The geometry is decomposed into two domains since it has onehole. The maximum and minimum radii for the two domains arelisted in Table 2. For the given bead height h of 2 mm, the beadwidth ranges from 4.42 mm to 7.76 mm through altering wire-feed rate and travel speed. Accordingly, the critical step-over dis-tances, *dmin and *dmax are 3.26 mm and 5.73 mm, respectively. Byapplying Eq. (6), the path planning variables n1 and n2 are bothsolved as 1. Consequently, the adaptive MAT path for the geometryis generated as path 1 and path 2 shown in Fig. 18. Red arrowsshow the start point (randomly set) and the deposition direction.

As the adaptive MAT path is generated, the step-over distances

Fig. 17. Schematic diagram of the ex

along the paths are progressively determined as shown in Figs. 19and 20. Step-over distances vary along the length of the depositionpath to accommodate the variation in thickness of the geometry.The wire-feed rates and travel speeds for the two paths are alsoshown in Figs. 19 and 20. For clarity, the minimum wire feed rateof 2.0 m/min and minimum travel speed of 0.4 m/min have beennormalised to 1.0. It can be seen that with the changing of step-over distance, both wire-feed rate and travel speed are changed toproduce the desired bead geometry suitable for the given range ofstep-over distance.

Using the generated paths and the welding process parameters,programme code for the robotic manipulator and the weldingpower source is automatically generated. The near-net shapecomponent was deposited with 15 layers, as shown in Fig. 21a.Fig. 21b shows the finished component after surface milling.

In order to check the potential internal defects, the finishedcomponent was cut into two pieces as shown in Fig. 22a. The crosssection A is parallel to the building direction and the cross section

perimental WAAM system [21].

Fig. 19. Calculated step-over distance (mm), normalised wire-feed rate, and normalised travel speed along the length of deposition path 1 (in mm).

Fig. 20. Calculated step-over distance (mm), normalised wire-feed rate, and normalised travel speed along the length of the deposition path 2 (in mm).

D. Ding et al. / Robotics and Computer-Integrated Manufacturing 39 (2016) 32–4240

B is perpendicular to the building direction. Close up of the crosssections are shown in Fig. 22b and c. It can be seen that the de-posited part is 100% solid and free of porosity.

5.2. Case study 2

Fig. 23 shows an example of a solid structure without holes.The geometry has a length of 100 mmwhile the width ranges from25 mm to 50 mm as described in Fig. 23. Table 3 provides thegeometrical parameters that were determined by the adaptiveMAT path planning algorithm. The deposition path comprised ofthree closed loops with varied step-over distances. The red arrowsin Fig. 23 indicate the start points and the travel directions. Foreach layer, deposition was performed from the centre towards theboundary of the geometry. Fig. 24 shows the step-over distance

Fig. 21. (a) Deposited near-net shape after 15 layers. (b

and the non-dimensional wire-feed rate and travel speed for theoutermost path.

Although there is a large variation of geometry width from25 mm to 50 mm, the component has been deposited throughusing only three closed paths with varying step-over distances.After deposition of 10-layers, the near-net shape component isshown in Fig. 25a. The track of the deposition path can be seenfrom the upper surface of the shape. Fig. 25b shows the finishedpart after post-process milling, free of internal voids or gaps.

5.3. Discussion

The two case studies have provided examples of componentshaving either thin-walled structures of varying thickness, or a solidstructure with a large variation in across-layer width. Both

) Finished part without voids after surface milling.

Fig. 22. Examination of the internal cross sections. (a) Cut parts of the case one; the cross section A is parallel to the building direction and the cross section B is per-pendicular to the building direction. (b) Close up of the cross section A. (c) Close up of the cross section B.

Fig. 23. Case 2, a solid structure without holes.

