You are on page 1of 15

J Intell Manuf

DOI 10.1007/s10845-015-1187-5

Optimization of laser brazing onto galvanized steel based on


ensemble of metamodels
Qi Zhou1 · Youmin Rong1 · Xinyu Shao1 · Ping Jiang1 · Zhongmei Gao1 ·
Longchao Cao1

Received: 1 November 2015 / Accepted: 22 December 2015


© Springer Science+Business Media New York 2016

Abstract Laser brazing (LB) provides a promising way to Keywords Laser brazing · Bead profile · Ensemble ·
join the galvanized steel in automotive industry for its sig- Metamodel · Genetic algorithm · Crimping butt
nificant advantages including high speed, small heat-affected
zone, and high welding seam quality. The process parame-
ters of LB have significant effects on the bead profile and Introduction
hence the quality of joint. Since the relationships between
the process parameters and bead profile cannot be expressed For the automobile industry, there are two main trends in
explicitly, it is impractical to determine the optimal process design of the car body. The first one is to improve corro-
parameters intuitively. This paper proposes an optimiza- sion resistance by using either aluminum or galvanized steel.
tion methodology by combining genetic algorithm (GA) and The other one is to reduce the total weight and at the same
ensemble of metamodels (EMs) to address the process para- time guarantee the structural strength by using light mate-
meters optimization of the bead profile in LB with crimping rials or appropriately selecting the thickness of sheet metal
butt. Firstly, Taguchi experimental design is adopted to gen- throughout the car body. With the superiority in terms of cor-
erate the experimental points. Secondly, the relationships rosion resistance, cost and hardness, galvanized steel might
between process parameters (i.e., welding speed, wire feed be more attractive in automotive industry than aluminum in
rate, gap) and the bead geometries are fitted using EMs based the future (Sinha et al. 2013; Colombo and Previtali 2014;
on the experimental data. The comparative results show that Zhao et al. 2012; Rong et al. 2015). However, the welding
the EMs can take advantage of the prediction ability of each stability of galvanized steel process by laser welding and
stand-alone metamodel and thus decrease the risk of adopting conventional fusion welding techniques (e.g., tungsten inert
inappropriate metamodels. Then, the GA is used to facili- gas arc welding (TIG), metal inert-gas welding (MIG) and
tate design space exploration and global optimum search. resistance welding) are weak because the evaporation tem-
Besides, the main effects and contribution rates of multiple perature (1179 K) is lower than that of the galvanized steel
process parameters on bead profile are analyzed. Eventually, (1823 K), which usually causes some porosity defects (Chen
the verification experiments are carried out to demonstrate et al. 2009; Qin et al. 2015; Mei et al. 2009; Ruggiero et al.
the effectiveness and reliability of the obtained optimal para- 2011). Fortunately, Laser brazing (LB), which combines the
meters. Overall, the proposed hybrid approach, GA–EMs, advantages of laser welding and brazing, provides a promis-
exhibits great capability of guiding the actual LB processing ing way to join the galvanized steel in automotive industry
and improving welding quality. for its significant advantages including but not limited to high
speed, small heat-affected zone, high welding seam quality
B Ping Jiang (Nasiri et al. 2013; Chen et al. 2011; Roos and Schmidt 2014).
jiangping@mail.hust.edu.cn It is very important to control quality of joint that is highly
1
relevant to weld bead profile (Zhao et al. 2012; Rong et al.
The State Key Laboratory of Digital Manufacturing
Equipment and Technology, School of Mechanical Science
2015). Generally, many process parameters, (e.g., wire feed
and Engineering, Huazhong University of Science & rate, weld speed, prescribed gap, laser power, etc.) have
Technology, Wuhan 430074, People’s Republic of China effects on the quality of joint. However, how to select appro-

123
J Intell Manuf

priate process parameters that can produce required seam methodology is applied to obtain optimum process parame-
quality for crimping butt of laser brazing on galvanized ters that can produce required seam quality for crimping butt
steel is rarely reported in automobile industry. Another sit- of laser brazing on galvanized steel. Firstly, a 3-factor and
uation is that the relationships between the input variables 5-level experimental program (L 25 35 ) on galvanized steel
and bead profile cannot be expressed explicitly (it is actually with IPG YLR-4000 fiber laser and wire filler CuSi3 has
a black-box model), which makes it impractical to deter- been conducted. Secondly, the advantages of ensemble of
mine the optimal process parameters intuitively. Besides, three stand-alone metamodels, including Kriging, support
trial and error methods often lead to sub-optimal solution vector regression (SVR), and RBF for fitting the relation-
and in some extent may result in a tremendous waste of ship between the process parameters and bead profile are
resources. A promising strategy is to utilize metamodels, discussed. Thirdly, the main effects and contribution rates
which are also called approximation models as they pro- of multiple process parameters on the bead profile are ana-
vides “model of model”, fitting the relationship between lyzed. Then, based on the EMs, genetic algorithm is utilized
the input process parameters and the output performance to solve the LB parameter optimization problem. Finally, the
for the purpose of welding design optimization. Ruggiero consistency between the obtained optimum and experiment
et al. (2011) investigated dissimilar full depth CO2 laser butt results are discussed.
welding of low carbon steel and austenitic steel AISI 316 The rest of this paper is organized as follows. In “Exper-
using RSM. Chaki et al. (2015) integrated Artificial Neural iment procedure” section, details of the experimental proce-
Networks (ANN) and non-dominated sorting genetic algo- dure of laser brazing onto galvanized steel are described. The
rithm (NSGA-II) approach for prediction and optimization proposed hybrid optimization methodology is presented in
of quality characteristics during pulsed Nd:YAG laser cut- “Background and proposed approach” section. In “Results
ting. Katherasan et al. (2014) used Artificial Neural Networks and discussions” section, the prediction ability of ensem-
(ANN) during the optimization of Flux cored arc welding ble of metamodel and stand-alone metamodel is compared.
process parameters to obtain desire processing output char- Meanwhile, the main effects of multiple process parameters
acteristics. Singh et al. (2014) predicted weld-bead geometry on the bead profile and verifications of consistency between
and hardness profile in ytterbium fiber laser welding of plain the obtained optimum and experiment results are presented.
carbon steel (DC05) using neural network (NN) and genetic Finally, in “Conclusion” section, the conclusion remarks and
algorithms (GA). Lim (2014) combined Kriging metamodel future work are given.
and GA for welding parameters optimization of TIG weld-
ing onto two thin pieces of titanium sheet metal. Gao et al.
(2013) used radial basis function neural network (RBFNN) to Experiment procedure
improve the performance of the Kalman Filtering (KF) algo-
rithm and constructed an adaptive filter to compensate for the Laser brazing specimen
errors of seam tracking. These practices do help designers
grasp the relationships between the input process parame- The workpiece used for welding was a 0.8 mm-thick gal-
ters and output performance and improve the efficiency of vanized DP590 steel. The welding joints were performed
welding design and optimization. However, the lack of suf- in a crimping butt configuration with specimen sizes of
ficient information describing the relationship between the 300 × 100 × 0.8 mm. To eliminate the interference from
process parameters and output performance makes it diffi- oxidation film and prevent the welding bead being polluted
cult for engineers to determine which metamodel is the best by oil, the specimen have been pretreated or degreased with
to be adopted. In addition, the accuracy of the constructed acetone before welding.
stand-alone metamodel largely depends on the current train-
ing sampling set and the problem properties at hand, so it Procedure of laser brazing
cannot be guaranteed that the selected metamodel will always
perform the best as another training data available or prob- Figure 1 demonstrates the laser brazing system used in this
lems are changed. work. The laser welder adopted here is IPG YLR-4000
To overcome the above mentioned shortcomings, this ytterbium-doped fiber laser. The maximum average power
paper proposes a hybrid optimization methodology that com- of the laser was 4000 W, which was a continuous laser. In
bines GA and ensemble of metamodel (EMs) for parameters this experiment, the power of the fiber laser was set to be
optimization of LB. The EMs is adopted to fit the relation- 3200 W. Laser was delivered through the optical fiber to the
ships between the process parameters and bead profile, which laser welder head, where a defocusing lens with the defocus
aims to take advantage of the prediction ability of stand-alone length of 45 mm was installed. Then the diameter of light
metamodels. The GA is used to facilitate design space explo- spot was enlarged to about 3.2 mm, which aimed to fully
ration and global optimum search. The proposed GA–EMs melt filler wire during the welding process. The wire feeding

