A Review of Data Mining Applications For

Expert Systems with Applications 38 (2011) 13448–13467
Contents lists available at ScienceDirect
Expert Systems with Applications

journal homepage: www.elsevier.com/locate/eswa
Review
A review of data mining applications for quality improvement

in manufacturing industry
Gülser Köksal a,⇑, Inci
_ Batmaz b, Murat Caner Testik c
a
Department of Industrial Engineering, Middle East Technical University, Inonu Blvd., Ankara 06531, Turkey
b
Department of Statistics, Middle East Technical University, Ankara 06531, Turkey
c
Department of Industrial Engineering, Hacettepe University, Ankara 06800, Turkey
a r t i c l e i n f o a b s t r a c t
Keywords: Many quality improvement (QI) programs including six sigma, design for six sigma, and kaizen require
Knowledge discovery in databases collection and analysis of data to solve quality problems. Due to advances in data collection systems
Data mining and analysis tools, data mining (DM) has widely been applied for QI in manufacturing. Although a few
Quality improvement review papers have recently been published to discuss DM applications in manufacturing, these only
Six sigma
cover a small portion of the applications for specific QI problems (quality tasks). In this study, an exten-
Design for six sigma
Quality description
sive review covering the literature from 1997 to 2007 and several analyses on selected quality tasks are
Prediction provided on DM applications in the manufacturing industry. The quality tasks considered are; product/
Classification process quality description, predicting quality, classification of quality, and parameter optimisation.
Parameter optimisation The review provides a comprehensive analysis of the literature from various points of view: data han-
Data mining software dling practices, DM applications for each quality task and for each manufacturing industry, patterns in
Manufacturing the use of DM methods, application results, and software used in the applications are analysed. Several
summary tables and figures are also provided along with the discussion of the analyses and results.
Finally, conclusions and future research directions are presented.
Ó 2011 Elsevier Ltd. All rights reserved.
1. Introduction The aim of this study is to extensively review DM applications

in manufacturing industries to handle QI problems that are of wide
Quality improvement (QI) of industrial products and processes interest to practitioners as well as to outline related future
requires collection and analyses of data to solve quality related research directions and challenges for researchers. Recently, some
manufacturing problems. While QI programs such as six sigma other reviews of DM applications in manufacturing and/or product
(6r), design for six sigma (DFSS), and kaizen keep encouraging design have also appeared in the literature (see Braha, 2001; Cho-
collection of data to attack quality problems, with advances in udhary, Harding, & Tiwari, 2008; Feng & Kusiak, 2006; Harding,
automation and computer systems, data from manufacturing pro- Shahbaz, Srinivas, & Kusiak, 2006; Kusiak, 2006; Pham & Afify,
cesses is becoming more and more available. Although traditional 2005; Wang, Tong, Eynard, Roucoules, & Matta, 2007a). However,
data analysis tools have been successfully used in improving qual- these reviews are concerned with DM applications in manufactur-
ity of products and processes, now better tools exist to mine mas- ing in general, and cover only a small sample of the papers that are
sive data sets collected through computerized systems in the reviewed in this study. Furthermore, their context is too broad to
industry. answer the following open research questions in the scope of our
Quality problems may involve several input and output vari- interest:
ables that are not easy to model and/or optimise. As Yang and
Trewn (2004) point out, data mining (DM) and knowledge discov- 1. What are the characteristics of data collected in these applica-
ery in databases (KDD) have been successfully used for solving QI tions? Which data preparation and preprocessing approaches
and control problems involving multivariate data in various stages are used?
of product/process life cycles. Yet, a review of such problem solving 2. How many DM applications have been published (within the
approaches is needed to guide practitioners. scope of this study) in different manufacturing industries and
how does this number change over time? Which manufacturing
⇑ Corresponding author. Tel.: +90 312 2102285; fax: +90 312 2104786. industries have applied DM frequently and for what purpose
E-mail addresses: koksal@ie.metu.edu.tr (G. Köksal), ibatmaz@metu.edu.tr
(quality task)? Which DM methods are commonly used for each
_ Batmaz), mtestik@hacettepe.edu.tr (M.C. Testik).
(I. quality task and manufacturing industry?
0957-4174/$ - see front matter Ó 2011 Elsevier Ltd. All rights reserved.
doi:10.1016/j.eswa.2011.04.063
G. Köksal et al. / Expert Systems with Applications 38 (2011) 13448–13467 13449
3. Do these applications contribute to the literature in terms of practices, discusses the application studies for each quality task
new methods, new methodologies, or successful implementa- and for each manufacturing industry, presents patterns of DM
tion of the existing ones? What are the reasons of selecting function and method usage, examines results of the applications,
these methods? What are the reported benefits and shortcom- and surveys the software used in these applications. Finally, con-
ings of these methods? What are the results of these clusions are presented and future research directions are
applications? discussed.
4. Which typical patterns or sequences of DM functions and meth-
ods are used in these applications to solve the QI problems?
5. Which software has been used in these applications? What are 2. Quality improvement tasks
the reasons of selecting these software products? What are the
reported benefits and shortcomings of the software? A classification of the QI and control tasks and the context of the
review are provided in the following. The classification is based on
The review of the literature covers the publications from 1997 the purpose of a study, and it does not necessarily yield disjoint
through 2007. Applications in the manufacturing sector in general classes, since the final purpose of a study can be achieved in a hier-
are considered and the publications are selected for the review archy of lower level purposes. In addition, the aim of this classifi-
based on the selection criteria provided in Table 1. cation is not to identify disjoint classes, but to present an
Each publication selected for review is analysed in detail to fill analysis of the DM applications considering different purposes in
out a table that summarises its basic characteristics which are of QI. Hence, before such a classification is presented, some back-
interest in this study as shown in Table 2. This table serves as ground on various quality initiatives and commonly used quality
the major source of information to answer the research questions. problem solving approaches is given as well as the QI and control
For this purpose, the applications are sorted according to a charac- activities that occur in a life cycle of a product.
teristic under consideration, and/or simple statistics are calculated As quality gained more and more importance over time, many
for those that satisfy a certain property. The complete table is quality initiatives and concepts have emerged. Inspection (100%),
available in Köksal, Batmaz, and Testik (2008). statistical quality control (SQC), total quality control (TQC), zero
Classification of DM functions and methods for a particular defects, total quality management (TQM), kaizen, ISO 9000 quality
application area is an important activity on its own. Even though standards, quality award programs (Malcolm Baldrige, European
DM functions are typically classified as data summarisation, classi- Quality Award and so on), 6r, DFSS, lean six sigma have been
fication, prediction, clustering and so on, these classes depend very among the most recognised ones (Fasser & Brettner, 2002; Mont-
much on the type of knowledge that can be discovered in dat- gomery, 2005). One apparent trend in the evolution of these initia-
abases (Choudhary et al., 2008; Wei, Piramuthu, & Shaw, 2003). tives is the emphasis on more proactive approaches and ‘‘upstream
Assigning the methods to the functions is also challenging, since (design) processes’’ (Kolarik, 1995; Montgomery, 2005; Phadke,
some DM methods serve multiple DM functions. In this study, 1989; Taguchi, Chowdhury, & Taguchi, 2000). Furthermore,
DM functions and methods used in the applications are classified emphasis on the bottom-line results in shorter time periods has
based on the desired knowledge to be discovered from the quality also gained importance and led to the 6r quality programs. These
data and the way they are used in the applications. programmes have received considerable recognition around the
In the following sections, a classification of QI and control tasks world due to their success, especially in improvement of ‘‘on-line’’
within the scope of this study is given first. Next, the KDD process, or downstream processes using the so called define-measure-ana-
relevant DM functions are briefly described. A comprehensive anal- lyse-improve-control (DMAIC) approach to reach six sigma (less
ysis of the reviewed literature from various points of view is pro- than 3.4 part per million (PPM) defectives) quality levels (Brady
vided subsequently. The analysis summarises data handling & Allen, 2006; Fasser & Brettner, 2002). Realising that such high
Table 1
Literature selection criteria and assessment methods used in the review.
Criteria Desired value Assessment

Contextual criteria
Purpose of the study Quality characterization and/or quality prediction and/or quality Analyzing the abstract, keywords; identifying types of quality tasks
classification and/or parameter optimisation involved, relevance of the cited references to the domain
Existence of an Yes Analyzing the abstract, keywords; identifying types of DM functions
industrial and methods involved, existence of case specific results, relevance of
application of a DM the cited references to the domain
approach
Industrial sector of the Manufacturing Checking if the type of product or process that the application is
application concerned with is covered by ISIC (2009) Section C.
Bibliographical criteria
Date of publication January 1997–December 2007 Checking if the publication’s date of print is within the desired time
period
Type of publication Full paper, conference proceeding, book, book chapter, thesis/ Analyzing the source of publication
dissertation (if a full paper or a conference proceeding is published
about the thesis later, then ignored)
Source of publication Covered by the databasesa searched by the keywordsb; references Reaching the sources by searching the selected databases, and
cited by the reviewed papers references of reviewed papers
a
Databases used: Science Direct, Taylor & Francis Online Journals, Academic Search Complete, Wiley Interscience Journals, Business Source Complete, IEL-IEEE/IEE
Electronic Library, Emerald Management Extra, MathSciNet, Oxford Online Journals, Springer Link, Web of Science – Science Citation Index Expanded, Ebrary, SIAM,
Dissertations and Thesis, Elsevier Electronic Books, Safari Books Online, Springer Book series, Springer Electronic Books, Book Index with Reviewers, ENGnet BASE, SCOPUS
b
Keywords used: quality, quality improvement, quality control, quality assurance, manufacturing, intelligent manufacturing, artificial intelligence, data mining, knowledge
discovery, machine learning, inductive learning, design, product design, process design, design optimisation, parameter optimisation, parameter design, process optimisation,
optimisation, prediction, classification, product characterization, process characterization, data analysis, clustering, association, failure analysis, yield improvement, six
sigma, neural network, genetic algorithm, Taguchi method, decision tree, fuzzy logic, regression
13450 G. Köksal et al. / Expert Systems with Applications 38 (2011) 13448–13467
quality levels cannot be achieved without developing strong de-
Brief description
signs of products or processes, DFSS programs have been put in
successful than
comments are
Also the other
of the results
unsuccessful,
(something).
practice, which follow a similar approach such as define-mea-
obtained, as
listed here
successful,
and more
Results sure-analyse-develop (optimise)-verify (DMADV) (Brady & Allen,
2006).
Many quality control and improvement activities take place
throughout the life cycle of a product as summarized in Table 3.
Software
Software
This summary is adapted from Phadke (1989), and it classifies

tools
used
used
the product development process into three phases: product and
manufacturing process design, manufacturing of the product, and
corresponding
data handling
usage of the product by the customer. In this study, we focus on
DM methods
DM methods
perform the
certain QI and control activities that occur during the first two
functions
used to
stages as explained below.

or DM
Three critical stages of product and manufacturing process de-

sign phase are: conceptual design of the products/processes, set-
corresponding
Data handling
data handling
ting values of their design parameters (i.e. parameter design, or

Types of the
quality task
involved in
design optimisation), and setting tolerances for the design param-

functions
functions
sequence
and DM
and DM
to each
eters (i.e. tolerance design). During the conceptual design of a

product the most appropriate architecture and technology for
achieving the desired functions of the product are selected. Then,
sequence
involved
the best settings are determined for the controllable product/pro-

Types of
Quality
quality
tasks
tasks
cess design parameters. At this parameter design stage, it is typical

the
in
to assume wide tolerances on uncontrollable noise factors, and

search for parameter settings insensitive to adverse effects of the
variables are
continuous,
number of
noise factors. If such settings cannot be found, then tolerances

ordinal as
each type
variables
them for
nominal,
Types of
given as
Output
well as
binary,
for variability of the product/process parameter values are selec-

output
tively reduced, and use of higher grade materials and sophisticated

control systems are considered for further reduction in the vari-
variables are
ability. In this study, we especially consider conceptual design

continuous,
number of
ordinal as
each type
variables
them for
nominal,
and parameter design stages, where DM has found widespread use.

Types of
given as
well as
binary,
Input
input
Major QI and control activities during the product manufactur-

ing phase are; determining factors that significantly affect quality,
modelling relationships between input and output characteristics
(train + test + verification)
Observational, Simulated.
Data source as one of the
given. Number of records

following: Experimental,
of quality, and predicting quality levels for a given set of input

If experimental, type of
the experiment is also
parameters. Such quality analyses provide invaluable feedback to

product/process design (or re-design) and other corrective actions
is also written.
for QI. Most common DM applications are expected to occur in

Data source
such activities due to ease of collecting relevant data, performing

the analysis, and interpreting the results. We also focus on such
QI activities within the scope of this study.
Some quality control and improvement activities are not con-
industry according
sidered in this review. These activities as well as the reasons of

excluding them are briefly explained in the following.
to ISIC (2009)
Code of the
application
When it comes to controlling quality of products, the oldest ap-

Section C
proach is to inspect all units produced and discard/repair the

defective ones. We have excluded the inspection activities in this
ISIC
study, since smart inspection systems utilizing DM techniques

Description of the data table for the reviewed application studies.
form a mature area that deserves a special attention on its own.

