micromachines
Article
Investigation on SMT Product Defect Recognition Based on
Multi-Source and Multi-Dimensional Data Reconstruction
Jiantao Chang , Zixuan Qiao
Citation: Chang, J.; Qiao, Z.; Wang,
Q.; Kong, X.; Yuan, Y. Investigation
on SMT Product Defect Recognition
Based on Multi-Source and
Multi-Dimensional Data
Reconstruction. Micromachines 2022,
13, 860. https://doi.org/10.3390/
mi13060860
Academic Editor:
Chang-Hwan Choi
Received: 11 May 2022
Accepted: 28 May 2022
Published: 30 May 2022
Publisher’s Note: MDPI stays neutral
with regard to jurisdictional claims in
published maps and institutional affil-
iations.
Copyright: © 2022 by the authors.
Licensee MDPI, Basel, Switzerland.
This article is an open access article
distributed under the terms and
conditions of the Creative Commons
Attribution (CC BY) license (https://
creativecommons.org/licenses/by/
4.0/).
Micromachines 2022, 13, 860. https://doi.org/10.3390/mi13060860
, Qibin Wang * , Xianguang Kong and Yunsong Yuan
The Key Laboratory of Electronic Equipment Structure Design, Ministry of Education, Xidian University,
Xi’an 710071, China; taocj@xidian.edu.cn (J.C.); qiaozx98@stu.xidian.edu.cn (Z.Q.);
kongxianguang@xidian.edu.cn (X.K.); yuanyunsong777@163.com (Y.Y.)
* Correspondence: qbwangpap@126.com
Abstract: The recognition of defects in the solder paste printing process significantly influences the
surface-mounted technology (SMT) production quality. However, defect recognition via inspection
by a machine has poor accuracy, resulting in a need for the manual rechecking of many defects and a
high production cost. In this study, we investigated SMT product defect recognition based on multi-
source and multi-dimensional data reconstruction for the SMT production quality control process in
order to address this issue. Firstly, the correlation between features and defects was enhanced by
feature interaction, selection, and conversion. Then, a defect recognition model for the solder paste
printing process was constructed based on feature reconstruction. Finally, the proposed model was
validated on a SMT production dataset and compared with other methods. The results show that the
accuracy of the proposed defect recognition model is 96.97%. Compared with four other methods, the
proposed defect recognition model has higher accuracy and provides a new approach to improving
the defect recognition rate in the SMT production quality control process.
Keywords: defect recognition; SMT production quality; solder paste printing; feature reconstruction
1. Introduction
Circuit boards are core parts that enable the functionality of many electronic products.
It can be said that electronic packaging technology is the driving force of SMT, and advances
in SMT also promote the continuous improvement of electronic packaging technology.
The performance of circuit boards is affected by the solder joints. Defects in the solder
paste printing process can lead to problems with the SMT production quality. Therefore,
it is essential to exercise control over the solder paste print quality during the process
of manufacturing circuit boards. The solder paste printing process in surface-mounted
technology (SMT) production is an essential part of the integrated circuit industry. The
solder paste printing process includes main three steps, as shown in Figure 1. For the
printing of solder paste, the squeegee moves horizontally in a specific direction to push the
solder paste to the other side of the stencil. The solder paste has a different distribution on
the solder pad owing to its flux. The distribution has a significant impact on the soldering
and quality of the PCB. The recognition of defects is critical to exercising control over the
solder paste print quality and ensuring an appropriate rate of qualified SMT products.
Two common approaches to recognizing defects in SMT are manual visual inspection and
automatic machine inspection.
In SMT production, manual visual inspection relies mainly on visual observation and
the expert’s personal experience and, consequently, can be time-consuming and laborious.
Moreover, some defects are invisible to the naked eye due to the solder joints being densely
distributed on the PCB. Thus, manual visual inspection requires a longer time and has
lower accuracy. Solder paste inspection (SPI) is most commonly used in the SMT inspection
and testing process to determine whether the height, volume, area, and offset of the
https://www.mdpi.com/journal/micromachines
Micromachines 2022, 13, 860 2 of 15

solder paste are within reasonable limits after the solder paste is printed on the PCB. Pre-
automated optical inspection (AOI) is applied to detect missing, offset, and misplaced
components before soldering. When soldering is finished, post-AOI is performed to
determine the quality of the soldering and, if defects are found, the PCB needs to be
rechecked and repaired manually. For exceptional cases, such as ball grid arrays, in which
AOI is inappropriate, X-ray inspection is required. A combination of SPI, AOI, in-circuit
tests, which are used to determine the electrical performance of components on a PCB, and
automated X-ray inspection can contribute to increasing the rate of qualified SMT products
and reducing rework costs. Although various machine inspection methods can be used to
control the quality of each processing section with a high degree of efficiency, there remain
some problems with SMT product defect recognition. Due to the high false alarm rate of
the inspection equipment, its accuracy relatively poor, and existing inspection methods
Micromachines 2022, 13, x FOR PEER REVIEW 2 of 16
cannot meet the requirements. Thus, many PCBs will need to be manually re-inspected by
the factory, resulting in high production costs.

Thesolder
Figure1.1.The
Figure solderpaste
pasteprinting
printingprocess.
process.

InSMT
In recent years, machine
production, manual learning
visualand deep learning
inspection relies methods
mainly on have been
visual widely applied
observation and
in defect recognition due to their ability to mine datasets for correlative
the expert’s personal experience and, consequently, can be time-consuming and laborious. features and be
scaled and generalized with less expertise and fewer manual operations.
Moreover, some defects are invisible to the naked eye due to the solder joints being Machine learning
methods,
densely such as relevance
distributed vector
on the PCB. machine
Thus, manual[1,2],visual
back-propagation neural networks
inspection requires [3,4],
a longer time
and convolutional neural networks (CNNs) [5,6], have been applied
and has lower accuracy. Solder paste inspection (SPI) is most commonly used in the SMT in defect recognition
for steel surfaces.
inspection and testingGuan et al. [6]
process proposed an
to determine algorithm
whether for the volume,
the height, recognition
area,ofand
defects in
offset
steel surfaces based on improved deep learning network models.
of the solder paste are within reasonable limits after the solder paste is printed on the PCB.
Machine vision, in combination with machine learning algorithms, has also been
Pre-automated optical inspection (AOI) is applied to detect missing, offset, and misplaced
applied in the semiconductor industry. The authors of [7] used a robust detection method
components before soldering. When soldering is finished, post-AOI is performed to de-
based on a vision attention mechanism and deep learning of a feature map to solve the
termine the quality of the soldering and, if defects are found, the PCB needs to be re-
problem of false and missed detection of casting defects in X-ray inspection. The result
checked and repaired manually. For exceptional cases, such as ball grid arrays, in which
showed that the false detection rate and the missed detection rate for casting defects are
AOI is inappropriate, X-ray inspection is required. A combination of SPI, AOI, in-circuit
both less than 4%. The authors of [8] used feature selection and a two-stage classifier for
tests, which are used to determine the electrical performance of components on a PCB,
solder joint inspection. The result showed that the proposed scheme is not only more
and automated X-ray inspection can contribute to increasing the rate of qualified SMT
efficient but also increases the recognition rate. Inspired by a visual background extraction
products and reducing rework costs. Although various machine inspection methods can
algorithm, the authors of [9] proposed an inspection method for IC solder joints with
be used to control the quality of each processing section with a high degree of efficiency,
an improved ViBe algorithm that considers the defect recognition problem as an object
there remain
detection some problems
problem in order to with SMT product
enhance defectofrecognition.
the accuracy Due toofthe
the classification high false
defects. The
alarm rate of the inspection equipment, its accuracy relatively poor,
results showed that the proposed method is universal, accurate, and easily debugged and existing inspec-
tion methods
compared withcannot
othermeet the methods.
existing requirements. Thus, many
The authors of [10]PCBs will need
employed to be manually
a method consisting
re-inspected by the factory, resulting in high production costs.
of a noise filter, defect clustering using the chameleon method, and model-based pattern
In recentThe
recognition. years, machine
authors of [11]learning
proposed and deep learningclustering
a model-based methodsalgorithm
have beenfor widely ap-
automatic
plied in defect
spatial defect recognition
recognition indue to their abilitywafers.
semiconductor to mineThedatasets for correlative
proposed algorithmfeatures
was used andto
be scaled and generalized with less expertise and fewer manual operations.
both find the defect and identify the pattern of each cluster from linear/curvilinear patterns, Machine
learning methods,
ellipsoidal patterns, such
andasring
relevance
spatialvector machine
patterns. [1,2], back-propagation
Furthermore, machine learning neural net-
has been
works [3,4], and convolutional neural networks (CNNs) [5,6], have been applied in defect
recognition for steel surfaces. Guan et al. [6] proposed an algorithm for the recognition of
defects in steel surfaces based on improved deep learning network models.
Machine vision, in combination with machine learning algorithms, has also been ap-
plied in the semiconductor industry. The authors of [7] used a robust detection method
based on a vision attention mechanism and deep learning of a feature map to solve the
Micromachines 2022, 13, 860 3 of 15