Table 3Information of Case 2 geometry and the resulting range of bead width. All units aremillimetres expect for n which is a dimensionless variable.

h Rmin Rmax wmin wmax *dmin *dmax n

2.5 16.15 21.69 5.90 10.28 4.35 7.59 3

D. Ding et al. / Robotics and Computer-Integrated Manufacturing 39 (2016) 32–42 41

geometries provide significant challenges for the WAAM process,which has several requirements that should be met for optimumresults:

Fig. 24. Calculated step-over distance (mm), normalised wire-feed rate, and no

1. Continuous deposits formed in closed loops to minimise start/stops that can give rise to height errors over multiple layers;

2. Continuous closed loop deposits at the component boundaries,to minimise the waste material that need to be removed byfinish machining;

3. No cross-over of weld paths that create large localised varia-tions in build height;

4. Adaptive step-over distance of the weld beads to avoid theformation of voids while producing a smooth upper surface foreach layer and simultaneously conforming to the componentboundaries;

5. Adaption of welding parameters to produce deposits of variablewidth at consistent height, to accommodate the adaptive step-over distance.

The automated path design, welding parameter selection,programme code generation, and final deposition of the above twocomponents indicates that the proposed adaptive MAT pathplanning algorithm is capable of generating continuous paths thatcan accommodate geometries with significant variations in wallthickness, while operating within the capabilities of the arc-wirewelding process. The resulting deposition using the generatedpath produces components that are internally void-free and havegood geometrical accuracy requiring minimal post-process mil-ling. Although not exercised in these case studies, there are si-tuations where the adaptive MAT process is not able to producepaths that conform to the component boundaries because of

rmalised travel speed along the length of the deposition path 3 (in mm).

Fig. 25. (a) Deposited near-net shape after 15 layers. (b) Finished part without voids after surface milling.

D. Ding et al. / Robotics and Computer-Integrated Manufacturing 39 (2016) 32–4242

limitations of the welding process to deposit beads of sufficientlyvariable width. In such cases, the path generation algorithm de-faults to using the non-adaptive MAT method. The component canstill be produced automatically, but the component surface is lessregular due to start/stops that must be performed at the boundary,and this would require more waste material to be removed toobtain the desired shape. This default situation is performed au-tomatically, and hence contributes to the ultimate goal of produ-cing a practical, computationally efficient and highly automatedWAAM system for industrial application.

6. Conclusions

This study has focused on the development of an automatedpath planning system for WAAM in order to free the operator fromthe tedious and error-prone steps of analysing the CAD model,planning the deposition path, and manually setting the weldingprocess parameters. As key steps in the automation system, pathplanning and weld parameter selection have been investigated indetail. The adaptive MAT path planning algorithm and the estab-lished bead models were tested through experimental results oftwo metal components. The results show that the WAAM systemhas the capability to achieve high quality, void-free componentsthat are geometrically accurate through properly designing theprocess variables. The developed bead models and adaptive MATalgorithm have been proven to be an effective programmingstrategy for WAAM.

Acknowledgements

The authors would like to thank Professor J. Norrish for hisexpertise in the welding field. This work is supported in part bythe State Scholarship Fund of the China Scholarship Council (No.2011684067).

References

[1] Douglas C. Hofmann, Scott Roberts, Richard Otis, Joanna Kolodziejska, R. PeterDillon, Jong-ook Suh, Andrew A. Shapiro, Zi-Kui Liu, John-Paul Borgonia, De-veloping gradient metal alloys through radial deposition additive manu-facturing, Sci. Rep. 4 (5357) (2014) 1–8.

[2] J.D. Spencer, P.M. Dickens, C.M. Wykes, Rapid prototyping of metal parts bythree-dimensional welding, Proc. Inst. Mech. Eng. 212 (1998) 175–182.

[3] Y. Song, S. Park, Experimental investigations into rapid prototyping of com-posites by novel hybrid deposition process, J. Mater. Process. Technol. 171

(2006) 35–40.[4] Y. Zhang, Y. Chen, P. Li, A.T. Male, Weld deposition-based rapid prototyping: a

preliminary study, J. Mater. Process. Technol. 135 (2003) 347–357.[5] S.H. Mok, G. Bi, J. Folkes, I. Pashby, Deposition of Ti–6Al–4V using a high power

diode laser and wire, Part I: investigation on the process characteristics, Surf.Coat. Technol. 202 (2008) 3933–3939.

[6] K.P. Karunakaran, S. Suryakumar, Pushpa Vishal, Akula Sreenathbabu, Low costintegration of additive and subtractive processes for hybrid layered manu-facturing, Robot. Comput. Integr. Manuf. 26 (2010) 490–499.

