Proceedings of the International MultiConference of Engineers and Computer Scientists 2011 Vol I,

IMECS 2011, March 16 - 18, 2011, Hong Kong

Feature Selection and Boosting Techniques to

Improve Fault Detection Accuracy in the
Semiconductor Manufacturing Process
Kittisak Kerdprasop and Nittaya Kerdprasop, Member, IAENG

 integrated circuit inside a package to form a ready-to-use

Abstract—Accurate and timely detection of any faults in the product, and testing of the finished products. A constant
semiconductor manufacturing process is an important issue for advancement in the semiconductor industry is due mainly to
process control engineers to ensure both productivity and persistent improvement of the wafer fabrication process.
reliability. Fault detection is a major step of process control
aiming at constructing a decision tool to help detecting as
The fabrication process consists of a series of steps to cover
quickly as possible any equipment or process faults in order to special material layers over the wafer surface. Wafers
maintain high process yields in manufacturing. Traditional reenter the same processing machines as each layer is
statistical based techniques such as univariate and multivariate successively covered. Some defects in this complicated
analyses have long been employed as a tool for creating model process can make the final products fail the test. Fault
to detect faults. Unfortunately, modern semiconductor detection and classification techniques [2], [3], [5-7], [11-
industries have the ability to produce measurement data
collected directly from sensors during the production process
15] applied to this critical manufacturing process can
and such highly voluminous data are beyond the capability of obviously improve product quality and reliability.
traditional process control method to detect fault in a timely In recent years, many manufacturing tools are equipped
manner. We thus propose the techniques based on the data with sensors to facilitate real-time monitoring of the
mining technology to automatically generate an accurate production process. These tool-state and production-state
model to predict faults during the wafer fabrication process of sensor data provide an opportunity for efficient control and
the semiconductor industries. In such process control context,
optimization. Unfortunately, such measurement data are so
the measurement data contain over 500 signals or features. The
feature selection technique is therefore a necessary tool to overwhelming that timely detection of any fault during the
extract the most potential features. Besides the feature production process is difficult. In this paper, we study the
selection method, we also propose a boosting technique to problem of accurate detection of fault states in the wafer
handle the imbalance situation of fail versus pass test cases. fabrication process. The dataset is donated by McCann et al
The experimental results support our assumption that choosing [10] and publicly available for re-experimentation.
the right features and boosting rare cases can considerably
improve detection accuracy of fault products and processes.
II. RELATED WORK
Index Terms—fault detection model, semiconductor Process control is crucially important to the
manufacturing process, feature selection, rare case boosting semiconductor industries that operate the multistage
manufacturing systems on the product scale of lesser 300
nanometers [12]. Modern technology in semiconductor
I. INTRODUCTION manufacturing enables real time process control with the

S EMICONDUCTOR manufacturing is a highly complex

production process composed of hundreds of steps. The
major processes in most semiconductor industries [1], [8]
measurement data obtained from the equipment sensors and
the final electrical test. With such high volume of data
recorded during the entire production process, effective
are in the following sequence: production of silicon wafers monitoring and optimal process control by investigating and
from pure silicon material, fabrication of integrated circuits analyzing these data are difficult work for process
onto the raw silicon wafers, assembly by putting the engineers. Traditional process control methodology like
univariate and multivariate control charts is no longer an
efficient method to control manufacturing systems with
Manuscript received December 8, 2010; revised January 20, 2011. This
hundreds of processing stages. Instead automatic and
work was supported in part by grants from the National Research Council
of Thailand, the Thailand Research Fund, and Suranaree University of advanced process control method is required.
Technology (via the Data Engineering and Knowledge Discovery (DEKD) Ison and colleagues [6], [7] proposed a decision tree
Research Unit). classification model to detect fault of plasma etch
Kittisak Kerdprasop is the director of DEKD research unit and the
associated professor of the school of computer engineering, Suranaree equipment. The model was built from the five sensor signal
University of Technology, 111 University Avenue, Nakhon Ratchasima data. Many researchers also studied the fault detection
30000, Thailand (phone: +66-(0)44-224349; fax: +66-(0)44-224602; e- problem during the etch process. Goodlin et al [3] proposed
mail: [email protected], [email protected]).
Nittaya Kerdprasop is the co-founder and principal researcher of the to build a specific control chart for detecting specific type of
DEKD research unit. She is an associate professor at the school of faults. They collected tool-state data directly from the
computer engineering, Suranaree University of Technology, Thailand (e- etcher. These data consist of 19 variables. The work of
mail: [email protected]).