extended to such applications as wood [12], the detection of cracks in tunnels [13] and oil
and gas pipelines [14], and 3D metal printing [15].
Inspired by machine learning and deep learning, there has been an exhaustive inves-
tigation of defect recognition in SMT, ranging from using traditional machine learning
methods to using a more modern deep learning approach. The authors of [16] used a
genetic algorithm to optimize the feature extraction region and a support vector machine
to classify the solder joint defect type. The authors of [17] utilized a fuzzy rule table to
represent a human inspector’s criteria in order to correct any misclassification of solder
joints by the neural network. The results showed that the proposed method is superior to
the neural network classification method alone in terms of the accuracy of its classification.
The authors of [18] proposed a dual-level defect detection algorithm, called PointNet, that
extracts point cloud features from defective solder paste pattern images automatically
obtained through solder paste inspection and performed defect detection at the micro-level
and the macro-level for the case where two or more defects can be observed in the image.
The experimental results showed that the proposed D3PointNet algorithm is robust to
changes in the sparsity and size of the DSPP image, and its exact match score was 10.2%
higher than that of the state-of-the-art CNN-based multi-label classification model on the
DSPP image dataset. The authors of [19] used a cascaded CNN to directly perform the
inspection task without low-level feature extraction and before learning to inspect SMT
solder joints. Comparison experiments showed that the proposed method achieved better
performance in SMT solder joint inspection than the state-of-the-art methods. Additionally,
a recent study [20] attempted to inspect and classify multiple types of defects that occur
in the printing stage through SPI datasets using multi-label classification for images of
various sizes. The experimental results on the customized SPI dataset as well as on the
VOC 2007 dataset demonstrated the excellent performance of the proposed network.
In summary, combinations of machine learning and deep learning methods have
enabled remarkable achievements in defect recognition in SMT production and most
of the studies are based on image data on solder paste printing obtained through ma-
chine vision or optical inspection methods for defect recognition. Moreover, the pieces of
equipment involved in the production and inspection stages of solder paste printing are
independent of each other, and the various types of data generated have multi-source and
multi-dimensional characteristics.
However, few scholars have looked at defect recognition from the perspective of multi-
source and multi-dimensional data from the solder paste printing process, resulting in a
low rate of recognition of defects in SMT products. SMT innovation has become a powerful
driver of the development of semiconductor and electronic manufacturing technology
and has a great impact on improvements in electronic packaging technology. Therefore,
in this study we investigate SMT product defect recognition based on multi-source and
multi-dimensional data reconstruction. Technologies in feature reconstruction and the
classification of defect types for resolving the problems described above are presented.
These are also the contributions of this paper, which can be summarized as follows:
(1) Feature interaction is adopted to mine for higher-order data between features. Then,
the data are transformed into high-dimensional grey-image data after the RF-based
feature selection method is applied to enhance the information between features;
(2) A CNN is adopted for defect recognition. The CNN can effectively extract and classify
the high-dimensional grey-image data obtained from the feature reconstruction.
The rest of the paper is organized as follows. Section 2 introduces the proposed
methodology. Sections 3 and 4 verify the method using an example and discuss the results
of the verification. Section 5 contains our conclusions.

2. Proposed Method
In order to improve the accuracy of defect recognition, we propose a defect recog-
nition method based on multi-source and multi-dimensional data reconstruction. The
proposed method consists of three parts, namely solder paste printing data collection and
 odology. Sections 3 and 4 verify the method using an example and discuss the results of
the verification. Section 5 contains our conclusions.

2. Proposed Method
Micromachines 2022, 13, 860 In order to improve the accuracy of defect recognition, we propose a defect recogni- 4 of 15
tion method based on multi-source and multi-dimensional data reconstruction. The pro-
posed method consists of three parts, namely solder paste printing data collection and
reconstruction, feature extraction and classification, and defect recognition. First, feature
reconstruction, feature extraction and classification, and defect recognition. First, feature
reconstruction is performed on the data collected from the PCB. The interaction of the one-
reconstruction is performed on the data collected from the PCB. The interaction of the one-
dimensional features is a fusion of raw data features that reflects the complex interaction
dimensional features is a fusion of raw data features that reflects the complex interaction
between features and benefits to enhance the information about features. The CNN and a
between features and benefits to enhance the information about features. The CNN and a
Softmax classifier are employed to construct a defect recognition model for solder paste
Softmax classifier are employed to construct a defect recognition model for solder paste
printingdefect
printing defectclassification.
classification.Finally,
Finally, each
each reconstructed
reconstructed solder
solder paste
paste printing
printing data
data point
point is
is classified as one of nine types of defects by the defect recognition
classified as one of nine types of defects by the defect recognition model. model.

2.1.Feature
2.1. FeatureReconstruction
Reconstruction
Featurereconstruction
Feature reconstructionincludes
includesfeature
featureinteraction,
interaction,random
randomforest
forest(RF)-based
(RF)-basedfeature
feature
selection,and
selection, and feature conversion,
conversion,as asshown
showninin Figure
Figure 2. First, eacheach
2. First, feature vectorvector
feature is multi-
is
plied by other
multiplied feature
by other vectors
feature to achieve
vectors the interaction
to achieve of features.
the interaction Then, RFThen,
of features. is applied
RF isto
calculate
applied to the importance
calculate scores of scores
the importance features
of and, based
features and,onbased
the RFonresult,
the RFseveral features
result, several
with low scores are eliminated. Finally, min–max normalization is used to
features with low scores are eliminated. Finally, min–max normalization is used to eliminate eliminate the
influence
the influenceof of
thethedimensionality
dimensionalityofofthethedata
data on
on the model. The The one-dimensional
one-dimensionalfeaturefeature
vectordata
vector dataare
aretransformed
transformedinto
intothe
theshape n ×n.n .
shapeofofn ×

Figure2.2.The
Figure Theflowchart
flowchartfor
forfeature
featurereconstruction.
reconstruction.

2.1.1. Feature Interaction

Feature interactions can characterize the more profound influences between defects.
The feature interaction of high-dimensional data is equivalent to a higher-dimensional data
space, where feature selection methods can be used to preferentially select features that
are useful for classification and eliminate irrelevant features. The extension of raw data
features to higher-order features can be used to take full advantage of the higher-order
correlation between the raw data [21]; e.g., the relationship between the cleaning speed and
the cleaning supply time characterizes the cleaning distance, and the relationship between
Micromachines 2022, 13, 860 5 of 15

the solder paste area and the solder paste height characterizes the solder paste volume on
higher orders. 
Feature interaction is performed on the original SMT training set X = x1 , x2 , · · · , x p .
All vectors are column vectors, and any two column vectors can be multiplied together to
obtain the feature interaction dataset X 0 = x1,2 , x1,3 , · · · , x p−1,p . The feature interaction
formula is as follows:
xi,j = xi × x j (1)
where xij represents the result of the interaction between feature xi and feature x j , and the
number of new feature vectors after the feature interaction is p(p − 1)/2.

2.1.2. Feature Selection

Random forest (RF) is an integrated algorithm that uses a decision tree as a base
learner with a bagging integration strategy. The advantages of RF for feature selection
are the ability to handle high-dimensional data and the speed of computation [22]. Thus,
RF is applied for the selection of new features after feature interaction. The process for
calculating the random forest importance score is as follows.
The feature importance score is expressed as a VIM, and the Gini index is expressed
as a GI. Assuming that there are m features X1 , X2 , · · · Xm , the importance of feature X j at
node m, that is, the Gini coefficient score V I M jGini for each feature X j can be calculated as:

Gini
V I M jm = GIm − GIl − GIr (2)

where GIl and GIr denote the Gini coefficients of the two new nodes after branching, and
GIm is expressed as:
|K | |K |
GIm = ∑ ∑
0
pmk pmk0 = 1 − ∑ p2mk (3)
k =1 k 6 = k k =1

where K represents the number of categories and pmk denotes the proportion of category k
that belongs to node m.
Finally, all the feature importance scores are normalized to

V I M jGini
V I M jGini = c (4)
∑ V I MiGini
i =1

c
where ∑ V I MiGini is the sum of the gain of all features and V I M jGini is the Gini coefficient
i =1
of feature X j .
Based on the RF importance scores, features whose Gini coefficient V I M jGini is greater
than the set feature importance threshold d are selected.

2.1.3. Feature Conversion

After feature selection, feature conversion is performed as shown in Figure 3. It is
assumed that the one-dimensional data have q feature vectors and a length of l. The data to
convert are: 
X = x1 , x2 , · · · , x j , · · · , x q (5)
where x j = [ x1j , x2j , · · · , xij , · · · , xlj ]. Once each row of the sample dataset yi = [ xi1 , xi2 , · · ·

, xij , · · · , xil has been reshaped, the input data of the CNN can be obtained.
 assumed that the one-dimensional data have q feature vectors and a length of l. The data
to convert are:
X = { x1 , x2 , , x j , , xq } (5)

where x j =  x1 j , x2 j ,, xij ,, xlj  . Once each row of the sample dataset
Micromachines 2022, 13, 860 6 of 15
yi =  xi1, xi 2 ,, xij ,, xil  has been reshaped, the input data of the CNN can be obtained.

Figure
Figure 3.
3. The
The proposed
proposed CNN
CNN framework.
framework.

