Available online at www.sciencedirect.

com

ScienceDirect
Procedia Manufacturing 11 (2017) 1809 – 1817

27th International Conference on Flexible Automation and Intelligent Manufacturing, FAIM2017,

27-30 June 2017, Modena, Italy

Stencil printing optimization using a hybrid of support vector

regression and mixed-integer linear programming
Nourma Khadera, Sang Won Yoona,*, and Debiao Lib
a
Department of Systems Science and Industrial Engineering, State University of New York, Binghamton, NY, 13905, USA
b
School of Economics and Management, Fuzhou University, Fuzhou, China

Abstract

This research proposes an optimization approach to enhance the stencil printing process (SPP) in surface mount printed circuit
board (PCB) assembly. Stencil printing behavior is affected by many variables including stencil design, solder paste composition,
squeegee speed and pressure, and other environmental conditions. In this research, support vector regression (SVR) model is
trained to capture the complex relationships among these variables, based on historical data. A mixed-integer linear programming
(MILP) model is proposed to minimize the total absolute predicted deviation of average volume transfer from target. The optimal
printing settings are retrieved for different sample problems with low computational cost.
©©2017
2017TheThe Authors.
Authors. Published
Published by Elsevier
by Elsevier B.V. B.V.
This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/).
Peer-review under responsibility of the scientific committee of the 27th International Conference on Flexible Automation and
Peer-review
Intelligent under responsibility of the scientific committee of the 27th International Conference on Flexible Automation and
Manufacturing.
Intelligent Manufacturing
Keywords: Surface mount technology; Stencil printing process; Support vector regression; Mixed-integer linear programming.

1. Introduction

Surface mount technology (SMT) is an essential method used to assemble different types of sophisticated
electronic devices. Stencil printing process (SPP), pick and place (P&P), and solder reflow are the main three
operations in surface mount assembly (SMA) as illustrated in Fig. 1. SPP involves depositing solder paste onto the

* Corresponding author. Tel.: +1 607 777 5935; fax: +1 607 777 4094.
E-mail address: yoons@binghamton.edu

2351-9789 © 2017 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/).
Peer-review under responsibility of the scientific committee of the 27th International Conference on Flexible Automation and Intelligent Manufacturing
doi:10.1016/j.promfg.2017.07.318
1810 Nourma Khader et al. / Procedia Manufacturing 11 (2017) 1809 – 1817

printed circuit board (PCB) pads by using a stencil and squeegee. In P&P operation, the surface mount components
(SMCs) are placed on top of PCB pads. A reflow oven is used to heat the assembled PCB to produce solder joints
with maintaining the original material characteristics.

Fig. 1. Surface mount assembly process (Retrieved from [1]).