123
J Intell Manuf

Fig. 1 Laser brazing system. a


Laser equipment and b Fiber optics
schematic of the laser brazing
setup
Nozzle Wire feeder
machine
Laser beam
Filler wire

Work piece
Worktable

(a) (b)

Table 1 Chemical
Element Si Al Fe Mn P Pb Sn Zn Cu
concentration of each element in
CuSi3 wire Percent 3.01 0.0042 0.0535 0.98 0.0072 <0.002 0.0248 <0.01 Balance

WT wire feed rate and gap. Considering these two circumstances


(Zhao et al. 2012; Rong et al. 2015): (a) low wire feed rate or
high welding speed are not good for diffusion of liquid filler
ELL ELR wire and may result in incomplete fusion, defects of filled
groove or even spatter (b) excessive wide gap may lead to
dent and collapse, the following bounds are used for process
WB
parameters:

2.6 m/ min ≤ W F ≤ 3.4 m/ min, 0.8 m/ min ≤ W S


≤ 1.6 m/ min, 0 mm ≤ G A P ≤ 0.8 mm.
Fig. 2 Schematic of the bead profile
The goal of data sampling is to minimize the influence
of errors in physical experiments to the response func-
machine and the laser welding header were installed on the tions, while allowing the designers to build metamodels
robot ABB IRC5 M2004. The filler wire was CuSi3 whose more efficiently (Zadeh et al. 2009). In this work, Taguchi
diameter was 1.6 mm. Table 1 shows the chemical compo- experimental design (Taguchi 1978) proposed by Taguchi is
sition in weight percent of the filler wire. The shielding gas adopted to generate the training points as it is able to provide
was not utilized during the welding. good uniform and space-filling sample points in the design
Figure 2 is the schematic of seam geometry, in which the space. For each input variables, five-levels with an interval
bead profile is illustrated clearly. WT and WB denote top of 0.2 units are considered, which showed in Table 2. As can
width and bottom width of the weld bead, respectively. ELL be seen in Table 2, the current problem is a three-factor and
and ELR represent the efficient length of connection on the five-level problem, hence, a L25 orthogonal array could be
left side and the right side, respectively. generated as demonstrated in Table 3. Table 3 illustrates the
total 25 experiments conducted for the optimization of the
Experiment design and results input parameters on the bead geometries.
According to the L25 orthogonal array listed in Table 3,
Generally, the process parameters of LB welding, e.g., weld- WS and WF were set in the ABB IRC5 M2004 robot and wire
ing speed (WS), wire feed rate (WF), gap (GAP), laser power, filler machine, and the feeler gauge were used to ensure the
torch angle, have influence on the bead profile. According to GAP of the crimping butt. The experiment results are listed
a great number of actual tests, primary investigations and in Table 3. In Table 3, the four output seam geometries of LB
surveys (Li et al. 2007; Tan et al. 2013; Cao et al. 2013; were measured on measurement software CSM1 after the
Lin et al. 2013), LB bead profile is mainly determined by transverse samples were cut from each weldment by electric
three welding process parameters including welding speed, cutting machine STDX600. To obtain a clear image of the