2.2. Defect
2.2. Defect Recognition
Recognition Model
Model Based
Based on on the
the CNN
CNN
TheCNN
The CNNproposed
proposedbyby[23][23] is used
is used forfor imageimage processing
processing and and pattern
pattern recognition.
recognition. The
CNN model includes sequentially cascaded input layers, convolutional layers,layers,
The CNN model includes sequentially cascaded input layers, convolutional and
and pool-
pooling layers, a fully connected layer consisting of mutually cascaded expanded layers,
ing layers, a fully connected layer consisting of mutually cascaded expanded layers, and
and a three-layer forward neural network, wherein the forward neural network includes
a three-layer forward neural network, wherein the forward neural network includes se-
sequentially cascaded hidden and output layers. According to the CNN framework shown
quentially cascaded hidden and output layers. According to the CNN framework shown
in Figure 3, the steps are as follows.
in Figure 3, the steps are as follows.
The reconstructed data are input into the CNN model, and convolution is performed
The reconstructed data are input into the CNN model, and convolution is performed
at a convolution kernel to obtain the convolutional feature. The output of the convolutional
at a convolution kernel to obtain the convolutional feature. The output of the convolu-
layer can be calculated as
tional layer can be calculated as S
!
Vjr = f ∑ V r −1
∗ Wijr + brj (6)
1 S i r −1 
V j = f  Vi ∗ Wijr + brj 
r
i = (6)
 i =1 
where S is the pooling filter size, * is the operator of the convolution, Vjr represents the j-th
where r −the
1 convolution, V r represents the
outputSfeature
is the pooling
map on filter size,
the r-th * is the operator
convolution layer, Vof
i denotes the i-th input
j feature map
on the (r − 1)-th convolution layer, W r is the convolutional r -1 kernel, which represents the
j-th output feature map on the r-th convolution ij layer, V i denotes the i-th input feature
i-th band of(r
the j-th pooling r
b j is theWbias
filter, layer, j-th feature map, f is anrepresents
map on the − 1)-th convolution r
ij
isof the
the convolutional andwhich
kernel, activation
function applied to the result, which is usually defined as a rectified linear unit (ReLU).
the i-th band of the j-th pooling filter, b rj is the bias of the j-th feature map, and f is an
The ReLU activation function is expressed as
activation function applied to the result, which is usually defined as a rectified linear unit
(ReLU). The ReLU activation function ReLU is( xexpressed
) = max(as x , 0) (7)
ReLU(x) = max(x,0)
The convolutional feature is maximally pooled through the pooling layer with(7)a
pooling
Thefilter. The specific
convolutional expression
feature for maximum
is maximally poolingthe
pooled through is pooling layer with a pool-
ing filter. The specific expression for maximum p pooling is
(m−1)×n+k
Pjm = max(p Vj ) (8)
k =1
Pjm = max(V j( m −1)× n + k ) (8)
k =1
where Pjm is the j-th output feature map of the m-th band pooling layer, and n and p are the
m
where
numberPof j
is the j-th output
sub-sampling feature
factors and map of the m-th
the pooling size,band pooling layer, and n and p are
respectively.
The one-dimensional
the number of sub-samplingexpansion
factors andof the
thepooling
second size,
pooled feature is performed via the
respectively.
expansion layer to obtain the
The one-dimensional one-dimensional
expansion of the second featurepooled
vector.feature
The one-dimensional
is performed viafeature
the
vector is used as the input into the three-layer forward neural network.
expansion layer to obtain the one-dimensional feature vector. The one-dimensional fea- For classification,
the Softmax
ture vector isregression function
used as the was selected
input into to be the
the three-layer last layer.
forward The network.
neural output of For
the Softmax
classifi-
regression
cation, function regression
the Softmax can be calculated
functionaswas
follows:
selected to be the last layer. The output of the
Softmax regression function can be calculated as follows:
e mi
Softmax(mi ) = r (9)
∑ e mi
i =1

where mi = [m1 , m2 , m3 , · · · , mr ] is the output of the fully connected layer and r is the
number of defect types.
The cross-entropy loss function was chosen to be the loss function used to measure
the difference between the predicted values and the true values of the model’s output.
During the training of the defect type classification model, the parameters need to be
 of defect types.
The cross-entropy loss function was chosen to be the loss function used to
the difference between the predicted values and the true values of the model’
During the training of the defect type classification model, the parameters need t
Micromachines 2022, 13, 860 7 of 15
stantly updated so that the value of the cross-entropy loss function becomes sma
cross-entropy loss function is defined as
N
constantly updated so that the value of1 the cross-entropy
( i ) loss function becomes
( i ) smaller.
L = −
N

(i )

i =1
(i )
The cross-entropy loss function is defined as [ y log yˆ + (1 − y ) log(1 − yˆ )]

N
1
y (ŷi )(i)represents
N i∑
where N represents L=− the total
[y(i) number
log ŷ(i) + (of1 −samples,
y(i) ) log(1 − )] the true
(10) label o
=1
x and takes a value from {0, 1}, and yˆ (i ) represents the probability that x has th
(i ) represents the true label of sample
sample
where label of 1.
N represents theThe
totalmodel
numberreachesof samples, an yoptimal state when the cross-entropy lo
x is
and takes a value from {0, 1}, and ŷ ( i ) represents the probability that output
x has theof
current
minimized by stochastic gradient descent [24]. The final the model is
sample label of 1. The model reaches an optimal state when the cross-entropy loss value
dicted value of the defect type corresponding to that sample.
is minimized by stochastic gradient descent [24]. The final output of the model is the
predicted value of the defect type corresponding to that sample.
2.3. The Proposed Defect Recognition Method
2.3. The Proposed Defect Recognition Method
The flowchart of the proposed method for SMT defect recognition based o
The flowchart of the proposed method for SMT defect recognition based on multi-
source
source andand multi-dimensional
multi-dimensional data reconstruction
data reconstruction is shown in is shown
Figure 4. in Figure 4.

Figure
Figure 4. The
4. The flowchart
flowchart for thefor the proposed
proposed defect recognition
defect recognition method. method.
The proposed method has five steps: data preparation, feature reconstruction, model
initialization, model training and optimization, and defect recognition.
Step 1: The dataset is divided into a training set and a test set.
Micromachines 2022, 13, 860 8 of 15

Step 2: Feature interaction, random forest (RF)-based feature selection, and feature
conversion are performed. Each feature vector is multiplied by other feature vectors to
achieve the interaction of features. Based on the RF result, several features with low scores
are eliminated. Finally, min–max normalization is used to eliminate the influence of the
dimensionality of the data on the model, and the one-dimensional feature vector data are
transformed into the shape of n × n.
Step 3: The CNN model includes a sequentially cascading input layer, the first convolu-
tional layer, the first pooling layer, the second convolutional layer, the second pooling layer,
and a fully connected layer consisting of mutually cascading expanded layers and a three-
layer forward neural network, where the forward neural network includes sequentially
cascading hidden and output layers.
Step 4: During the training of the defect recognition model, the parameters need to be
constantly updated so that the value of the cross-entropy loss function becomes smaller.
When the value of the cross-entropy loss function reaches its minimum or lies within an
acceptable range, and a reasonable number of iterations have been run, the model reaches
an optimal state.
Step 5: The optimal model is saved, and the model outputs the result of the defect
recognition operation.