Description of the
process involved
product and/or
For example, a branch of chemistry, called chemometrics, has

in the study
grown into a well-established data collection and analysis method

Product or
in areas such as multivariate calibration and pattern recognition in

process
the chemical process industry (Yang & Trewn, 2004). In general,

automated visual inspection (AVI) has become inevitable in mod-
ern manufacturing (Newman & Jain, 1995). In AVI, many conven-
Description of the
major purpose of
tional image processing functions have been employed, however,

much recent work has also focused on incorporating techniques
the study
from the area of artificial intelligence (AI) and DM into the systems.
Purpose
Pham and Alcock (2003) provide a good coverage of many of these

applications.
Other important QI and control activities that we have excluded
information for
the application
are quality monitoring and process control. Process industries and

discrete parts manufacturing industries have had a long history of
Reference
these activities that aim to reduce variability. While quality mon-

study
itoring tries to reduce variability by detection and removal of

Table 2
assignable causes, process control is based on the idea of process

ID
compensation and regulation to reduce variability. Although they

Table 3
Quality control and improvement activities and methods (adapted from Phadke, 1989).
Product development stage Quality control and improvement Examples of traditional methods used
activity
Product design, manufacturing process Concept design QFD, Pugh’s concept selection, TRIZ, technological forecasting
design Parameter design Design of experiments, ANOVA, response surface modelling and analysis, regression,
optimisation
Tolerance design Statistical tolerancing, cost analysis
Manufacturing Inspection/Screening Pattern recognition, automated inspection
Quality analysis ANOVA, regression, classification, clustering
Process control Feedback control, feed forward control, manual adjustments
Quality monitoring Statistical process control (control charts), Principal component analysis
Customer usage Warranty and repair/replacement Replacement analysis
consist of important DM applications through on-line process Classification of quality: Classifying a quality characteristic of
monitoring and automatic process control, they make use of spe- interest for nominal, binary or ordinal outputs (such as defects).
cial methods collected under temporal data mining due to ‘time’ For a given set of input parameters, predicting the class of the
being an important variable of the models. Furthermore, the liter- quality output.
ature on these topics is too voluminous to be included in the scope Similar to predicting quality, this quality task is typically per-
of this review. For example, reviews of artificial neural networks formed at the Analyse stage of both DMAIC and DMADV cycles.
(ANNs) and wavelet transforms for quality monitoring are pro- Parameter optimisation: Based on the learned characteristics of
vided in Zorriassatine and Tannock (1998) and Ganesan, Das, and the cases yielding high quality, finding optimal levels of pro-
Venkataraman (2004), respectively. Similarly, a review paper on cess/product parameters that consistently yield target quality
process control using ANNs is provided in Hussain (1999). These performance. Note that this quality task is also performed at
reviews indicate the widespread use of DM in the area. Yet, there the Improve or Develop (Optimise) – Verify stages.
are many other DM methods in the literature, such as principal
component analysis (PCA), partial least squares (PLS), support vec- 3. The knowledge discovery in databases process and data
tor machines (SVM), and decision trees (DTs), that have found mining
applications in quality monitoring and process control. A separate
review of these might also be beneficial for future studies. Fayyad, Piatetsky-Shapiro, and Smyth (1996) define KDD as
The final phase of product development process considers the ‘‘the nontrivial process of identifying valid, novel, potentially use-
usage of the product by the customers. Since the product is in ful and ultimately understandable patterns in data’’. It consists of
the hands of the customer at this phase, quality perception can the following main steps: (i) data preparation, (ii) data preprocess-
be improved through repair or replacement of the product under ing, (iii) DM, (iv) evaluation and interpretation, (v) implementa-
warranty, and other services beyond warranty. We have also tion. Note that DM is a step in the KDD process that consists of
excluded DM applications involving QI and control at after-sales- applying data analysis and discovery algorithms. DM tasks can be
services, which typically utilise text mining approaches due to classified into two groups (Han & Kamber, 2006): descriptive and
the textual nature of the associated data. predictive. These tasks can be accomplished by using various
In summary, the QI and control activities we choose to consider methods based on DM functions, used to specify the types of pat-
in this study involve the following tasks, which we refer to as qual- terns to be mined. These functions include summarisation (charac-
ity tasks from this point on: terisation), clustering, association, classification, prediction and so
on. Below, the steps of KDD process are described briefly.
Description (or characterisation) of product and process quality:
Quality of products or processes can be defined or characterised 3.1. Data preparation
by performing the following tasks:
– Identifying attributes/variables, which affect quality DM functions mainly utilise available data sources such as data
significantly. warehouses, marts, databases or files for gathering data (Pyle,
– Ranking the attributes/variables based on their significance. 1999). In applications, the data sources are first located, accessed,
– Identifying how low, medium and high yielding products are and integrated. Next, selected data is put into a tabular format in
naturally grouped in data, and finding the most probable caus- which instances and variables take place in rows and columns,
ative factor(s) that discriminates between low and high yielding respectively (Giudici, 2003). If the data set built is very large, a rep-
products. resentative reduced data set can be obtained by sampling. In cer-
These activities are typically performed at earlier stages (define- tain situations all data may not be readily available for mining,
measure-analyse) of DMAIC and DMADV for quality analysis where data farming process can help us ‘‘defining features that
and product/process design. are the most appropriate for DM’’ (Kusiak, 2006). Then, one may
Predicting quality: When quality output is a real valued variable, collect necessary data recording the feature values directly from
developing models that relate input characteristics of quality to real time or experimental observations, or indirectly from simula-
the output, and using such models to predict what the resulting tion results.
quality characteristic value will be for a given set of input
parameter values. This quality task is typically performed at 3.2. Data preprocessing
the Analyse stage of both DMAIC and DMADV cycles dealing
with real valued quality output characteristics. The prediction Real-world data is generally dirty, incomplete and inconsistent.
models developed can be used later in Improve or Design (Opti- Redundancies may also occur due to integration of data from var-
mise) – Verify stages, or directly in the Control stage. ious sources. The main purpose of this step is to handle these kinds
problems to improve the data quality. In addition, transforming prediction methods can be categorised into the following groups
and reducing data can help to improve the accuracy and efficiency (Dunham, 2003): statistical based (S-based) methods use classical
of DM function(s). Basic data preprocessing techniques are as fol- techniques which depend on statistical theory, and thus, provide
lows (see Giudici, 2003; Pyle, 1999; Witten, 2005): statistical inference. DT-based algorithms construct DTs which
look like a flowchart using a top-down recursive approach. DT-
Data cleaning involves techniques for filling in missing values, based algorithms automatically generate rules having ‘if-then’ type
smoothing out noise, handling outliers, detecting and removing structures. ANN-based algorithms, on the other hand, consist of a
redundant data. set of connected input–output units each having a weight, which
Data transformation puts the data into appropriate forms for is updated by a learning algorithm such as backpropagation (BP)
mining when necessary. used with gradient descent (GD) optimisation technique, or the ra-
Data reduction is applied to reduce the data set to be mined. dial basis function (RBF). In addition, different classification meth-
While ‘dimension reduction’ technique eliminates unnecessary ods can be combined to improve the model accuracy. Following a
attributes, ‘data compression’ and ‘numerosity reduction’ tech- successful implementation of a classification/prediction function,
niques provide other forms of reduced data representations. or as a stand-alone DM function, optimisation can be performed
Discretisation, a form of data reduction, reduces the number of to determine the settings of factors (or design parameters) that
levels of an attribute by collecting and replacing low-level con- yield desired responses.
cepts with high-level concepts.
3.4. Evaluation and interpretation
3.3. Data mining
The KDD process described above tries to uncover previously
DM methods can be categorised based on various criteria. In unknown structures that may reside in data. Depending on the
this study, we classify them according to the ‘types of knowledge data sets and research objectives, one can start at any step and con-
mined’ (DM functionalities) such as clustering, association, classifi- tinue with the others as long as there are research questions to be
cation, and so on for achieving descriptive/predictive DM tasks answered. Besides, one may try several methods for describing or
(Dunham, 2003). modelling process/product quality. Therefore, assessing the utility
and reliability, and then, interpreting the information discovered in
3.3.1. Descriptive data mining the modelling should be the final stage in a DM process (Giudici,
Descriptive DM involves exploration of patterns and relation- 2003). Evaluation of the DM methods to reach a final decision re-
ships that may exist in data. Basic descriptive functions are sum- quires a comparison of results obtained from various DM methods
marisation, clustering, association rule generation and sequence using several measures including accuracy, time and resource
discovery (Dasu & Johnson, 2003; Giudici, 2003). These are also requirements. It is obvious that to obtain reliable results, knowl-
used for data exploration before a classification/prediction and/or edge extracted should be evaluated and interpreted correctly
optimisation DM function is implemented. In this review, sequence (Dunham, 2003).
discovery is not observed due to the nature of problems within the
scope of the study.
3.5. Implementation
Summarisation is the presentation of general characteristics of
a data set. Basic approaches are OLAP and attribute-oriented induc-
The KDD process described above provides tools for better
tion. There are numerous statistical methods available for data
understanding the relations in quality data. The final KDD step in-
summarisation (Giudici, 2003). Descriptive statistics and graphical
volves implementation of the results obtained into the QI related
displays can effectively describe univariate data. For bivariate data,
decisions of the industry (Giudici, 2003). This is typically per-
additional methods (e.g. correlation analysis (CA), scatter plots)
formed informally. Kusiak (2006) proposes a framework for more
can be used to determine the relationship that may exist between
structured and transparent decision making based on decision
the variables. For describing multivariate data, however, depen-
making constructs called decision tables, decision maps, atlases,
dency and association measures as well as multidimensional
and library.
graphs such as scatter plot matrix and Andrews curves are needed
(Martinez & Martinez, 2002).
Clustering is the process of grouping data into classes of similar 4. Analysis of the applications
objects. The similarity among objects is usually measured by dis-
tance measures. Major distance-based clustering methods can be In the following, we examine and discuss the reviewed litera-
organized in two categories (Han & Kamber, 2006). Partitioning ture from various points of view. Data handling or more specifi-
methods classify the data into k parts in such a way that observa- cally data preparation and data preprocessing before performing
tions in each part are closely related to each other. Hierarchical the DM functions in the quality tasks are discussed first. Next,
methods group the data into a tree of clusters by either using bot- DM applications in each of the quality tasks defined earlier are cat-
tom-up (agglomerative) or top-down (divisive) approaches. egorised with an emphasis on the manufacturing industries. Find-
Besides, there are others classified as density-based, grid-based ings of these applications are presented subsequently, and then
and model-based methods. some DM patterns found in these applications are explained. Final-
Association tries to identify groups of items that occur together. ly, software tools used in these applications are examined.
Assuming that a database consists of a set of records which con-
tains a set of items, most algorithms accomplish the association 4.1. Data handling
task in two steps: finding frequent item sets, and then generating
interesting if-then rules (Hand, Mannila, & Smyth, 2001). Data preparation and data preprocessing are two essential steps
that should be taken before performing any DM function. Never-
3.3.2. Predictive data mining theless, these steps are either briefly explained or considerably less
Predictive modelling can be accomplished by performing classi- emphasis is given to them in presenting the applications. Below,
fication or prediction functions to forecast future values of categor- we provide some analyses of the literature regarding data prepara-
ical or continuous type data, respectively. Major classification/ tion and preprocessing steps.
4.1.1. Preparation of quality data niques applied for that purpose, analysis of variance (ANOVA), sub-
In most of the QI applications reviewed (37% of all data sets), jective evaluation (SE) and ANN being the most noticeable among
data were collected through statistically designed experiments them. In addition, data compression techniques such as PCA to ob-
(DOE). Experimental design settings used in these applications tain manageable sizes of data are also common for data reduction.
are provided in Table 4. The number of records in the experimental Finally, discretisation techniques used in the preprocessing of data
data sets ranged from 9 to 1323. Observational data sets are also are provided in Table 8. These were applied extensively on contin-
common in the reviewed applications (28% of all data sets) and uous type of data, especially before DT-based methods were used,
these were either obtained through online sensor measurements and mostly SE method was preferred.
or from historical logs. The number of records in the online-obser- Depending on the analysis of literature reviewed above, we can
vational data sets ranged from 27 to 16,381. On the other hand, say that quality data can be either too small or too large in size. Be-
historical-observational data sets (20% of all data sets) were typi- sides, it may contain missing, outlying, inconsistent and incom-
cally obtained from two different databases, one for production plete observations (see Table 5). In addition, quality data may
and the other for quality data. The number of records in such have several mixed type (discrete and continuous) of input and
observational data sets ranged from 27 to 58,076. Alternatively, output variables (Köksal et al., 2008). Due to enormous number
process simulations and simulated data sets (15% of all data sets) of input variables, correlations among them are frequently ob-
also existed and these were used to test the performance of the served, and usually are handled by using dimension reduction
algorithms used. The largest number of records observed in the re- techniques (see Table 7). Moreover, quality data is usually imbal-
viewed applications is 250,000 and this belongs to a simulated data anced in terms of the number of records contained for defective/
set. nondefective items.
4.1.2. Preprocessing of quality data 4.2. Analysis of DM applications according to quality tasks
Data cleaning techniques used in the reviewed applications are
presented in Table 5. Most of the cleaning techniques were applied The number of DM applications for the covered quality tasks
to the historical-observational type data sets. In the applications, has been increasing for the last 10 years (see Fig. 1). A total of
data transformations were also considered in the data preprocess- 127 application studies (presented in 130 papers) within the scope
ing step. As can be seen in Table 6, normalisation of the data was of this review have been found. In this section, these applications
the most common transformation used. Preprocessing of data are reviewed according to the quality tasks performed, giving spe-
through data reduction was also common in the reviewed studies cial attention to the manufacturing industries involved.
(see Table 7). Especially, dimension reduction is observed to be a An analysis of the usage of quality tasks (shown in Fig. 2) indi-
widely used form of data reduction, and there is a long list of tech- cates that predicting quality is the most frequently performed task
Table 4
Design types used in the QI literature.
Design type QI applications

2- and 3-level full Brinksmeier et al. (1998), Lin and Wang (2000), Chen and Ramaswamy (2002), Feng and Wang (2003), Jiao et al. (2004), Shi et al. (2004),
factorials Feng and Wang (2004), Kurtaran et al. (2005), Karnik et al. (2007) and Erzurumlu and Oktem (2007)
Fractional factorials Feng and Wang (2002), Yang and Tsai (2002), Ip et al. (2003), Kim et al. (2003), Shi et al. (2004), Feng and Wang (2004) and Yang et al.
(2005)
Taguchi orthogonal Tay and Butlar (1997), Lin and Wang (2000), Shen et al. (2007), Xue et al. (2005), Krimpenis et al. (2006), Ozcelik and Erzurumlu (2006),
designs Hung (2007), Lee and Dornfeld (2007), and Teng and Hwang (2007)
D-optimal design Kim et al. (2001)
One factor at a time Wang et al. (2006) and Riverol and Cooney (2007)
experiments
Table 5
Data cleaning techniques used in the QI literature.
Data to be cleaned Techniques QI applications

Missing values Deleted Li et al. (2003b), Bakır et al. (2006), Huang et al. (2006), Batmaz (2007), Hsu and Chien (2007) and
Chien et al. (2007)
Replaced with a representative value (e.g. Cool et al. (1997), Batmaz (2007) and Maimon and Rokach (2001)
mean)
Unspecified Kang et al. (1999) and Ho et al. (2006)
Outliers Removed Dhond et al. (2000), Zhai et al. (2002), Bakır et al. (2006), Batmaz (2007), Zhou et al. (2006) and
Karim et al. (2006)
Unspecified Kang et al. (1999)
Incomplete data Filtered out using OLAP techniques Ho et al. (2006)
Filtered out using low filters Rallo et al. (2002)
Deleted Chien et al. (2006)
Inconsistent data Treated by filters and removed Zhou et al. (2006)
Misspelled data Eliminated Chien et al. (2007)
Noise/unwanted Wang (2007)
values
Unspecified Unspecified Shahbaz et al. (2006)
Table 6
Transformations applied in the QI literature.
Transformation QI applications
Normalisation Cool et al. (1997), Lewis and Ransing (1997), Tsai et al. (1999), Cherian et al. (2000), Suneel et al. (2002), Vasudevan et al. (2002),
Rallo et al. (2002), Li et al. (2003a), Li et al. (2003b), Zuperl and Cus (2003), Shi et al. (2004), Perzyk et al. (2005), Hou et al. (2003), Ho
et al. (2006), Yin and Yu (2006), Karim et al. (2006), Zhou et al. (2006), Zhang et al. (2007), Karnik et al. (2007), Lee and Dornfeld
(2007), Chen et al. (2007) and Hung (2007)
Smoothing Li et al. (2003b)
Logarithm Umbrello et al. (2007)
Transforming categorical data to Karim et al. (2006)
numerical
Centring data Tsai et al. (1999)
Adjusting abnormal data Dhond et al. (2000)
Unspecified Shahbaz et al. (2006)
Table 7
Data reduction techniques used in the QI literature.
Data reduction Methods QI applications

approach
Dimension reduction MLR Ali and Chen (1999), Tseng et al. (2005) and Yin and Yu (2006)
ANOVA Lin et al. (2000), Chiang et al. (2002), Shi et al. (2004), Ozcelik and Erzurumlu (2006), Krimpenis et al. (2006) and Lee and
Dornfeld (2007)
DT Ali and Chen (1999), and De Abajo et al. (2004)
CA Lian et al. (2002), and De Abajo et al. (2004)
ANN Kang et al. (1999), Han et al. (1999), Dhond et al. (2000), Rallo et al. (2002) and De Abajo et al. (2004)
RST Kusiak (2000), Zhai et al. (2002), Huang et al. (2006) and Chang et al. (2007)
PLS Zhou et al. (2006), Chong et al. (2007) and Achiche et al. (2007)
SE Liao et al. (1999a), Liao et al. (1999b), De Abajo et al. (2004), Yin and Yu (2006) and Chong et al. (2007)
GA De Abajo et al. (2004) and Zhang et al. (2007)
GP Li et al. (2006)
Experimentation Braha and Shmilovici (2002)
GSA and ANN Yin and Yu (2006)
Specialized Ordieres Meré et al. (2004) and Vosniakos et al. (2007)
techniques
Unspecified Lu (2001)
Data compression PCA Chang and Jiang (2002), Lian et al. (2002), Skinner et al. (2002), Li et al. (2003b) and Brudzewski et al. (2006)
WT Brudzewski et al. (2006)
SE Brudzewski et al. (2006)
Desirability Li et al. (2003a) and Hung (2007)
function
Table 8
Discretisation methods used in QI literature.
Cases and methods QI applications