123
J Intell Manuf

Table 2 Welding parameters and their levels SVR are constructed to estimate the bead geometries of the
Parameters Units Notations Factor levels galvanized steel welded using LB. The obtained ensemble of
metamodels will be used in the LB optimization process if
1 2 3 4 5
their accuracy achieved.
Welding speed m/min WS 0.8 1.0 1.2 1.4 1.6 The remainder of this section describes the proposed LB
Wire feed rate m/min WF 2.6 2.8 3.0 3.2 3.4 process parameters optimization approach in more details as
Gap mm GAP 0.0 0.2 0.4 0.6 0.8 follows: “Stand-alone metamodeling methods” section gives
the background of the three related stand-alone metamodel-
ing techniques, “Ensemble of metamodels” section presents
Table 3 Taguchi L25 design matrix and experimental results the details of how to combining multiple metamodels in
No. WS WF GAP WT WB ELL ELR a weighted-sum formulations, “Genetic algorithm” section
(m/min) (m/min) (mm) (mm) (mm) (mm) (mm) gives the main operations of GA and “Steps for the pro-
posed optimization approach” section presents a step-by-step
1 0.8 2.6 0.0 3.43 2.48 4.12 0.61
description of the proposed LB process parameters optimiza-
2 0.8 2.8 0.2 2.55 2.07 3.61 0.95
tion approach.
3 0.8 3.0 0.4 2.37 1.79 4.07 1.26
4 0.8 3.2 0.6 3.32 1.98 4.20 1.03
Stand-alone metamodeling methods
5 0.8 3.4 0.8 3.10 2.26 3.80 1.25
6 1.0 2.6 0.2 2.38 2.29 3.84 3.11
Generally, the relationship between a vector of input vari-
7 1.0 2.8 0.4 2.56 1.93 3.71 0.96
ables x and corresponding output values Y can be expressed
8 1.0 3.0 0.6 2.88 1.93 4.00 0.93
as:
9 1.0 3.2 0.8 2.62 1.86 3.88 1.24
10 1.0 3.4 0.0 2.61 1.86 3.82 0.83
Y = fˆ(x, β) + ε (1)
11 1.2 2.6 0.4 2.73 2.53 3.70 0.85
12 1.2 2.8 0.6 2.59 1.72 3.63 1.28
where fˆ(·) is the approximation model, β represents the
13 1.2 3.0 0.8 3.02 2.94 3.65 1.01
vector of coefficients, ε denotes a stochastic factor. Metamod-
14 1.2 3.2 0.0 2.36 1.33 3.42 1.03
eling technologies differ with each other as to their choices
15 1.2 3.4 0.2 2.46 1.68 3.64 1.11
of approximation models and stochastic process expressions.
16 1.4 2.6 0.6 1.91 0.92 3.24 1.41
In this paper, three related metamodeling methods, includ-
17 1.4 2.8 0.8 2.37 1.76 3.62 1.30 ing Kriging, radial basis function (RBF) and support vector
18 1.4 3.0 0.0 2.10 1.33 3.50 1.09 regression (SVR), are described.
19 1.4 3.2 0.2 1.92 1.60 3.41 1.12
20 1.4 3.4 0.4 2.07 1.81 3.37 1.27
Kriging
21 1.6 2.6 0.8 2.40 1.24 3.50 1.35
22 1.6 2.8 0.0 3.44 1.25 2.23 1.01 Kriging metamodels is an interpolative Bayesian metamodel-
23 1.6 3.0 0.2 1.92 1.443 3.31 1.12 ing technique. It was originated from geo-statistical and used
24 1.6 3.2 0.4 2.16 1.12 3.26 1.24 by Sacks et al. (1989) for predicting the unknown response
25 1.6 3.4 0.6 2.39 2.07 2.41 0.86 at sample points. Kriging treats the observed response as a
combination of a global model and local deviations:

bead profile, all welding seam have been initially rubbed fˆ(x) = p(x) + Z (x) (2)
by different molybdenum sandpaper and cleaned by acetone
before measurement. where p(x) is a known polynomial function, Z (x) is the real-
ization of a stochastic process with mean zero and nonzero
Background and proposed approach covariance. The nonzero covariance of Z (x) is given by:
     
The goal of the proposed approach is to determine the optimal COV Z (xi ) , Z x j = σ 2 R R xi , x j (3)
process parameters of LB by combining the GA and e within
limited experiment resource. The approach starts out with where R is the correlation matrix. R(xi , x j ) is the correla-
generating an initial sample points and conducting experi- tion function between two sample points xi and x j . When
ment at these points. Based on the experiment data, ensemble the Gaussian correlation function is employed, it can be cal-
of three stand-alone metamodels including Kriging, RBF, and culated by:

123
J Intell Manuf


K  2 where, x is a vector of design variables, m is the number of
R(θ ) = exp − θk xik − x kj (4) sample points, xi is the ith sample point. x − xi  denotes
k=1
the Euclidean distance between the design variable and the
where K demotes the dimensions of design space and θk sample points given by:
are the unknown correlation parameters to be determined.
Because Kriging is an interpolative Bayesian metamodeling, x − xi  = (x − xi )T (x − xi ) (11)
the model will have no mean square error (MSE) at all sam-
ple points. If the MSE is minimized, the predictor fˆ (x) for φ(·) is a radial basis function. Commonly used radial basis
unobserved points is expressed as: functions include: (1) bi-harmonic, φ (r ) = r (2) cubic,
φ (r ) = (r + c)3 ; (3) thin-plate
√ spline,φ (r ) = r 2 log(r ); (4)

fˆ (x) = β̂ + rT (x) R−1 f − β̂p (5) multiquadric, φ (r ) = r 2 + c2 ; (5) inverse-multiquadric,
 
φ (r ) = √ 21 2 ; (6) Gaussian, φ (r ) = exp −αr 2 , α > 0.
r +c
where f is the column vector of length m that contains the λi are the weight coefficients of the liner combinations can
sample data of the responses, and p is a column vector of be obtained by:
length m that is filled with ones when p(x) is taken as a con-  −1
stant. r T (x) is the correlation vector between an unobserved λi = T  + T yi (12)
point x and the sample points.
   T where, are all zero except for the regularization parameters
rT (x) = R x, x 1 , R x, x 2 , · · ·, R x, x N (6) along its diagonal.  is the design matrix can be expressed
as:
The scalar β̂ is estimated using the following equation: ⎛ ⎞
φ(x1 − x1 ) φ(x1 − x2 ) · · · φ(x1 − xm )
⎜ φ(x − x ) φ(x − x ) · · · φ(x − x ) ⎟
β̂ = (pT R−1 p)−1 pT R−1 f ⎜ 2 1 2 2 2 m ⎟
(7) =⎜
⎜ .. .. .. .. ⎟

⎝ . . . . ⎠
The estimated variance of the output model can be calculated φ(xm − x1 ) φ(xm − x2 ) · · · φ(xm − xm )
by:
(13)
(f − β̂p)T R−1 (f − β̂p) Support vector regression
σ̂ 2 = (8)
N
Support vector regression (SVR) comes from the theory of
The unknown correlation parameters θk are founded using
support vector machines, but adds the capability to approx-
maximum likelihood estimation can be formulated as Wang
imate black box functions. Commonly used SVR is ε-SVR
et al. (2015):
which aims to find a function that has at most ε deviation
 