ISBN: 978-988-18210-3-4 IMECS 2011

ISSN: 2078-0958 (Print); ISSN: 2078-0966 (Online)
Proceedings of the International MultiConference of Engineers and Computer Scientists 2011 Vol I,
IMECS 2011, March 16 - 18, 2011, Hong Kong

Spitzlsperger and colleagues [11] was also based on the

statistical method. They adopted the multivariate control
chart method to maintain changes in the mean and standard
deviation coefficients by remodeling technique.
Recent interest in fault detection has been shifted toward
the non-parametric approaches. He and Wang [5] proposed
to use the k-nearest neighbor rule for fault detection.
Verdier and Ferreira [14], [15] also applied the k-nearest
neighbor method, but they proposed to use the adaptive
Mahalanobis distance instead of the traditional Euclidean
distance. Tafazzoli and Saif [13] proposed a combined
support vector machine methodology for process fault
diagnosis. Ge and Song [2] applied support vector data to
the principal component analysis method to detect process
abnormalities. Fig. 1 Framework of proposed method and research study
Most work on fault detection methods has studied the
process control problem with a few features of tool-state The dataset is in a form of matrix; rows represent each
and process-state measurement data. McCann and his team observation or instance and columns represent features
[9] proposed a rather different setting in which the which are values recorded from each sensor. The steps in
measurement data from the wafer fabrication process our proposed method for creating an accurate model to
contain as much as 590 features. With such abundant detect fault case from highly imbalance data with numerous
features or variables, feature selection techniques [4] are features are as follows:
obviously necessary in order to improve both the prediction
and the computational performances. Data Cleaning Phase
In this paper, we also analyze the wafer fabrication data (1) Investigate data observed from each sensor, i.e. data
[10] collected from 590 sensors with the last feature is a in each column. If the data appear to be a single
label stating pass or fail state. The observed data contain value, then remove that feature.
1,463 pass cases with only 104 fail cases. In this work not (2) Count in each column the „not available‟ or missing
only a feature selection method for extracting the post values. If data are missing more than 55%, then
discriminative sensors is proposed, but also a boosting remove that feature.
technique is devised to deal with highly imbalance between Feature Selection Phase
the pass and fail cases. (3) Apply two statistical based feature selection
techniques: chi-square and principal component
III. FAULT DETECTION TECHNIQUE analysis (PCA), and save the result as two separate
The SECOM dataset [10] contains 1567 examples taken datasets.
from a wafer fabrication production line. Each example is a (4) Apply an information theoretical based technique:
vector of 590 sensor measurements plus a label of pass/fail gain ratio, and save the result in a separate dataset.
test. Among the 1567 examples, there are only 104 fail (5) Apply the following cluster-based feature selection
cases which are labeled as positive (encoded as 1), whereas technique, called MeanDiff:
much larger amount of examples pass the test and are (5.1) Clustering data into two clusters (fail cluster
labeled as negative (encoded as -1). The imbalance of pass and pass cluster)
and fail examples in addition to the large number of (5.2) Compare value differences in every feature of
metrology data obtained from hundreds of sensors make this the fail cluster mean and the pass cluster mean
dataset a difficult one to accurately analyze. It is thus our (5.3) Ranking features in descending order according
main focus to devise a method based on data mining to the magnitude of mean differences
techniques to build an accurate model for fault detection. computed in step 5.2, and output the ranked
The framework of our study is presented in Fig.1. features
Feature selection techniques in our study are ranging Case Boosting Phase
from simply removing features with a constant value and (6) Separate data obtained from step 2 into two datasets:
features containing numerous missing values (more than train data and test data. Each data set maintains the
55% of values are missing), to statistical based analysis such same proportion of pass and fail cases.
as chi-square and principal component analysis (PCA) and (7) Pumping the fail cases in the train data by duplicating
information theoretical based such as gain ratio. We also the fail cases to be the same amount as the pass
devise a cluster based technique call MeanDiff to analyze cases.
discrimination power of each feature. On the model building Model Building Phase
phase, we apply four methods to induce the fault-detection (8) Build a prediction model with decision tree, naïve
model namely decision tree, naïve Bayes, logistic Bayes, k-nearest neighbor, and logistic regression
regression, and k-nearest neighbor. algorithms.
(9) For datasets from steps 3-5, evaluate model accuracy

ISBN: 978-988-18210-3-4 IMECS 2011

ISSN: 2078-0958 (Print); ISSN: 2078-0966 (Online)
Proceedings of the International MultiConference of Engineers and Computer Scientists 2011 Vol I,
IMECS 2011, March 16 - 18, 2011, Hong Kong

with 10-fold cross validation technique. Dataset from

step 7 is evaluated with the test set.