Cases where discretisation is Kang et al. (1999), Maimon and Rokach (2001), Braha and Shmilovici (2002), Georgilakis and Hatziargyriou (2002), De Abajo et al.
used before DT-based (2004), Chien et al. (2006), Wang (2007), Hsu and Chien (2007), Shao et al. (2007) and Wang (2007)
methods
Discretisation SE Kusiak and Kurasek (2001), Maimon and Rokach (2001), Braha and Shmilovici (2002), De Abajo et al. (2004), Hou and Huang (2004)
methods DT Georgilakis and Hatziargyriou (2002)
DA Tseng et al. (2005)
k-means Chang et al. (2007)
CC Chien et al. (2006)
OD Kusiak (2000)
Unspecified Kang et al. (1999), Last and Kandel (2001), Zhai et al. (2002), Perzyk et al. (2005), Huang et al. (2006), Hsu and Chien (2007), Shao
et al. (2007) and Wang (2007)
(42%). Applications involving classification of quality (25%) and manufacturers for those (30%). Plastics (7%), basic metals (7%), coke
parameter optimisation (23%) are also considerable in number. and refined petroleum (4%) and chemical products (4%) manufac-
However, description of quality is performed relatively less (10%). turing industries observed significantly less usage of DM methods.
A classification of the reviewed applications according to the Moreover, motor vehicles, trailers, food, beverages, textile, paper
manufacturing industry codes of ‘‘International Standard Industrial and paper products, nonmetallic mineral products, wood products
Classification (ISIC) of All Economic Activities, Revision 4, Section C except furniture, pharmaceutical products and electrical equip-
– Manufacturing’’ is provided in Fig. 3. Many DM applications in ment manufacturing industries had rare (2–0.8%) DM applications.
the literature regarding the QI activities appear to be from the me- For many other industries, DM applications within the scope
tal product manufacturing industries (with operations such as have not been observed. These include tobacco products, wearing
forming, welding, casting and machining) (37%), as well as com- apparel, leather and related products, printing and reproduction
puter and electronic product manufacturing industries and device of recorded media, machinery and equipment (other than motor
Fig. 1. Number of the application papers observed in the literature.
Fig. 2. Number of times DM functions are used for quality tasks in the application studies.
vehicles), and furniture manufacturing industries, which can be patterns from wafer bin map to improve yield. Association analysis
considered as potential application areas. was also used for identification of root-cause machine set (Chen,
Tseng, & Wang, 2005). In order to infer possible causes of faults
4.2.1. Description of product and process quality and manufacturing process variations in semiconductor manufac-
Product and process quality description is generally the first turing, Chien, Wang, and Cheng (2007) considered k-means clus-
stage of a QI study in an industrial application. In the computer tering. Furthermore, hierarchical clustering was used for
and electronic product manufacturing industry, self organizing describing the relationships between machines and wafer yield
maps (SOMs) have been used to identify critical poor yield factors rates (Hu & Su, 2004), for determining the cause of low quality wa-
(Gardner & Bieker, 2000) and a variant of SOM has been used to im- fers (Skinner et al., 2002), and for improving yield (Baek, Jeong, &
prove yield (Karim, Halgamuge, Smith, & Hsu, 2006) in wafer man- Han, 2005).
ufacturing. Kang, Choe, and Park (1999) also considered SOM for Metal product manufacturing industry applications also include
generating better operating conditions in semiconductor manufac- several product and process quality description studies, where
turing. On the other hand, Tseng, JothiShankar, and Wu (2004a), product/process attributes or variables affecting quality were
Tseng, JothiShankar, Wu, Xing, and Jiang (2004b) utilised rough determined, ranked based on significance, or grouped based on
set theory (RST) to identify the features that produce solder ball similarities. In packaging manufacturing De Abajo, Diez, Lobato,
defects and affect the quality in printed circuit board (PCB) manu- and Cuesta (2004) developed a tinplate quality diagnostic model,
facturing. Moreover, rule induction has been utilized to select attri- where SE, genetic algorithm (GA), SOM, and DT were used. Ali
butes that represent the quality of motherboard assemblies and Chen (1999) used multiple linear regression (MLR) and DT in
(Huang, Li, & Peng, 2006) and for eliminating waste by improving an injection moulding process for QI. For a cutting process, Chang
the quality of wafers in integrated circuit manufacturing (Kusiak, and Jiang (2002) applied PCA for dimension reduction, and for a hot
2000). Bertino, Catania, and Caglio (1999) considered DT and asso- rolling process Cser et al. (2001) used SOM for clustering. ANN was
ciation analysis for determining possible causes of faulty wafers, used by Dhond, Gupta, and Vadhavkar (2000) in a steel mill blast
whereas Hsu and Chien (2007) considered DT and ANN to extract furnace process and by Han, Han, and Liu (1999) in a pressurized
cess. For a robust design of very-large scale integration (VLSI)