N ln(σ̂ 2 )+ln|R| from the targets of the training inputs (Clarke et al. 2005).
max () = − 2 (9) For the linear regression case, ε-SVR can be depicted as:
s.t.  > 0
fˆ(x) = w · x + b (14)
where  denotes the vector of θk ,and both σ̂ and R are the
function of .
where w · x is the dot product between w and x. Another
aims of SVR is to make the fˆ(x) to be as flat as possible.
Radial basis function
Flatness in this sense means a small w in Eq. 14. Hence, we
solve the optimization problem described in the following
Radial basis function (RBF) is a type of neural network
equation:
employing a hidden layer of radial units and an output layer
of linear units. Let x = {x1 , x2 . . . xm } be a set of sampling 1
points. Y = {y1 , y2 . . . ym } are the response function values min |w|2
2
at data locations. The RBF metamodels can be specified as a s.t. yi − w · xi  − b ≤ ε (15)
line combination of some RBFs with weight coefficients in
Eq. 10: w · xi  + b − yi ≤ ε


m An assumption made in Eq. 15 is that the prediction errors
fˆ(x) = λi φ(x − xi ) (10) at all sample points are smaller than ε. However, this is not
i=1 always the case and two slack factors can be incorporated

123
J Intell Manuf

into the original optimization problem to yield a modified moid and Inhomogeneous polynomial. Then the SVR meta-
formulation (Zhou et al. 2015b): models for nonlinear regression becomes:

1  l 
l
 
min |w|2 + C ξi + ξi∗ fˆ(x) = (δi − δi∗ )k xi · x j + b (20)
2
i=1 i=1
s.t. yi − w · xi  − b ≤ ε + ξi (16)
Ensemble of metamodels
w · xi  + b − yi ≤ ε + ξi∗
ξi , ξi∗ ≥ 0 Comparative studies of accuracy of metamodels have demon-
strated that different metamodels perform well under differ-
where C is a constant value used to balance the flatness and ent conditions (Zhou et al. 2015b; Wang and Shan 2007;
the degree of the tolerated errors. Acar 2015). Arbitrarily selecting a metamodel to fit the rela-
According to Lagrangian theory and the Karush–Kuhn– tionship between the input parameters and output responses
Tucker (KKT) condition, the optimization problem can be may increase the risk of adopting an inappropriate meta-
written in the following dual form: model when considering the following two realities: (a) the
nature of internal structure for the input parameters and
1 
l
   output responses is not a priori and the resource to obtain
max − δi − δi∗ δ j − δ ∗j xi · x j information describing the relationship between input vari-
2
i, j=1
ables and the response values is limited (b) the accuracies of

l
  l
  constructed metamodel depend on the current training data,
−ε δi + δi∗ + yi δi − δi∗ and a different metamodel type may become more accurate
i=1 i=1
(17)
than the selected one with a new available data set. There-

l
  fore, as an alternative to select a stand-alone metamodel, this
s.t. δi − δi∗ = 0
paper fits the relationship between the process parameters
i=1
and bead geometries using EMs, which combine the sepa-
(δi − δi∗ ) ∈ [0, C] rate stand-alone metamodels aiming to take advantage of the
prediction ability of each stand-alone metamodels. Three dif-
The weight vector and corresponding linear regression are ferent stand-alone metamodeling techniques are considered
then obtained through the following expressions: here, they include: Kriging, RBF and SVR. The most com-
monly used EMs constructed method is the weighted average

l
  
l
    ensemble, which is defined as:
w= δi − δi∗ xi , fˆ(x) = δi − δi∗ xi · x j + b
i=1 i=1

3
(18) fˆen (x) = w j fˆj (x) (21)
j=1
Another benefit of the above dual form is that nonlinear
regression can also be used by replacing the dot product where, fˆ(x) is the predicted responses obtained from the
of input vector with kernel functions (Clarke et al. 2005). ensemble metamodel, fˆj (x)( j = 1, 2, 3) are the predic-
Appling the kernel functions in Eq. 18, we obtain: tions obtained using the Kriging, RBF and SVR metamodels,
respectively. w j ( j = 1, 2, 3) are weight factors that deter-
1 
l
   mine the relative contribution of the three metamodels in
max − δi − δi∗ δ j − δ ∗j k xi · x j
2 ensemble. To obtain unbiased response estimations, a con-
i, j=1
straint is posed on weight factors as follows:

l
  
l
 

−ε δi + δi + yi δi − δi∗
(19) 
3
i=1 i=1
wj = 1 (22)

l
  j=1
s.t. δi − δi∗ = 0
i=1 The weight factors can be determined such that the meta-
 
δi − δi∗ ∈ [0, C] model with high accuracy have large weight factors and vice
  versa (Acar and Rais-Rohani 2009; Zhou et al. 2011). In
where k xi · x j denotes the kernel functions. Commonly this paper, the weight factors are determined by minimizing
used kernel functions include: polynomial, Gaussian, Sig- the generalized mean square Leave-one-out (LOO) errors

123
J Intell Manuf

Fig. 3 Explanation of the


corresponding relationship
between the design variables
and a chromosome

(GMSE LOO ) of the ensemble metamodels. LOO method is 20 in a chromosome, corresponds to a design variable. The
one of the cross validation methods that can be used for corresponding relationship between the design variables and
assessing the accuracy of a metamodel (Aute et al. 2013). The chromosome can be depicted in Fig. 3.
GMSE LOO for the ensemble of metamodel can be expressed The main operations of GA include selection, crossover
as: and mutation, which will be introduced in more details as
follows.
m 
 2
GMSE LOO = fˆen,−i (xi ) − f (xi ) /m (23)
i=1 Selection

where f (xi ) denotes the actual response value at xi , Selection operator is a process of choosing the more fit indi-
fˆen,−i (xi ) denotes the prediction of the response value for viduals from the current population according to their fitness
xi using the ensemble of metamodel created based on the values. Commonly, the fitness value assigned to each individ-
current sampling set with the sample point xi moved out. m ual according to its design objective function. The selected
is the total number of the sample points. probability is in proportion to its fitness value, that is, the
Combining the weighted-sum formulations of the ensem- greater the individual fitness value is, the more likely it is
ble metamodels in Eq. 21 and error prediction metric in selected. The selected probability of the pth individual is
Eq. 23, a constrained minimization problem is established calculated by:
to select the optimum weight factors. It can be expressed as
follows: 
Na
Pp = f p / fq (25)
find w j q=1
⎛ ⎞2

m 
3

min GMSE LOO (w j ) = ⎝ w j fˆj (xi )− f (xi )⎠ m where Na is the total number of individuals in the current
i=1 j=1
population, f p and f q are the fitness value of the pth and
qth individual, respectively. The basic operators of GA are