Printing defects such as an inadequate deposition volume of solder paste can lead to PCB failure and substantial
repairing and reworking costs. It is widely recognized that 60% of soldering defects in PCB assembly are associated
with SPP [1-4]. Furthermore, 40% of production time is related to troubleshooting the problems resulted from poor
printing application [5]. To prevent from solder joint failures, it is necessary to control the amount of solder paste
and ensure that it is close to the target transfer. Target value indicates 100% volume transfer, which is the ratio
between the actual volume of solder paste deposit and the volume of the corresponding aperture. Short or excessive
depositions lead to different printing defects such as bridges, shifting, slumping, and incompleteness, which can
significantly cause PCB failure. Thus, it becomes necessary to optimize the variables that affect the stencil printing
behavior to 1) increase the quality of prints and first-pass yields, 2) supply the other SMA stages with conforming
products, 3) minimize the rework and repair costs, and 4) ultimately reduce the associated manufacturing costs.
Many researchers investigated the different influential control variables on stencil printing (e.g., product
configuration, squeegee speed and pressure, solder paste composition, stencil design, etc.) to improve the printing
performance [6-14]. The consistency of solder paste pattern is affected by paste particle size, the rheology and
distribution of paste, and the change in temperature and viscosity during printing [15-17]. Snap-off distance was
found to impact the solder paste deposition significantly in [18]. Aperture design was declared as an important factor
to achieve a proper paste volume [19]. The importance of stencil thickness and aperture area in SPP was highlighted
in [14]. Squeegee pressure and speed were also indicated as controllable variables that improve the SPP performance
[15]. In this research, 13 features are selected to train the prediction model based on the literature [4, 14-17, 20-22]
and the domain knowledge.
In the literature, there are mainly three common data-driven methodologies that are currently applied to enhance
the soldering quality and attain defect-free SMA line. Design of experiments (DOE) is extensively used in this
domain to investigate the relationships between the SPP variables and solder paste deposition. Moreover, DOE
analysis is utilized to statistically identify the variables that affect the printing outputs significantly. This approach is
mainly followed by response surface optimization or Taguchi based methods to retrieve the optimal printing
conditions [4, 6, 23-27]. Statistical process control (SPC) method is widely used to monitor the significant
characteristics that impact the soldering quality. SPC system improves the printing quality by controlling the
variability in the process and early detection of defects [28, 29]. Lastly, artificial intelligence (AI) approaches
including neural networks, rough-set algorithm, and neuro-fuzzy are utilized to acquire knowledge on the essential
variables that affect the performance of any SMA operation [1, 28, 30]. Generally, these AI approaches are followed
by metaheuristic optimization to retrieve the optimal/near optimal printing conditions [1].
In this research, a new approach is proposed to optimize the stencil printing parameters, which combines data
 Nourma Khader et al. / Procedia Manufacturing 11 (2017) 1809 – 1817 1811

mining and mathematical programming. Data mining algorithm specifically support vector regression (SVR) is used
to model the complex relationships among the SPP variables that affect the solder paste volume transfer. A mixed-
integer linear programming (MILP) model is formulated that includes the trained SVR prediction model as its cost
function. The objective of the proposed model is to minimize the total absolute predicted deviation of average
volume transfer from target (100%) for a new PCB configuration of aperture shapes and sizes. Solder paste transfer
of 100% is optimal but still transfer percentage between spec limits 50-150% or 70-130% or 80-120% are
acceptable in SMT industry. The spec limits are included in the optimization model to ensure an average transfer
within the limits. The inputs of the optimization model are information about the stencil layout or PCB configuration
(aperture shapes, orientations, and sizes). Thus, the optimal printing parameters can be retrieved for different PCB
configurations, as long as the prediction model is trained for the associated aperture shapes and sizes.
The rest of this paper is organized as follows: Section 2 presents the data used to conduct the study, SVR model,
and the MILP model formulations; The experimental results are shown in Section 3; The conclusions and future
work of this research are provided in Section 4.

2. Methodology

2.1. Data description

The data is collected from printing several PCBs using a stencil layout that is specifically designed for
experimentation and analyses. The stencil consists of a large number of aperture shapes and sizes to cover most of
the components used in electronics devices. An electronic product’s PCB may consist of a subset of the studied
aperture shapes and sizes. Circular (open area between 0.041/63.617 to 0.456/706.858 mm/mil) and rectangular
(open area between 0.314/486 to 6.452/10000 mm/mil) aperture shapes are the focus of this study because they are
the most commonly used in SMT applications and are the troublemakers in SPP. The array or collection of
apertures/openings/pads of certain shape and size combination is called a device. There are total of 38 devices
included in this study (10 circular, 14 rectangular placed horizontally, and 14 rectangular placed vertically) where
each device consists of 16 apertures (4
4) array. The number of printing runs is 576 which result in the collection
of 21,888 records (576

38). To train the SVR model, 13 features are selected which are described in Table 1. The
SPP features are the inputs of the SVR prediction model and the output is the average volume transfer for each
device.

Table 1. Brief description of each feature.