circuits, Ilumoka (1998) also considered ANN. Furthermore, Yang,
Tsai, and Yeh (2005) used ANN to propose a quality control system
for the solder stencil printing process, and Shi, Schillings, and Boyd
(2004) used ANN for modelling nonlinear cause-effect relation-
ships in PCB manufacturing, whereas general regression neural
networks (GRNNs) were used in Tseng (2006) for predictions in a
spot welding process. Prediction applications involve other meth-
ods, too. Ip, Kwong, and Wong (2003) used fuzzy regression (FR)
to model transfer moulding for microchip encapsulation of elec-
tronic packages. For a silicon wafer manufacturing process and in
order to model the relationships between surface roughness and
cutting parameters in turning operations, Jiao, Lei, Pei, and Lee
Fig. 3. Number (percentage) of the application studies with respect to ISIC (2004) considered MLR and fuzzy adaptive network (FAN). MLR,
manufacturing activities. (10 – Manufacture of food products, 11 – Manufacture parametric and nonparametric time series analysis (TSA), ANN,
of beverages, 13 – Manufacture of textiles, 16 – Manufacture of wood and of and case-based reasoning (CBR) were used to predict the manufac-
products of wood and cork, except furniture, 17 – Manufacture of paper and paper
turing process in plastic optic fibres manufacturing (Kim & Lee,
products, 19 – Manufacture of coke and refined petroleum products, 20 –
Manufacture of chemicals and chemical products, 21 – Manufacture of basic 1997). To specify the failure cause and to improve the yield, Mieno
pharmaceutical products and pharmaceutical preparations, 22 – Manufacture of et al. (1999) employed DT. Skinner et al. (2002) considered MLR,
rubber and plastics products, 23 – Manufacture of other nonmetallic mineral generalized linear model (GLZ), and DT to determine the quality
products, 24 – Manufacture of basic metals, 25 – Manufacture of fabricated metal or yield of wafers. DT was also used to clarify the correlation be-
products, except machinery and equipment, 26 – Manufacture of computer,
tween yield and various wafer parametrical data (Tsuda, Shirai,
electronic and optical products, 29 – Manufacture of motor vehicles, trailers and
semi-trailers, U-unspecified). Takagi, & Take, 2000) and for prediction in integrated circuit man-
ufacturing (Maimon & Rokach, 2001). Furthermore, Last and Kan-
cylinder manufacturing. Lewis and Ransining (1997) employed del (2001) considered info-fuzzy network (IFN) for predicting line
Bayesian belief network (BBN) and proposed a network architec- yield and flow times, and Li, Huang, and Wu (2006) considered ge-
ture as a manufacturing diagnostic method in a pressure die cast- netic programming (GP) for improving the product yield in semi-
ing process. RST was considered in Chang, Liu, Fan, and Chang conductor manufacturing.
(2007) for electrical discharge machining and in Tseng, Leeper, Metal product manufacturing industries observed many DM
Banda, Herren, and Ford (2004c) for surface roughness in a machin- applications for prediction. Feng and Wang (2002, 2003) consid-
ing process. CA, maximal tree method (MTM), and PCA were uti- ered nonlinear regression (NLR) and computational ANN (Comput-
lized by Lian, Lai, Lin, and Yao (2002) to control dimensional NN) together with NLR to develop an empirical model for surface
deviation of body-in-white in the assembly of sheet metal prod- roughness. Sung, Kim, Xue, Kim, and Cha (2007) and Xue et al.
ucts. Tseng, Kwon, and Ertekin (2005) used MLR and rule induction (2005) used FR to predict the bead geometry in robotic welding
for deriving rules for the process variables that affect the quality process. On the other hand, MLR, NLR, and ANN were considered
characteristics in machining processes. in Kim, Son, and Yarlagadda (2003) to present models for predict-
There are also some product and process quality description ing top-bead width in a welding process. Ali and Chen (1999) also
studies in the other industries. ANOVA was used to determine considered MLR, NLR, and ANN, but to present models for quality
the attributes of quality for a polymerisation process in silicon-fil- characteristics in an injection moulding process. Moreover, MLR
ler manufacturing (Chiang, Su, Li, & Huang, 2002) and for warpage and ANN were used to estimate mechanical properties of galva-
of thin shell plastic parts in a plastic injection moulding process nized steel (Ordieres Meré, González Marcos, González, & Lobato
(Ozcelik & Erzurumlu, 2006). For a better glass coating quality, Li, Rubio, 2004), to model a cutting process (Chang & Jiang, 2002),
Feng, Sethi, Luciow, and Wagner (2003b) applied PCA to define to adjust influential factors of quality in steel plate production
process attributes. SOM for process variables was used in a low- (Deng & Liu, 2002), to construct a prediction model for surface
density polyethylene process by Rallo, Ferre-Gine, Arenas, and Gir- roughness in turning operation (Lin & Wang, 2000), to construct
alt (2002), and a variant of SOM was used in industrial conveyor a prediction model for surface roughness of end milling (Tsai, Chen,
belt manufacturing by Karim et al. (2006). Yin and Yu (2006) con- & Lou, 1999), and to forecast hot metal temperature in a steel mill
sidered SE, MLR, and grey superior analysis (GSA) in order to select blast furnace. ANN together with NLR was considered for model-
important variables in worsted spinning process. In the petro- ling a grinding process in Brinksmeier, Tönshoff, Czenkusch, and
chemical industries, Zhou, Xiong, Zhang, and Xu (2006) considered Heinzel (1998). Erzurumlu and Oktem (2007) considered ANN to-
PLS in an ethylene pyrolisis process, and Liao, Zhan, and Mount gether with response surface methodology (RSM) to develop a
(1999a,1999b) considered an algorithm for inducing modular rules model to predict surface roughness value error on mould surfaces,
(PRISM) for identifying failure mechanisms of mechanical compo- whereas Ho, Lee, Chen, and Ho (2002) applied ANN, fuzzy neural
nents. In order to minimize loss of product during system change- network (FNN), and neural fuzzy inference systems (ANFIS) to
over in a refinery fluid catalytic cracking process, Sebzalli and predict surface roughness in a turning process. ANN also had many
Wang (2001) considered PCA and fuzzy c-means (FCM) clustering. other prediction applications; in a metal inert gas welding process
Neuro fuzzy logic (FL) and DT were considered in Shao, Rowe, and (Meng & Butler, 1997; Tay & Butler, 1997), in steel welding
York (2007) for pharmaceutical product formation. processes (Cool, Bhadeshia, & Mackay, 1997; Vasudevan,
Murugananth, & Bhaduri, 2002; Vasudevan, Rao, Venkatraman,
4.2.2. Predicting quality Jayakumar, & Raj, 2005), in a laser welding process (Olabi, Casalino,
Several computer and electronic product manufacturing indus- Benyounis, & Hashmi, 2006), in a shot welding process (Hamedi,
try applications serving the goal of predicting quality have been Shariatpanahi, & Mansourzadeh, 2007), in powder metal parts
published. Chen, Lee, Deng, and Liu (2007) employed ANN in a manufacturing (Cherian, Midha, & Pipe, 2000), in a knurling pro-
chemical vapour deposition process to establish a predictor of cess (Feng & Wang, 2004), in a steel casting process (Perzyk,
product quality, and Li, Su, and Chiang (2003a) used ANN to deter- Biernacki, & Kochanski, 2005), in a pressure die casting process
mine parameter settings in a silicon compound manufacturing pro- (Krimpenis, Benardos, Vosniakos, & Koukouvitaki, 2006), in a
computer numeric control (CNC) turning process (Suneel, Pandle, & ing process. Kusiak (2002) used RST to derive associations among
Date, 2002), in oil methods manufacturing (Aloudat, 2006), in control parameters and the product quality in metal forming pro-
stainless still drilling process (Karnik, Gaitonde, & Davim, 2007), cess. RST was also employed together with fuzzy set theory (FST)
and in a hard machining process (Umbrello, Ambogio, Filice, & (Hou & Huang, 2004) and together with ANN (Hou, Liu, & Lin,
Shivpuri, 2007). 2003) in conveyor belts manufacturing for rule induction.
To develop prediction models for quality, other industries also Classification applications also have been observed in other
made use of DM methods. ANNs were used in all of these applica- industries to improve quality. DT was used in ultra precision man-
tions. Plastics manufacturing industries utilised ANN, especially in ufacturing (Huang & Wu, 2005). Jemwa and Aldrich (2005) consid-
plastic injection moulding processes (Kurtaran, Ozcelik, & ered an integration of SVM and DT for a continuous stirred tank
Erzurumlu, 2005; Ozcelik & Erzurumlu, 2006; Sadeghi, 2000; Shen, reactor. ANN were considered in diagnosis of the causes of pre-
Wang, & Li, 2007). Chemical process industries have a variety of stressed concrete damages (Tam, Tong, Lau, & Chan, 2004), and
ANN applications; in a thermal spray process (Guessasma, Salhi, in predicting the quality of injected plastic parts (Sadeghi, 2000).
Montavon, Gougeon, & Coddet, 2004), in a plasma spraying process Sarimveis, Doganis, and Alexandridis (2006) considered ANN and
(Wang, Fang, Zhao, & Zeng, 2007b), in a low density polyethylene FST for classifying the product quality in paper manufacturing.
process (Rallo et al., 2002), in an oxidative dehydrogenation of pro- Achiche, Baron, Balazinski, and Benaoudia (2007) characterized
pane process (Holena & Baerns, 2003), in an ethylene pyrolysis pro- wood chip quality online to optimise a thermo mechanical pulp
cess (Zhou et al., 2006), in a diamond coating like carbon process process through use of fuzzy decision support system (FDSS) based
(Ho, Lau, Lee, Ip, & Pun, 2006), and in a silicon compound manufac- on GA learning.
turing process (Li et al., 2003a). Other industry applications include
a reaming process (Mathews & Shunmugam, 1999), a laser sinter- 4.2.4. Parameter optimisation
ing process (Wang, Wang, Zhao, & Liu, 2006), a canned foods pro- Computer and electronic product manufacturing industries
cess (Chen & Ramaswamy, 2002), a beer fermentation process made use of ANN for simultaneously optimising multiple re-
(Riverol & Cooney, 2007), a worsted spinning process (Yin & Yu, sponses in an ion implantation process (Hsieh & Tong, 2001). GA
2006), and a particle board manufacturing process (Cook, Ragsdale, was used for parameter optimisation of multiple quantum well
& Major, 2000). Also for a glass coating process, Li et al. (2003b) avalanche photodiodes (Kim, Oh, Lee, Lee, & Yun, 2001). Taguchi
used both ANN and DT. method (TM) was considered in Teng and Hwang (2007) to develop
more accurate predictions of the package warpage in electronic
4.2.3. Classification of quality packages manufacturing. To optimise multiple quality responses
Classification of a quality characteristic or predicting the class of in a surface mount electronic assembly operation, Lu and Antony
a quality output is also a common task in DM applications. Com- (2002) utilised fuzzy rule based TM. Hung (2007) employed TM
puter and electronic product manufacturing industries exploited in combination with ANN and GA to optimise wire bond design
several DM methods with the purpose of classification. RST rule parameters. Moreover, RSM was used for robust design of VLSI cir-
induction was used in integrated circuit manufacturing (Kusiak, cuits (Ilumoka, 1998) and for optimising the manufacturing
2000), in PCB printed circuit board manufacturing (Kusiak & Kur- parameters in light-emitted diode (LED) packaging processes,
asek, 2001), and in motherboard assembly (Huang et al., 2006). where association rules were also used, for dynamic parameter
Furthermore, Yang and Tsai (2002) employed neuro-fuzzy systems adjustments (He, Li, & Qi, 2007).
(NFSs) in a surface mount technology assembly process and Metal product manufacturing industries employed GA for opti-
Georgilakis and Hatziargyriou (2002) utilized DT, ANN, and entro- misation of a grinding process (Brinksmeier et al., 1998), optimisa-
py network (EN) for increasing the classification success rate of tion of spot welding parameters (Hamedi et al., 2007; Tseng, 2006),
transformer iron losses. In an effort to improve wafer cleaning pro- for determination of cutting parameters in machining operations
cess Braha and Shmilovici (2002) used SOM, DT, ANN, and combi- (Cus & Balic, 2003), for obtaining optimal process conditions in
nation of multiple classifiers (CMC). To analyse final product pressure die casting (Krimpenis et al., 2006), and for optimising
quality ANN was used in Chen et al. (2007). DT was used in Kang burr size in drilling stainless steel (Karnik et al., 2007). Chong,
et al. (1999) to generate better operating conditions and to im- Albin, and Jun (2007) used patient rule induction method (PRIM)
prove the yield (Baek et al., 2005; Chien, Li, & Jeang, 2006; Chien based rule induction for determining optimal settings of process
et al., 2007; Li et al., 2006). Moreover, DT was also used for drop variables in a steel making process. ANN was used to control a
test analysis of portable electronic products (Zhou, Nelson, Xiao, hot rolling mill (Cser et al., 2001), to optimise a metal inert gas
Tripak, & Lane, 2001) and for classification in integrated circuit welding process (Meng & Butler, 1997; Tay & Butler, 1997), to opti-
manufacturing (Maimon & Rokach, 2001; Rokach & Maimon, mise cutting parameters in a turning process (Zuperl & Cus, 2003),
2006). Lu (2001) proposed ways of first reducing the massive data to optimise a steel manufacturing process (Liu, Tang, Fan, & Deng,
sets into smaller size data, then using traditional methods such as 2004), and to optimise a sintering process (Zhang, Xie, & Shen,
ANN for identifying and classifying semiconductor and electronics 2007). On the other hand, Lin and Wang (2000) used simulated
process quality problems. annealing (SA) for a surface roughness and error of roundness
Classification applications in the metal product manufacturing study in turning operations, Olabi et al. (2006) used TM for opti-
industries were not as many as those of prediction. DTs were used mising a laser welding process, Lee and Dornfeld (2007) used TM
in an aluminium coating process (Baek, Kim, & Kim, 2002), in in a cutting process. In optimising electrical discharge machining
assembly of sheet metal products (Lian et al., 2002), in casting pro- processes, Lin, Wang, Yan, and Tarng (2000) and Lin, Lin, and Ko
cesses (Bakır et al., 2006), and in assembly of automobiles (Wang, (2002) used fuzzy based TM. Furthermore, Lou and Huang (2003)
2007). Furthermore, Shahbaz, Srinivas, Harding, and Turner (2006) employed FL in an automotive coating process.
considered DT and association rules in a fan blade manufacturing Other studies optimising process parameters for use in QI stud-
process. ANN had also some applications; in hot rolling (Cser et ies are as follows: GA was used to optimise process parameters of a
al., 2001), in tinplate manufacturing (De Abajo et al., 2004), in pres- particle board manufacturing process (Cook et al., 2000), to deter-
sure die casting (Krimpenis et al., 2006), and in a cutting process mine the optimal values of process parameters in plastic injection
(Lee & Dornfeld, 2007). On the other hand, Perzyk et al. (2005) em- moulding (Kurtaran et al., 2005; Shen et al., 2007), to minimize
ployed naïve Bayesian classifier (NBC) in a steel casting process warpage of thin shell plastic parts (Ozcelik & Erzurumlu, 2006),
and Tseng et al. (2005) employed RST rule induction in a machin- and to determine parameter settings in a silicon compound manu-
facturing process (Li et al., 2003a). Other optimisation methods vantages. S-based methods used in the classification/predict-
used are, ANN for optimising multiple quality characteristics of a ing quality tasks take the advantage of statistical theory
polymerization process (Chiang et al., 2002) and for optimising which leads to ‘statistical inference’. Nevertheless, the theory
process parameter values in a wax model rapid prototyping pro- they are based on requires distributional assumptions, which
cess (Vosniakos, Maroulis, & Pantelis, 2007). For a slider manufac- may be hard to be validated for multidimensional data. This
turing process, Ho et al. (2006) also considered ANN together with may necessitate the help of a human expert. In contrast to
fuzzy rule sets (FRS), and expert systems (ESs). Holena and Baerns the S-based methods, computational methods provide ‘com-
(2003) considered sequential quadratic programming (SQP) for putational inference’ without expressing any probability. In
approximating the dependency of propane yield on catalyst com- spite of these shortcomings, it is easier with computational
position. MLR and nonlinear programming (NLP) for optimising a methods to develop automatic models without a significant
textiles batch dyeing process was also considered in Köksal, Smith, human intervention.
Fathi, Lu, and McGregor (1998). Other methods that are found to be more successful than
MLR are FAN (Jiao et al., 2004), RST (Tseng et al., 2005),
4.3. Reported performance of the DM methods TSA, CBR (Kim & Lee, 1997), and GSA combined with ANN
(Yin & Yu, 2006). Findings of the other comparative studies
The reviewed applications can be categorised into three groups for prediction related to the ANN can be listed as follows:
according to the performance of the DM methods reported by ANN with multi-layer perceptron (MLP) and BP outper-
them: forms RBF (Hamedi et al., 2007), but requires long com-
puting time.
1. In the majority of the studies, one or more of the method(s) ANN with multiple inputs is more successful than ANN
are used to accomplish the aimed quality task(s). In such with a single input (Mathews & Shunmugam, 1999).
studies, usually successful results are reported. These typi- According to Guessasma et al. (2004) for an ANN, the
cally involve a DM method such as ANN (e.g. Cool et al., most adequate architecture, learning paradigm, transfer
1997; Suneel et al., 2002; Yang et al., 2005), DT (e.g. Mieno function and the error function are the multilayer normal
et al., 1999; Tsuda et al., 2000), NLR (e.g. Feng & Wang, feedforward quick propagation, sigmoid functions and the
2002), FR (e.g. Ip et al., 2003; Xue et al., 2005), Bayesian neu- MSE, respectively.
ral network (BNN) (e.g. Vasudevan et al., 2002) or IFN (e.g. The most apparent finding based on a comparative study
Last & Kandel, 2001) for predicting quality; DT (e.g. Chien for classification related to the ANN is stated by Perzyk
et al., 2006; Kang et al., 1999), probabilistic neural network et al. (2005) as follows:
(PNN) (e.g. Tam et al., 2004), RST (e.g. Kusiak & Kurasek, NBC, an S-based method improved through Bayesian clas-
2001) or NFS (e.g. Yang & Tsai, 2002) for classifying quality; sifier (BC) by assuming class conditional independence,
TM (e.g. Chang et al., 2007; Teng & Hwang, 2007), GA (e.g. performs better than ANN.
Cus & Balic, 2003) or ANN (e.g. Hsieh & Tong, 2001) for In addition to ANN, performance of DT is also compared to
parameter optimisation; apriori (e.g. Da Cunha, Agard, & those of other methods. DT is found to be more successful
Kusiak, 2006) and an agglomerative clustering method (e.g. than MLR/GLZ for prediction/classification (Bakır et al.,
Hu & Su, 2004) for quality description. 2006; Skinner et al., 2002). The commonly used DT meth-
Alternatively, a method is used for achieving several quality ods (such as C4.5, C5.0) can deal with missing and contin-
tasks. For example, DT is used both for describing and classi- uous type data. In addition, these methods generate if-
fying quality (e.g. Baek et al., 2005); ANN is used for both then type of rules that are not provided by ANN methods.
predicting quality and parameter optimisation (e.g. Chiang However, they also suffer from ‘overfitting’ that may
et al., 2002) and also used for both classifying and predicting occur due to modelling the noise in the training data.
quality (e.g. Sadeghi, 2000). Other methods compared to DT for classification and their
In some other studies, several DM methods are used together respective performances are stated as follows:
to accomplish the quality tasks. For example, while ANN is GP is more successful than DT (Li et al., 2006).
used for predicting quality as a part of a parameter optimisa- Attribute decomposition approach (ADA), NBC, and DT
tion study, GA (e.g. Karnik et al., 2007; Shen et al., 2007), TM have the best, better and the worst performances, respec-
(e.g. Olabi et al., 2006), RSM (e.g. Ilumoka, 1998) or SQP (e.g. tively (Maimon & Rokach, 2001).
Holena & Baerns, 2003) is used for finalising the optimisa- Breadth-oblivious-wrapper (BOW) is more successful
tion. In another application, after SOM is used for classifying than NBC and DT (Rokach & Maimon, 2006).
quality, ANN is used for parameter optimisation (e.g. Cser et All DT algorithms mentioned above perform well for small
al., 2001). to medium sized databases. There are also algorithms
2. In another group of studies, several well-known DM meth- developed particularly for large databases such as SLIQ
ods are used as alternatives for accomplishing the same and SPRIT, which might be used in future QI studies.
function and their performances are compared. The most fre- Other findings of comparative studies for prediction can be
quently used DM method in this group is the ANN. Its perfor- summarised as follows:
mance is mostly compared to the performance of the MLR is more successful than SVM for prediction (Brudzew-
classical statistical modelling method MLR. These compari- ski, Kesik, Kolodziejczyk, Zborowska, & Ulaczyk, 2006).
son studies indicate that ANNs are more successful than FNN outperforms polynomial network (PN) and ANFIS for
MLR modelling for prediction (e.g. Deng & Liu, 2002; Kim prediction (Ho et al., 2002).
et al., 2003; Rallo et al., 2002). Better performance of ANNs Findings of comparative studies related to optimisation
can naturally be observed in multidimensional data since are as follows:
these are powerful tools in modelling nonlinear relationships ANN with BP learning algorithm is more accurate, but
(Fu, 1994). Yet, there is another study where both ANN and more time consuming than ANN with RBF (Zuperl & Cus,
MLR perform equally well (Dhond et al., 2000). We should 2003).
note here that both S-based and computational methods, PRIM-based rule induction is more robust compared to
such as ANN and DT, have their own advantages and disad- the functional approach (Chong et al., 2007).
Grey relational analysis (GRA) is more straightforward quality (39%). It is used relatively less for classification of qual-
than fuzzy-based TM method (Lin, Lin, & Ko, 2002). ity (22%). Besides, there is only one study (2%) that utilises the
ANN (Gaussian, RBF, GD) is successful, but quite time con- clustering function for parameter optimisation.
suming (Tay & Butler, 1997). Optimisation is only used in performing the quality task param-
eter optimisation.
3. In a relatively smaller group of studies, either a new method
Association is the most rarely used DM function (1%) compared
is developed to obtain a better performing one (e.g. Baek et
to the others, and this is used for both description and classifi-
al., 2002; Liao et al., 1999a; Liao et al., 1999b) or two or more
cation of quality.
known methods are used in combination to improve the per-
formance. Findings related to the combined methods are
Table 9 presents the number of times each DM method is used
given as the following:
to perform a DM function for a quality task in either a stand-alone
FST and RST combined is more successful than RST alone
or a combined manner. If a DM method used as a part of a com-
for classification (Hou & Huang, 2004).
bined/hybrid approach to accomplish a DM function indirectly
SE is more successful than spatial statistics integrated
serving to the eventual goal of the study, this method is counted
with adaptive resonance theory neural network (ARTNN)
in parentheses in the table. With the help of this type of display,
for clustering (Hsu & Chien, 2007).
one can easily see which (how many) DM functions (methods)
Support vector classifier (SVC) and DT combined is more
are involved in performing each quality task directly or indirectly.
rapid and SVC has better generalisation properties for
Table 9 does not contain the summarisation DM function, since
classification (Jemwa & Aldrich, 2005).
it is explicitly stated in only two of the applications (Chien et al.,
GSA in combination with ANN is more successful than
2007; Ordieres Meré et al., 2004) where summarisation techniques
MLR for prediction (Yin & Yu, 2006).
such as descriptive statistics and scatter plot matrix are used to ex-
CMC (boosting and stacked generalisation) is more suc-
plore the data.
cessful than the performances of individual methods DT,
Regarding the description of quality task, it can be observed
ANN and SOM for classification (Braha & Shmilovici,
from Table 9 that clustering, especially by SOM, PCA, agglomera-
2002).
tive clustering, and k-means (e.g. Batmaz, 2007; Sebzalli & Wang,
Hybrid DT and ANN classifier (HDTNNC) performs better
2001; Skinner et al., 2002) and classification, by DT and RST (e.g.
in terms of optimal time and accuracy compared to EN,
Bertino et al., 1999; Huang et al., 2006) were relatively common.
DT and ANN for classification (Georgilakis & Hatziargy-
Furthermore, this task is generally followed by a classification of
riou, 2002).
quality task using, for example DT or SVM (e.g. Baek et al., 2005;
Fuzzy-rule based TM is found successful for optimisation
Brudzewski et al., 2006), or by a predicting quality task using
(Lu & Antony, 2002).
ANN or DT (e.g. Chiang et al., 2002; Chien et al., 2007). Predicting
ANN used in combination with GA (e.g. Hamedi et al., 2007;
quality and classification of quality are the most frequent quality
Krimpenis et al., 2006; Vosniakos et al., 2007), used in com-
tasks observed in the literature. Typical classification/prediction
bination with TM (Olabi et al., 2006) and used in combina-
methods are generally used correspondingly for the tasks classifica-
tion with RSM (Ilumoka, 1998) produce good results.
tion/predicting quality (e.g. Baek et al., 2002; Bakır et al., 2006, etc./
Cherian et al., 2000; Deng & Liu, 2002, etc.). However, it is observed
In spite of the long list of DM methods used for different DM
that clustering and/or association methods are also used for classi-
functions, there are still some methods not found a use in QI appli-
fying/predicting quality (Chien et al., 2007; Shahbaz et al., 2006;
cations. Among these, robust regression (RR) and multivariate
Skinner et al., 2002). Additionally, some other applications used
adaptive regression splines (MARS) worth mentioning, since they
clustering methods with classification methods for classifying qual-
are found to be successful, especially for modelling complex rela-
ity. More specifically; Brudzewski et al. (2006) used a clustering
tionships (Kartal, 2007).
method to determine the number of classes in data and then used
Knowledge obtained by DM applications might be hard to inter-
this number in the SVM classification method. Hsu and Chien
pret and put in use, especially by industry people (Harding,
(2007) and Lian et al. (2002) used a clustering method to determine
Shahbaz, Srinivas, & Kusiak, 2006; Kusiak, 2006; Wang et al.,
special patterns (groups) in the data, and then used DT methods to
2007a). However, this has not been a concern raised strongly by
identify the root causes that lead to the identification of these
the applications, since most of these applications were made by
groups. Tsai, Chiu, and Chen (2005) used a clustering algorithm in
‘‘experts’’ as can be deduced from the authors’ academic identities.
relation with CBR classification method to decrease processing time
of the classification, and Sarimveis et al. (2006) used FCM to allocate
4.4. Patterns of DM function and method usage in quality tasks the input data into different clusters and then an ANN structure is
automatically formulated by assigning each cluster to a hidden
The ordering (or hierarchy) of DM functions as well as DM layer node. Clustering methods had also some use with prediction
methods used to perform a quality task is referred to as ‘patterns’. methods for the predicting quality task. For instance, Ordieres
Different patterns followed in the applications are analysed and Meré et al. (2004) determined nonhomogenous groups by cluster-
findings are presented in this section. ing methods and then developed different models such as MLR
An analysis of DM function usage for each quality task (Fig. 2) and ANN, for each group separately, whereas Rallo et al. (2002) bur-
indicates the following: ied a clustering method into the prediction methods (i.e. fuzzy pre-
dictive adaptive resonance theory neural network (Fuzzy
Prediction is the most common (44%) DM function used in the ARTMAPNN), dynamic radial basis function neural network (Dy-
applications. namic RBFNN)). On the other hand, some DM methods were used
Classification is the second most commonly (25%) used DM in combination for classification of quality. For example, ANN and
function, and it is typically used (86%) for classification of DT (e.g. Georgilakis & Hatziargyriou, 2002), DT and SVM (Jemwa
quality. It is rarely considered for description of quality (8%), & Aldrich, 2005), RST and linear programming (LP) (Kusiak, 2000),
predicting quality (4%), and for parameter optimisation (1%). and RST and FST (Hou & Huang, 2004) were used in combination.
Clustering is used in 15% of all of the used DM functions, and it Some patterns have also been observed in the parameter opti-
is commonly used for quality description (37%) and predicting misation studies. For instance, several studies directly used ANN
Table 9
Distribution of reviewed application studies with respect to DM functions and quality tasks involved.
Quality task
DM functions Approach Method Description of quality Classification of quality Predicting quality Parameter optimisation
Clustering S-Based PCA (1) 1 (1) (2)
NN-Based SOM (1) 1 (1)
Others Agglomerative 2 1 (4)
BIRCH (1)
FCM (1) (2) (2)
GSOM 1
k-means 2 (1) (3)
Modified k-means 1 (2)
MTM 1 (1)
PAM 1 (2)
Rule Induction (1)
SE (1) (1)
Single linkage 1
Spatial Stats integrated with ART ANN (1) (1)
Unspecified 1 (3)
Association Others Apriori 1
Unspecified 2 1
Classification S-Based GLZ 2
NBC 3
DT-Based C4.5 1 9
C5.0 1
CART 1 1
ID5R 1
Statistical Batch Based 1
Unspecified 1 7 (1)
NN-Based BNN 2
LVQ 1 (1)
MLP with BP (GD) 5
MLP with BP(LM) 1
PNN 1
RBF 2
SOM 1
Others ADA 1
Boosting 1
BOW 1
CBR 1
DNN-Based 1
DT and ANN 2
EN 1
FDSS based on GA 1
FST and RST 1
FRS 1
GA and RST 1
GP (1)
MIG 1
NFS 1
PRISM 1 2
RST 2 6
RST and LP 1
SVM 2
SVM and DT 1
TM and ANN 1
Prediction S-Based ANOVA 1
BBN 1 1
MLR 16 (1)
NLR 6 (1)
RSM 2
TSA 1
DT-Based CART 4
ID3 1
Unspecified 2
NN-Based ADNN 1
BNN 4
ComputNN 1
Fuzzy ARTMAPNN 1
GRNN 3 (2)
MANN 1 (1)
MLP with BP(GD) 29 (2)
MLP with BP (SA) 1
MLP with BP(CG) 1
MLP with BP(LM) 3 (3)
Table 9 (continued)
Quality task
DM functions Approach Method Description of quality Classification of quality Predicting quality Parameter optimisation
MLP with BP (unspecified) (6)
PN 1
RBF 5 (2)
Recurrent NN Feedforward BP 1
Unspecified 6 (3)
Others AN 1 (1)
ANFIS 1
BANN 1
CBR 1
FAN 1
FNN 1
FR 3
FST (1)
FST and ANN 1
FRS 1
GP 1
IFN 1
NFS 1
SO and MLP 1
TM and ANN 1
Optimisation S-Based RSM 1
TM 3
NN-Based ADNN 1
MLP with BP 4
RBF 2
Unspecified 3
Others ANN, Fuzzy Rule Sets and ES 1
ANN and GA 13
ANN and TM 1
TM, ANN and GA 1
TM and FL 3
RSM and MANN 1
Fuzzy Rule Sets 1
GA 1
GA and NLR 1
GA and FST 1
FMM 1
GRA 1
NLP and MLR 1
NLP and ANN 1
PRIM-Based rule induction 1
SA and ANN 1
SQP 1
for parameter optimisation (e.g. Hsieh & Tong, 2001; Meng & But- spreadsheets, databases, statistical software, DM software, special
ler, 1997; Zhang et al., 2007), while some others considered ANN purpose software, and high-level languages.
for predicting the fitness function values in optimisation by GA Well known statistical software packages; MINITAB™, SAS™,
(e.g. Brinksmeier et al., 1998; Hamedi et al., 2007; Kim et al., SPSS™ and STATISTICA™ were preferred for implementing S-based
2001; Ozcelik & Erzurumlu, 2006). In addition, ANNs were also methods such as MLR, GLZ, ANOVA and PCA. Similarly, for data
used for predicting the response variable in optimisation by RSM preparation and preprocessing, some studies preferred the well
(Ilumoka, 1998), SA (Lin & Wang, 2000), NLP (Lou & Huang, known spreadsheet-application Excel™, database management
2003), and TM (Hung, 2007; Olabi et al., 2006). Similarly, it was ob- system MS ACCESS™ or ORACLE™. However, the other studies
served that TM for parameter optimisation was either used alone mostly utilised the general purpose software MATLAB™, high level
(e.g. Chang et al., 2007; Teng & Hwang, 2007) or combined with languages such as C/C++, or various special purpose software pack-
FL (e.g. Lin, Wang, Yan, & Tarng, 2000; Lu & Antony, 2002). ages. DM software packages such as SPSS Clementine™ and SAS/
The use of more efficient clustering methods such as CLARA, or EM™ were only used in a few of the applications.
CLARANS for handling large data sets has not been observed in the MATLAB™ was used in various applications for clustering, predic-
reviewed applications. Likewise, advanced clustering methods, tion, classification and optimisation, due to its appropriateness for
such as density-based (e.g. DBSCAN, DENCLUE), grid-based (e.g. developing domain specific solutions, supported by several open
STING, WaveCluster), and combined methods (e.g. CLIQUE that source toolboxes. Even though scalability might not be achieved
adopts both grid- and density-based methods) have not been used. with MATLAB™ (Kwon, Omitaomu, & Wang, 2008), the sizes of data
This might be due to the fact that these advanced methods require in the QI applications were not too large to present this as an issue.
more elaborate approaches. In some applications of DM methods, e.g. as ANN and DT, the
programs were developed from scratch by using high level lan-
guages such as C/C++ for flexibility in handling and analysing the
4.5. Software used for mining QI data data according to the particular methodologies followed by the
authors.
Different computational tools shown in Table 10 have been Special purpose software, on the other hand, may be beneficial
used with various purposes in mining quality data. These include and more affordable, if the quality problem is clearly defined and
Table 10
Software used for accomplishing DM tasks in the QI literature.
Type of software Name of software Used by Used for

