Journal of Manufacturing Processes 70 (2021) 543–552

Contents lists available at ScienceDirect

Journal of Manufacturing Processes

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

Machine learning based layer roughness modeling in robotic

additive manufacturing
Ahmed Yaseer, Heping Chen *
Ingram School of Engineering, Texas State University, Texas 78666, USA

A R T I C L E I N F O A B S T R A C T

Keywords: Wire Arc Additive Manufacturing (WAAM) is a manufacturing technique that deposits metal layer upon layer to
Random Forest manufacture 3D parts based on welding processes. Most researchers considered weld bead width, height, and
Multilayer Perceptron penetration as the characteristic performance in WAAM. However, layer roughness is also important because it
Machine learning
affects the machining cost and mechanical properties of fabricated parts. If the roughness of a deposited layer can
Weaving path
Surface roughness
be reduced, less machining will be required, and material wastage will be reduced. Reduced layer roughness will
Wire Arc Additive Manufacturing (WAAM) also enable better bonding between adjacent layers. Hence, the deposition of weld beads with minimized
Laser scanning roughness demands great attention. A few researchers who tried to investigate roughness in WAAM used straight
Industrial robot paths for material deposition, but the investigation of the weaving path, which has a great potential to reduce
layer roughness, has not been investigated well. The main contribution of this paper is about successfully
implementing two machine learning methods to accurately model surface roughness in WAAM using a weaving
path: Random Forest and Multilayer Perceptron (MLP) which is also known as Artificial Neural Network (ANN).
Both methods are effective for modeling and predicting the layer roughness for a given set of robotic WAAM
parameters, but Random Forest gave better results than MLP in terms of accuracy and computational time.

1. Introduction
WAAM is an additive manufacturing process to fabricate a part. It has
the prospect of revolutionizing the production processes in industries
like aerospace, automotive, oil and gas, nuclear, shipbuilding, heavy
equipment, etc. WAAM has been proven to be more efficient than the
previously used methods like Electron Beam Additive Manufacturing
(EBAM) and Wire Laser Additive Manufacturing (WLAM), for producing
large metal parts [1–7]. Dimensional stability and precision during
making complex parts is the major focus of the state-of-the-art WAAM
technology and this paper will investigate the layer roughness in order
to achieve better dimensional control in WAAM.
The dimensional accuracy and efficiency of WAAM largely depend
on the deposition pattern of metal, called deposition path and WAAM
process parameters. Various path planning methods, such as zigzag,
raster and Medial Axis Transformation (MAT), have been used in
WAAM, and all these methods are based on straight pass deposition.
With these methods, where the traditional overlapped single pass beads
are used, the bead top surface is not very flat and layer height is not
uniform, which makes it difficult for multilayer deposition. When
multiple beads are deposited side by side in WAAM, the produced layers
have high roughness that needs to be minimized by machining. With the
weaving path, if the roughness is taken into consideration, flat layers
with uniform height will be produced. Surface roughness in WAAM also
affects the contact and joining [8] between two adjacent layers, and thus
impacts mechanical strength between two layers. To solve this problem,
researchers came up with the idea of weaving path planning method to
produce beads with better flatness than other methods. However, the
flatness of the beads produced by the weaving path should be quantified
by measuring the bead roughness, so that the amount of machining and
material wastage can be reduced. Hence, there is a scope of research for
producing layers with minimized roughness with the weaving path
method. Moreover, at increased travel speed, humping [9–11] is also a
big problem for single-pass beads. With the proposed weaving method,
the humping effect will also greatly be reduced. Besides this, pits are also
generated at the borders of two adjacent beads while depositing multiple
overlapping beads with the weaving path due to heat accumulation
[12]. Impurities and bubbles are also removed by the weaving arc,
reducing inclusions and porosity [13].
To minimize the material usage of WAAM, the produced part should

* Corresponding author.
E-mail address: hc15@txstate.edu (H. Chen).