3. Experiment
3.1. Dataset Description
The experimental dataset used in this study is comprised of actual SMT production
and inspection data, in which a variety of individual boards are involved (Table 1). The
package types and package distributions contained in different individual boards are not
the same. Different boards have different parameters, such as size, shape, devices, and
welding temperature. Different packages may cause different types of defects. In order
to increase the generalizability of the proposed model, data on multiple products were
selected. The dataset is composed of process parameters, printing process status parameters,
and inspection results on a total of 35,756 defect samples. A description of the physical
meaning of the data is given in Table 1. The data that need to be collected during the
solder paste printing stage include PCB properties, printing process parameters, printing
process status parameters, and SPI results. The PCB properties include the PCB length,
width, and height. The printing process parameters include squeegee speed, squeegee
pressure, cleaning speed, work separation speed, work separation distance, worktable
printing height offset, squeegee separation speed, and squeegee separation distance. The
printing process status parameters include average pressure, minimum pressure, maximum
pressure, cleaning supply time, automatic cleaning, and automatic cleaning count. The
SPI results include the solder paste volume, solder paste area, and solder paste height.
The real labels of each sample were tagged by SPI or a manual inspection. The manual
inspection was performed after SPI and contained three detection fields, namely test-result,
warn-result, and confirm-result. Test-result represents the SPI machine result, and it may
be different from the manual inspection result. Warn-result represents the set value of
the alarm result. Confirm-result represents the final defect as determined by the manual
inspection that was performed after SPI. When the results of the three fields were consistent,
the test-result data were used as the inspection result. When the results of the three fields
were inconsistent, the solder paste on the pad was suspected to have a problem and, thus,
needed to be manually re-checked. The confirm-result data from the manual re-check were
selected as the inspection result. The processing of these inspection result labels can both
help to correct defects that cannot be detected by SPI machines and reduce the number of
manual re-inspections and labor costs.
Micromachines 2022, 13, x FOR PEER REVIEW 9 of 16
Micromachines 2022, 13, x FOR PEER REVIEW
Table 9 of 16
Table 1.
Micromachines 2022, 13, x FOR PEER REVIEW1. Description
Description of
of the
the physical
physical meaning
meaning of
of the
the data.
data. 9 of 16
Micromachines 2022, 13, x FOR PEER
Table
REVIEW
1. Description of the physical meaning of the data.
Abbrevia-
Table
Micromachines 2022, 13, x FOR PEER REVIEW1. Description of the physical meaning of the data.
Abbrevia- 99 of
of 16
16
Type
Micromachines Parameter
Type 2022, 13, x FOR PEER REVIEW
Parameter Unit Meaning
Unit of the physical meaning
Meaning Sample
Sample 9 of 16
Abbrevia-
tion
Table 1. Description
tion of the data.
Type 2022, 13,
Micromachines Parameter
x FOR PEER REVIEW
Abbrevia-
Table Unit of the physical meaning
1. Description Meaning
of the data. Sample 9 of 16
Type Parameter Table tion Unit of the physical meaning
1. Description Meaning
of the data. Sample
Micromachines 2022, 13, 860 tion
Abbrevia-
Table 1. Description
Description of the
the physical
physical meaning
meaning of the
the data.
data.
9 of 15
Type Parameter Table Abbrevia-
1. Unit of Meaning
of Sample
Type PCB
PCB Length
Parameter
Length tion
L
Abbrevia-
Table L mm
Unit
mm
1. Description Length
Length
of the physical meaning of
Meaning
of
of the
the
the PCB.
PCB.
data. Sample
Type Parameter
PCB Length Table tion
Abbrevia- Unit
L1. Description
mm of the physical meaning Length Meaning
of thethedata.
PCB. Sample
Abbrevia-
tion of
Type
Type Parameter
PCB Length Abbrevia-
Parameter L Unit
mm
Unit Length Meaning
of the PCB.
Meaning Sample
Sample
tion
Table
tion1. Description of the physical meaning of the data.
Type Parameter
PCB Length Abbrevia-L Unit
mm Length Meaning
of the PCB. Sample
Type Parameter
PCB Length tion
L Unit
mm Length Meaning
of the PCB. Sample
PCB
PCBTypeproper- PCB
proper- Parameter Length tion
L
Abbreviation mm Unit Length of the PCB.
Meaning PCB. Sample
PCB Width
PCB Length
Width WW mm
mm Width
Width of of the
PCBtiesproper- PCB
ties PCB Length L
L mm
mm Length
Length of the
of the PCB.
the PCB.
PCB.
PCB proper- PCB PCB Width W mm Width of the PCB.
ties PCB Length
Width L
W mm
mm Length of
Width of the PCB. the PCB.
PCB ties
proper- PCB PCBLength L mm mm Length of the PCB.
PCB proper- PCB Width L
W mm Width Length
of the of the
PCB. PCB.
Length
PCB ties
proper- PCB Width W mm Width of the PCB.
PCBties
PCB proper- PCB
proper-
ties
PCB Height
PCB
Width WH mm
mm Width of
Height of the
the PCB.
PCB.
PCB Height
PCB Width
Width W
H
W mm
mm
mm
Height
Width of
Width
of the
of the PCB.
the
PCB.
PCB.
PCB PCBproper- PCB Height
ties H mm Height of the PCB.
ties
PCB ties
proper- PCB PCB Width
PCBWidth
Height W
WH mm
mm mm Width
HeightWidthof the
of the PCB.
of thePCB.PCB.
properties PCB Width W mm Width of the PCB.
ties PCB Height H mm Height of the PCB.
PCB Height H mm Height of the PCB.
Squeegee
PCB Height
Squeegee H mm Movement
Movement Height speed
speed of ofthe
of thesqueegee
the PCB.
squeegee during
during
PCB Height SS
SS
H mm/s
mm/s
mm Height of the PCB.
Squeegee
PCB Speed
Height
Speed H mm Movement speed
Height of
printing.
of
printing.the
the squeegee
PCB. during
PCB Height
Squeegee SS
H mm/smm Movement speed Height
of of the PCB. during
PCB Height
Speed H
SS mm
mm/s Height ofthe
printing.thesqueegee
PCB.
PCB Speed
Squeegee
Height H mm Movement Height speed printing.
of
ofthe thesqueegee
PCB. during
Squeegee SS mm/s Movement speed of the squeegee during
Speed
Squeegee SS mm/s Movement speedbyprinting.
of the
the squeegee
squeegee during during
Squeegee
SpeedPres-
Squeegee Pres- SS mm/s Pressure
Pressure exerted
exerted printing.
by
Squeegee
Squeegee SP kg Movement
Movement speed
speed of the
the squeegee
squeegee during
during
Speed
Squeegee
Squeegee sure
sure Pres- SP
SS
SS
SS
kg Movement
mm/smm/sPressure exertedprinting.
mm/s by the squeegee during
of
printing.
speed
printing.
the
of squeegee
the squeegee during
during
Squeegee
Speed
Squeegee
Speed Pres- SP kg Movement
Pressure speed
exerted of the
printing.
by the
printing. squeegee
squeegee during
during
Speed
sure
Squeegee SS
SP mm/s
kg Movement speed printing.
printing.
of the squeegee during
Speed
Squeegee sure Pres- SS mm/s Pressure exertedprinting. by the squeegee during
printing.
Squeegee
SpeedPres- SP kg Pressure exertedprinting. by the squeegee during
Squeegee sure Pres- SP kg Pressure exerted printing.
by the squeegee during
Cleaning
Squeegee
sure Pres-
Cleaning SP kg Pressure exerted
printing. by the squeegee during
Squeegee
Squeegee sure Pres- SP
CS
CS mm/s
mm/s kg Pressure
Stencil
Pressure
Stencil exerted
wiping
exerted
wiping by the
speed
by
printing.
speed theduring
squeegee
squeegee
during during
cleaning.
during
cleaning.
Cleaning
Pressure
Speed
Speed SP
SP kg
kg
printing.
Squeegee sure Pres-
Cleaning CS mm/s Pressure exertedprinting.
Stencil wiping by theduring
speed squeegee during
cleaning.
Squeegee sure
Speed Pres- SP
CS kg
mm/s Stencil
Pressure wiping
exerted printing.
speed
by the during
squeegee cleaning.
during
sure
Speed
Cleaning SP kg printing.
Cleaning
sure
Cleaning CS mm/s Separation
Stencil
Separation speed
wiping of
speed the
printing.
speedspeed worktable
during
of the duringworktable and
and the
cleaning. the
Speed
Cleaning CS
CS mm/smm/s Stencil
Separation wiping
Stencil wiping
speed of speed
the during
worktable cleaning.
cleaning.
and the
Speed
Speed
Cleaning CS mm/s stencil
stencil while
Stencil
while the
wiping
the worktable
speed
worktable duringis
is used
used as
cleaning.
as aa plat-
plat-
Work
Work Separa-
Cleaning
Speed
Separa- Separation speed of the worktable and the
CS
WSS mm/s
mm/s form Stencil
stencilto while wiping
the speed
worktable during
is usedcleaning.
as a plat-
WorkCleaning
Speed
tion Separa-
Speed CS
WSS form
stencilto support
Stencil
while
Separation wiping
support speed
thespeed
thethe
PCB,
PCB,
worktable
of
and
theofduring
and thecleaning.
isthe
used
worktable
PCB
PCB as
and
and
and
a and
plat-
the
the
the
Printing pro- Work Speed
tion Speed
Separa-
Cleaning CS
WSS mm/s stencil
form Stencil
to
Separation
are wiping
Separation
support speed
separated the speed
speed
PCB,
ofat thethe during
the
and
worktable
end ofcleaning.
worktable
the PCB
the and and the
printing the
Printing pro- tion Speed
Speed WSS
CS stencil
mm/s form toare
Stencil
stencil separated
support
while
Separation thethe
wiping
speed atthe
PCB,
speed
worktable
of the andend