Feature Variable name Description
Stencil thickness X1 Thickness of the stencil used to control paste volume
Stencil type X2 Stencil manufacturing way
Clamp X3 Stencil is fixed with/without clamp
Snap-off distance X4 Distance between the stencil and PCB
Separation speed X5 Speed assigned to separate the stencil and the PCB
Squeegee speed X6 Speed of squeegee moving along the stencil surface
Squeegee pressure X7 Pressure applied on stencil by squeegee
Powder particle size X8 Ball size of solder paste
Powder metal load X9 Metal content in solder paste
Printing direction X10 Direction of squeegee relative to the PCB
Aperture shape X11 Geometric shape of the stencil aperture
Aperture orientation X12 Aperture physical location on stencil
Aperture size X13 Diameter for circular and width by length for rectangular

The first 10 variables shown in Table 1 are the printing parameters that require optimization. A printing setting is
a combination of these variables. The last 3 variables in Table 1 represent the stencil layout that is designed for a
specific PCB. The stencil aperture has the exact dimensions of the PCB pad which indicates 1:1 area ratio. The
stencil layout related features are included in training the SVR model to capture the effects of different aperture
shapes, orientations, and sizes on average volume transfer. In the optimization phase, the PCB design or stencil
layout (aperture shapes, orientations, and sizes) is known and it is required to find the optimal printing setting for
such layout.
1812 Nourma Khader et al. / Procedia Manufacturing 11 (2017) 1809 – 1817

2.2. Support vector regression

Support vector (SV) algorithm is the nonlinear generalization of the Generalized Portrait algorithm, which was
developed in the sixties in Russia. SV is firmly founded on the statistical learning theory or the VC theory which
was developed by Vapnik and Chervonenkis over the last three decades [31, 32]. Statistical learning theory provided
an effective framework for both classification and regression tasks when different features are involved. The result
of this framework was the support vector machines (SVM) which works by minimizing a constrained quadratic
problem. The objective function of the optimization problem is the loss function with a regulation term. SVM types
are support vector classification (SVC) and SVR [33]. SVR achieves a generalized performance by minimizing the
generalized error bound instead of only minimizing the training observed error as in traditional regression methods.
SVR is known of its capability to solve complex nonlinear regression problems by mapping the input features into a
high dimensional space wherein they are linearly correlated with the output. These mapping functions are also called
kernel functions including linear, polynomial and Gaussian functions.
Stencil printing behavior exhibits complex relationships, which make SVR a suitable approach to use. The linear
kernel function is used in this study to simply present the SVR prediction equation and integrate it in the proposed
optimization model. In addition, linear kernel function converges quickly with providing good results. Future work
will consider incorporating polynomial and kernel functions in optimization. SVR is mainly trained to represent the
cost function of the MILP model. The general representation of the obtained SVR function is shown in Equation 1.

f ( X ) = X ′β + b (1)

where
is a set of features,
is a linear combination of training examples and
is the bias. There are categorical
and continuous features under the study. Snap-off distance, separation speed, and squeegee speed and pressure are
the only continuous variables. Some of the categorical variables have two categories while others have three. Thus,
there is a coefficient for each category in the categorical variable, and only one coefficient for each continuous
variable as shown in Table 2. Only the coefficients related to printing parameters are shown in Table 2. Note that
there are other coefficients related to stencil layout features (aperture shape, orientation, and size) retrieved from the
SVR model.

Table 2. Trained SVR model coefficients.

Factor Category Betas

3 -3.8503
Stencil thickness (mil) 4 5.8719
5 -2.0216
Electroformed -5.5503
Stencil type
Laser-cut 5.5503
No 0.6311
Clamp
Yes -0.6311
Gap N/A 0.6054
Separation speed N/A 3.5026
Printing speed N/A 2.3246
Printing pressure N/A 0.9728
T3 -0.9359
Powder particle size
T4 0.9359
87.25% (T4) or 87.5%(T3) 0.6631
Metal load (weight %) 88.25% (T4) or 88.5%(T3) -1.1844
89.25% (T4) or 89.5%(T3) 0.5213
Backward 1.4675
Printing direction
Forward -1.4675

2.3. Mixed-integer linear programming model

The objective of the optimization model is to find the optimal printing setting for any given PCB configuration or
stencil layout that is defined by aperture shapes, orientations and sizes. Thus, avoiding the waste that results from
 Nourma Khader et al. / Procedia Manufacturing 11 (2017) 1809 – 1817 1813

trial and error to determine the optimal printing parameters in actual manufacturing setting. The list of notations
used in deriving the objective function and constraints for the MILP model is shown in Table 3.