https://doi.org/10.1016/j.jmapro.2021.08.056
Received 5 July 2021; Received in revised form 19 August 2021; Accepted 28 August 2021
Available online 14 September 2021
1526-6125/© 2021 The Society of Manufacturing Engineers. Published by Elsevier Ltd. All rights reserved.
A. Yaseer and H. Chen
be very close to its actual dimension and machining should be mini­
mized. Traditional weld bead cross-sections are shaped like a parabola,
which requires more machining. Moreover, porosity and delamination
will also occur due to high surface roughness for multilayer deposition
[14]. Thus, the top surface of the weld bead should be as flat as possible
by reducing the layer roughness.
Aldalur et al. [15] compared the WAAM-manufactured walls
deposited by a weaving path and a straight-pass overlapped path and
found that both strategies showed similar mechanical properties (ten­
sile, hardness, impact), but the weaving path showed more homoge­
neous properties in both horizontal and vertical directions because of
the homogeneity of the beads. They also found that the weaving path
produced better surface flatness and hence material utilization was
increased because it needed less machining. It was also found that with
the traditional straight-pass overlapped path, a cooling time was needed
after each bead was deposited. But, with the weaving path, there was no
overlap of beads, and that's why, no cooling time was needed, and
hence, it needed less manufacturing time.
Khan et al. [16] designed and developed a barrel finishing machine
for improving the surface finish of parts produced by WAAM where the
part was submerged in a barrel filled with dry or wet abrasives and the
barrel was rotated with the part inside it. Some researchers used a hybrid
manufacturing process combining WAAM with milling for a better sur­
face finish [17,18]. But they predicted the surface roughness after
milling from the input milling parameters. Li et al. [19] predicted the
roughness of parts produced by WAAM after machining from the input
machining parameters using the 2nd order regression model and Tian
et al. [20] did this using Response Surface Methodology. Laser polishing
after machining for improving surface quality has also been investigated
too [21]. However, if the surface roughness of WAAM can be modeled
and predicted, the milling process might be eliminated or reduced to a
large amount, and this is more important for achieving efficiency and
reducing material wastage.
A few researchers tried to relate the wall roughness of thin-walled
parts produced by WAAM but only Xia et al. [14] considered the
roughness of a layer. In [22], the average surface of the arc-laser hybrid
WAAM system was measured at mm-level. They kept the gas metal arc
welding (GMAW) parameters constant and changed the laser power to
model the surface roughness along the layer stacking direction. They
also suggested that overflowing and collapsing of the liquid weld pool
and incomplete fusion lowers the arc stability and thus decreases the
surface quality. Xiong et al. [23] investigated the influence of process
parameters on wall roughness of thin-walled deposition (multiple layers
with a single pass) and suggested that a lower wire feed speed associated
with a lower travel speed gives the minimum surface roughness. They
also found that increasing the wire feed speed greatly decreases the
surface roughness because the overflow of the molten pool increases
along with the wire feed speed due to the effect of arc forces. However,
since we will use the weaving path in this research, the parameters,
suitable for the single pass beads might not be suitable for the weaving
method, and we also have to consider some extra parameters like
weaving amplitude and wavelength. In [24], lateral surface waviness of
thin walls produced with single pass was investigated by mainly
changing the wire feed speed and travel speed.
The surface roughness of WAAM was predicted using the process
parameters in [14]. Three approaches are used-Adaptive neuro-fuzzy
inference system (ANFIS), Extreme Machine Learning (ELM), and Sup­
port Vector Regression (SVR), to predict the surface roughness from the
process parameters including- travel speed, wire feed speed, and bead
step-over distance. They found that surface roughness decreased with
increasing step-over distance for multi-bead deposition with a straight
pass. Travel speed and wire feed speed showed a non-linear relationship
with surface roughness. They collected 27 sets of experimental data and
used the K-fold cross-validation for machine learning implementation.
Among the three machine learning methods, ANFIS gave the best results.
The three machine learning methods used in their research have also
544
Journal of Manufacturing Processes 70 (2021) 543–552
been used several times by other researchers for predicting various
welding performance characteristics [25–31].
To characterize the roughness, various sensors have been used like
arc sensor, ultrasonic sensor, electromagnetic sensor, spectrum sensor,
force sensor, vision sensor, etc. Among these sensors, the vision sensor
and laser scanner are most effective [32]. There are two types of surface
roughness measurement techniques namely contact type and non-
contact type, and non-contact type methods have recently been found
advantageous.
Bead width, height, and sometimes penetration of the weld beads,
have long been considered as the main performance characteristics to
evaluate the final shape and strength of the fabricated 3D part. The weld
bead, height, and penetration depend on the input welding parameters
such as wire feed speed (WFS), travel speed, contact to work distance
(CTWD), arc voltage, current, and gas flow rate. In WAAM, researchers
have used various Design of Experiments (DOE) methods [33,34] such
as Central Composite Design and Response Surface Methodology, to
design the experiment matrix in the welding process involving the
lowest possible number of experiments. After the experiments are done
according to the designed experiment matrix, the results are evaluated
and the welding parameters with the strongest influence on the desired
weld bead characteristics are identified using the regression analysis or
Analysis of Variance (ANOVA) table. 2nd order regression [35–38] and
3rd order regression [39,40] have also been used to model the input and
output relationship of the WAAM process. For modeling the single bead
geometry of thin-walled objects deposited with weaving path, Ma et al.
[36,37] used the 2nd order regression model where the bead width or
height was expressed as a function of the process parameters such as
travel speed, wire feed speed, weaving wavelength, and weaving
amplitude. Multilayer Perceptron (MLP) or Artificial Neural Network
(ANN) methods were used to predict various weld bead characteristics
from the input process parameters [35,41–43]. Since the welding pro­
cess is very complex and highly nonlinear, MLP can predict the weld
bead geometry more accurately than regression models [35,44–46].
Therefore, in recent years, MLP has been widely used by researchers for
weld bead modeling in WAAM. MLP [47–50] is adopted to model and
predict the surface roughness of parts produced by milling, however, to
our knowledge, it has not been applied to model the roughness produced
by WAAM.
Some researchers also used tree-based methods such as Decision Tree
[51,52] and Random Forest [51,53–55] in arc welding and WAAM [56].
They deposited a single layer with a straight pass and predicted bead
width and height from input parameters including current, cladding
voltage, deposition velocity, and initial interlayer temperature.
Random Forest, known for its suitability for modeling non-linear
processes, will be investigated in this paper. The performance of
Random Forest will be compared with that of MLP for predicting the
layer roughness for a given set of input parameters such as travel speed,
wire feed speed, weaving amplitude, and wavelength. This paper makes
the following contributions:
• Modeling layer roughness deposited using a robotic weaving path
which has not been done before;
• Comparing the effectiveness of Random Forest and MLP for layer
roughness modeling in robotic WAAM.
For conducting the experiments, the WAAM process parameters
including welding parameters (travel speed and wire feed speed) and
weaving parameters (weaving wavelength and weaving amplitude) are
selected. After a layer is deposited, the layer roughness is measured after
collecting the point cloud data using a laser scanner. Then modeling
methods based on Random Forest and MLP are implemented to predict
the surface roughness of a deposited layer for a given set of WAAM
process parameters and the experimental results are analyzed and
compared.
A. Yaseer and H. Chen
2. Proposed methods
The objective of this research is to predict the roughness using the
WAAM process parameters. The WAAM process model can be expressed
using,
y = f (X) + ε
where y is the layer roughness; X is a vector of WAAM process param­
eters including travel speed, wire feed speed, weaving amplitude, and
weaving wavelength; ε is noise.
Because WAAM is complex and the relationship between y and X is
non-linear, it is difficult to build a physical model to predict the
roughness given a set of WAAM process parameters. To predict the
roughness using the WAAM process parameters, we propose a method
based on Random Forest in this research. Random Forest is suitable for
modeling non-linear systems because the decision trees are formed
based on yes-no questions to aggregate different types of data into a
group.
To construct a roughness prediction model, datasets must be
collected first. Suppose the following datasets are obtained:
y = [y1 , y2 , …, yn ]
X = [X1 , X2 , …, Xn ]
where y and X are the vectors of output and input respectively; yi is the ith
collected roughness data and Xi the ith set of WAAM process parameters;
n is the number of collected datasets. The datasets will be divided into
training and testing datasets for modeling and testing a model con­
structed using Random Forest respectively.
The procedures for the construction of a machine learning model are
shown in Fig. 1.
As shown in Fig. 1, the loss function should be defined which can be
defined using the mean squared error (MSE)-
1∑ n
( )2
floss = yi − ypi
n i=1
where yi is the actual roughness value of the ith collected dataset and ypi
is the corresponding predicted roughness values using the machine
learning model. If the value of the loss function of the testing dataset is
not acceptable, the model configuration is changed until the value is
Fig. 1. Construction procedures of a machine learning model.
Journal of Manufacturing Processes 70 (2021) 543–552
acceptable.
2.1. Modeling based on Random Forest
Random Forest is a supervised and ensemble machine learning
technique that consists of many decision trees as its building blocks. As a
(1)
supervised machine learning technique, random forest learns to build
the relationship between the input X and output y in the training stage.
Once the training is done, we test the model using the testing datasets.
Random forests can model complex and highly non-linear relationships.
It is also generally not affected by outliers and noise, and also a faster
process [57].
The steps of implementing Random Forest are shown in Fig. 2.
As shown in Step 1, each decision tree considers a random subset of
the total training datasets with replacement, which is called bagging or
bootstrap aggregating. According to Fig. 3, for bagging, row sampling
from the total training datasets is done randomly for each tree, and thus,
datasets D1, D2, …., Dm are the selected randomly from the total training
dataset D. Hence, here row sampling means that several rows from the
training datasets containing the input WAAM process parameters (wire
feed speed, travel speed, weaving wavelength, weaving amplitude) and
(2) the output variable (roughness) are selected to make each of the datasets
D1, D2, …., Dm. Also, D1, D2, …., Dm all have the same dimension. This
(3) bagging technique not only increases the diversity of trees in the forest
but also reduces overfitting. The predicted roughness is made by aver­
aging the output of the individual decision or regression tree.
As shown in Step 2, the decision tree initiates and propagates to­
wards generating a final prediction by asking yes/no or true/false
questions at each node. The starting point of the decision tree is called
the root node; the intermediate nodes are called internal nodes; and the
final nodes providing the prediction or outcome are called leaf nodes.
The root node is formed with respect to the right feature for splitting the
node. The splitting is done based on the features that create the highest
information gain (IG), thus making the possible purest child nodes. For
each node, the impurity (I) is calculated using
(4)
1 ∑N
I= (yi − y)2 (5)
N i=1
where
Fig. 2. Steps of Random Forest.
545
A. Yaseer and H. Chen Journal of Manufacturing Processes 70 (2021) 543–552