3
performed on the basis of the fitness function evaluation
s.t. wj = 1
i=1
in maximization problem. Roulette Wheel Selection (RWS)
(24) method is used in the selection operation in which a random
between 0 and 1 is generated equally and regarded as pointer
Since the optimization problem in Eq. 24 is not neces- for identifying the individual to be selected.
sary convex, evolutionary algorithms are preferred to be
selected to solve this problem. In this study, genetic algo- Crossover
rithms (Coello Coello 2000) is used and the constraint is
processed using penalty function. The details of the opera- Crossover is an operation which exchanges partial structure
tions of GA are presented in the following section. of the two parent individuals to generate new individuals. In
this paper, the classical single point crossover, in which one
Genetic algorithm cross point is selected randomly on each of the two selected
parent chromosomes and then strings exchange and recom-
GA is one of the most efficient population-based algorithms bine at the cross point to produces the children, is adopted.
and it has been successfully applied in welding optimiza- An illustration of the crossover operation is shown in Fig. 4:
tion in recent years (Wang et al. 2013; Islam et al. 2015;
Tan and Hou 2009). Commonly used encoded modes in GA Mutation
include binary coding method, floating-point coding method,
and symbolic coding method. In this paper, a binary ver- Mutation is an operation in which one or more sites on a
sion of the GA is adopted. Each substring, with its length is selected chromosomal string are converted from 0 to 1 or vice

123
J Intell Manuf

Fig. 4 An illustration of the


crossover operation

Step 6 Check the termination criteria. Two stopping


conditions are considered in this paper: (1) the relative
distance between the optimal process parameters of two
successive iterations is below 1e − 3 (2) the pre-specified
maximum number of iterations is reached.
Fig. 5 An illustration of the mutation operation Step 7 Output the optimum process parameters and verify
their validity via real-life experiments.

versa. The chromosome is selected randomly based on a low


mutation probability whose value is generally between 0.01 Results and discussions
and 0.1. An example of the mutation operation is illustrated
in Fig. 5: Comparison and discussion of metamodel predictions

Steps for the proposed optimization approach Accuracy measures

The flowchart of the proposed laser optimization approach is Two different accuracy metrics are adopted to assess the accu-
demonstrated in Fig. 6. The details steps are as follows: racies of each metamodel: (1) root mean square error (RMSE)
(2) relative maximum absolute error (RMAE). RMSE reveals
the overall accuracy of the metamodel, while the RMAE
Step 1 Define the laser brazing optimization problem reflects the local accuracy of the metamodel (Zhou et al.
including objective function, design variables and ranges. 2015a). The lower the values of RMSE/RMAE, the more
In this paper, the objective is to obtain the maximum WT, accurate the metamodel. Expressions of these two accuracy
WB, ELL and ELR. Three variables include WS, WF and metrics are defined as follows:
GAP, with ranges are 2.6 m/ min ≤ W F ≤ 3.4 m/ min,
0.8 m/ min ≤ W S ≤ 1.6 m/ min, 0 mm ≤ G A P ≤   
RMAE = max  yi − ŷi  /yi , i = 1, . . . , N
0.8 mm. 

Step 2 Adopt Taguchi method for Design of Experiment 1  N
(DOE) and conduct the experiment at the generated sam- RMSE =  (yi − ŷi )2 (26)
N
ple points to obtain the output responses. i=1
Step 3 Construct ensemble of metamodels for the output
responses. This step can be divided into two steps: (1) where N represents the total number of test points; yi is the
built three stand-alone metamodels (Kriging, RBF, and actual response at test point i; ŷi is the predicted response at
SVR) for the response values (2) minimize the general- test point i.
ized mean square LOO errors to find the optimum weight
factors by solving the optimization problem in Eq. 24. Comparison and discussion of metamodel predictions
Step 4 Check the accuracy criterion. Check whether the
desired level of accuracy is achieved or not. If yes, the To calculate the two accuracy metrics, additional 7 randomly
obtained ensemble metamodels can be used for LB opti- validation experiments are carried out under new test condi-
mization, otherwise, go back to Step 2. In this paper, tions, but are within the design domain defined earlier. The
relative maximum absolute error (RMAE) is used as the accuracy results of different metamodeling approaches are
accuracy metric. summarized in Table 4. In Table 4, the best results for each
Step 5 Implement GA to obtain the optimum process model type are printed in bold, while the worst results for each
parameters. During the GA optimization, the fitness val- model type are printed in bolditalic. Three intuitive conclu-
ues assigned to the populations are according to predicted sions can be drawn from Table 5, i.e., (1) The metamodels
value using ensemble of metamodels. constructed by EMs nerve be the worst one among the four

123
J Intell Manuf

Fig. 6 Flowchart for the


proposed approach

Table 4 Comparison of
Bead profile Error metrics Kriging RBF SVR EMs
accuracies of stand-alone and
ensemble of metamodels for WT RMAE 0.3914 0.4333 0.3988 0.3553
bead profile
RMSE 0.7401 0.8268 0.8978 0.6958
WB RMAE 1.7943 1.6579 1.6311 1.5755
RMSE 0.8124 1.0905 0.8359 0.7862
ELL RMAE 0.5541 0.4727 0.6272 0.5456
RMSE 0.9100 0.8899 0.8585 0.8081
ELR RMAE 0.6249 0.6423 0.6382 0.605
RMSE 0.8808 0.939 1.0549 0.8625

metamodeling approaches for all the bead geometries, while can take advantage of the prediction ability of the stand-alone
other three stand-alone metamodels are the worst at least for metamodels (3) The accuracy performance of the three stand-
one output response, which indicates that the proposed EMs alone metamodels (Kriging, RBF, SVR) constructed for bead
metamodeling can decrease the risk of adopting an inappro- geometries are close. Fig. 7 demonstrates the 3D surface of
priate metamodel (2) For most test cases (3/4), metamodels the ensemble of metamodels for WT, WB, ELL and ELR.
constructed by EMs are demonstrated to be the bests in terms In Fig. 7, the red points represent the sample points used to
of both local and global accuracy, illustrating that the EMs construct the EMs.