Table 3. List of notations.

Indices
c Circular aperture size,












d Device of specific shape, orientation, and size,












h Horizontal rectangular aperture size,












i Printing feature with two categories,












j Category number for each i,









k Printing feature with three categories,












l Category number for each k,









m Continuous printing feature,












n Number of original features
v Vertical rectangular aperture size,












Parameters
b Bias of SVR model
LSL, USL Customer lower and upper spec limits for acceptable average volume transfer, respectively









Coefficient for variable i and category j, variable k and category l, and variable c, respectively



Coefficients for circular and rectangular shapes, respectively





Coefficients for no orientation, horizontal, and vertical, respectively








Coefficients for circular size, horizontal rectangular size, and vertical rectangular size, respectively

Threshold to determine the accepted deviation from target (100%) for each device
Variables


 ,

 1 if certain category is selected from a feature with two and three categories, respectively; 0 otherwise


Value for continuous stencil printing feature

Cost function

The objective of this model is to minimize the total absolute predicted deviation of average volume transfer from
100% across circular, and horizontal and vertical rectangular devices. A general form of the objective function is
presented in Equation (2).

D V H C
∑ fd ( X1, X 2 ,..., X n ) − 100 = ∑ ∑ ∑ f ( X1, X 2 ,..., X n ) − 100 . (2)
d =1 v =1h =1c =1
Note that
is the prediction function obtained from training SVR model. To illustrate further, the prediction
function is expanded as shown in Equation (3) and the absolute value function is removed for simplification.

V H C⎛ I J K L M ⎞
∑ ∑ ∑ ⎜⎜⎜ ∑ ∑ βij xij + ∑∑ βkl xkl + ∑ βm xm ⎟⎟⎟ +
v =1h =1c =1⎝ i =1 j =1 k =1l =1 m =1 ⎠
C H V V H C V H C
∑ (γ + θ + ρc ) + ∑ (δ + τ + σ h ) + ∑ (δ + ϑ + ϕv ) + ∑ ∑ ∑ b − ∑ ∑ ∑100.
c =1 h =1 v =1 v =1h =1c =1 v =1h =1c =1 (3)

The constants mainly represent the coefficients of SVR prediction model that are related to the shape, orientation
and size of circular, horizontal and rectangular devices. They are constants because the information about the stencil
layout or PCB configuration are given and input to the model. The generalization of Equation (3) is shown in
Equation (4).
I J K L M
(V + H + C ) ∑ ∑ βij xij + ∑ ∑ β kl xkl + ∑ β m xm +
i =1 j =1 k =1l =1 m =1
C H V
(4)
Cγ + Cθ + ∑ ρc + Hδ + Hτ + ∑σ h + Vδ + Vϑ + ∑ ϕv + (V + H + C )b − (V + H + C )100.
c =1 h =1 v =1
1814 Nourma Khader et al. / Procedia Manufacturing 11 (2017) 1809 – 1817

To simplify further, the sum of constants is denoted as
as shown in Equation (5).

⎛ I J K L M ⎞
(V + H + C )⎜⎜ ∑ ∑ βij xij + ∑∑ β kl xkl + ∑ β m xm ⎟⎟ + ω. (5)
⎜ i =1 j =1 k =1l =1 m=1 ⎟
⎝ ⎠
The constant terms do not affect the feasible region or optimal solution, thus, it is not included in the objective
function. However, the calculations of these constants are important for the constraints and the evaluation of the
objective function value.

Constraints

Equations (6 & 7) indicate that only one category is selected for each feature that consists of two or three
categories, respectively. For instance, printing direction has two values (backward or forward) and Equation (6)
determines that only one direction should be selected. Similarly, stencil thickness could take three values (3, 4, 5)
mil, and Equation (7) ensures only one value is considered in the optimal solution. The continuous features under
the study are bounded as shown in Equation (8), to represent the actual parameters’ ranges of the printer machine.
For instance, the printer under the study has a range of separation speed between 0.1 and 20 mm/s. Thus, any value
beyond this range cannot be tested using this specific printer and is considered as infeasible. Equations (9 & 10)
guarantee that the predicted average volume for each device is within the customer spec limits. Volume transfer of
100% is preferred, however, transfer percentage between 50-150% or 70-130% or 80-120% is acceptable in SMT
industry. Equation (11) ensures to get an optimal solution that not only provides an average volume transfer between
spec limits for all the devices, but also that attains the minimum absolute deviation from 100% for each device. The
absolute value function in Equation (11) is formulated as two linear expressions. Equation (12) indicates the integer
variables.