duringistheofcleaning.
used
worktable the
PCB as printing
and and
aused
plat-
thethe
Printing pro- Work Work
tion Speed stencil arethe stencil
separated while the
at the end worktable
of the is
printing
cess
cess parame-
parame- Speed
Separa- stencil while
Separation the worktable
speed process.
process.
of the is used
worktable as
and a plat-
the
Printing pro- Work Separa-
Separation WSS
WSS mm/smm/sstencil
form
stenciltoareasseparated
support
while
Separation thethe
a platform
speed at
PCB,
to
worktable
of thethe and
supportend
istheof PCB
the
used
worktable the
PCB,
as printing
and and
aand
plat-
thethe
cessPrinting
parame-
ters
ters tion
Work Speed
Separa- WSS mm/s form to support
Separation speedthe process.
PCB,
of the and the
worktable PCB and and atthe
the
cess parame-
Printing pro- Work Speed
tion Speed WSS mm/s formstencil
stencil
stencil while
the
are PCB the
separated
while
to support worktable
and
thethe the
process.
at
worktable
PCB, stencil
the and is
end used
are
istheof
used the
PCB as
separated
as a plat-
printing
a plat-
and the
ters pro- Work
Printing Separa- Separation
Separation
stencil speed
distance ofof the
the worktable
worktable and
and the
the
cess
process
ters
parame-
Printing pro-
Separa-
tion Speed WSS
WSD
WSS mm/s
mm
mm/s
stencil
form
form
stencil toare
Separation
to
separated
while
support
support
are
the
separated
end
distance
the
the
ofatthe
of
worktable
the PCB,
process.
PCB,
at
the
the
end
printing
and
and
end the
ofprocess.
worktable
isthe
used
of
theas
PCB
PCB
the
printing
and
and
and
printing
the
a plat-the
the
parameters
cess parame- Work
tion Separa-
tionDistance
Speed WSD mm Separation
stencil while distance
thethe of the worktable
worktable
stencil.
process. isthe
used asand the
a plat-
Printing pro- tion
tionWorkSpeed
Distance WSS
WSD mm/s
mm form
stencil to support
are
Separation separated
distance stencil.
PCB,
at
of the andend of
worktable PCB
the and
printing
and the the
ters pro- Work
Printing Separa- stencil
cess parame-
ters
Printing
tionDistance
tion Speed
Separation
pro- Work WSD
WSS
WSD mm mm
mm/s form
stencil toare separated
Separation
support
are separated
at theand
theprocess.
distance
stencil.
PCB,
at the
end
of the
endthe
the stencil.
ofworktable
the printing
of PCB and the
the printing
cess
cess parame-
parame-
ters tion
tion Distance
Speed
Separa- Separation distance process.
stencil.
of
and
process. the worktable and the
Printing pro- Work
cess parame- Separa-
Distance WSD mm stencilFor
For different
are separated
Separation
different stencil
distance
stencil atthicknesses,
of end of the
the worktable
thicknesses,
process. the work-
printing
the and
work- the
ters
ters Worktable
tion
Work Distance
Separa-
Worktable WSD mm Separation distance stencil.
ofthicknesses,
the worktable and the
cess parame- tion Distance WSD mm bench For
bench different
needs
needsFor
stencil
different
different separation
process.
differentstencil.
separation the work-
distances
distances to
to
ters Work
Work
tion Separa- WPHO
Worktable
Printing
Separa-
Distance
Worktable mm Separation
For different
Separation distance
stencil
distance ofstencil
of
stencil.the
thicknesses,
the worktable
worktable
thicknesses, the and the the
work-
and the
ters Printing
Worktable WPHO
WSD mm bench
ensure needs
that different
the
workbench solder
needs separation
paste
differenton distances
the PCB
separation is to
Work
tion
Height Separa-
Distance
Printing
PrintingOffset WSD
WPHO mm
mm bench ensure
Separation that the
distance
needs different
For different solder
stencil of
stencil.thepaste on
worktable
separation distances
thicknesses, the PCB
and
the work-to is
the
tion
Height
Work Distance
Offset
Printing
Separa- WPHO
Worktable WSD
WPHO mm mm ensure
mm Separation that
distances
Forcompletely
different the
distanceto
stencilstencil.
solder
ensure paste
that
ofthicknesses,
released the from on
the
worktable
the the
solder PCB
the and paste
work-
stencil. is
the
tion Distance
Height
Height Offset
Worktable WSD mm bench Forcompletely
ensure that
needs
different the released
different
stencilstencil.
solder from
paste
separation
thicknesses, thethe
on stencil.
thePCB
distances
work- isto
Height
tion Offset
Distance
Printing WPHO mm bench on
completely
needs thedifferent
PCB is completely
released
stencil. from
separation released
the stencil.
distances from to
Worktable
Offset For different
completely stencil
released thicknesses,
from on thethe thePCBwork-
stencil.
Printing
Worktable WPHO mm bench ensure
For thatdifferent
different
needs the solder
stencil the paste
stencil.
thicknesses,
separation the work-
distances isto
Squeegee
Height
Worktable
Printing
Squeegee Sep-
Offset
Sep- WPHO mm bench Separation
ensure
Separation
For different
needs speed
thatdifferent
the
speed
stencil of
solder the squeegee
paste
ofthicknesses,
the on the
squeegee
separation and
PCB
and
the work-
distances the
theisto
Height Offset
Worktable
Squeegee
Printing SSS
SSS
WPHO mm/s
mm bench
mm/s completely
ensure needs released
thatdifferent
the solder from
paste
separation thethe
on stencil.
PCB
distances isto
Squeegee
aration
Printing
Height
aration Sep-
Speed
Offset
Speed WPHO mm Separation
For different
completely speed
Separationstencil of
released
speed the
stencil. squeegee
thicknesses,
stencil. from
of the the theandwork-
stencil.
squeegee the
and
Squeegee
Worktable
Separation Sep- WPHO SSS
SSS mm/smm/s bench
ensure needs
Separationthatdifferent
speed of separation
thereleased
solder the paste on
squeegee distances
the PCBthe
and isto
Printing
Height Offset
aration Offset
Speed SSS mm
mm/s bench ensure that
completely
needs the solder
different stencil.
the paste
from
separation on
the
stencil.on the the PCB
stencil.
distances is to
Height
Speed
aration
Squeegee Speed
Printing Sep- WPHO mm ensure
Separationthat
completely the
speed solder
released
stencil.
of thefrompaste
from the
squeegee PCB
stencil.
and the is
Height Offset
Squeegee Sep- SSS mm/s completely
ensure
Separation that the released
speed solder
of thefrompaste the
on
squeegee stencil.
the PCB
and the is
Squeegee
Height Offset
aration
Squeegee Sep-
Speed
Sep- SSS mm/s completely
Separation released
speed stencil.
ofofthe the stencil.
squeegee and the
aration
Squeegee Speed SSS mm/s Separation
completely
Separation distance
released
distance stencil.
of the
thefromsqueegee
the
squeegee and
stencil.
andthe the
the
Squeegee
aration
Squeegee
aration
aration Sep-
Dis-
Sep-
Speed
Dis- SSD
SSD mm
mm Separation
Separation speed
Separation
speed of the
distance
of the of
stencil. squeegee
the squeegee
squeegee and
and and
the
Squeegee
SeparationSep- SSS
SSD
SSS mm/s
mm/s mm Separation distance of
stencil. the squeegee and the
Squeegee
aration
aration Sep-
Dis-
Speed
tance SSD mm Separationdistance
Separation speedstencil.
ofof
thethe
stencil. squeegee
squeegeeand
stencil.
the andthe the
aration
aration
Squeegee Speed
tance
Distance Dis-
Sep- SSS
SSD mm/s
mm Separation speed stencil.
stencil.
of the squeegee and the
aration
Squeegee Speed
tance Sep- SSS mm/s Separation distance stencil.
of the squeegee and the
stencil.
xtance
aration
aration
Micromachines 2022,Squeegee
13, FOR Dis-
Speed
Sep-
PEER SSD
REVIEW mm Separation distance of the squeegee and the
stencil. 10 of 16
Printing pro- aration
Printing pro- Squeegee Dis-
Sep- SSD mm stencil.
Separation distance of the squeegee and the
Average
Squeegee
aration
Average
Average Pres-
tanceDis-
Sep-
Pres- SSD mm stencil.
Printing
cess
cess status
statuspro-
pa-
pa- tance AP
AP
AP kg
kg kg Separation
Separation distance
Average
distance
Average Average of
squeegee
of
stencil. the squeegee
the
squeegee
squeegee squeegee
pressure.
pressure.
pressure. and the
and the
Printing
Printing Squeegee
aration
pro- Average sure
Pressure Sep-
Dis-
Pres- SSD mm
cess status
rameters pa- aration
tance
sure
Average Dis-
Squeegee Dis-
Pres-
Sep- SSD
AP mm
kg Separation distance
Average of
stencil.
squeegee the squeegee
pressure. and the
rameters
cess process pa- aration
statuspro-
Printing tance
sure SSD
AP mm
kg Separation Average
distance stencil.
squeegee pressure. and the
of the squeegee
rameters
Printing pro- Averagetance
sure
aration Pres-
Dis- SSD mm stencil.
status tance
rameters
cess statuspro-
Printing pa- Average
Minimum Pres- AP kg Average squeegee stencil. pressure.
parameters
cess statuspro-
Printing pa- Average surePres- MinP
tance
Minimum AP kg kg
kg Average
Minimum squeegee
squeegee pressure.
pressure.
rameters
Printing
cess statuspro-
pa- Average sure MinP
AP kg Minimum
Average squeegee squeegee pressure.
pressure.
rameters Pressure
Pressure
Average surePres-
Pres-
Printing
cess statuspro-
pa- AP kg Average squeegee pressure.
rameters
cess statuspro-
Printing pa- Average surePres- AP kg Average squeegee pressure.
cess status pa- Average
rameters surePres- AP kg Average squeegee pressure.
rameters
cess status pa- sure AP kg Average squeegee pressure.
rameters sure
rameters Maximum
MaxP kg Maximum squeegee pressure.
Pressure