decreases the variance.

For testing, the testing datasets are used in all the decision trees and
the average of the predicted values of all the decision trees is used as the
final prediction of the random forest.

2.2. Modeling based on Multilayer Perceptron (MLP)

Since MLP or ANN can deal with highly complex systems, they have
been widely used in multiple input multiple output (MIMO) systems. An
MLP network is composed of an input layer, single or multiple hidden
Fig. 3. The bagging process. layers, and an output layer. Fig. 6 shows the connections of each hidden
neuron between inputs and outputs. The efficiency of a neural network
∑
N depends on properly choosing the number of hidden layers, the number
y= yi (6) of neurons in each hidden layer, activation functions at each layer, and
i=1
optimizers.
The MLP algorithm works on the feed-forward back propagation
where N is the number of datasets in the node.
approach, as shown in Fig. 7. This means that, in every cycle or epoch,
Fig. 4 shows a node, its impurity, and the number of datasets in that
there is one forward calculation and one backward calculation.
node. Here, roughness is 0.574 which is the predicted roughness of all
According to step 1 shown in Fig. 7, the weights are initialized first.
the 25 samples in this node and is calculated using Eq. (6).
Let us consider that x1, x2, x3, x4 are the 4 input WAAN process pa­
The datasets of this node are then divided into two subsets based on
rameters and w1k, w2k, w3k, w4k are the associated weights, as shown in
the true and false conditions. To generate the child nodes from the node
Fig. 8. Since we have 10 neurons in the hidden layer, k = 1–10. W4k
shown in Fig. 4, the information gain is calculated using
denotes the weight of the 4th input neuron and the kth hidden neuron.
NL NR The weight vector between the input and the kth hidden neuron is
IG = IDP − ID − ID (7)
NP L NP R expressed as:

where IDP, IDL, and IDR are the impurities of the parent node, left child W = [w1k , w2k , w3k , w4k ] (8)
node, and right child node, respectively, which can be calculated using th
The net input to the k hidden neuron netk is expressed as.
Eq. (5). NP, NL, and NR are the number of datasets in the parent node, left ∑4
child node, and right child node respectively. netk = wik xi (9)
i=1
If the node in Fig. 4 is the parent node, there are three possible child
nodes for each of the true and false conditions using the other three The output of kth hidden neuron hk is expressed as.
parameters- wire feed speed, weaving wavelength, and weaving hk = fact (netk ) (10)
amplitude. The impurities of the three parameters are calculated and the
information gains are computed. For example, in the true condition, where fact is the activation function.
there are 17 samples, and the three possible child nodes are obtained as The activation function ReLU (Rectified Linear Unit) is generally
shown in Fig. 5. In both Fig. 4 and Fig. 5, the values of the WAAM used in the hidden layer as shown in Fig. 9.
process parameters and roughness are the scaled (within 0–1) values. Fig. 10. shows an output neuron. For the output layer, uk denotes the
The child node is chosen using the maximum IG. Hence, for example, weight between the kth hidden neuron (k = 1, …, 10) and the output
the node in Fig. 5 (b) is chosen and the average value of some of the wire neuron (since we have only one output, roughness, as shown in Fig. 6).
feed speed data of this node is calculated to be 0.375. The datasets are The weight vector between the hidden and output layer is expressed as:
then split based on the criterion of wire feed speed ≤0.375. So, for
making a decision tree, each node is iteratively split until the set U = [u1 , u2 , u3 , ……,u10 ] (11)
threshold is reached. The net input to the only output neuron neto is expressed as:
After the tree starts propagating, in step 3, a decision is made ∑10
whether to stop the propagation or not. This decision is made by taking neto = k=1
uk hk (12)
into consideration any of the three scenarios: (1) the tree has reached its
maximum set depth; (2) there is the set minimum number of samples in The output of the only output neuron, roughness (Ra), can be
the leaf node; (3) the information gain after splitting is not greater than expressed as:
the set minimum information gain. If any of these 3 conditions are not Ra = fact (neto ) (13)
met, then step 2 is again initiated by splitting the node with respect to a
feature giving the highest information gain. Linear or Identity activation function is generally used in the output
In step 4, the prediction of a tree is made by finding the sample mean layer. As shown in Fig. 11, the output layer takes inputs, multiplies them
of a node according to Eq. (6). with weights of the output neurons, and finally gives the output.
Because many datasets are generated using the bagging technique, In step 2, the weight matrices W and U are updated by back­
many decision trees will be generated. In step 5, for implementing the propagation such that the prediction error becomes minimum. For doing
random forest, all the trees are run in parallel, and the outcome of the this, the mean squared error of the output neuron is calculated by
independent decision trees are averaged, Thus, random forest gives comparing the actual roughness with the predicted output (yp) as
better performance than an individual decision tree. Moreover, it 1 ∑N 1( )2
E⋅ = yai − ypi (14)
N i=1 2