J
∑ xij = 1 ∀i (6)
j =1
L
∀k (7 )
∑ xkl = 1
l =1
lb ≤ xm ≤ ub ∀m (8 )
f d ( X1, X 2 ,..., X n ) ≤ USL ∀d (9 )
f d ( X1, X 2 ,..., X n ) ≥ LSL ∀d (10)
f d ( X1, X 2 ,..., X n ) − 100 ≤ ζ ∀d (11)
xij , xkl ∈ {0,1} ∀i, j , ∀k , l (12)
3. Experimental results and analysis

The MILP model is tested for a PCB configuration that consists of several circular and rectangular devices. This
problem requires finding the optimal setting for printing a PCB that has 22 devices (9 circular, 6 horizontal
rectangular, and 7 vertical rectangular). The optimal printing settings are obtained for spec limits 50-150% and 70-
130%. However, no feasible solution attained for spec limits 80-120%. The smallest spec limits’ range that has an
optimal feasible solution is 80-122%. The retrieved optimal printing settings and the corresponding objective
function values using the proposed MILP model are shown in Table 4.
 Nourma Khader et al. / Procedia Manufacturing 11 (2017) 1809 – 1817 1815

Table 4. Optimal printing setting.

Spec. limits X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 Obj.
(mil) (mm) (mm/s) (mm/s) (bar) (weight %) value
50-150% 4 Electroformed No -0.5 0.1 83.38 19 T3 87.5% Forward 216
70-130% 4 Electroformed No -0.5 0.1 83.38 19 T3 87.5% Forward 216
80-122% 4 Laser-cut No -0.5 0.1 25 7.18 T3 87.5% Forward 266

The optimal solutions obtained for spec limits 50-150% and 70-130% are the same because the model not only
satisfies the spec limits’ condition but also ensures that the obtained solution has the minimum absolute predicted
deviation from 100% for each device. For 80-122% spec limits, the only printing parameters’ values that changed
are stencil type, and squeegee speed and pressure, which indicates to their significance on changing the solder paste
volume transfer. The objective function value represents the total absolute predicted deviations from 100% target
transfer for all the devices in PCB configuration. Although 80-122% represents the smallest range of spec limits, the
objective function value is the highest compared to other spec limits. The reason is that the optimization model tries
to find a solution that satisfies the spec limits’ condition, specifically for small devices, which may result in larger
deviations for other devices. To explain further, the obtained optimal solutions are evaluated by retrieving the
predicted average transfer for each of the sizes using the trained SVR model as shown in Table 5.

Table 5. Average transfer based on optimal printing setting.

Shape Size Open area Average transfer Average transfer
(mil) (mm) 50-150% / 70-130% 80-122%
C 9 0.041 75.00 80.00
C 10 0.051 80.43 85.43
C 11 0.061 87.84 92.84
C 12 0.073 89.42 94.42
C 13 0.086 92.21 97.21
C 14 0.099 95.16 100.16
C 20 0.203 105.56 110.56
C 25 0.317 111.37 116.37
C 30 0.456 115.11 120.11
H 54 x 9 0.314 105.34 110.34
H 72 x 9 0.418 107.55 112.55
H 90 x 9 0.523 106.72 111.72
H 72 x 12 0.557 107.85 112.85
H 90 x 15 0.871 109.21 114.21
H 120 x 15 1.161 109.45 114.45
V 72 x 9 0.418 110.40 115.40
V 60 x 10 0.387 107.56 112.56
V 80 x 10 0.516 108.11 113.11
V 90 x 9 0.523 109.58 114.58
V 100 x 10 0.645 106.69 111.69
V 150 x 15 1.452 109.50 114.50
V 200 x 50 6.452 106.19 111.19

In case of 80-122% spec limits, the MILP model tries to find a solution that satisfies spec limits’ constraint for
size C9, however this solution affects the deviations for the rectangular devices and leads to larger objective
function value.