123
J Intell Manuf

Fig. 7 The EMs for four different bead geometries. a Ensemble of metamodels for WT, b ensemble of metamodels for WB, c ensemble of
metamodels for ELL and d ensemble of metamodels for ELR

Main effects and contribution rates analysis of different WS is the most significant factor associated with the ELL.
parameters on bead profile Large WS will bad for the ELL. As shown in Fig. 8d, a low
magnitude of WF and a large value of WS are more likely
The main effect of a factor on response is defined as the to lead to a satisfied ELR. Meanwhile, the effect of GAP on
average response of all the tests for a factor at a certain level. ELR is not obvious.
Analysis of variance (ANOVA) is an effective method used Besides the main effects of the input process parameters
for obtaining the main effects of factors on responses (Tamrin on output performance, the contribution rates of the process
et al. 2014). In this subsection, ANOVA is adopted to analyze parameters and their interactions contribution rates of bead
the main effects of LB process parameters (i.e., GAP, WF geometries are also analyzed. Fig. 9 demonstrates the contri-
and WS) on bead geometries including WT, WB, ELL and bution rates of the process parameters to bead geometries.
ELR. The main effects of each process parameter on bead The positive contribution rates indicates that their corre-
geometries are plotted in Fig. 8. It can be concluded in Fig. 8 sponding output response will increase with an increase in
that the bead geometries are directly related to the GAP, WF the discussed process parameters; otherwise decreases. The
and WS. From Fig. 8a, it is clear that as the GAP increase and magnitude of the bars demonstrated its degree of importance
the WS decease the WT will increase. A similar conclusion to the bead geometries. As illustrated in Fig. 9a, the WS has
can also be made from Fig. 8b that as the GAP increase and the most significant negative contribution rate to the WT,
the WS decease the WB will increase. While, for the WF, followed by the interaction contribution rate of WF and WS
its effects on WT and WB are just the opposite. As can be and the interaction contribution rate of the GAP and WS. The
seen in the Fig. 8c, GAP and WF slightly affect ELL, while GAP has the most positive contribution rate to WT, which

123
J Intell Manuf

Fig. 8 Main effects of each process parameters on bead geometries. a Main effects on WT, b main effects on WB, c main Effects on ELL and d
main effects on ELR

indicates that it will be very effective to improve WT by Optimization results and verification experiments
increasing GAP. As shown in Fig. 9b, it can be observed that
the interaction of the GAP and WS has the most significant In this section, GA is adopted to solve the LB parameter
positive contribution rates to the WB, while the WF and WS optimization problem. The settings for the GA in the opti-
have negative contribution rates to the WB. It can be seen in mization problems are listed in Table 5. Four independent
Fig. 9c that the WS is the most significant factor associated optimization problems are implemented firstly. They include
with the ELL, which indicates that decrease the WS will be maximize WT, WB, ELL, and ELR. Fitness values of them in
very helpful to improve the magnitude of ELL. Meanwhile, GA optimization process are predicted by ensemble of meta-
it can also be concluded that the ELL can be improved by models constructed in the previous steps. In this paper, the
increasing the GAP and WF. Fig. 9d demonstrates that three computational platform with a 3.30 GHz Intel(R) Core(T M)
process parameters have negative contribution rates to ELR, i3 CPU and 4 GB RAM is used. Fig. 10 demonstrates the
among which the effect of WF on ELR is the most obvi- iteration curves of each output parameters. The optimal para-
ous. The interaction contribution rates of two of them have meters, the corresponding outputs and the execution time for
positive contribution rates to ELR. solving each optimization problem are listed in Table 6.

123
J Intell Manuf

Fig. 9 Contribution rates of the process parameters to bead geometries. parameters have negative contribution rate to the output performance.
a Contribution rate of WT, b Contribution rate of WB, c Contribu- The symbol ∩ indicates the interaction contribution rate to the output
tion rate of ELL and Contribution rate of ELR. The blue bars denote performance for different combination types of manufacture parameters
that the manufacture parameters have positive contribution rate to the (Color figure online)
output performance, while the red bars denotes that the manufacture

Table 5 Settings of GA in the LB optimization problem According to the discussion in “Main effects and contri-
Parameters Population Max Crossover Mutation bution rates analysis of different parameters on bead profile”
size iterations probability probability section, the effects of process parameters on ELL resem-
ble that of ELR, and the effects of process parameters on
Values 40 500 0.7 0.02
WT are similar to that of WB, the verification experiments
for Optimal process parameters of WT, ELR, and SBG are
carried out to verify the accuracy of the optimal results. The
It is important to point out that the qualities of the weld bead geometries of three verification experiments are demon-
bead in laser brazing are influenced by all the bead geometries strated in Fig. 11. As shown in Fig. 11a, the optimum value
including WT, WB, ELL, and ELR. In fact, since the mechan- of WT is 3.13 mm compared with the experiment result of
ical property are jointly affected by these four geometry WT is 3.50 mm, with the relative error is 10.57 % attributing
sizes, the optimal process parameters of these bead geome- to some experiment errors and measurement errors coming
tries will be difficult to separately guide the laser brazing from torch angle, defocus length, fluctuation of process, etc.
process. Hence, the sums of the four bead geometries are The experiment value of ELR is 1.88 mm shown in Fig 11b,
defined to make it possible for actual welding process para- which is almost the same as the optimal result. As illustrated
meter optimization. The sums of bead geometries (SBG) can in Fig 11c, the optimum value of SBG is 11.20 mm compared
be expressed as follows: with the experiment result of SBG is 10.76 mm, with the rel-
ative error between the experiments result and optimal result
SBG = WT + WB + ELL + ELR (27) is about 5 %, which indicates that the prediction accuracy
of proposed ensemble of metamodels can meet the need of
In Eq. 27, we assume that the WT, WB, ELL, and ELR have actual LB process optimization.
equal effects on the quality of Laser brazing. The optimal To demonstrate the quality of bead shape, the
results for SBG are also demonstrated in Table 6. obvious defects of bead shape for the above optimal