where N is the total number of training data.

The errors are propagated backward subsequently from layer to layer
for updating the respective weights. The gradient descent algorithm is
used to find the global minima (the lowest error) by differentiating the
Fig. 4. A node of a decision tree. error function with respect to the weight of the corresponding neuron so

546
A. Yaseer and H. Chen Journal of Manufacturing Processes 70 (2021) 543–552

Fig. 5. Possible child nodes.

Fig. 6. MLP architecture.

Fig. 9. ReLU activation function.

Fig. 10. The only neuron of the output layer.

Fig. 7. Steps of MLP.

Fig. 11. Linear activation function.

This process of updating weights continues until the pre-set number

of epochs has been reached or the acceptable error has been satisfied.
Then the updated weight matrix is used in the next epoch for the forward
calculation according to step 1.
Fig. 8. A neuron of the hidden layer.
During the training process, calculation in the forward direction is
done to predict the roughness, and then errors are propagated in the
that the global minima can be found. The weights of the hidden layer are backward direction to the subsequent layers. Here an epoch is composed
updated using: of a forward calculation and a back propagation. As shown in step 3,
dE once the preset number of epochs is reached, the training process is
Δwik = − α (15) terminated.
dwik
The MLP model is very reliable in modeling nonlinear processes.
Similarly, the weights of the output layer are updated using: Determining the number of neurons in the hidden layer is very impor­
dE tant. The number of neurons in the hidden layer is generally optimized
Δuk = α (16) according to the final mean square error (MSE).
duk

where α is the learning rate.

547
A. Yaseer and H. Chen
3. Experiments and results
To validate the proposed methods, a robotic wire arc welding process
is developed to perform a WAAM process.
3.1. Experiment procedure
An ABB 6-axis robot (model IRB140) is used to control the welding
torch along a weaving path as shown in Fig. 12.
The experimental setup is shown in Fig. 13 and the robot is
controlled using its controller (model-IRC5). The welding machine is
Thermal Arc PowerMaster 500 and the wire feeder is Miller r-115 with
1.2 mm diameter steel wire (ER70S-6). This wire contains higher level of
Si and Mn than other carbon steel wires, enabling them to weld steels
with a moderate amount of rust. Also, the presence of Si provides more
puddle and wetting action. The shielding gas with a combination of 90%
Argon and 10% CO2 is utilized with a gas flow rate 31.78 ft3/h. Direct
current reverse polarity (DCRP) is used for connecting the welding
machine with the MIG gun and substrate. A 3D laser scanner (COGNEX
DS 1300R) is applied along with its controller (model- VC5) to obtain the
point cloud data for measuring the surface roughness. The spray transfer
mode of GMAW is implemented to reduce spatter and produce quality
weld. Robot programming is done using the robot teach pendant
(Fig. 13) to move the robot to the desired positions. The program for
moving the robot with the MIG gun along the weaving path is written
using the ABB robotic programming language RAPID with the software
RobotStudio.
According to [37], with the same wire feed speed and weaving speed,
the bead width increases and height decreases with increasing weaving
amplitude and thus makes a flatter surface. In this research, numerous
trial experiments were done and we found that with increasing travel
speed and wire feed speed using weaving path, bead cross-section tends
to be parabolic instead of flat. At higher wire feed speed, it also causes
more spatter. Considering these issues, we chose the parameters for the
experiments whose ranges are given in Table 1.
The Taguchi Method is one of the most used Design of Experiments
(DOE) methods in welding to design the experiment matrix. Using
Minitab, the L25 Taguchi orthogonal array was designed which has 4
parameters with 5 levels. The 5 levels of each parameter were chosen
from its range shown in Table 1. Each of the 25 experiments of the
Taguchi L25 array was conducted twice, hence a total of 50 experiments
were performed. All these 50 datasets were used to train and validate
(using cross-validation) the proposed Random Forest and MLP methods.
In order to test the proposed methods, another 22 experiments were
conducted. The values of the parameters of these 22 experiments were
chosen randomly (not using DOE) based on the range shown in Table 1.
3.2. Roughness measurement
After a part is fabricated, the part surface was scanned using a
COGNEX laser scanner, and the point cloud data were collected using
Fig. 12. A weaving path with three parameters: weaving amplitude, weaving
wavelength and travel speed.
548
Journal of Manufacturing Processes 70 (2021) 543–552
Fig. 13. Experimental setup.
Table 1
Input Process Parameters.
Parameters Ranges
Wire feed speed (m/min) 1.52–2.28
Travel speed (mm/s) 3–5
Weaving wavelength (mm) 2–4
Weaving amplitude (mm) 6–12
the Cognex Designer.
Among various definitions of surface roughness such as average
roughness, root-mean-square (RMS) roughness, and maximum peak-to-
valley roughness, used in industries, average roughness is the most
effective index of product quality [47] which is used in this research.
Some unique observations were found in this research related to surface
roughness. As shown in Fig. 14(a), the surface of the part is smooth but
not flat; the surface is flat but not smooth in Fig. 14(b). To reduce the
surface roughness, the deposited surface should be smooth as well as
flat. Our approach in this research for measuring the surface roughness,
as shown in Fig. 15, considers both these aspects. Some other fabricated
parts are shown in Fig. 14(c) and Fig. 14(d).
According to Fig. 15, the point cloud data were collected after
scanning a part using a COGNEX laser scanner. Fig. 16 shows the
experimental setup. Since some irregularities are generally found at the
starting and ending areas, we only consider the middle portion of the
deposited bead as the region of interest (ROI) for roughness measure­
ment. Processing the point cloud data, plane fitting, and roughness
calculation were done using MATLAB.
As shown in Fig. 17, the 1st plane was fitted to the point cloud data
using Matlab. But if roughness is calculated with respect to this 1st fitted
plane, it would produce a high roughness value (because many of the
beads not being flat yet smooth) which is not a true representation of
roughness. So, the 2nd plane was fitted using all the points that were
situated above the 1st plane. After that, the average surface roughness,
Ra was calculated with respect to the 2nd fitted plane using,
1 ∑N
|axi + byi + czi + 1|
Ra = √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ (17)
N i=1 a2 + b2 + c2
where a, b and c are the coefficients of the fitted 2nd plane and (xi, yi, zi)
are the coordinates of the ith points in the point cloud data. N is the total
number of points above the 1st fitted plane. Roughness values for all of
the 72 fabricated parts (50 for training/validation and 22 for testing)
were obtained with a range of 0.21–0.52 mm.
Some pictures of produced flat and smooth weaving beads with small
A. Yaseer and H. Chen Journal of Manufacturing Processes 70 (2021) 543–552