Spreadsheet- Excel™ Shahbaz et al. (2006) and Chien et al. (2007) Data preparation, data preprocessing
application or EDM
Database MS ACCESS™ Liao et al. (1999a) and Liao et al. (1999b)
management ORACLE™ Shahbaz et al. (2006)
system
Statistical MINITAB™ Feng and Wang (2002), Skinner et al. (2002), Feng and Wang (2003), Tseng S-Based methods such as MLR, GLZ,
software et al. (2005), Ozcelik and Erzurumlu (2006) and He et al. (2007) NLR, ANOVA, PCA, Taguchi Design,
package SAS™ Skinner et al. (2002), Kim et al. (2003), Shi et al. (2004) and Chien et al. etc.
(2007)
SPSS™ Tsai et al. (1999)
STATISTICA™ Tseng et al. (2005)
DM software SPSS Clementine™ Wang (2007), Bakır et al. (2006) and Batmaz (2007) Data preprocessing, classification,
package (currently IBMTM SPSSTM prediction
Modeler)
TANAGRA Da Cunha et al. (2006) Association
SAS/EM™ Deng and Liu (2002) Prediction
General purpose MATLAB™ Brudzewski et al. (2006) Data preprocessing
software Batmaz (2007 Clustering
Kim et al. (2001), Holena and Baerns (2003), Kim et al. (2003), Kurtaran Prediction
et al. (2005), Krimpenis et al. (2006), Ozcelik and Erzurumlu (2006), Yin and
Yu (2006) and Erzurumlu and Oktem (2007)
Jemwa and Aldrich (2005) Classification
Ip et al. (2003), Xue et al. (2005), Ho et al. (2006), Sung et al. (2007), Shen Optimisation
et al. (2007) and Vosniakos et al. (2007)
Special purpose BrainMaker™ Feng and Wang (2003),), Feng and Wang (2004) ANN
software NeuroShell™ Predictor Chen and Ramaswamy (2002)
NeuralWorks™ Predict Cook et al. (2000)
NeuralWorks™ Professional Hsieh and Tong (2001), Yang et al. (2005)
II/Plus
FuzzyTECH™ Yang and Tsai (2002) Neurofuzzy applications
FormRules™ Shao et al. (2007)
Fuzzy-Flou Achiche et al. (2007) FDSS
Mine Set™ Bertino et al. (1999) Association
Q-Yield™
Rosetta Huang et al. (2006) Clustering
c5.0™ Shao et al. (2007) DT
ATLaS Kim et al. (2001) Simulation
ProCAST™ Krimpenis et al. (2006)
DataTools De Abajo et al. (2004) Data preprocessing/classification
High-level C/C++ Han et al. (1999), Tsai et al. (1999), Suneel et al. (2002), Shi et al. (2004), Various purposes such as ANN, DT
language Tseng et al. (2005) and Shahbaz et al. (2006) etc.
Visual Basic™ Liao et al. (1999a), Liao et al. (1999b), Perzyk et al. (2005) and Vasudevan Classification, prediction
et al. (2005)
JAVA™ Chen et al. (2005) Association
understood. For example, BrainMaker™ is a popular ANN tool 5. Conclusions and future research directions
which can be used with different data sources. It provides an op-
tional package based on GA to assist users to find the best network. This paper presents the results of a comprehensive review of
NeuralWorks™ and NeuroShell™ are the other special purpose soft- DM applications in manufacturing industries to the selected QI
ware preferred for ANN modelling and analysis. For association problems. Based on these results, the following concluding re-
analysis, Mine Set™ and Q-Yield™; for fuzzy applications, Fuzzy- marks can be stated, each remark addressing a research question
TECH™, FormRules™ and Fuzyy-Flou; for clustering, Rosetta, and of this study listed in the first section:
for DT applications C5.0 were used. Furthermore, some special pur-
pose software such as ProCAST™ and ATLaS allowed collecting data 1. Data characteristics and data handling: An interesting observa-
through simulation of the experimental conditions. These data tion regarding the DM applications in the literature is the
were then analysed using some other appropriate software for amount of records used in the studies. Although some studies
the chosen DM methods. reported applications using over 10,000 records, in many cases
Use of commercial or open source DM software seems to be lim- the number of records was less than 1000, typically collected
ited compared to the general purpose and special purpose ones. through statistically designed experiments. Even though effec-
Among them, SPSS Clementine™, SAS/EM™ and TANAGRA were tiveness of most DM methods depends on the size of the data
used. SPSS Clementine™ and SAS/EM™ are the most complete pack- set, this has not been reported as a concern in the applications
ages and both dominate the market (Dunham, 2003; Goebel & based on small data. Furthermore, in considerable number of
Gruenwald, 1999; Haughton et al., 2003; Rexer Analytics, 2008). the cases, separately stored production and quality data had
Their limited use observed within the scope of this review can be to be combined to obtain an appropriate data set. Manufactur-
attributed to their being less affordable compared to general pur- ing organisations with well established and integrated data col-
pose and special purpose software packages, as well as their limi- lection systems benefit to a larger extend from DM.
tations in handling the quality problems in an unconventional way, In many of the studies reviewed, data preprocessing and the use
which might be preferred by expert users. of some descriptive DM methods are not reported at all, even
though the success of a DM study depends heavily on the suc- 4. Patterns of DM function and method usage: DM methods are
cess in preprocessing of the data. The comprehensive summary either used as stand-alone or more interestingly as combined
we provide on data handling practices can be used as a refer- (hybrid). Prediction is the most frequently used DM function
ence for future studies. observed in the literature, alone or as a prior/integral part of
2. DM applications according to quality tasks and manufacturing optimisation. In most of the applications ANNs were used for
industries: It has been observed that there is an increasing trend prediction purposes as well as MLR in considerable number of
in the use of DM algorithms for QI. This can be attributed to the the others. DTs or ANNs, on the other hand, were preferred
availability of massive data sets in some manufacturing for classification. For optimisation, several studies directly used
domains and the need to improve processes with the intense ANNs while some others considered ANN for predicting the fit-
competition in the industries. ness function values in optimisation by GA.
Analysis of the literature also indicates that DM applications 5. Software tools: Among various computational tools used in the
within the context of this study were mostly encountered in reviewed applications well known statistical software packages
the metal, computer and electronic products manufacturing such as MINITAB™, SAS™, SPSS™ and STATISTICA™ were pre-
industries, and relatively less observed in plastics, glass, paper, ferred for statistical-based methods. Similarly, well known
food processing and chemical manufacturing industries. DM spreadsheet-applications (Excel™) and database management
applications can find more application domains in the other systems (MS ACCESS™, ORACLE™) were utilised for data han-
industries. However, technical training requirements to suc- dling. The other studies mostly preferred the general purpose
cessfully implement a DM study and costs of software may be software MATLAB™, high level languages such as C/C++, or var-
two major factors that prevent widespread use of the methods. ious special purpose software packages, for their flexibility and
In the literature regarding product and process quality descrip- adoptability.
tion, extracted knowledge from these studies were frequently 6. Other conclusions and directions for future studies: An important
used to accomplish the eventual goals such as quality classifica- necessity for industrial applications is speed in learning and
tion, prediction, and/or optimisation. Many applications that developing the DM models and their solutions, due to continu-
involve predicting quality task were again from the metal, com- ously changing customer and technical requirements in manu-
puter and electronic products manufacturing. Furthermore, facturing industries. 6r projects have been largely successful
ANN was the most widely used DM method in those domains. mainly because of successful training of the industry people
Applications involving classification of quality were not as for the use of the methods with the help of affordable and user
many as those in the predicting quality category. ANN and DT friendly software. For more effective and wide-spread use of the
were common methods in those applications. Parameter opti- DM approaches, DM software with comparable qualities needs
misation tasks were generally performed after predictive mod- to be developed. This, in turn, necessitates availability of more
els were developed for a quality output. These tasks mostly robust, easy to learn and implement data handling and DM
aimed for finding optimal levels of process/product parameters approaches for QI problems. Such software should also have
that consistently yield target quality. ANN, GA, and TM were the capability to help users to select the most appropriate
commonly used in parameter optimisation. methods for the problem, and to interpret the results obtained
3. Reported performance of the DM methods: Performance compar- from the applications. This study can guide researchers and
isons of a few DM methods with each other, as well as with tra- software producers in their effort to develop/further improve
ditional statistical methods, are also common in the literature. their methods and tools by providing them with critical infor-
With a relatively small number of records and variables, tradi- mation on typical characteristics of QI data collected, neces-
tional statistical methods may still provide valuable informa- sary/most preferred DM functions and methods, and expected
tion. results.
In many of the comparative studies, for prediction ANN, DT,
FAN, RST, TSA, CBR and GSA used in combination with ANN
were found more successful compared to MLR. On the other Acknowledgements
hand, for classification of quality, NBC, GP, BOW, SVM, ADA
and CMC outperformed ANN or DT. A few studies comparing This study is supported by Turkish Scientific and Technological
performances of ANN and DT indicated that ANN performed Research Institute through the contract 105M138 and by Middle
better than DT for both prediction and classification. Moreover, East Technical University through the contract BAP-2006-07-02-
findings support the fact that individual classifiers’ accuracy can 06. The authors thank both project teams especially to Fatma
be improved by combined/hybrid use of different DM methods. Güntürkün and Berna Bakır for their contributions to finding and
Furthermore, for parameter optimisation, ANN used in combi- classifying the relevant survey papers.
nation with GA were found as successful.
In QI studies, robustness (to noise in data, mixed types of vari-
ables and complexity) and ease of modelling are important con- Appendix A. Abbreviations
sidering that the practitioners are typically not experts in data
handling and DM. Even though there exist potentially useful
approaches such as RR and MARS, which are strong in these ADA attribute decomposition approach
respects, they have not found a use in the reviewed literature, ADNN adaptive neural networks
yet. Although appeared only in a few studies, FR also seems to AI artificial intelligence
be a promising approach for modelling quality data based on AN abductive network
measurements involving uncertainty or semantic scale usage ANFIS neural fuzzy inference systems
due to SE involved in their collection procedure. ANN artificial neural network
Interpretation and usage of the knowledge obtained by DM ANN-based artificial neural network based
applications might be hard. However, the application studies, ANOVA analysis of variance
mostly performed by academic experts have not given this
problem due consideration. (continued on next page)
ARTNN adaptive resonance theory neural MIG modified breath-first search of an