4. Conclusions and future work

This research focused on optimizing SPP which is considered a critical operation in SMA because it contributes
to 60% of soldering defects. Different stencil aperture shapes and sizes are considered in this study, specifically
circular and rectangular. Several SPP controllable variables are included in this study such as stencil thickness and
type, squeegee speed and pressure, and others. SVR model was trained to capture the complex relationships between
these variables and the average volume transfer of solder paste. The prediction model served as the cost function for
the proposed optimization model. To retrieve the optimal printing setting under a set of constraints, an MILP model
1816 Nourma Khader et al. / Procedia Manufacturing 11 (2017) 1809 – 1817

is formulated and solved using MATLAB optimization toolbox. The objective of the MILP model is to minimize the
total deviations from target transfer efficiency (100%) for all the devices. Furthermore, it ensures that the average
volume transfer for each device is within the spec limits specified by the customer. The proposed approach that
combines data mining and mathematical modeling shows promising results in improving the stencil printing
behavior in actual manufacturing setting. Future work will involve modeling polynomial and Gaussian kernel
functions of SVR as a nonlinear MIP optimization model.

References

[1] Tsai, TN. Modeling and optimization of stencil printing operations: A comparison study. Computers & Industrial Engineering
2008;54(3):374-389.
[2] He, D, Ekere, NN, & Currie, MA. The behavior of solder pastes in stencil printing with vibrating squeegee. IEEE Transactions on
Components, Packaging, and Manufacturing Technology: Part C 1998;21(4):317-324.
[3] Pan, J. Modeling and process optimization of solder paste stencil printing for micro-BGA and fine pitch surface mount assembly (Doctoral
dissertation, Lehigh University) 2000.
[4] Tsai, TN. Development of an integrated reflow soldering control system using incremental hybrid process knowledge. Expert Systems with
Applications 2005;28(4):681-692.
[5] Amir, D. Expert system for SMT assembly. In Proceedings of the Surface Mount International Conference and Exposition-Technical Program
1994;691-699.
[6] Lofti, A, & Howarth, M. Industrial application of fuzzy systems: Adaptive fuzzy control of solder paste stencil printing. Information Sciences
1998;107(1-4):273-285.
[7] Pan, J, Tonkay, GL, Storer, RH, Sallade, RM, & Leandri, DJ. Critical variables of solder paste stencil printing for micro-BGA and fine-pitch
QFP. IEEE Transactions on Electronics Packaging Manufacturing 2004;27(2):125-132.
[8] Mangin, CH. Where quality is lost on SMT boards. Circuits Assembly 1991;63-64.
[9] Haslehurst, L, & Ekere, NN. Parameter interactions in stencil printing of solder paste. Journal of Electronics Manufacturing 1996;6(04):307-
316.
[10] Durairaj, R, Nguty, TA, & Ekere, NN. Critical factors affecting paste flow during the stencil printing of solder paste. Soldering & Surface
Mount Technology 2001;13(2):30-34.
[11] Manessis, D, Patzelt, R, Ostmann, A, Aschenbrenner, R, & Reichl, H. Technical challenges of stencil printing technology for ultra fine pitch
flip chip bumping. Microelectronics Reliability (2004);44(5):797-803.
[12] Jackson, G J, Hendriksen, MW, Kay, RW, Desmulliez, M, Durairaj, RK, & Ekere, NN. Sub process challenges in ultra fine pitch stencil
printing of type-6 and type-7 Pb-free solder pastes for flip chip assembly applications. Soldering & Surface Mount Technology
2005;17(1):24-32.
[13] Durairaj, R, Ramesh, S, Mallik, S, Seman, A, & Ekere, N. Rheological characterisation and printing performance of Sn/Ag/Cu solder pastes.
Materials & Design 2009;30(9):3812-3818.