Fig. 14. Various types of weaving beads: (a) Bead is smooth but not flat; (b) Bead is flat but not smooth; (c) Bead is both flat and smooth; (d) Bead is neither flat
nor smooth.

Fig. 15. Steps of roughness measurement.

Fig. 17. Plane fitting from point cloud.

roughness (Ra) values together with the process parameters-wire feed

speed (WFS), travel speed (TS), ratio of wire feed speed to travel speed
(R), weaving wavelength (WL), and weaving amplitude (WA) are shown
in Fig. 18.
From Fig. 18, we can conclude that for producing both flat and
smooth weaving beads with minimized roughness, the wavelength
should be larger and the ratio of wire feed speed to travel speed should
be around 7. It was investigated in [14] that the relationships between
the roughness and the WAAM process parameters such as the wire feed
speed and travel speed are non-linear (for overlapped straight-pass
Fig. 16. Laser scanning setup. deposition). In this research we found that, for weaving path, the rela­
tionship between weaving amplitude and roughness is also non-linear as
shown in Fig. 19. Random Forest and MLP are capable of modeling non-
linear systems.

549
A. Yaseer and H. Chen Journal of Manufacturing Processes 70 (2021) 543–552

where, Zmax and Zmin are the maximum and minimum values of the raw
datasets, respectively.
K-fold cross-validation (10-fold in this research) was also used. The
total training datasets were divided into 10 subsets. In each iteration, 9
subsets were used for training and the remaining 1 subset was used for
validation. Since there are 50 sets of data in the training datasets, each
subset has 5 sets of data. Thus, in each fold, 45 sets of data were used for
training and the rest 5 sets of data were used for validation. Both 5-fold
and 10-fold cross validations (the 2 most popular values of K in K-fold
cross-validation) were used in this research. It was found that 10-fold
cross-validation gave better results than 5-fold cross-validation did.
In MLP, various activation functions like sigmoid, ReLU, and tanh
were used in the hidden layer, and the linear activation function was
used in the output layer. Various optimizers such as adam, sgd, and
adabound were used in MLP, but the state-of-the-art adabound opti­
mizer gave the best results with the 4–10-1 structure. The hyper­
parameters giving the best results in Random Forest and MLP are shown
in Table 2.
After cross-validation, the developed Random Forest and MLP
models, trained and validated using the 50 datasets, were tested using
the 22 testing datasets.

3.4. Prediction results

The prediction graph of Random Forest for the 50 cross-validated

training datasets is shown in Fig. 20. We can see that the data points
scattered closely around the 45◦ line which indicate that the error is
small. The mean squared error is 0.00052 and the average error per­
centage (also known as mean absolute percentage error (MAPE)) is
5.64% (maximum 12.58% and minimum 0.47%).
The prediction graph of Random Forest for the 22 testing datasets is
shown in Fig. 21. The mean squared error is 0.00063 and the average
Fig. 18. Weaving beads and their parameters and roughness values.
error percentage is 5.77% (maximum 13.74% and minimum 1.25%).
The prediction graph of MLP for the 50 cross-validated training
datasets is shown in Fig. 22. The best MLP architecture is 4-10-1 with a
minimum MSE of 0.00052 and an average error percentage of 4.94%
(maximum 17.10% and minimum 0.23%).
The prediction graph using MLP for the 22 testing datasets is shown
in Fig. 23. The mean squared error is 0.00087 and the average error
percentage is 7.37% (maximum 20.32% and minimum 0.80%).
In both Random Forest and MLP, the validation (cross-validated)
MSE and MAPE were better than the testing MSE and MAPE. For the 50
training (cross-validated) datasets, even though the MAPE of MLP is
slightly better than that of Random Forest, but the maximum error rate
of MLP (17.10%) is larger than that of Random Forest (12.58%). The
MSE is about the same for both methods. Hence, for the 50 training
datasets, we can conclude that Random Forest provides better results
than MLP.
For the 22 testing datasets, Random Forest has better results than
Fig. 19. Relationship between weaving amplitude and roughness. MLP for all the performance indices. Moreover, Random Forest took less
time compared to MLP since Random Forest has few hyperparameters.
3.3. Implementation
Table 2
Python libraries NumPy and Pandas have been used for processing Machine Learning hyperparameters.
the data. MLP was implemented using the Python library Keras and Machine Leaning Method Hyperparameter Value
Random Forest using the Python library ScikitLearn. Random Forest No. of trees 100
Before training the network, the data were scaled (using Min­ Maximum depth 6
MaxScaler in Python) to ensure that each welding parameter has an Maximum features None
identical effect on the network. Hence, all training and testing datasets Maximum samples split None
Maximum samples leaf None
were scaled into the closed interval [0,1]. The scaled value (Zscaled) for
No. of hidden layer 1
each raw input or output (Z) was calculated using- No. of neurons in hidden layer 10
z − zmin Activation function in hidden layer ReLU
(18) MLP
Zscaled = Activation function in output layer Linear
zmax − zmin
Optimizer Adabound
Learning rate 0.001