network interest graph
BANN boosting and bagging (bootstrap MLP multi-layer perceptron
aggregation) neural network MLR multiple linear regression
BBN Bayesian belief network MSE mean squared error
BC Bayesian classifier MTM maximal tree method
BIRCH balanced iterative reducing and NBC naïve Bayesian classifier
clustering using hierarchies NFS neuro-fuzzy system
BNN Bayesian neural network NLP nonlinear programming
BOW breadth-oblivious-wrapper NLR nonlinear regression
BP backpropagation OD object decomposition
CA correlation analysis OLAP on-line analytical processing
CART classification and regression tree PAM partitioning around medoids
CBR case-based reasoning PCA principal components analysis
CC composite classifier PLS partial least squares
CG conjugate-gradient PN polynomial network
CMC combination of multiple classifiers PNN probabilistic neural network
ComputNN computational artificial neural PPM part per million
network PRIM patient rule induction method
DA discriminant analysis PRISM an algorithm for inducing modular
DFSS design for six sigma rules
DM data mining QFD quality function deployment
DMADV define-measure-analyse-develop QI quality improvement
(optimise)-verify RBF radial basis function
DMAIC define-measure-analyse-improve- RR robust regression
control RSM response surface methodology
DNN-based decision neural network based RST rough set theory
DOE design of experiments SA simulated annealing
DT decision trees S-based statistical based
DT-based decision tree based SE subjective evaluation
Dynamic RBFNN dynamic radial basis function neural SO self organization
network SOM self organizing map
EN entropy network SQC statistical quality control
ES expert system SQP sequential quadratic programming
FAN fuzzy adaptive network 6r six sigma
FCM fuzzy c-means SVC support vector classifier
FDSS fuzzy decision support system SVM support vector machines
FL fuzzy logic TM Taguchi method
FMM fuzzy min-max TQC total quality control
FNN fuzzy neural network TQM total quality management
FR fuzzy regression TSA time series analysis
FRS fuzzy rule sets WT wavelet transform
FST fuzzy set theory
Fuzzy ARTMAPNN fuzzy predictive adaptive resonance
theory neural network References
GA genetic algorithm
GD gradient descent Achiche, S., Baron, L., Balazinski, M., & Benaoudia, M. (2007). Online prediction of
pulp brightness using fuzzy logic models. Engineering Applications of Artificial
GLZ generalized linear model Intelligence, 20(1), 25–36.
GP genetic programming Ali, O. G., & Chen, Y. (1999). Design quality and robustness with neural networks.
GRA grey relational analysis IEEE Transactions on Neural Networks, 10(6), 1518–1527.
Aloudat, M. (2006). A methodology for quality conformance analysis and prediction.
GRNN general regression neural network Thesis (PhD). University of Houston.
GSA grey superior analysis Baek, D. H., Jeong, I. J., & Han, C. H. (2005). Application of data mining for improving
GSOM growing self organizing map yield in water fabrication system. In O. Gervasi et al. (Eds.). Computational
science and its applications – ICCSA 2005, 9-12 May 2005 Singapore (Vol. 3483,
HDTNNC hybrid decision tree and artificial pp. 222–231). Berlin: Springer-Verlag.
neural network classifier Baek, J. G., Kim, C. O., & Kim, S. S. (2002). Online learning of the cause-and-effect
IFN info-fuzzy network knowledge of a manufacturing process. International Journal of Production
Research, 40(14), 3275–3291.
KDD knowledge discovery in databases _ A., Köksal, G., & Özdemirel, N. E.
Bakır, B., Batmaz, I., Güntürkün, F. A., Ipekci, I.
LM Levenberg–Marquart (2006). Defect cause modeling with decision tree and regression analysis. In
LP linear programming Proceedings of XVII international conference on computer and information science
LVQ learning vector quantization and engineering, 8–10 December 2006. Cairo: World Enformatica Society (Vol. 16,
pp. 266–269).
MAE mean absolute error Batmaz, I. (2007). Data mining applications on manufacturing data: A casting
MANN modular artificial neural network quality improvement case. In H. O. Ayhan & I. Batmaz (Eds.), Recent advances in
MARS multivariate adaptive regression statistics (pp. 197–206). Ankara, Turkey: TUIK.
Bertino, E., Catania, B., & Caglio, E. (1999). Applying data mining techniques to wafer
splines
manufacturing. Principles of Data Mining and Knowledge Discovery, 1704, 41–50. Erzurumlu, T., & Oktem, H. (2007). Comparison of response surface model with
Brady, J. E., & Allen, T. T. (2006). Six sigma literature: A review and agenda for future neural network in determining the surface quality of moulded parts. Materials
research. Quality and Reliability Engineering International, 22, 335–367. and Design, 28(2), 459–465.
Braha, D., & Shmilovici, A. (2002). Data mining for improving a cleaning process in Fasser, Y., & Brettner, D. (2002). Management for quality in high-technology
the semiconductor industry. IEEE Transactions on Semiconductor Manufacturing, enterprises. New York: Wiley-Interscience.
15(1), 91–101. Fayyad, U., Piatetsky-Shapiro, G., & Smyth, P. (1996). The KDD process for extracting
Braha, D. (Ed.). (2001). Data mining for design and manufacturing: Methods and useful knowledge from volumes of data. Communication of ACM, 39(11), 27–34.
applications. Dordrecht: Kluwer Academic Publishers. Feng, C. X. J., & Wang, X. F. (2003). Surface roughness predictive modeling: Neural
Brinksmeier, E., Tönshoff, H. K., Czenkusch, C., & Heinzel, C. (1998). Modeling and networks versus regression. IIE Transactions on Design and Manufacturing, 35(1),
optimization of grinding processes. Journal of Intelligent Manufacturing, 9(4), 11–27.
303–314. Feng, C. X. J., & Kusiak, A. (2006). Data mining applications in engineering design,
Brudzewski, K., Kesik, A., Kolodziejczyk, K., Zborowska, U., & Ulaczyk, J. (2006). manufacturing and logistics. International Journal of Production Research, 44(14),
Gasoline quality prediction using gas chromatography and FTIR spectroscopy: 2689–2694.
An artificial intelligence approach. Fuel, 85(4), 553–558. Feng, C. X. J., & Wang, X. (2002). Development of empirical models for surface
Chang, D. S., & Jiang, S. T. (2002). Assessing quality performance based on the on- roughness prediction in finish turning. International Journal of Advanced
line sensor measurements using neural networks. Computers and Industrial Manufacturing Technology, 20(5), 348–356.
Engineering, 42(2–4), 417–424. Feng, C. X. J., & Wang, X. F. D. (2004). Data mining techniques applied to predictive
Chang, P. C., Liu, C. H., Fan, C. Y., & Chang, H. C. (2007). Data clustering and fuzzy modeling of the knurling process. IIE Transactions, 36(3), 253–263.
neural network for sales forecasting in printed circuit board industry. In Fu, L. M. (1994). Neural networks in computer intelligence. New York: McGraw-Hill.
Proceedings of the 2007 IEEE symposium on computational intelligence and data Ganesan, R., Das, T. K., & Venkataraman, V. (2004). Wavelet-based multiscale
mining (CIDM 2007) (pp. 107–113). Honolulu, HI: IEEE. statistical process monitoring: A literature review. IIE Transactions, 36(9),
Chen, C. R., & Ramaswamy, H. S. (2002). Modeling and optimization of variable 787–806.
retort temperature (VRT) thermal processing using coupled neural networks Gardner, M., & Bieker, J. (2000). Data mining solves tough semiconductor
and genetic algorithms. Journal of Food Engineering, 53(3), 209–220. manufacturing problems. In S.J. Proceedings of the sixth ACM SIGKDD
Chen, W. C., Lee, A. H. I., Deng, W. J., & Liu, K. Y. (2007). The implementation of international conference on knowledge discovery and data mining, KDD 2000
neural network for semiconductor PECVD process. Expert Systems with (pp. 376–383), 20–23 August, Boston. New York: ACM.
Applications, 32(4), 1148–1153. Georgilakis, P., & Hatziargyriou, N. (2002). On the application of artificial
Chen, W. C., Tseng, S. S., & Wang, C. Y. (2005). A novel manufacturing defect intelligence techniques to the quality improvement industrial processes. In
detection method using association rule mining techniques. Expert Systems with Proceedings of the second hellenic conference on AI: Methods and applications of
Applications, 29(4), 807–815. artificial intelligence , 11–12 April, Thessaloniki, Greece. In I. P. Vlahavas & C. D.
Cherian, R. P., Midha, P. S., & Pipe, A. G. (2000). Modeling the relationship between Spyropoulos (Eds.). Lecture notes in computer science (Vol. 2308, pp. 473–484).
process parameters and mechanical properties using bayesian neural networks London: Springer-Verlag.
for powder metal parts. International Journal of Production Research, 38(10), Giudici, P. (2003). Applied data mining: Statistical methods for business and industry.
2201–2214. New York: J. Wiley.
Chiang, T. L., Su, C. T., Li, T. S., & Huang, R. C. C. (2002). Improvement of process Goebel, M., & Gruenwald, L. (1999). A survey of data mining and knowledge
capability through neural networks and robust design: A case study. Quality discovery software tools. SIGKDD Explor. Newsl., 1(1), 20–33.
Engineering, 14(2), 313–318. Guessasma, S., Salhi, Z., Montavon, G., Gougeon, P., & Coddet, C. (2004). Artificial
Chien, C., Li, H., & Jeang, A. (2006). Data mining for improving the solder bumping intelligence implementation in the APS process diagnostic. Materials Science and
process in the semiconductor packaging industry. Intelligent Systems in Engineering B, 110(3), 285–295.
Accounting, Finance and Management, 14(1–2), 43–57. Hamedi, M., Shariatpanahi, M., & Mansourzadeh, A. (2007). Optimizing spot welding
Chien, C. F., Wang, W. C., & Cheng, J. C. (2007). Data mining for yield enhancement in parameters in a sheet metal assembly by neural networks and genetic
semiconductor manufacturing and an empirical study. Expert Systems with algorithm. Proceedings of the Institution of Mechanical Engineers Part B-Journal
Applications, 33(1), 192–198. of Engineering Manufacture, 221(7), 1175–1184.
Chong, I. G., Albin, S. L., & Jun, C. H. (2007). A data mining approach to process Han, J., & Kamber, M. (2006). Data mining: Concepts and techniques. San Francisco:
optimization without an explicit quality function. IIE Transactions, 39(8), Morgan Kaufmann Publishers.
795–804. Han, L., Han, L., & Liu, C. (1999). Neural network applied to prediction of the failure
Choudhary, A. K., Harding, J. A., & Tiwari, M. K. (2008). Data mining in stress for a pressurized cylinder containing defects. International Journal of
manufacturing: A review based on the kind of knowledge. Journal of Pressure Vessels and Piping, 76(4), 215–219.
Intelligent Manufacturing, published online, ISSN:0956-5515. Hand, H., Mannila, H., & Smyth, P. (2001). Principles of data mining. Cambridge,
Cook, D. F., Ragsdale, C. T., & Major, R. L. (2000). Combining a neural network with a Mass: MIT Press.
genetic algorithm for process parameter optimization. Engineering Applications Harding, J. A., Shahbaz, M., Srinivas, S., & Kusiak, A. (2006). Data mining in
of Artificial Intelligence, 13(4), 391–396. manufacturing: A review. Journal of Manufacturing Science and Engineering-
Cool, T., Bhadeshia, H. K. D. H., & Mackay, D. J. C. (1997). The yield and ultimate Transactions of ASME, 128(4), 969–976.
tensile strength of steel welds. Materials Science and Engineering A, A223(1–2), Haughton, D., Deichmann, J., Eshghi, A., Sayek, S., Teebagy, N., & Topi, H. (2003). A
186–200. review of software packages for data mining. The American Statistician, 57(4),
Cser, L., Gulyas, J., Szucs, L., Horvath, A., Arvai, L., & Baross, B. (2001). Different kinds 290–309.
of neural networks in control and monitoring of hot rolling mill. In Proceedings He, S.-G., Li, L., & Qi, E.-S. (2007). Study on the quality improvement of injection
of the 14th international conference on industrial and engineering applications of molding in LED packaging processes based on DOE and data mining. In
artificial intelligence and expert systems: engineering of intelligent systems, IEA/AIE Proceedings of the wireless communications, networking and mobile computing
2001, 4–7 June 2001 Budapest Hungary. In L. Monostori, J. Vancza, & M. Ali (Eds.). WICom 2007, 21–25 September 2007 Shanghai China (pp. 6625–6628). IEEE Con.
Lecture notes in computer science (Vol. 2070, pp. 791–796). London: Springer- Pub.
Verlag. Ho, G. T. S., Lau, H. C. W., Lee, C. K. M., Ip, A. W. H., & Pun, K. F. (2006). An intelligent
Cus, F., & Balic, J. (2003). Optimization of cutting process by GA approach. Robotics production workflow mining system for continual quality enhancement.
and Computer Integrated Manufacturing, 19(1–2), 113–121. Intelligent Journal of Advanced Manufacturing Technology, 28(7–8), 792–809.
Da Cunha, C., Agard, B., & Kusiak, A. (2006). Data mining for improvement of Ho, S., Lee, K., Chen, S., & Ho, S. (2002). Accurate modeling and prediction of surface
production quality. International Journal of Production Research, 44(18/19), roughness by computer vision in turning operations using an adaptive
4027–4041. neurofuzzy inference system. International Journal of Machine Tools and
Dasu, T., & Johnson, T. (2003). Exploratory data mining and data cleaning. New York: Manufacture, 42(13), 1441–1446.
Wiley-Interscience. Holena, M., & Baerns, M. (2003). Feedforward neural networks in catalysis – A tool
De Abajo, N., Diez, A. B., Lobato, V., & Cuesta, S. R. (2004). ANN quality diagnostic for the approximation of the dependency of yield on catalyst composition, and
models for packaging manufacturing: an industrial data mining case study. In R. for knowledge extraction. Catalysis Today, 81(3), 485–494.
Kohavi et al. (Eds.). KDD-2004: Proceedings of the tenth ACM SIGKDD international Hou, T. H., & Huang, C. C. (2004). Application of fuzzy logic and variable precision
conference on knowledge discovery and data mining, 22–25 August 2004 Seattle rough set approach in a remote monitoring manufacturing process for diagnosis
Washington (pp. 799–804). New York: ACM Press. rule induction. Journal of Intelligent Manufacturing, 15(3), 395–408.
Deng, B., & Liu, X. (2002). Data mining in quality improvement. SUGI27: Proceedings Hou, T. H., Liu, W. L., & Lin, L. (2003). Intelligent remote monitoring and diagnosis of
of the twenty-seventh annual SASÒ users group international conference, 14–17 manufacturing processes using an integrated approach of neural networks and
April 2002 Orlando, Florida [online]. Available from: http://www2.sas.com/ rough sets. Journal of Intelligent Manufacturing, 14(2), 235–239.
proceedings/sugi27/Proceed27.pdf Accessed 23.03.2008. Hsieh, K. L., & Tong, L. I. (2001). Optimization of multiple quality responses
Dhond, A., Gupta, A., & Vadhavkar, S. (2000). Data mining techniques for optimizing involving qualitative and quantitative characteristics in IC manufacturing using
inventories for electronic commerce. In S.J. proceedings of the sixth ACM SIGKDD neural networks. Computers in Industry, 46(1), 1–12.
international conference on knowledge discovery and data mining, KDD 2000 (pp. Hsu, S. C., & Chien, C. F. (2007). Hybrid data mining approach for pattern extraction
480–486), 20–23 August. Boston. New York: ACM. from wafer bin map to improve yield in semiconductor manufacturing.
Dunham, M. H. (2003). Data mining introductory and advanced topics. New Jersey: International Journal of Production Economics, 107(1), 88–103.
Prentice Hall/Pearson Education. Hu, C., & Su, S. (2004). Hierarchical clustering methods for semiconductor
manufacturing data. Proceedings of the IEEE international conference on
networking, sensing and control, 21–23 March 2004 Taipei, Taiwan (Vol. 2, Lee, S. H., & Dornfeld, D. A. (2007). Prediction of burr formation during face milling
pp. 1063–1068). IEEE. using an artificial neural network with optimized cutting conditions.
Huang, C., Li, T., & Peng, T. (2006). Attribute selection based on rough set theory for Proceedings of the Institution of Mechanical Engineers Part B-Journal of
electromagnetic interference (EMI) fault diagnosis. Quality Engineering, 18(2), Engineering Manufacture, 221(12), 1705–1714.
161–171. Lewis, R. W., & Ransing, R. S. (1997). A semantically constrained bayesian network
Huang, H., & Wu, D. (2005). Product quality improvement analysis using data for manufacturing diagnosis. International Journal of Production Research, 35(8),
mining: A case study in ultra-precision manufacturing industry. In Fuzzy 2171–2187.
systems and knowledge discovery, proceedings of the second international Li, T. S., Huang, C. L., & Wu, Z. Y. (2006). Data mining using genetic programming for
conference on fuzzy systems and knowledge discovery, FSKD 2005, 22–29 August construction of a semiconductor manufacturing yield rate prediction system.
2005 Changsha China. In L. Wang & Y. Jin (Eds.). Lecture notes in computer science Journal of Intelligent Manufacturing, 17, 355–361.
(Vol. 3614, pp. 577–580). Berlin: Springer. Li, T. S., Su, C. T., & Chiang, T. L. (2003a). Applying robust multi-response quality
Hung, Y. H. (2007). Optimal process parameters design for a wire bonding of ultra- engineering for parameter selection using a novel neural-genetic algorithm.
thin CSP package based on hybrid methods of artificial intelligence. Computers in Industry, 50(1), 113–122.
Microelectronics International, 24(3), 3–10. Li, M., Feng, S., Sethi, I. K., Luciow, J., & Wagner, K. (2003b). Mining production data
Hussain, M. A. (1999). Review of the applications of neural networks in chemical with neural network and CART. In Proceedings of the third IEEE international
process control — Simulation and online implementation. Artificial Intelligence conference on data mining (ICDM 2003), 19–22 November 2003 Melbourne FL
in Engineering, 13(1), 55–68. (pp. 731–734). IEEE.
Ilumoka, A. A. (1998). A modular neural network approach to microelectronic Lian, J., Lai, X. M., Lin, Z. Q., & Yao, F. S. (2002). Application of data mining and
circuit yield optimization. Microelectronics Reliability, 38(4), 571–580. process knowledge discovery in sheet metal assembly dimensional variation
International Standard Industrial Classification of All Economic Activities, Revision 4, diagnosis. Journal of Materials Processing Technology, 129(1–3), 315–320.
Section C – Manufacturing (10 February (2009). http://unstats.un.org/unsd/cr/ Liao, T. W., Zhan, Z. H., & Mount, C. R. (1999a). An integrated database and expert
registry/regcst.asp?Cl=27, United Nations. system for failure mechanism identification: Part I – Automated knowledge