We assess the model performance based on the four

metrics: true positive rate (TP rate or recall), precision, F-
measure, and false positive rate (FP rate or false alarm). The
computation methods of these metrics are given in Fig. 2
(TP = true positive, FP = false positive, FN = false negative,
TN = true negative).
Predicted class
TP
Class=1 Class= -1 TP rate, or Recall 
TP  FN
(fail) (pass)
TP
Precision 
Actual class

Class= 1 TP FN TP  FP Fig. 3 TP rate of fault detection models on different feature

2TP selection methods
F - measure 
2TP  FP  FN
Class= -1 FP TN FP
FP rate 
FP  TN

Fig.2 Fail/pass classification matrix and performance computation

IV. EXPERIMENTAL RESULTS

A. Feature Selection Technique Comparison

We use the WEKA software [16] to perform a series of
experiments. The first part of our study aims at selecting
principal features that show the most discrimination power
of differentiating fail cases from pass cases. In the cleaning
step, we remove 137 features that contain a single value and Fig. 4 Precision of fault detection models on different feature
selection methods
lots of missing values. From the remaining 454 features, we
select the best 168 features (to maintain around 95% of
variances) by means of principal component analysis
(PCA), Chi-square test, gain ratio computation, and our own
MeanDiff method. The fault detection models are then
derived from each feature selected data. We want the model
that shows the highest values of TP rate, precision, and F-
measure, but the lowest value in FP rate. The experimental
results on the four model measurement metrics are shown in
Figs. 3-6.
For this specific data domain, it can be noticed that
feature selection can considerably improve the accuracy of
fault detection models. The proposed MeanDiff method
contributes the most to decision tree model, whereas the Fig. 5 F-measure of fault detection models on different feature
gain ratio method is the best feature selection method for the selection methods
naive Bayes and logistic regression model building
approaches. The k-nearest neighbor method (in which k was
set to be one on our experiments because it yields the best
result) needs a cleaned dataset without any other feature
selection facility. If model comprehensibility is a major
concern, the model built from a MeanDiff feature selected
data with a decision tree approach is the most appropriate
one. It is worth mentioning here that for such a large
number of features like this application the neural network
and support vector machine approaches consume so much
memory that they cannot run to completion. Among the four
model building methods, naïve Bayes model can detect fault
cases at the success rate as high as 90%, but the false alarm
(FP rate) is also as high as 80% as well. We compare the TP Fig. 6 FP rate of fault detection models on different feature
rate versus the FP rate of each model and provide the result selection methods
in Fig. 7.

ISBN: 978-988-18210-3-4 IMECS 2011

ISSN: 2078-0958 (Print); ISSN: 2078-0966 (Online)
Proceedings of the International MultiConference of Engineers and Computer Scientists 2011 Vol I,
IMECS 2011, March 16 - 18, 2011, Hong Kong

propose the idea of separating the SECOM dataset into a

train set and a test set. The test set contains 468 instances in
which 59 instances are fail test and 409 are pass test. The
train set contains 45 instances of fail test and 1054 of pass
test. We then duplicate the number of fail test in the training
data to be 1096 instances. The fault detection models are
built from this rare case boosting training dataset. The
models are then evaluated their classification performances
by the separated test dataset. The classification error
matrices of models built from the four different learning
methods are given in Fig. 8 and the performance criteria are
summarized in Table 1. The boosted true positive rate,
precision, F-measure, and the lower false positive rate of
each model are also graphically provided in Fig. 9.

k-Nearest Neighbor Predicted class

Class=1 (fail) Class= -1 (pass)

Actual class
Class= 1 58 1

Class= -1 98 311

Logistic regression Predicted class

Class=1 (fail) Class= -1 (pass)

Actual class

Class= 1 59 0

Class= -1 137 272

Naïve Bayes Predicted class

Class=1 (fail) Class= -1 (pass)

Actual class

Class= 1 44 15

Class= -1 144 265

Decision Tree Predicted class

Class=1 (fail) Class= -1 (pass)