[7] J. Xiong, G. Zhang, Z. Qiu, Y. Li, Vision-sensing and bead width control of asingle-bead multi-layer part: material and energy savings in GMAW-basedrapid manufacturing, J. Clean. Prod. 41 (2013) 82–88.

[8] D. Ding, Z. Pan, D. Cuiuri, H. Li, Wire-feed additive manufacturing of metalcomponents: technologies, developments and future interests, Int. J. Adv.Manuf. Technol. 81 (2015) 465–481, Online available.

[9] P.S. Almeida, S. Williams, Innovative process model of Ti–6Al–4V additive layermanufacturing using cold metal transfer (CMT), in: Proceedings of the 21stAnnual International Solid Freeform Fabrication Symposium, University ofTexas Austin, TX, USA, 2010.

[10] J. Ding, P. Colegrove, J. Mehnen, S. Ganguly, P.M. Sequeira Almeida, F. Wang,S. Williams, Thermo-mechanical analysis of wire and arc additive layermanufacturing process on large multi-layer parts, Comput. Mater. Sci. 50(2011) 3315–3322.

[11] R. Dwivedi, R. Kovacevic, Automated torch path planning using polygon sub-division for solid freeform fabrication based on welding, J. Manuf. Syst. 23(2004) 267–280.

[12] G.Q. Jin, W.D. Li, L. Gao, An adaptive process planning approach of rapidprototyping and manufacturing, Robot. Comput. Integr. Manuf. 29 (2013)23–38.

[13] Y. Song, S. Park, D. Choi, H. Jee, 3D welding and milling: part I – a direct ap-proach for freeform fabrication of metallic prototypes, Int. J. Mach. ToolsManuf. 45 (2005) 1057–1062.

[14] F. Martina, J. Mehnen, S.W. Williams, P. Colegrove, F. Wang, Investigation of thebenefits of plasma deposition for the additive layer manufacture of Ti-6Al-4V,J. Mater. Process. Technol. 212 (2012) 1377–1386.

[15] R. Martukanitz, J. Hollingsworth, Taking the next step in additive manu-facturing, Weld. J. 40 (2014) 40–43.

[16] N. Kapustka, I.D. Harris, Exploring arc welding for additive manufacturing oftitanium parts, Weld. J. 40 (2014) 32–35.

[17] A. Heralic., Towards full automation of robotized laser metal-wire deposition(Ph.D. thesis), Chalmers University of Technology, Goteborg, Sweden, 2009.

[18] F. Bonaccorso, L. Cantelli, G. Muscato, An arc welding robot control for a shapemetal deposition plant: modular software interface and sensors, IEEE Trans.Ind. Electron. 58 (2011) 3126–3132.

[19] D. Ding, Z. Pan, D. Cuiuri, H. Li, A practical path planning methodology for wireand arc additive manufacturing of thin-walled structures, Robot. Comput.Integr. Manuf. 34 (2015) 8–19.

[20] D. Ding, Z. Pan, D. Cuiuri, H. Li, A tool-path generation strategy for wire and arcadditive manufacturing, Int. J. Adv. Manuf. Technol. 73 (2014) 173–183.

[21] D. Ding, Z. Pan, D. Cuiuri, H. Li, A multi-bead overlapping model for roboticwire and arc additive manufacturing (WAAM), Robot. Comput. Integr. Manuf.31 (2015) 101–110.

[22] S. Suryakumar, K.P. Karunakaran, A. Bernard, U. Chandrasekhar,N. Raghavender, D. Sharma, Weld bead modelling and process optimization inhybrid layered manufacturing, Comput.-Aided Des. 43 (2011) 331–344.

[23] J. Xiong, G. Zhang, H. Gao, L. Wu, Modeling of bead section profile and over-lapping beads with experimental validation for robotic GMAW-based rapidmanufacturing, Robot. Comput. Integr. Manuf. 29 (2013) 417–423.

[24] J.H. Kao, F.B. Prinz, Optimal motion planning for deposition in layered man-ufacturing, Proc. DETC (1998) 13–16.