123
J Intell Manuf

Fig. 10 Iteration curves of different bead geometries using GA. a The optimal curve of WT, b the optimal curve of WB, c the optimal curve of
ELL and d the optimal curve of ELR

Table 6 optimal process parameters and corresponding bead geome- WT and ELR. Hence, optimizing the sums of the four bead
tries geometries is more suitable for actual LB process parameter
Bead Optimum WS WF GAP Computational optimization.
geometries (mm) (m/min) (m/min) (mm) times (s)

WT 3.13 0.80 3.2 0.60 48.3


WB 2.36 1.18 3.0 0.80 41.8
Conclusion
ELL 4.15 0.80 2.6 0.80 44.0
In this work, a hybrid optimization methodology that com-
ELR 1.83 1.00 2.6 0.20 40.2
bines GA and EMs is proposed for parameters optimization
SBG 11.20 0.95 2.6 0.21 154.2
of laser brazing. The proposed methodology has been suc-
cessfully applied to improve the quality of the weld joint
results are obtained by using metallographic microscope through optimizing the bead geometry about crimping butt
LWD300LMDT at 500x, shown in the Fig 12, As illustrated of the LB. In addition, the main effects and contribution rates
in Fig 12, the obvious defects for the optimal results are of multiple process parameters on bead geometries are ana-
few in number. Besides, the defect for the optimal results of lyzed. The following conclusions are drawn from the above
SBG is apparently smaller than those in optimal results of investigation:

123
J Intell Manuf

Fig. 11 Verification of bead geometries for the optimal results. a Experiment values of the optimal result for WT, b experiment values of the
optimal result for ELR and c experiment values of the optimal result for SBG

Fig. 12 Obvious defects of bead shape for the optimal results. a Bead shape of the optimal result for WT, b bead shape of the optimal result for
ELR and c bead shape of the optimal result for SBG

1. In terms of local and overall accuracy, the prediction putational simulation model can be taken as a low fidelity
capability of the ensemble of metamodels significantly model that is used to reflect the most prominent charac-
outperforms the stand alone metamodel. Arbitrarily teristics of the LB bead geometries at a considerably less
selecting metamodels to approximate the relationship computationally demanding, while, a small number of real-
between the input LB process parameters and output life experiments are used to guarantee the accuracy.
responses increases the risk of adopting an unsatisfactory
metamodel. Thus, the ensemble of metamodels accords Acknowledgments This research has been supported by the National
Basic Research Program (973 Program) of China under Grant No.
with “insurance policy” mode and hold great potential in 2014CB046703, the National Natural Science Foundation of China
approximating these black box problems. (NSFC) under Grant Nos. 51505163, 51421062 and 51323009, and the
2. By ANOVA, it is found that (a) WS has the most signifi- Fundamental Research Funds for the Central Universities, HUST: Grant
cant positive contribution rate to the WT, WB and ELL. No. 2014TS040. The authors also would like to thank the anonymous
referees for their valuable comments.
Meanwhile, a small value of WS is more likely to lead to
satisfied bead geometries (b) GAP has the most positive
contribution rate to WT, WB and ELL, while its effect
on ELR is not obvious (c) WF is the main factor that
References
affects the ELR and decrease the WF will very helpful to
improve the magnitude of ELR. Acar, E. (2015). Effect of error metrics on optimum weight factor
3. The process parameters of LB onto galvanized steel selection for ensemble of metamodels. Expert Systems with Appli-
are optimized by GA–EMs. Verification experiments cations, 42(5), 2703–2709.
have been carried out according to the obtained opti- Acar, E., & Rais-Rohani, M. (2009). Ensemble of metamodels with
optimized weight factors. Structural and Multidisciplinary Opti-
mal process parameters, which verify the reliability of mization, 37(3), 279–294.
the obtained optimal process parameters. Overall, the Aute, V., Saleh, K., Abdelaziz, O., Azarm, S., & Radermacher, R.
proposed hybrid approach, GA–EMs exhibits great capa- (2013). Cross-validation based single response adaptive design of
bility of guiding the actual LB processing and improving experiments for Kriging metamodeling of deterministic computer
simulations. Structural and Multidisciplinary Optimization, 48(3),
welding quality. 581–605.
Cao, R., Yu, G., Chen, J. H., & Wang, P. (2013). Cold metal trans-
fer joining aluminum alloys-to-galvanized mild steel. Journal of
As part of future work, different fidelity models will be incor- Materials Processing Technology, 213(10), 1753–1763.
porated into LB process parameter optimization to make a Chaki, S., Bathe, R. N., Ghosal, S., & Padmanabham, G. (2015). Multi-
trade-off between high accuracy and low expense, i.e., com- objective optimisation of pulsed Nd: YAG laser cutting process