[14] Pan, J., Tonkay, GL, Storer, RH, Sallade, RM, & Leandri, DJ. Critical variables of solder paste stencil printing for micro-BGA and fine-
pitch QFP. IEEE Transactions on Electronics Packaging Manufacturing 2004;27(2):125-132.
[15] Amalu, EH, Ekere, NN, & Mallik, S. Evaluation of rheological properties of lead-free solder pastes and their relationship with transfer
efficiency during stencil printing process. Materials & Design 2011;32(6):3189-3197.
[16] Bao, X., Lee, N. C., Raj, R. B., Rangan, K. P., & Maria, A. Engineering solder paste performance through controlled stress rheology analysis.
Soldering & Surface Mount Technology 1998;10(2), 26-35.
[17] Nguty, TA, & Ekere, NN. The rheological properties of solder and solar pastes and the effect on stencil printing. Rheologica Acta
2000;39(6):607-612.
[18] Mannan, SH, Ekere, NN, Ismail, I, & Lo, EK. Squeegee deformation study in the stencil printing of solder pastes. IEEE Transactions on
Components, Packaging, and Manufacturing Technology: Part A 1994;17(3):470-476.
[19] Whitmore, M, MacKay, C, & Hobby, A. Plastic stencils for bottom-side chip attach. Electronic Packaging And Production 1997;37(13).
[20] Owczarek, JA, & Howland, FL. A study of the off-contact screen printing process. II. Analysis of the model of the printing process. IEEE
Transactions on Components, Hybrids, and Manufacturing Technology 1990;13(2):368-375.
[21] Ekere, NN, Lo, EK, & Mannan, SH. Process modelling maps for solder paste printing. Soldering & Surface Mount Technology 1994;6(2):4-
11.
[22] Markstein, HW. Controlling the variables in stencil printing. Electronic Packaging and Production 1997;37(3).
[23] Fidan, I, Kraft, RP, Ruff, LE, & Derby, SJ. Designed experiments to analyze the solder joint quality output of a SMD remanufacturing
system. In Electronics Manufacturing Technology Symposium, 1996, Nineteenth IEEE/CPMT (pp. 422-429). IEEE.
[24] Montgomery, DC. Keats, JB, Perry, LA, Thompson, JR, & Messina, WS. Using statistically designed experiments for process development
and improvement: an application in electronics manufacturing. Robotics and Computer-Integrated Manufacturing 2000;16(1):55-63.
[25] Tsai, TN. Improving the fine-pitch stencil printing capability using the Taguchi method and Taguchi fuzzy-based model. Robotics and
Computer-Integrated Manufacturing 2011;27(4):808-817.
[26] Huang, JC. Reducing solder paste inspection in surface-mount assembly through Mahalanobis–Taguchi analysis. IEEE Transactions on
Electronics Packaging Manufacturing 2010;33(4):265-274.
 Nourma Khader et al. / Procedia Manufacturing 11 (2017) 1809 – 1817 1817

[27] Lau, FK, & Yeung, VW. A hierarchical evaluation of the solder paste printing process. Journal of Materials Processing Technology
1997;69(1-3):79-89.
[28] Barajas, LG, Egerstedt, MB, Kamen, EW, & Goldstein, A. Stencil printing process modeling and control using statistical neural networks.
IEEE Transactions on Electronics Packaging Manufacturing 2008;31(1):9-18.
[29] Villalobos, JR, Muñoz, L, & Gutierrez, MA. Using fixed and adaptive multivariate SPC charts for online SMD assembly monitoring.
International Journal of Production Economics 2005;95(1):109-121.
[30] Yang, T, Tsai, TN, & Yeh, J. A neural network-based prediction model for fine pitch stencil-printing quality in surface mount assembly.
Engineeing Applications of Artificial Intelligence 2005;18(3):335-341.
[31] Vapnik, V. Pattern recognition using generalized portrait method. Automation and Remote Control 1963;24:774-780.
[32] Vapnik, V, & Chervonenkis, A. A note on one class of perceptrons. Automation And Remote Control 1964;25(1):103.
[33] Basak, D, Pal, S, & Patranabis, DC. Support vector regression. Neural Information Processing-Letters and Reviews 2007;11(10):203-224.