Ip, K. W., Kwong, C. K., & Wong, Y. W. (2003). Fuzzy regression approach to acquisition. Engineering Failure Analysis, 6(6), 387–406.
modeling transfer moulding for microchip encapsulation. Journal of Materials Liao, T. W., Zhan, Z. H., & Mount, C. R. (1999b). An integrated database and expert
Processing Technology, 140(1–3), 147–151. system for failure mechanism identification: Part II – The system and
Jemwa, G. T., & Aldrich, C. (2005). Improving process operations using support performance testing. Engineering Failure Analysis, 6(6), 407–421.
vector machines and decision trees. AICHE Journal, 51(2), 526–543. Lin, C. L., Lin, J. L., & Ko, T. C. (2002). Optimization of the EDM process based on the
Jiao, Y., Lei, S., Pei, Z. J., & Lee, E. S. (2004). Fuzzy adaptive networks in machining orthogonal array with fuzzy logic and grey relational analysis method.
process modeling: Surface roughness prediction for turning operations. International Journal of Advanced Manufacturing Technology, 19(4), 271–277.
International Journal of Machine Tools and Manufacture, 44(15), 1643–1651. Lin, J. L., Wang, K. S., Yan, B. H., & Tarng, Y. S. (2000). Optimization of the electrical
Kang, B. S., Choe, D. H., & Park, S. C. (1999). Intelligent process control in discharge machining process based on the Taguchi method with fuzzy logics.
manufacturing industry with sequential processes. International Journal of Journal of Materials Processing Technology, 102(1–3), 48–55.
Production Economics, 583–590. Lin, W. S., & Wang, K. S. (2000). Modelling and optimization of turning processes for
Karim, M. A., Halgamuge, S., Smith, A. J. R., & Hsu, A. L. (2006). Manufacturing yield slender parts. International Journal of Production Research, 38(3), 587–606.
improvement by clustering. In Neural information processing: 13th international Liu, X., Tang, H., Fan, Z., & Deng, B. (2004). Quality improvement modeling and
conference proceedings Part III, 3–6 October 2006 Hong Kong China. In I. King et al. practice in baosteel based on KIV-KOV analysis. In Advances in web-age
(Eds.). LNCS (Vol. 4234, pp. 526–534). Berlin: Springer. information management: 5th international conference, WAIM 2004, 15–17 July
Karnik, S. R., Gaitonde, V., & Davim, J. P. (2007). Integrating Taguchi principle with 2004 Dalian, China. In Q. Li, G. Wang, & L. Feng (Eds.). Lecture Notes in Computer
genetic algorithm to minimize burr size in drilling of AISI 316L stainless steel Science (Vol. 3129, pp. 720–725). Berlin: Springer-Verlag.
using an artificial neural network model. Proceedings of the Institution of Lou, H. H., & Huang, Y. L. (2003). Hierarchical decision making for proactive quality
Mechanical Engineers Part B-Journal of Engineering Manufacture, 221(12), control: System development for detect reduction in automotive coating
1695–1704. operations. Engineering Applications of Artificial Intelligence, 16(3), 237–250.
Kartal, E. (2007). Metamodeling complex systems using linear and nonlinear regression Lu, J. C. (2001). Methodology of mining massive data sets for improving
methods. Unpublished Master of Science Thesis, Middle East Technical manufacturing quality/efficiency. In D. Braha (Ed.), Data mining for design and
University, Ankara, Turkey. manufacturing: Methods and applications (pp. 255–288). Dordrecht: Kluwer
Kim, E. S., Oh, C., Lee, S., Lee, B., & Yun, I. (2001). Modeling and optimization of Academic Publishers.
process parameters for GaAs/AlGaAs multiple quantum well avalanche Lu, D. W., & Antony, J. (2002). Optimization of multiple responses using a fuzzy-rule
photodiodes using genetic algorithms. Microelectronics Journal, 32(7), 563–567. based inference system. International Journal of Production Research, 40(7),
Kim, I., Son, J., & Yarlagadda, P. K. D. V. (2003). A study on the quality improvement 1613–1625.
of robotic GMA welding process. Robotics and Computer Integrated Maimon, O., & Rokach, L. S. (2001). Data mining by attribute decomposition with
Manufacturing, 19(6), 567–572. semiconductor manufacturing case study. In D. Braha (Ed.), Data mining for
Kim, S. H., & Lee, C. M. (1997). Nonlinear prediction of manufacturing systems design and manufacturing (pp. 311–336). Dordrecht: Kluwer Academic
through explicit and implicit data mining. Computer and Industrial Engineering, Publishers.
33(3-4), 461–464. Martinez, W. L., & Martinez, A. R. (2002). Computational statistics handbook with
Kolarik, W. J. (1995). Creating quality: Concepts, systems, strategies, and tools. New MATLAB. Boca Raton: Chapman and Hall.
York: McGraw-Hill. Mathews, P. G., & Shunmugam, M. S. (1999). Neural-network approach for
Köksal, G., Smith, W. A., Jr., Fathi, Y., Lu, J. C., & McGregor, R. (1998). A case study in predicting hole quality in reaming. International Journal of Machine Tools &
off-line quality control: Characterization and optimization of batch dyeing Manufacture, 39(5), 723–730.
process design. International Journal of Technology Management: Special Issue on Meng, T. K., & Butler, C. (1997). Solving multiple response optimisation problems
Total Quality Management, 16(4/5/6), 358–382. using adaptive neural networks. International Journal of Advanced Manufacturing
Köksal, G., Batmaz, I., _ & Testik, M. C. (2008). Data mining processes and a review of Technology, 13(9), 666–675.
their applications for product and process quality improvement in manufacturing Mieno, F., Sato, T., Slubnya, Y., Odagiri, K., Tsuda, H., & Take, K. (1999). Yield
industry. Technical Report No: 08-03, Industrial Engineering Department, improvement using data mining system. In Semiconductor manufacturing
Middle East Technical University, Ankara. conference proceedings, 1999 IEEE international symposium on, 11–13 October
Krimpenis, A., Benardos, P. G., Vosniakos, G. C., & Koukouvitaki, A. (2006). Santa Clara CA (pp. 391–394). IEEE.
Simulation-based selection of optimum pressure die-casting process Montgomery, D. (2005). Introduction to statistical quality control. Hoboken, NJ: John
parameters using neural nets and genetic algorithms. International Journal of Wiley.
Advances Manufacturing Technology, 27(5–6), 509–517. Newman, T. S., & Jain, A. K. (1995). A survey of automated visual inspection.
Kurtaran, H., Ozcelik, B., & Erzurumlu, T. (2005). Warpage optimization of a bus Computer Vision, Graphics and Image Processing, 61(2), 231–262.
ceiling lamb base using neural network model and genetic algorithm. Journal of Olabi, A. G., Casalino, G., Benyounis, K. Y., & Hashmi, M. S. J. (2006). An ANN and
Materials Processing Technology, 169(2), 314–319. Taguchi algorithms integrated approach to the optimization of CO2 laser
Kusiak, A., & Kurasek, C. (2001). Data mining of printed-circuit board defects. IEEE welding. Advances in Engineering Software, 37(10), 643–648.
Transactions on Robotics and Automation, 17(2), 191–196. Ordieres Meré, J. B., González Marcos, A., González, J. A., & Lobato Rubio, V. (2004).
Kusiak, A. (2000). Decomposition in data mining: An industrial case study. IEEE Estimation of mechanical properties of steel strip in hot dip galvanising lines.
Transactions on Electronics Packaging Manufacturing, 23(4), 345–354. Ironmaking and Steelmaking, 31(1), 43–50. 8.
Kusiak, A. (2002). A data mining approach for generation of control signatures. Ozcelik, B., & Erzurumlu, T. (2006). Comparison of the warpage optimization in the
Journal of Manufacturing Science and Engineering-Transactions of the ASME, plastic injection molding using ANOVA, neural network model and genetic
124(4), 923–926. algorithm. Journal of Materials Processing Technology, 171(3), 437–445.
Kusiak, A. (2006). Data mining: Manufacturing and service applications. Perzyk, M., Biernacki, R., & Kochanski, A. (2005). Modeling of manufacturing processes
International Journal of Production Research, 44(18/19), 4175–4191. by learning systems: The naive bayesian classifier versus artificial neural
Kwon, Y., Omitaomu, O. A., & Wang, G. N. (2008). Data mining approaches for networks. Journal of Materials Processing Technology, 164–165, 1430–1435.
modeling complex electronic circuit design activities. Computers & Industrial Phadke, M. S. (1989). Quality engineering using robust design. Englewood Cliffs, NJ:
Engineering, 54, 229–241. Prentice-Hall.
Last, M., & Kandel, A. (2001). Data mining for process and quality control in the Pham, D. T., & Afify, A. A. (2005). Machine-learning techniques and their
semiconductor industry. In D. Braha (Ed.), Data mining for design and applications in manufacturing. Proceedings of the Institution of Mechanical
manufacturing (pp. 207–234). Dordrecht: Kluwer Academic Publishers. Engineers Part B-Journal of Engineering Manufacture, 219(5), 395–412.
Pham, D. T., & Alcock, R. J. (2003). Smart inspection systems: Techniques and Tseng, T. L., Leeper, T., Banda, C., Herren, S. M., & Ford, J. (2004c). Quality assurance in
applications of intelligent vision. Amsterdam, Boston: Academic Press. machining process using data mining. IERC 2004, Houston, TX, 6.
Pyle, D. (1999). Data preparation for data mining. San Francisco, CA: Morgan Tsuda, H., Shirai, H., Takagi, O., Take, R. (2000). Yield analysis and improvement by
Kaufmann Publishers. reducing manufacturing fluctuation noise. In Proceedings of International
Rallo, R., Ferre-Gine, J., Arenas, A., & Giralt, F. (2002). Neural virtual sensor for the Symposium on Semiconductor Manufacturing (ISSM 2000), 26-28 September
inferential prediction of product quality from process variables. Computers and Tokyo. 249-252.
Chemical Engineering, 26(12), 1735–1754. Umbrello, D., Ambogio, G., Filice, L., & Shivpuri, R. (2007). An ANN approach for
Rexer Analytics (2008). Available from: http://www.rexeranalytics.com/Data- predicting subsurface residual stresses and the desired cutting conditions
Miner-Survey-Results-2008.html. during hard turning. Journal of Materials Processing Technology, 189(1–3),
Riverol, C., & Cooney, J. (2007). Estimation of the ester formation during beer 143–152.
fermentation using neural networks. Journal of Food Engineering, 82(4), Vasudevan, M., Rao, B. P. C., Venkatraman, B., Jayakumar, T., & Raj, B. (2005).
585–588. Artificial neural network modeling for evaluating austenitic stainless steel and
Rokach, L., & Maimon, O. (2006). Data mining for improving the quality of Zircaloy-2 welds. Journal of Materials Processing Technology, 169(3), 396–400.
manufacturing: A feature set decomposition approach. Journal of Intelligent Vasudevan, M., Murugananth, M., & Bhaduri, A. K. (2002). Application of bayesian
Manufacturing, 17(3), 285–299. neural network for modeling and prediction of ferrite number in austenitic
Sadeghi, B. H. M. (2000). A BP-neural network predictor model for plastic injection stainless steel welds. In H. Cerjak & H. K. D. H. Bhadeshia (Eds.), Mathematical
molding process. Journal of Materials Processing Technology, 103(3), 411–416. modeling of weld phenomena-VI (pp. 1079–1099). London: The Institute of
Sarimveis, H., Doganis, P., & Alexandridis, A. (2006). A classification technique based Materials.
on radial basis function neural networks. Advances in Engineering Software, Vosniakos, G. C., Maroulis, T., & Pantelis, D. (2007). A method for optimizing process
37(4), 218–221. parameters in layer-based rapid prototyping. Proceedings of the Institution of
Sebzalli, Y. M., & Wang, X. Z. (2001). Knowledge discovery from process operational Mechanical Engineers Part B-Journal of Engineering Manufacture, 221(8),
data using PCA and fuzzy clustering. Engineering Applications of Artificial 1329–1340.
Intelligence, 14(5), 607–616. Wang, K. (2007). Applying data mining to manufacturing: the nature and
Shahbaz, M., Srinivas, S., Harding, J. A., & Turner, M. (2006). Product design and implications. Journal of Intelligent Manufacturing, 18, 487–495.
manufacturing process improvement using association rules. Proceedings of the Wang, K., Tong, S., Eynard, B., Roucoules, L., & Matta, N. (2007a). Review on
Institution of Mechanical Engineers Part B-Journal of Engineering Manufacture, application of data mining in product design and manufacturing. In Lei, J., Yu, J.,
220(2), 243–254. Zhou, S. (Eds.), Proceedings of the fourth international conference on fuzzy systems
Shao, Q., Rowe, R. C., & York, P. (2007). Comparison of neurofuzzy logic and decision and knowledge discovery (FSKD 2007), 24–27 August 2007Haikou Hainan, China.
trees in discovering knowledge from experimental data of an immediate release IEEE (pp. 613–618).
tablet formulation. European Journal of Pharmaceutical Sciences, 31(2), 129–136. Wang, L., Fang, J. C., Zhao, Z. Y., & Zeng, H. P. (2007b). Application of backward
Shen, C., Wang, L., & Li, Q. (2007). Optimization of injection molding process propagation network for forecasting hardness and porosity of coatings by
parameters using combination of artificial neural network and genetic plasma spraying. Surface and Coating Technology, 201(9–11), 5085–5089.
algorithm method. Journal of Materials Processing Technology, 183(2–3), Wang, R., Wang, L., Zhao, L., & Liu, Z. (2006). Influence of process parameters on part
412–418. shrinkage in SLS. Intelligent Journal of Advanced Manufacturing Technology, 33(5–
Shi, X., Schillings, P., & Boyd, D. (2004). Applying artificial neural networks and 6), 498–504.
virtual experimental design to quality improvement of two industrial processes. Wei, C. P., Piramuthu, S., & Shaw, M. J. (2003). Knowledge discovery and data
International Journal of Production Research, 42(1), 101–118. mining. In C. W. Holsapple (Ed.). Handbook on Knowledge Management (Vol. 2,
Skinner, K. R., Montgomery, D. C., Runger, G. C., Fowler, J. W., McCarville, D. R., pp. 157–189). Berlin: Springer-Verlag.
Rhoads, T. R., et al. (2002). Multivariate statistical methods for modeling and Witten, I. H. (2005). Data mining: Practical machine learning tools and techniques.
analysis of wafer probe test data. IEEE Transactions on Semiconductor Boston, MA: Morgan Kaufman.
Manufacturing, 15(4), 523–530. Xue, Y., Kim, I. S., Son, J. S., Park, C. E., Kim, H. H., Sung, B. S., et al. (2005). Fuzzy
Suneel, T. S., Pandle, S. S., & Date, P. P. (2002). A technical note on integrated product regression method for prediction and control the bead width in the robotic arc-
quality model using artificial neural networks. Journal of Materials Processing welding process. Journal of Materials Processing Technology, 164, 1134–1139.
Technology, 121(1), 77–86. Yang, K., & Trewn, J. (2004). Multivariate statistical methods in quality management.
Sung, B. S., Kim, I. S., Xue, Y., Kim, H. H., & Cha, Y. H. (2007). Fuzzy regression model New York: McGraw-Hill.
to predict the bead geometry in the robotic welding process. ACTA Metallurgica Yang, T., & Tsai, T. (2002). Modeling and implementation of a neurofuzzy system for
Sinica, 20(6), 391–397. surface mount assembly defect prediction and control. IIE Transactions, 34(7),
Taguchi, G., Chowdhury, S., & Taguchi, S. (2000). Robust engineering. New York: 637–646.
McGraw-Hill. Yang, T., Tsai, T., & Yeh, J. (2005). A neural network-based prediction model for fine
Tam, C. M., Tong, T. K. L., Lau, T. C. T., & Chan, K. K. (2004). Diagnosis of prestressed pitch stencil printing quality in surface mount assembly. Engineering
concrete pile defects using probabilistic neural networks. Engineering Structures, Applications of Artificial Intelligence, 18(3), 335–341.
26(8), 1155–1162. Yin, X., & Yu, W. (2006). The virtual manufacturing model of the worsted yarn based
Tay, K. M., & Butler, C. (1997). Modelling and optimizing of a MIG welding process – on artificial neural networks and grey theory. Applied Mathematics and
A case study using experimental designs and neural networks. Quality and Computation, 185(1), 322–332.
Reliability Engineering International, 13(2), 61–70. Zhai, L., Khoo, L., & Fok, S. (2002). Feature extraction using rough set theory and
Teng, S. Y., & Hwang, S. J. (2007). Predicting the process induced warpage of genetic algorithms—An application for the simplification of product quality
electronic packages using the P–V–T–C equation and the Taguchi method. evaluation. Computers and Industrial Engineering, 43(4), 661–676.
Microelectronics Reliability, 47(12), 2231–2241. Zhang, J. H., Xie, A. G., & Shen, F. M. (2007). Multi-objective optimization and
Tsai, Y. H., Chen, J. C., & Lou, S. (1999). An in-process surface recognition system analysis model of sintering process based on BP neural network. Journal of Iron
based on neural networks in end milling cutting operations. International and Steel Research, International, 14(2), 01–05.
Journal of Machine Tools and Manufacture, 39(4), 583–605. Zhou, C., Nelson, P. C., Xiao, W., Tripak, T. M., & Lane, S. A. (2001). An intelligent data
Tsai, C. Y., Chiu, C. C., & Chen, J. S. (2005). A case-based reasoning system for PCB mining system for drop test analysis of electronic products. IEEE Transactions on
defect prediction. Expert Systems with Applications, 28, 813–822. Electronics Packaging Manufacturing, 24(3), 222–231.
Tseng, H. Y. (2006). Welding parameters optimization for economics design using Zhou, Q., Xiong, Z., Zhang, J., & Xu, Y. (2006). Hierarchical neural network based
neural approximation and genetic algorithm. International Journal of Advanced product quality prediction of industrial ethylene pyrolysis process. In
Manufacturing Technology, 27(9/10), 897–901. Proceedings of the third international symposium on neural networks, advances
Tseng, T. B., Kwon, Y., & Ertekin, Y. M. (2005). Feature-based rule induction in in neural networks (ISNN 2006). In J. Wang et al. (Eds.). Lecture Notes in Computer
machining operation using rough set theory for quality assurance. Robotics and Science, 28 May – 1 June, 2006 Chengdu China (Vol. 3973, pp. 1132–1137). Berlin:
Computer-integrated Manufacturing, 21(6), 559–567. Springer-Verlag.
Tseng, T. L., JothiShankar, M. C., & Wu, T. (2004a). Quality control problems in Zorriassatine, F., & Tannock, J. D. T. (1998). A review of neural networks for
printed circuit board manufacturing – An extended rough set theory approach. statistical process control. Journal of Intelligent Manufacturing, 9(3), 209–224.
Journal of Manufacturing Systems, 23(1), 56–72. Zuperl, U., & Cus, F. (2003). Optimization of cutting conditions during cutting by
Tseng, T. L., JothiShankar, M. C., Wu, T., Xing, G., & Jiang, F. (2004b). Applying data using neural networks. Robotics and Computer-Integrated Manufacturing, 19(1–
mining approaches for defect diagnosis in manufacturing industry. IERC 2004, 2), 189.
Houston, TX, 7.