123
J Intell Manuf

using integrated ANN–NSGAII model. Journal of Intelligent Man- Ruggiero, A., Tricarico, L., Olabi, A. G., & Benyounis, K. Y. (2011).
ufacturing, 1–16. doi:10.1007/s10845-015-1100-2. Weld-bead profile and costs optimisation of the CO2 dissimilar
Chen, W., Ackerson, P., & Molian, P. (2009). CO2 laser welding of laser welding process of low carbon steel and austenitic steel
galvanized steel sheets using vent holes. Materials and Design, AISI316. Optics and Laser Technology, 43(1), 82–90.
30(2), 245–251. Sacks, J., Welch, W. J., Mitchell, T. J., & Wynn, H. P. (1989). Design
Chen, S., Li, L., Chen, Y., Dai, J., & Huang, J. (2011). Improving and analysis of computer experiments. Statistical Science, 4(4),
interfacial reaction nonhomogeneity during laser welding–brazing 409–435.
aluminum to titanium. Materials and Design, 32(8), 4408–4416. Singh, A., Cooper, D. E., Blundell, N. J., Pratihar, D. K., & Gibbons, G.
Clarke, S. M., Griebsch, J. H., & Simpson, T. W. (2005). Analysis of J. (2014). Modelling of weld-bead geometry and hardness profile
support vector regression for approximation of complex engineer- in laser welding of plain carbon steel using neural networks and
ing analyses. Journal of Mechanical Design, 127(6), 1077–1087. genetic algorithms. International Journal of Computer Integrated
Coello Coello, C. A. (2000). Use of a self-adaptive penalty approach for Manufacturing, 27(7), 656–674.
engineering optimization problems. Computers in Industry, 41(2), Sinha, A. K., Kim, D. Y., & Ceglarek, D. (2013). Correlation analysis of
113–127. the variation of weld seam and tensile strength in laser welding of
Colombo, D., & Previtali, B. (2014). Laser dimpling and remote welding galvanized steel. Optics and Lasers in Engineering, 51(10), 1143–
of zinc-coated steels for automotive applications. The Interna- 1152.
tional Journal of Advanced Manufacturing Technology, 72(5–8), Taguchi, G. (1978). Performance analysis design. International Journal
653–663. of Production Research, 16, 521–530.
Gao, X., Zhong, X., You, D., & Katayama, S. (2013). Kalman filter- Tamrin, K. F., Nukman, Y., Sheikh, N. A., & Harizam, M. Z. (2014).
ing compensated by radial basis function neural network for seam Determination of optimum parameters using grey relational analy-
tracking of laser welding. IEEE Transactions on Control Systems sis for multi-performance characteristics in CO2 laser joining of
Technology, 21(5), 1916–1923. dissimilar materials. Optics and Lasers in Engineering, 57, 40–47.
Islam, M., Buijk, A., Rais-Rohani, M., & Motoyama, K. (2015). Tan, Z., & Hou, D. (2009). Improve accuracy of laser beam width
Process parameter optimization of lap joint fillet weld based on measurement using a genetic algorithm. Optics and Lasers in Engi-
FEM–RSM–GA integration technique. Advances in Engineering neering, 47(11), 1091–1096.
Software, 79, 127–136. Tan, C. W., Li, L. Q., Chen, Y. B., Mei, C. X., & Guo, W. (2013).
Katherasan, D., Elias, J. V., Sathiya, P., & Haq, A. N. (2014). Simulation Interfacial microstructure and fracture behavior of laser welded-
and parameter optimization of flux cored arc welding using arti- brazed Mg alloys to Zn-coated steel. The International Journal of
ficial neural network and particle swarm optimization algorithm. Advanced Manufacturing Technology, 68(5–8), 1179–1188.
Journal of Intelligent Manufacturing, 25(1), 67–76. Wang, X., Chen, H., Liu, H., Li, P., Yan, Z., Huang, C., et al. (2013). Sim-
Li, X., Lawson, S., Zhou, Y., & Goodwin, F. (2007). Novel technique ulation and optimization of continuous laser transmission welding
for laser lap welding of zinc coated sheet steels. Journal of Laser between PET and titanium through FEM, RSM, GA and experi-
Applications, 19(4), 259. ments. Optics and Lasers in Engineering, 51(11), 1245–1254.
Lim, N. K. (2014). Optimization of TIG weld geometry using a Kriging Wang, D., DiazDelaO, F. A., Wang, W., & Mottershead, J. E. (2015).
surrogate model and Latin Hypercube sampling for data genera- Full-field digital image correlation with Kriging regression. Optics
tion. Long Beach: California State University. and Lasers in Engineering, 67, 105–115.
Lin, J., Ma, N., Lei, Y., & Murakawa, H. (2013). Shear strength of CMT Wang, G. G., & Shan, S. (2007). Review of metamodeling techniques in
brazed lap joints between aluminum and zinc-coated steel. Journal support of engineering design optimization. Journal of Mechanical
of Materials Processing Technology, 213(8), 1303–1310. Design, 129(4), 370–380.
Mei, L., Chen, G., Jin, X., Zhang, Y., & Wu, Q. (2009). Research on Zadeh, P. M., Toropov, V. V., & Wood, A. S. (2009). Metamodel-based
laser welding of high-strength galvanized automobile steel sheets. collaborative optimization framework. Structural and Multidisci-
Optics and Lasers in Engineering, 47(11), 1117–1124. plinary Optimization, 38(2), 103–115.
Nasiri, A. M., Chartrand, P., Weckman, D. C., & Zhou, N. Y. (2013). Zhao, Y., Zhang, Y., Hu, W., & Lai, X. (2012). Optimization of laser
Thermochemical analysis of phases formed at the interface of a welding thin-gage galvanized steel via response surface method-
Mg alloy-Ni-plated steel joint during laser brazing. Metallurgical ology. Optics and Lasers in Engineering, 50(9), 1267–1273.
and Materials Transactions A, 44(4), 1937–1946. Zhou, X. J., Ma, Y. Z., & Li, X. F. (2011). Ensemble of surrogates
Qin, G., Su, Y., Meng, X., & Fu, B. (2015). Numerical simulation on with recursive arithmetic average. Structural and Multidiscipli-
MIG arc brazing-fusion welding of aluminum alloy to galvanized nary Optimization, 44(5), 651–671.
steel plate. The International Journal of Advanced Manufacturing Zhou, Q., Shao, X., Jiang, P., Cao, L., Zhou, H., & Shu, L. (2015a).
Technology, 78(9), 1917–1925. Differing mapping using ensemble of metamodels for global
Rong, Y., Zhang, Z., Zhang, G., Yue, C., Gu, Y., Huang, Y., et al. (2015). variable-fidelity metamodeling. CMES: Computer Modeling in
Parameters optimization of laser brazing in crimping butt using Engineering and Sciences, 106(5), 323–355.
Taguchi and BPNN-GA. Optics and Lasers in Engineering, 67, Zhou, Q., Shao, X., Jiang, P., Zhou, H., & Shu, L. (2015b). An adap-
94–104. tive global variable fidelity metamodeling strategy using a support
Roos, C., & Schmidt, M. (2014). Remote laser welding of zinc coated vector regression based scaling function. Simulation Modelling
steel sheets in an edge lap configuration with zero gap. Physics Practice and Theory, 59, 18–35.
Procedia, 56, 535–544.

123

You might also like