550
A. Yaseer and H. Chen
Fig. 20. Roughness prediction graph of Random Forest for cross-validation.
Fig. 21. Roughness prediction graph of Random Forest for the testing datasets.
Fig. 22. Roughness prediction graph of MLP for cross-validation.
Fig. 23. Roughness prediction graph of MLP for the testing datasets.
Even though MLP is well known for its good prediction results for non-
linear problems [58], Random Forest outperforms MLP especially with
smaller datasets. In this paper, several examples were presented and the
results demonstrate that Random Forest outperforms MLP.
551
Journal of Manufacturing Processes 70 (2021) 543–552
4. Conclusion
The paper investigates the layer roughness for robotic WAAM using a
weaving path. Random Forest and MLP models were developed that
were able to predict the layer roughness given a set of robotic WAAM
parameters. Experiments were performed and the results show that both
methods are effective in modeling the relationship between robotic
WAAM parameters and layer roughness, however, the modeling method
using Random Forest achieves better results than that using MLP. The
developed Random Forest model was able to predict the weld bead
roughness from the given input parameters with an average error rate
below 6% for both validation and testing. Since Random Forest is not
computationally expensive, it can be extended for online roughness
modeling for quality control purposes.
Declaration of competing interest
The authors declare that they have no known competing financial
interests or personal relationships that could have appeared to influence
the work reported in this paper.
References
[1] Ding D, Pan Z, Cuiuri D, Li H. Wire-feed additive manufacturing of metal
components: technologies, developments and future interests. Int. J. Adv. Manuf.
Technol. 2015;81(1):465–81. https://doi.org/10.1007/s00170-015-7077-3. /10/
01 2015.
[2] Mehnen J, Ding J, Lockett H, Kazanas P. Design for Wire and Arc Additive Layer
Manufacture. Int. J. Prod. Dev. 2011;19(01/01). https://doi.org/10.1007/978-3-
642-15973-2_73.
[3] Coykendall J, Cotteleer M, Holdowsky J, Mahto M. 3D Opportunity in Aerospace
and Defense: Additive Manufacturing Takes Flight. Deloitte University Press; 2014.
[4] Uziel A. Looking at large-scale, Arc-based Additive Manufacturing. Weld. J. 2016;
95(4):42–6. 04//. [Online]. Available, http://libproxy.txstate.edu/login?url=http
://search.ebscohost.com/login.aspx?direct=true&db=egs&AN=113833486&site
=ehost-live&scope=site.
[5] Huang S, Liu P, Mokasdar A, Liang H. Additive manufacturing and its societal
impact: a literature review. Int. J. Adv. Manuf. Technol. 2012;67(07/01). https://
doi.org/10.1007/s00170-012-4558-5.
[6] Jafari D, Vaneker T, Gibson I. Wire and arc additive manufacturing: Opportunities
and challenges to control the quality and accuracy of manufactured parts.
Materials & Design 2021;202:109471. https://doi.org/10.1016/j.
matdes.2021.109471. ISSN 0264-1275.
[7] Valentin M, Arnaud C, Kling R. Additive Manufacturing by Wire Based Laser Metal
Deposition (SPIE LASE). SPIE; 2019.
[8] Xiong J, Li Y-J, Yin Z-Q, Chen H. Determination of surface roughness in wire and
arc additive manufacturing based on laser vision sensing. Chinese Journal of
Mechanical Engineering 2018;31(12/01):74. https://doi.org/10.1186/s10033-
018-0276-8.
[9] Wu CS, Hu ZH, Zhong LM. Prevention of humping bead associated with high
welding speed by double-electrode gas metal arc welding. Int. J. Adv. Manuf.
Technol. 2012;63(5):573–81. https://doi.org/10.1007/s00170-012-3944-3. 11/01
2012.
[10] Soderstrom E, Mendez P. Humping mechanisms present in high speed welding. Sci.
Technol. Weld. Join. 2006;11(5):572–9. https://doi.org/10.1179/
174329306X120787. 09/01 2006.
[11] Ye D, Wu D, Hua X, Xu C, Wu Y. Using the multi-wire GMAW processes for
controlling the formation of humping. Welding in the World 2017;61(4):649–58.
https://doi.org/10.1007/s40194-017-0458-5. 06/01 2017.
[12] Xu X, Ding J, Ganguly S, Diao C, Williams S. Preliminary investigation of building
strategies of maraging steel bulk material using wire + arc additive manufacture.
J. Mater. Eng. Perform. 2019;28(2):594–600. https://doi.org/10.1007/s11665-
018-3521-5. 02/01 2019.
[13] Li Z, Ma G, Zhao G, Yang M, Xiao W. Weave bead welding based wire and arc
additive manufacturing technology. In: Wang S, Price M, Lim MK, Jin Y, Luo Y,
Chen R, editors. Recent Advances in Intelligent Manufacturing, Singapore. Springer
Singapore; 2018. p. 408–17.
[14] Xia C, Pan Z, Polden J, Li H, Xu Y, Chen S. Modelling and prediction of surface
roughness in wire arc additive manufacturing using machine learning. J. Intell.
Manuf. 2021. https://doi.org/10.1007/s10845-020-01725-4. 01/23 2021.
[15] Aldalur E, Veiga F, Suárez A, Bilbao J, Lamikiz A. Analysis of the wall geometry
with different strategies for high deposition wire arc additive manufacturing of
mild steel. Metals 2020;10(7). https://doi.org/10.3390/met10070892.