Cleaning Sup-
CST s The time taken to clean the stencil.
ply Time
Micromachines 2022, 13, x FOR PEER REVIEW 10 of 16
Micromachines 2022, 13, x FOR PEER REVIEW 10 of 16
Micromachines 2022,
Micromachines 2022, 13,
13, xx FOR
FOR PEER
PEER REVIEW
REVIEW 10 of
10 of 16
16
Micromachines 2022, 13, x FOR PEER REVIEW 10 of 16
Micromachines 2022, 13, x FOR PEER REVIEW 10 of 16
Minimum
Micromachines 2022, Minimum
13, 860 MinP kg Minimum squeegee pressure. 10 of 15
Pressure MinP kg Minimum squeegee pressure.
Minimum
Minimum
Pressure
Minimum MinP
MinP kg
kg Minimum squeegee
Minimum squeegee pressure. pressure.
Pressure
Pressure MinP kg Minimum squeegee pressure.
Minimum
Pressure
Maximum MinP
Table 1. Cont. kg Minimum squeegee pressure.
Pressure MaxP kg Maximum squeegee pressure.
Maximum
Pressure
Type Parameter MaxP kg Unit Maximum squeegee pressure.
Pressure Abbreviation
Maximum
Maximum MaxP kg Maximum squeegee
Meaning
squeegee pressure. pressure.
Sample
Maximum
Pressure MaxP kg Maximum
Pressure MaxP kg Maximum squeegee pressure.
Maximum
Pressure
Maximum
Cleaning Sup- MaxP
MaxP kg kg Maximum Maximum squeegee
squeegee pressure.
pressure.
Pressure
Pressure CST s The time taken to clean the stencil.
Cleaning Sup-
ply Time
Cleaning Sup- CST s The time taken to clean the stencil.
Cleaning Sup-
ply Time CST s The time
time taken taken to to clean
clean the the stencil.
stencil.
Cleaning
Cleaning
ply Time
TimeSup- CST s The
ply CST s The time taken to clean the stencil.
Supply
Cleaning
ply TimeSup- CST s The time taken to clean the stencil.
Printing CST s The cleaning
time taken to clean
process
Time
Automatic
ply Time Different speeds willthe stencil.differ-
produce
Automatic AC s Different
status Cleaning AC s ent cleaning effects on the stencil.differ-
cleaning speeds will produce
Automatic
Automatic
Cleaning Different
Different cleaning
cleaning
entDifferent
cleaningcleaning speeds
speeds will
will
effects speeds
on the produce
produce
stencil. differ-
differ-
parameters Automatic
Automatic AC
AC ss Different cleaning speeds will will
produce produce
differ-
Cleaning
Cleaning AC s ent cleaning
entdifferent
cleaning effects on
effectseffects the
on theonstencil. stencil.
Cleaning
Automatic AC s Different
Different automatic cleaning cleaning countthesettings
stencil.
Cleaning AC s ent cleaning
cleaning speeds
effects willthe
on produce
stencil. differ-
Automatic
Cleaning ACC \ Different
will produce automatic
different cleaning
cleaning count
effectssettings
on the
Automatic ent cleaning
Different automatic effects
cleaningon the stencil.
count settings
Automatic
Cleaning
Automatic ACC \ Different
will produce automatic
different
Different cleaning
cleaning
automatic count
effects
cleaning settings
on
count the
Count
Automatic Different automatic stencil.
cleaning count settings
Cleaning
Cleaning
Cleaning
Count ACC
ACC
ACC \
\ will
\will produce
produce
settings different
different willstencil. cleaning
cleaning
produce different effects
effects on the
on
cleaning the
Automatic
Cleaning
Count
Count ACC \ Different
will produce automatic
different cleaning
cleaning
stencil.
effects on the count
effects
stencil. settings
on the
Count stencil.
Solder Paste
Cleaning
Count ACC \ Ratio
will of the measured
produce different value of effects
cleaning
stencil. solder on pastethe
Solder Paste SPV \ Ratio of the measured value of solder paste
Volume
Count
SolderPaste SPV \ volume to the theoretical
stencil. value.
Solder
Solder
VolumePaste Ratio of
Ratio of Ratio
the measured
the
volume measured
of tothe themeasured value value
value
theoretical of solder
of solder
value. paste
paste
of solder
PastePaste
Solder SPV
SPV
SPV \
\ \ Ratio of the measured value of solder paste
Volume
Volume SPV \ volume
paste volume
volume to the theoretical
to the theoretical
to the theoretical value.
value. value.
Volume
Solder Paste Ratio volume
of the measured value of solder
Volume SPV \ to the theoretical value. paste
Solder Paste
Volume Ratio volume
of the measured to the theoreticalvalue of solder value. paste
SPI results Solder Paste SPA \ Ratio of the measured value of solder paste
SPI Area
Solder area
Ratioto ofthe the theoretical
measured value value.
of solder
SPIresults
results Solder
Solder
AreaPaste
Paste
Paste Area
SPA
SPA \ Ratio of
\ Ratio ofarea
the measured
the measured
to the value of
value
theoretical
paste area to the theoretical value.
ofvalue.
solder paste
solder paste
SPI results
SPI results Solder Paste SPA
SPA \
\ Ratio of the measured value of solder paste
SPI results Area
Area SPA \ area to
area to the the theoretical
theoretical value. value.
Solder
AreaPaste Ratio ofarea the measured
to the value ofvalue.
theoretical solder paste
SPI results SolderPaste
Solder SPA \ Ratio ofarea the measured value value ofvalue.
solder paste
Area SPH \ Ratio to of the
the theoretical
measured of solder
Solder
PastePaste SPH \ Ratio of the measured value of solder paste
Height SPH \ height
paste to heightthe theoretical
to the theoretical value. value.
Solder
Height
Solder Paste
Paste
Height Ratio of
Ratio of the measured
the
height measured
to the theoretical value of
value ofvalue.
solder paste
solder paste
Solder Paste SPH
SPH \
\ Ratio of the measured value of solder paste
Height
Height SPH \ height to
height to the the theoretical
theoretical value. value.
Solder Paste
Height Ratio of the measured
height to the value of value.
theoretical solder paste
SPHAAsummary \
Height summaryofofthe theexact number
height
exact number to the ofoftheoretical
samples
samplesbelonging value. totoeach
belonging eachdefect
defecttypetypeand andthe the
labels Aissummary
shown inofTable
the exact
2. A number
total of of samples
70% of the belonging
raw data wereto each
used defect
as the type
trainingand set,
the
labels Aisis shown inofTable
summary 2. A total
the exact
exact number of 70% of
of samplesthe rawbelonging
samples data weretoused each asdefect
the training
type set, the
and and
labels
and A
the summary
shown inof
remaining the
Table
30% of2.the number
A total
raw dataof 70% of
were of used
the raw belonging
as data
the test to each
were
set. useddefect
as thetype
training and the
set,
the
labels remaining
A summary
isremaining
shown in 30%of of
the
in Table the
Tableof2. raw
exact
2.the data
number
A total
totaldata were
of 70% of
70% used
samples
ofused as
the raw the
raw test
belonging
data set.
wereto each defect
used as
as the type
the training and
training set, the
set,
labels
and theis shown 30% A raw of were of the data
as the testwere
set. used
labels
and Aissummary
the shown inof30%
remaining the of
Table
30% exact
2.the
A raw
the number
total
raw dataof 70% of
were samples
ofusedthe raw
used belonging
as data
the testwereto each
set. useddefect
as thetype
training and set,
the
and
Table the
Table2.is
remaining
2.Description
Description of
ofthe
of
theraw
raw data.
data.
data were as the test set.
labels
and the
Table
shown in
remaining of
2. Description
Table
30% theof
2.
raw
A total
thedata. of 70% of
raw data were used as the test set. the raw data were used as the training set,
and
Tablethe remaining of
2. Description
Description 30%
of theofrawthedata.
raw data were used as the
data. test set.
Number
Table 2.
Defect Type
the raw
Abbreviation Label NumberofofTrain- Number Number of Test
of
Defect Type Abbreviation
Table 2. Description of the raw data. Label Number
ing of
Samples Train- Number
Samples of Test
Defect Type Abbreviation Label Training Number Samples
of Train-
Train- NumberTest Samples
Number of Test
Test
TableDefect
2. Description
Type of the raw data. +
Abbreviation Label Number
ing Samplesof Samples of
Large
Defect
Large Volume
Type
Volume Vol + Vol
Abbreviation 0 Label 0 Number
ing 1389
1389 of
SamplesTrain- Number 596
Samples596of Test
Large
Defect Volume
Type − Vol
Abbreviation 01 ing Samples
1389 Samples
596
1 Label
+
Small
Small Volume
Volume Vol Vol−+ Number
ing543 543of Train- Number
Samples 232
Samples232of Test
Large
Defect
Large
Small Volume
Type
Volume + Vol
Abbreviation
Vol +−+ Label
01 0 1389
1389
543 596
596
232
Large Area
LargeVolume
Area Area Area + 2 2 531531
ing Samples 228
228
Samples
Large
Small
Small Volume
No Solder NS Area
Vol−+−−+
Vol 3 1201 1389
543
543
634
596
232
232
271
Large
No Area
Solder NS 3 531
634 228
271
Large
Small
Large
Small
Volume
Volume
Area
Area Area
Vol
Vol
−Area
+
− + 4
0
1 2 1389
2561543
531 596
232
228
1098
Large
No Area
Solder Area
NS ++ 23 531
634 228
271
SmallVolume
XSmall
Large Area
Area Area
X+ AreaVol −− 14 2561
543 1098
232
Positive
No Solder
No
Small
Offset
Solder
Area +
NS −+
NS
Area
5
3423 5663531
634
634
2561
228
2427
271
271
1098
XY Positive
Large
Positive
No SolderOffset
Area
Offset Y X
Area
NS
+ + 6 25 5663
818531 2427
228
350
X
Y
Small
Small
Positive Area
Area
Offset Y− Area
Area
X + −−
−
4534 634
2561
2561
5663
271
1098
1098
2427
Y Negative
Positive
No Offset
Offset
Solder
Small Area Y
NS
Area
+ 7 3 6 3871818
634 1659
350
271
5645 2561 1098
−
X Positive
X Positive Offset
Offset H+ X X++++ 5663
5663 2427
2427
YY
X
High Height
Negative
Small
Positive Offset
Area
Offset
Y
Y+−+ −
Area
X
8
547 818
9019
3871
2561
5663
350
3866
1659
1098
2427
Y
Y Positive
Y Negative Offset
Offset Y
Y+−+ 678 6 818
818
3871 350
350
1659
High
X Negative
PositiveHeight
Offset HX−++−−+ 567 9019
5663 3866
2427
YYY Positive
Negative
High
Offset
Offset
Offset
Height H
Y
Y + 78
818
3871
3871
9019
350
1659
1659
3866
Y When the
Positive solder paste volume
Offset Y +− (SPV) value6 is less than 818 70%, the volume is 350 too small
Y Negative
High
High Offset
Heightand when the H
Height HY +
++ 7
88is lessthan 3871
9019 the volume 3866
9019 1659
3866
(Small When
Y NegativeVolume)
the solder
Offset paste volume SPV
Y+ value
(SPV) is
value greater than 150%,
70%, the volume is
is toosmall
too large
High Height
When the solder paste volume H −
(SPV) value 87is less than 70%, 3871
9019 the volume
1659
3866
is too small
(Large
(Small Volume).
Volume) A solder
and when the SPV paste area (SPA)
value is greaterthat is greater
than 150%,than 140% is too
the volume is3866large (Large
too large
High
When Height
the solder paste volume H+ (SPV) value 8is less 9019
than150%,
70%, the volume is too small
(Small
Area),
(Large Whenandthe
Volume)an solder
Volume). SPAand
A thatpaste
when
solder volume
the
is paste
less SPV
than
area (SPV)
value
70%(SPA) isvalue
istoogreater
thatsmall isisless than
than
(Small
greater 70%,
Area).
than the
the
140% volume
When
is is
thelarge
too too
solder small
too(Large
large
paste
(Small
(Small
(Large When the
Volume)
has noVolume)
Volume). solder
solder pasteand
and paste
when
whenpaste
A solder
volume volume
the
the SPV
SPV
areavalue
(SPV), (SPV)
value
no(SPA) value
is
isthat
solder greater
greater
paste is less than
than
than(SPA),
is greater
area 70%,
150%,
150%,
than 140% the
the volume
thenovolume
or is is
is
too large
solder too
too small
large
too(Large
paste large
height
(Small
(Large When the solder
Volume)
Volume). and
A solderpastepaste
when
solder volume
the area(SPV)
SPV value
(SPA) value
isthat
greater
that is
is less than150%,
than
greater 70%,140%
than the volume
the is too
too is too(Large
large small
large
(Large
(SPH),Volume).
a No Solder A defect paste
will area
occur. If(SPA)
the center is greater
point of the than 140%
solder is
paste large
deviates (Large
by more
(Small Volume)
(Large Volume). and
A whenpaste
solder the SPV area value
(SPA) isthat
greater is than 150%,
greater than the volume
140% is too is too(Large
large large
than 20% in the positive direction of the X-axis, an X Positive Offset defect will occur. If the
(Large Volume). A solder paste area (SPA) that is greater than 140% is too large (Large
center point of the solder paste deviates by more than 20% in the positive direction of the Y
axis, a Y Positive Offset defect will occur. If the center point of the solder paste is offset by
 Micromachines 2022, 13, 860 11 of 15