A Review of Data Mining Applications For

Uploaded by

Copyright:

Available Formats

A Review of Data Mining Applications For

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

A Review of Data Mining Applications For

Uploaded by

Copyright:

Available Formats

Expert Systems with Applications 38 (2011) 13448–13467

Contents lists available at ScienceDirect

Expert Systems with Applications

A review of data mining applications for quality improvement

1. Introduction The aim of this study is to extensively review DM applications

Criteria Desired value Assessment

quality levels cannot be achieved without developing strong de-

This summary is adapted from Phadke (1989), and it classifies

stages as explained below.

Three critical stages of product and manufacturing process de-

ting values of their design parameters (i.e. parameter design, or

design optimisation), and setting tolerances for the design param-

eters (i.e. tolerance design). During the conceptual design of a

the best settings are determined for the controllable product/pro-

cess design parameters. At this parameter design stage, it is typical

to assume wide tolerances on uncontrollable noise factors, and

noise factors. If such settings cannot be found, then tolerances

for variability of the product/process parameter values are selec-

tively reduced, and use of higher grade materials and sophisticated

ability. In this study, we especially consider conceptual design

and parameter design stages, where DM has found widespread use.

Major QI and control activities during the product manufactur-

given. Number of records

of quality, and predicting quality levels for a given set of input

parameters. Such quality analyses provide invaluable feedback to

for QI. Most common DM applications are expected to occur in

such activities due to ease of collecting relevant data, performing

sidered in this review. These activities as well as the reasons of

When it comes to controlling quality of products, the oldest ap-

proach is to inspect all units produced and discard/repair the

study, since smart inspection systems utilizing DM techniques

form a mature area that deserves a special attention on its own.

For example, a branch of chemistry, called chemometrics, has

grown into a well-established data collection and analysis method

in areas such as multivariate calibration and pattern recognition in

the chemical process industry (Yang & Trewn, 2004). In general,

tional image processing functions have been employed, however,

Pham and Alcock (2003) provide a good coverage of many of these

are quality monitoring and process control. Process industries and

these activities that aim to reduce variability. While quality mon-

itoring tries to reduce variability by detection and removal of

assignable causes, process control is based on the idea of process

compensation and regulation to reduce variability. Although they

Design type QI applications

Data to be cleaned Techniques QI applications

Data reduction Methods QI applications

Cases and methods QI applications

Fig. 1. Number of the application papers observed in the literature.

cess. For a robust design of very-large scale integration (VLSI)

Type of software Name of software Used by Used for

ARTNN adaptive resonance theory neural MIG modified breath-first search of an

You might also like