[16] Khan AU, et al. Development of barrel finishing machine to improve surface finish
of the wire arc additive manufactured parts. Procedia CIRP 2020;91:330–5. 2020/
01/01, https://doi.org/10.1016/j.procir.2020.02.184.
[17] Nagamatsu H, Sasahara H, Mitsutake Y, Hamamoto T. Development of a
cooperative system for wire and arc additive manufacturing and machining. Addit
A. Yaseer and H. Chen
Manuf 2020;31:100896. https://doi.org/10.1016/j.addma.2019.100896. 01/01/
2020.
[18] Alonso U, Veiga F, Suárez A, Artaza T. Experimental investigation of the influence
of wire arc additive manufacturing on the machinability of titanium parts. Metals
2020;10(1). https://doi.org/10.3390/met10010024.
[19] Li F, Chen S, Shi J, Tian H, Zhao Y. Evaluation and optimization of a hybrid
manufacturing process combining wire arc additive manufacturing with milling for
the fabrication of stiffened panels. Appl. Sci. 2017;7(12). https://doi.org/10.3390/
app7121233.
[20] T. Hongyu, L. Zhenyang, L. Fang, and C. Shujun, "Predictive modeling of surface
roughness based on response surface methodology after WAAM," in 2019
International Conference on Electronical, Mechanical and Materials Engineering
(ICE2ME 2019), 2019/03 2019: Atlantis Press, 47–50, doi: https://doi.org/10.
2991/ice2me-19.2019.11. [Online]. Available: doi:https://doi.org/10.2991
/ice2me-19.2019.11.
[21] Thakur VS, et al. Laser polishing of wire arc additive manufactured SS316L. In:
Shunmugam MS, Kanthababu M, editors. Advances in additive manufacturing and
joining, Singapore. Springer Singapore; 2020. p. 127–35.
[22] Zhang Z, Sun C, Xu X, Liu L. Surface quality and forming characteristics of thin-
wall aluminium alloy parts manufactured by laser assisted MIG arc additive
manufacturing. International Journal of Lightweight Materials and Manufacture
2018;1(2):89–95. https://doi.org/10.1016/j.ijlmm.2018.03.005. 06/01/ 2018.
[23] Xiong J, Li Y, Li R, Yin Z. Influences of process parameters on surface roughness of
multi-layer single-pass thin-walled parts in GMAW-based additive manufacturing.
J. Mater. Process. Technol. 2018;252:128–36. https://doi.org/10.1016/j.
jmatprotec.2017.09.020. 02/01/ 2018.
[24] Yehorov Y, da Silva L João, Scotti A. Balancing WAAM production costs and wall
surface quality through parameter selection: a case study of an Al-Mg5 alloy
multilayer-non-oscillated single pass wall. Journal of Manufacturing and Materials
Processing 2019;3(32). https://doi.org/10.3390/jmmp3020032. 04/16.
[25] Foorginejad A, Azargoman M, Mollayi N. Application of the extreme learning
machine for modeling the bead geometry in gas metal arc welding process 2020;09
(28).
[26] Huang Y, Yang D, Wang K, Wang L, Fan J. A quality diagnosis method of GMAW
based on improved empirical mode decomposition and extreme learning machine.
J. Manuf. Process. 2020;54:120–8. https://doi.org/10.1016/j.
jmapro.2020.03.006. 06/01/ 2020.
[27] ULAS OA Mustafa, GURGENC Turan, Cihan ÖZEL. A new approach for prediction
of the wear loss of PTA surface coatings using artificial neural network and basic,
kernel-based, and weighted extreme learning machine. Friction 2020;8(6):
1102–16. https://doi.org/10.1007/s40544-017-0340-0. 12-05 2020.
[28] Liang R, Yu R, Luo Y, Zhang Y. Machine learning of weld joint penetration from
weld pool surface using support vector regression. J. Manuf. Process. 2019;41:
23–8. https://doi.org/10.1016/j.jmapro.2019.01.039. 05/01/ 2019.
[29] Mollayi N, Eidi M. Application of multiple kernel support vector regression for
weld bead geometry prediction in robotic GMAWProcess. International Journal of
Electrical and Computer Engineering 2018;8:2310–8. 08/01, 10.11591/ijece.v8i4.
pp2310-2318.
[30] Girinath B, Shanmugam NS, Sankaranarayanasamy K. Weld bead graphical
prediction of cold metal transfer weldment using ANFIS and MRA model on Matlab
platform. Simulation 2019;95(8):725–36. https://doi.org/10.1177/
0037549718809162.
[31] Subashini L, Vasudevan M. Adaptive neuro-fuzzy inference system (ANFIS)-based
models for predicting the weld bead width and depth of penetration from the
infrared thermal image of the weld pool. Metall. Mater. Trans. B 2012;43(1):
145–54. https://doi.org/10.1007/s11663-011-9570-x. 02/01 2012.
[32] Ma Z. Vision-sensing and bead width control of the first layer of multi-layered part
in GTAW based additive manufacturing. University of Wollongong; 2017.
[33] Xiong J, Zhang G, Qiu Z, Li Y. Vision-sensing and bead width control of a single-
bead multi-layer part: material and energy savings in GMAW-based rapid
manufacturing. J. Clean. Prod. 2013;41:82–8. 02/01, https://doi.org/10.1016/j.
jclepro.2012.10.009.
[34] Almeida P, Williams S. Innovative Process Model of Ti–6Al–4V Additive Layer
Manufacturing Using Cold Metal Transfer (CMT). 2010. p. 25–36.
[35] Xiong J, Zhang G, Hu J, Wu L. Bead geometry prediction for robotic GMAW-based
rapid manufacturing through a neural network and a second-order regression