more than 20% in the negative direction of the Y axis, a Y Negative Offset defect will occur.
If the solder paste height (SPH) is greater than 170%, the height is too high (High Height).

3.2. Experimental Protocols

Using the grid search method [25], the values of the CNN parameters were calculated
and are shown in Figure 3. Sample data of size 13 × 13 were input into the CNN model,
and convolution was performed at a convolution kernel size of 5 × 5 to obtain the first
convolution feature of size 13 × 13@16, where 16 represents three mapping layers, and the
padding value was set to 2. Each mapping layer had an image size of 13 × 13. The first
convolutional feature was then maximally pooled on the first pooling layer with a pooling
filter size of 3×3 to obtain the first pooling feature of size 6 × 6@16, and the stride value
was set to 2. The first pooling feature was convolved on the second convolutional layer at a
convolutional kernel size of 3 × 3 to obtain a second convolutional feature of size 4 × 4@32.
The second convolutional feature was maximally pooled on the second pooling layer at a
pooling filter size of 3 × 3 to obtain a second pooling feature of size 2 × 2@32, the stride
value was set to 2, and the padding value was set to 1. The one-dimensional expansion
of the second pooled feature was performed on the expansion layer to obtain the one-
dimensional feature vector. The one-dimensional feature vector was used as the input
into the three-layer forward neural network. The output of the three-layer forward neural
network was activated using the ReLU activation function. The final output was the
predicted value of the defect type corresponding to the sample.

4. Results and Discussion

4.1. Results of CNN-Based Defect Recognition
First, the experimental data went through feature reconstruction. According to
Section 2.1, each feature vector was multiplied by other feature vectors. After feature
interaction, RF was employed to select 169 essential features from 210 interaction features
since features with RF importance scores less than 0.00045 were deleted. The RF scores of
interaction features are shown in Figure 5. Figure 5 shows that the solder paste height (SPH)
Micromachines 2022, 13, x FOR PEER REVIEW 12 of 16
influenced almost all of the main RF importance scores of feature interactions. Subsequently,
min–max normalization was utilized to eliminate the influence of the dimensionality of the
data. Before feature extraction, the 169 remaining interaction features were transformed
into a 13 × 13 matrix.

Figure 5. RF importance scores of interaction features.

Figure 5. RF importance scores of interaction features.

The training data were input into the CNN model for model training after feature
conversion.
The confusion matrix of the proposed method for the test set is shown in Figure 6.
The data in the diagonal column represent the proportion of samples that were clustered
Micromachines 2022, 13, 860 12 of 15
Figure 5. RF importance scores of interaction features.

The training
The training data
data were
were input
input into
into the
theCNN
CNNmodelmodelfor formodel
model training
trainingafter feature
after fea-
conversion.
ture conversion.
Theconfusion
The confusionmatrix
matrixof ofthe
theproposed
proposed method
method forfor the
the test
test set
set isis shown
shown in in Figure
Figure 6.6.
Thedata
The dataininthethediagonal
diagonalcolumn
columnrepresent
representthe theproportion
proportionofofsamples
samplesthat thatwere
wereclustered
clustered
togetherwithout
together withoutbeing
beingmisclassified.
misclassified. The
The results
results indicate
indicate thatthat a few
a few of theof the samples
samples in la-
in labels
+ , YX
Xbels + ,+,and Y− were
Y+, and Y− were mistaken
mistaken forfor other
other defect
defect types,
types, but
but mostofofthe
most thesamples
samplesunder
underthethe
different conditions were clustered together without being misclassified
different conditions were clustered together without being misclassified by the proposed by the proposed
method.The
method. Theresults
resultsindicate
indicatethat
thatthe
theproposed
proposedmethodmethodcan caneffectively
effectivelyrecognize
recognizemost
mostofof
thefeatures
the featuresofofthe thenine
ninetypes
typesofofdefects
defectsininsolder
solderpaste
pasteprinting.
printing.

Figure6.6.Confusion
Figure Confusionmatrix
matrixofofthe
theCNN
CNNprediction
predictionresults.
results.

Micromachines 2022, 13, x FOR PEER REVIEW

Thevalidation
The validationaccuracy
accuracyand
andloss
lossof
ofthe
theproposed
proposedmethod
methodfor
forfeature 13and
featureextraction
extraction of
and16
classificationare
classification areshown
shownininFigure
Figure7.7.

Figure7.7. Diagram
Figure Diagram of
of the
theCNN
CNNmodel’s
model’svalidation
validationaccuracy
accuracyand
andloss.
loss.

4.2.
4.2. Method
Method Comparison
Comparison andand Evaluation
Evaluation Index
Index
Other
Other methods
methods werewere also
also applied
applied onon the
thesame
samedataset
datasetto toelucidate
elucidatethe
theeffectiveness
effectiveness
of
of the proposed CNN-based defect recognition method. Using the same neural network
the proposed CNN-based defect recognition method. Using the same neural network
model
model structures,
structures, aa deep
deep neural
neural network
network (DNN)
(DNN) [14],
[14], aa deep
deep belief
belief net
net (DBN)
(DBN) [16],
[16], and
and
sparse
sparseauto
autoencoder
encoder (SAE) [17][17]
(SAE) werewere
also also
introduced. The deep
introduced. The neural
deep network approaches
neural network ap-
were all tested
proaches were over 3000 over
all tested iterations five times,five
3000 iterations and the mean
times, classification
and the accuracyaccu-
mean classification was
used for the
racy was comparison.
used The parameters
for the comparison. were initialized
The parameters with a weight
were initialized with matrix that
a weight had
matrix
athat
mean of 0.5 and a standard deviation of 0.5. The traditional machine
had a mean of 0.5 and a standard deviation of 0.5. The traditional machine learninglearning model
model support vector machine (SVM) [18] was introduced in order to demonstrate the
advantages of the CNN in solving multi-classification problems.
(1) CNN without feature reconstruction (FR): The same training set was utilized for the
CNN with the same structure without feature reconstruction. The learning rate and
Micromachines 2022, 13, 860 13 of 15

support vector machine (SVM) [18] was introduced in order to demonstrate the advantages
of the CNN in solving multi-classification problems.
(1) CNN without feature reconstruction (FR): The same training set was utilized for the
CNN with the same structure without feature reconstruction. The learning rate and
the batch size were 0.01 and 64, respectively, which are also the values that were used
for the proposed CNN. Then, the trained model was directly applied to the test set.
(2) DNN: The architecture of the standard DNN is 169-128-64-32-9. The learning rate was
0.001, and the batch size was 64.
(3) DBN: The architecture of the standard DBN is 169-128-64-32-9. The learning rate and
the batch size were 0.01 and 64, respectively.
(4) SAE: The architecture of the standard SAE is 169-128-64-32-9. The learning rate and
the batch size were 0.01 and 64, respectively.
(5) SVM: The RBF kernel was applied. The penalty factor and the radius of the kernel
function were 0.1 and 0.5, respectively.
After validating each model five times, the mean and standard deviation of each model
were obtained. As shown in Figure 8, the model proposed in this paper has the best stability
and accuracy. Table 3 lists and compares the various defect recognition methods. The
weighted average recognition accuracies of the proposed CNN with feature reconstruction
and without feature reconstruction are 96.97% and 82.17%, respectively. The weighted
Micromachines 2022, 13, x FOR PEER REVIEW
average recognition accuracies of the CNN, DNN, DBN, SAE, and SVM are 96.97%, 91.08%,14 of 16

89.88%, 91.66%, and 74.84%, respectively. The results prove that, compared with DNN,
DBN, SAE, and SVM, the proposed model is more effective at recognizing defects in solder
paste printing.

Figure 8. The results of different models.

Figure
Table 3.8.AThe results of between
comparison differentthe
models.
proposed model and other models.

Table 3. A comparison between the proposed model

Models and otherAverage
Weighted models.Recognition Accuracy (%)
CNN with FR 96.97
Models
CNN without FR Weighted Average Recognition
82.17 Accuracy (%)
DNN
CNN with FR 96.97 91.08
CNN withoutDBN
FR 82.17 89.88
SAE 91.66
DNN SVM 91.08 74.84
DBN 89.88
SAE 91.66
SVM 74.84

Figure 9 displays the accuracy curves of the defect recognition results for the nine
types of defects using the proposed method and the other five methods. The weighted
average recognition accuracies of the CNN with feature reconstruction and without fea-
Micromachines 2022, 13, 860 14 of 15

Figure 9 displays the accuracy curves of the defect recognition results for the nine
types of defects using the proposed method and the other five methods. The weighted
average recognition accuracies of the CNN with feature reconstruction and without feature
reconstruction range from 91.56% to 100% and from 61.43% to 99.7%, respectively. The
weighted average recognition accuracies of the CNN, DNN, DBN, SAE, and SVM range
from 91.56% to 100%, from 59.2% to 100%, from 81.14% to 100%, from 45.94% to 100%,
and from 66.57% to 100%, respectively. The defect recognition accuracies of the other five
methods are low and vary remarkably. The curve difference shown in Figure 9 and the
Micromachines 2022, 13, x FOR PEER REVIEW 15 of 16
comparison of the quantitative results shown in Table 3 illustrate the high accuracy of the
proposed defect recognition method.

Figure9.9.Results
Figure Resultsof
ofdifferent
differentmodels.
models.