analysis. J. Intell. Manuf. 2014;25(1):157–63. https://doi.org/10.1007/s10845-
012-0682-1. 02/01 2014.
[36] Ma G, Zhao G, Li Z, Xiao W. A path planning method for robotic wire and arc
additive manufacturing of thin-walled structures with varying thickness. IOP
552
Journal of Manufacturing Processes 70 (2021) 543–552
Conference Series: Materials Science and Engineering 2019;470:012018. https://
doi.org/10.1088/1757-899X/470/1/012018.
[37] Ma G, Zhao G, Li Z, Yang M, Xiao W. Optimization strategies for robotic additive
and subtractive manufacturing of large and high thin-walled aluminum structures.
Int. J. Adv. Manuf. Technol. 2019;101(5):1275–92. https://doi.org/10.1007/
s00170-018-3009-3. 04/01 2019.
[38] Suryakumar S, Karunakaran KP, Bernard A, Chandrasekhar U, Raghavender N,
Sharma D. Weld bead modeling and process optimization in hybrid layered
manufacturing. Comput. Aided Des. 2011;43(4):331–44. 2011/04/01, https://doi.
org/10.1016/j.cad.2011.01.006.
[39] Liang Z, et al. Characteristics of metal droplet transfer in wire-arc additive
manufacturing of aluminum alloy. Int. J. Adv. Manuf. Technol. 2018;99. https://
doi.org/10.1007/s00170-018-2604-7. 11/01.
[40] Martina F, Mehnen J, Williams SW, Colegrove P, Wang F. Investigation of the
benefits of plasma deposition for the additive layer manufacture of Ti–6Al–4V.
J. Mater. Process. Technol. 2012;212(6):1377–86. https://doi.org/10.1016/j.
jmatprotec.2012.02.002. 06/01/ 2012.
[41] Ding D, Pan Z, Cuiuri D, Li H. Process planning for robotic wire and arc additive
manufacturing. Conference on Industrial Electronics and Applications (ICIEA)
2015:2000–3. https://doi.org/10.1109/ICIEA.2015.7334441.
[42] Li F, Chen S, Wu Z, Yan Z. Adaptive process control of wire and arc additive
manufacturing for fabricating complex-shaped components. Int. J. Adv. Manuf.
Technol. 2018;96:1–9. 04/01, https://doi.org/10.1007/s00170-018-1590-0.
[43] Ding D, Pan Z, Cuiuri D, Li H, Duin S van, Larkin N. Bead modelling and
implementation of adaptive MAT path in wire and arc additive manufacturing.
Robot. Comput. Integr. Manuf. 2016;39:32–42. https://doi.org/10.1016/j.
rcim.2015.12.004. 06/01/ 2016.
[44] Ismail MIS, Okamoto Y, Okada A. Neural network modeling for prediction of weld
bead geometry in laser micro-welding. Advances in Optical Technologies 2013;
2013. https://doi.org/10.1155/2013/415837. 12/09.
[45] Dutta P, Pratihar D. Modeling of TIG welding process using conventional
regression analysis and neural network-based approaches. J. Mater. Process.
Technol. 2007;184:56–68. 04/01, https://doi.org/10.1016/j.jmatprotec.2006.
11.004.
[46] Zhao D, Wang Y, Liang D, Ivanov M. Performances of regression model and
artificial neural network in monitoring welding quality based on power signal.
J Mater Res Technol 2020;9(2):1231–40. https://doi.org/10.1016/j.
jmrt.2019.11.050. 03/01/ 2020.
[47] Hossain MS, Ahmad N. Artificial intelligence based surface roughness prediction
modeling for three dimensional end milling 2012;45. 09/01.
[48] Yazdi MR Soleymani, Khorram A. Modeling and optimization of milling process by
using RSM and ANN methods. International Journal of Engineering and
Technology 2010;2:474–80. 01/01, https://doi.org/10.7763/IJET.2010.V2.167.
[49] Aykut Ş. Surface roughness prediction in machining castamide material using ANN.
Acta Polytechnica Hungarica 2011;8:2011–21. 01/01.
[50] Rai R, Shettigar A, Rao SS, Shriram. Development of a surface roughness prediction
system for machining of hot chromium steel (AISI H11) based on artificial neural
network. J. Eng. Appl. Sci. 2010;5:53–9. 01/01.
[51] Sumesh A, Rameshkumar K, Mohandas K, Babu RS. Use of machine learning
algorithms for weld quality monitoring using acoustic signature. Procedia
Computer Science 2015;50:316–22. 2015/01/01, https://doi.org/10.1016/j.pr
ocs.2015.04.042.
[52] Seo BW. Machine learning for prediction of arc length for seam tracking in tandem
welding. Journal of Welding and Joining 2020;38(3):241–7. 06, https://doi.
org/10.3345/kjp.2020.38.3.241.
[53] Zhu H. Deep Learning-Based Classification of Weld Surface Defects. 2019.
[54] Zhang Z, Yang Z, Ren W, Wen G. Random forest-based real-time defect detection of
Al alloy in robotic arc welding using optical spectrum. J. Manuf. Process. 2019;42:
51–9. https://doi.org/10.1016/j.jmapro.2019.04.023. 06/01/ 2019.
[55] Chen Z. Visual Sensing and Defect Detection of Gas Tungsten Arc Welding. 2019.
[56] Deng J, Xu Y, Zuo Z, Hou Z, Chen S. Bead geometry prediction for multi-layer and
multi-bead wire and arc additive manufacturing based on XGBoost. In: Chen S,
Zhang Y, Feng Z, editors. Transactions on Intelligent Welding Manufacturing,
Singapore. Springer Singapore; 2019. p. 125–35.
[57] Breiman L. Random Forests. Mach. Learn. 2001;45(1):5–32. https://doi.org/
10.1023/A:1010933404324. 10/01 2001.
[58] Roßbach P. Neural networks vs. random forests – does it always have to be deep
learning? Frankfurt School of Finance & Management 2018.