5.5.Conclusions
Conclusions
In
Inthis
thisstudy,
study,we
weinvestigated
investigatedSMT
SMTproduct
productdefect
defectrecognition
recognitionbased
basedononmulti-source
multi-source
and
and multi-dimensional data reconstruction in order to increase the low defectrecognition
multi-dimensional data reconstruction in order to increase the low defect recognition
rate
rateand
andaddress
addressthe
theinsufficient
insufficientutilization
utilization of
ofmulti-source
multi-source solder
solder printing
printing data
data in
in SMT
SMT
production. Our conclusions are as follows.
production. Our conclusions are as follows.
(1)
(1) Features
Features correlated
correlated to
to defects
defects were
were enhanced
enhanced by by feature
feature reconstruction
reconstructionusingusingfeature
feature
interaction, feature
featureselection,
selection,and
andfeature
feature conversion.
conversion. Compared
Compared withwith the CNN
the CNN with-
without
out feature reconstruction, the performance metric of recognition accuracy waswas
feature reconstruction, the performance metric of recognition accuracy im-
improved
proved byby 14.80%.
14.80%.
(2)
(2) Results
Results from
from the
the experiment
experiment showshow that
that the
the accuracy
accuracy ofof the
the proposed
proposed defect
defect recogni-
recogni-
tion
tion model with feature reconstruction is 96.97% and the standard deviation is
model with feature reconstruction is 96.97% and the standard deviation is 1.43.
1.43.
Compared
Comparedwith withfour
fourother methods,
other methods,the proposed
the proposeddefectdefect
recognition model model
recognition has higher
has
accuracy and better stability.
higher accuracy and better stability.
Thus,
Thus, the
the proposed defect
defect recognition
recognitionmethod
methodbased
basedononmulti-source
multi-source andand multi-
multi-di-
dimensional data reconstruction has the potential to facilitate the widespread
mensional data reconstruction has the potential to facilitate the widespread application of application
of defect
defect recognitionininthe
recognition theSMT
SMTproduction
productionquality
qualitycontrol
control process.
process. SMT,
SMT, as
as aa new
new type
typeofof
electronic
electronicpackaging
packagingtechnology,
technology, has
haspenetrated
penetrated every industry
every industryandandevery field.
every TheThe
field. defect
de-
recognition method
fect recognition proposed
method in this
proposed paper
in this has has
paper a broad range
a broad of application
range prospects
of application in
prospects
such fields as electronics manufacturing and aerospace engineering.
in such fields as electronics manufacturing and aerospace engineering.

Author Contributions:Methodology,
AuthorContributions: Methodology,Q.W.;
Q.W.;writing—original
writing—originaldraft
draftpreparation,
preparation,J.C.;
J.C.;visualization,
visualization,
Y.Y.;
Y.Y.;writing—review
writing—reviewand andediting,
editing,Z.Q.;
Z.Q.;supervision
supervisionand
andfunding
fundingacquisition,
acquisition,X.K.
X.K.All
Allauthors
authorshave
have
read
readand
andagreed
agreedtotothe
thepublished
publishedversion
versionofofthe
themanuscript.
manuscript.
Funding: This
Funding: This work
work was
was supported
supported by
by the
the Key
KeyResearch
Researchand
andDevelopment
Development Program
Program of
ofShaanxi
Shaanxi
Province,
Province, China (Grant Nos. 2020ZDLGY07-08 and 2020ZDLGY07-04) and the NationalKey
China (Grant Nos. 2020ZDLGY07-08 and 2020ZDLGY07-04) and the National KeyR&D
R&D
Program
ProgramofofChina
China(Grant
(GrantNo.
No.2019YFB1705404).
2019YFB1705404).
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: Not applicable.
Conflicts of Interest: The authors declare no conflicts of interest.
Micromachines 2022, 13, 860 15 of 15

Institutional Review Board Statement: Not applicable.

Informed Consent Statement: Not applicable.
Data Availability Statement: Not applicable.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Hou, J.; Xia, K.; Yang, F. Strip Steel Surface Defects Recognition Based on SOCP Optimized Multiple Kernel RVM. Math. Probl.
Eng. 2018, 2018, 9298017.
2. Cui, D.; Xia, K. Dimension Reduction and Defect Recognition of Strip Surface Defects Based on Intelligent Information Processing.
Arabian J. Sci. Eng. 2018, 43, 6729–6736. [CrossRef]
3. Li, X.; Zhang, J. Defect Recognition of Cold Rolled Plate Shape Based on RBF-BP Neural Network. In Proceedings of the 10th
World Congress on Intelligent Control and Automation (WCICA), Beijing, China, 6–8 July 2012; pp. 496–500.
4. Wang, Z.-S. Research on Image Detection and Recognition for Defects on the Surface of the Steel Plate Based on Magnetic Flux
Leakage Signals. In Proceedings of the 29th Chinese Control and Decision Conference (CCDC), Chongqing, China, 28–30 May
2017; pp. 6139–6144.
5. Yi, L.; Li, G.; Jiang, M. An End-to-End Steel Strip Surface Defects Recognition System Based on Convolutional Neural Networks.
Steel Res. Int. 2017, 88, 176–187. [CrossRef]
6. Guan, S.; Lei, M.; Lu, H. A Steel Surface Defect Recognition Algorithm Based on Improved Deep Learning Network Model Using
Feature Visualization and Quality Evaluation. IEEE Access 2020, 8, 49885–49895. [CrossRef]
7. Cai, N.; Lin, J.; Ye, Q.; Wang, H.; Weng, S.; Ling, B.W.K. A New IC Solder Joint Inspection Method for an Automatic Optical
Inspection System Based on an Improved Visual Background Extraction Algorithm. IEEE Trans. Compon. Packag. Manuf. Technol.
2016, 6, 161–172.
8. Zhao, X.; Cue, L. Defect pattern recognition on Nano/Micro integrated circuits wafer. In Proceeding of the 3rd IEEE International
Conference on Nano/Micro Engineered and Molecular Systems, Sanya, China, 6–9 January 2008; pp. 519–523.
9. Yuan, T.; Kuo, W. A model-based clustering approach to the recognition of the spatial defect patterns produced during semicon-
ductor fabrication. IIE Trans. 2007, 40, 93–101. [CrossRef]
10. Chen, N.; Men, X.; Hua CWang, X.; Han, X.; Chen, H. Research on Edge Defects Image Recognition Technology Based on Artificial
Neural Network. In Proceedings of the 13th IEEE Conference on Industrial Electronics and Applications, Wuhan, China, 31
May–2 June 2018; pp. 1929–1933.
11. Protopapadakis, E.; Voulodimos, A.; Doulamis, A.; Doulamis, N.; Stathaki, T. Automatic crack detection for tunnel inspection
using deep learning and heuristic image post-processing. Appl. Intell. 2019, 49, 2793–2806. [CrossRef]
12. Tripicchio, P.; Camacho-Gonzalez, G.; D’Avella, S. Welding defect detection: Coping with artifacts in the production line. Int. J.
Adv. Manuf. Technol. 2020, 111, 1659–1669. [CrossRef]
13. Zhang, H.; Yu, X. Research on oil and gas pipeline defect recognition based on IPSO for RBF neural network. Sustain. Comput.-Infor.
2018, 20, 203–209. [CrossRef]
14. Ye, D.; Hong, G.S.; Zhang, Y.; Zhu, K.; Fuh, J.Y.H. Defect detection in selective laser melting technology by acoustic signals with
deep belief networks. Int. J. Adv. Manuf. Technol. 2018, 96, 2791–2801. [CrossRef]
15. Dou, J.; Xu, C.; Jiao, S.; Li, B.; Zhang, J.; Xu, X. An unsupervised online monitoring method for tool wear using a sparse
auto-encoder. Int. J. Adv. Manuf. Technol. 2020, 106, 2493–2507. [CrossRef]
16. Song, J.-D.; Kim, Y.-G.; Park, T.-H. SMT defect classification by feature extraction region optimization and machine learning. Int. J.
Adv. Manuf. Technol. 2019, 101, 1303–1313. [CrossRef]
17. Ko, K.W.; Cho, H.S. Solder joints inspection using a neural network and fuzzy rule-based classification method. IEEE Trans.
Electron. Packag. Manuf. 2000, 23, 93–103.
18. Park, J.-M.; Yoo, Y.-H.; Kim, U.-H.; Lee, D.; Kim, J.H. D3PointNet: Dual-Level Defect Detection PointNet for Solder Paste Printer
in Surface Mount Technology. IEEE Access 2020, 8, 140310–140322. [CrossRef]
19. Cai, N.; Cen, G.; Wu, J.; Li, F.; Wang, H.; Chen, X. SMT Solder Joint Inspection via a Novel Cascaded Convolutional Neural
Network. IEEE Trans. Compon. Packag. Manuf. Technol. 2018, 8, 670–677. [CrossRef]
20. Park, J.-Y.; Hwang, Y.; Lee, D.; Kim, J.H. MarsNet: Multi-Label Classification Network for Images of Various Sizes. IEEE Access
2020, 8, 21832–21846. [CrossRef]
21. Jakulin, A. Machine Learning Based on Attribute Interactions. Ph.D. Thesis, Faculty of Computer and Information Science,
University of Ljubljana, Ljubljana, Slovenia, 2005; pp. 56–68.
22. Breiman, L. Random Forests. Mach. Learn. 2001, 45, 5–32. [CrossRef]
23. Lecun, Y.; Bottou, L.; Bengio, Y.; Haffner, P. Gradient-based learning applied to document recognition. Proc. IEEE 1998,
86, 2278–2324. [CrossRef]
24. Diederik, K.; Jimmy, B.; Adam, A. Method for Stochastic Optimization. In Proceedings of the 3rd International Conference for
Learing Representations, San Diego, CA, USA, 7–9 May 2015.
25. Bergstra, J.; Bengio, Y. Random search for hyper-parameter optimization. J. Mach. Learn Res. 2012, 13, 281–305.