Actual class

Class= 1 59 0

Class= -1 66 343

Fig. 8 Classification error matrices of fault detection models

TABLE I
FAULT DETECTION MODEL ASSESSMENT
Fig. 7 TP rate versus FP rate comparison of each fault detection
model k-Nearest Logistic Naïve Decision
Neighbor Regression Bayes Tree

B. Rare Case Boosting Results TP rate 0.983 1.0 0.746 1.0

For the specific problem of fault detection, the number of FP rate 0.24 0.335 0.352 0.161
fail test is very few (104 instances in the SECOM dataset)
Precision 0.372 0.301 0.234 0.472
comparative to the number of pass test (1463 instances). It
is therefore a difficult task to build automatically the F-measure 0.54 0.463 0.356 0.641
accurate model that can detect such rare cases. We thus

ISBN: 978-988-18210-3-4 IMECS 2011

ISSN: 2078-0958 (Print); ISSN: 2078-0966 (Online)
Proceedings of the International MultiConference of Engineers and Computer Scientists 2011 Vol I,
IMECS 2011, March 16 - 18, 2011, Hong Kong

The high true positive but low false positive of the

decision tree model make it a good candidate for automatic
generation of the fault detection model to be used in the
semiconductor manufacturing process. The fault detection
model in a form of decision tree is given in Fig.10. The top
level of the decision tree is on the left hand side in which
the value from sensor number 511 is the first parameter to
be considered. The normal state (encoded as -1) is expected
if the value of sensor 511 is less than or equal 28.3784. The
fault state is to be detected when the following sensor values
are reported: S511 > 28.3784, S470 > 4.3751, S16 >
401.1307, S472 > 4.4751, S51 ≤ 646.9073, S4 > 905.1501,
S188 > 11.54, S431 > 3.8926, S439 > 28.6219, S495 >
1.3638 S56 > 2875, S548 > 398.552, S178 ≤ 0.448, S29 ≤
73.4556, S578 ≤ 16.4303, S474 ≤ 27.9511, and S39 ≤
86.3506. Other prediction rules can be interpreted in the
same manner.

S511 <= 28.3784: Predict -1

S511 > 28.3784
| S470 <= 4.3751: Predict -1
| S470 > 4.3751
| | S16 <= 423.3311
| | | S16 <= 401.1307: Predict -1
| | | S16 > 401.1307
| | | | S472 <= 4.4751: Predict -1
| | | | S472 > 4.4751
| | | | | S51 <= 646.9073
| | | | | | S4 <= 905.1501: Predict -1
| | | | | | S4 > 905.1501
| | | | | | | S188 <= 11.54: Predict -1
| | | | | | | S188 > 11.54
| | | | | | | | S431 <= 3.8926: Predict -1
| | | | | | | | S431 > 3.8926
| | | | | | | | | S439 <= 28.6219: Predict -1
| | | | | | | | | S439 > 28.6219
| | | | | | | | | | S495 <= 1.3638
| | | | | | | | | | | S56 <= 2875
| | | | | | | | | | | | S548 <= 398.552: Predict -1
| | | | | | | | | | | | S548 > 398.552
| | | | | | | | | | | | | S178 <= 0.448
| | | | | | | | | | | | | | S29 <= 73.4556
| | | | | | | | | | | | | | | S578 <= 16.4303
| | | | | | | | | | | | | | | | S474 <= 27.9511
| | | | | | | | | | | | | | | | | S39 <= 86.3506: Predict 1
| | | | | | | | | | | | | | | | | S39 > 86.3506: Predict -1
| | | | | | | | | | | | | | | | S474 > 27.9511: Predict 1
| | | | | | | | | | | | | | | S578 > 16.4303
| | | | | | | | | | | | | | | | S161 <= 614: Predict -1
| | | | | | | | | | | | | | | | S161 > 614: Predict 1
| | | | | | | | | | | | | | S29 > 73.4556
| | | | | | | | | | | | | | | S414 <= 25.0931: Predict -1
| | | | | | | | | | | | | | | S414 > 25.0931: Predict 1
| | | | | | | | | | | | | S178 > 0.448
| | | | | | | | | | | | | | S273 <= 19.8922: Predict 1
| | | | | | | | | | | | | | S273 > 19.8922: Predict -1
| | | | | | | | | | | S56 > 2875
| | | | | | | | | | | | S28 <= 7.373: Predict -1
| | | | | | | | | | | | S28 > 7.373: Predict 1
Fig. 9 High increases in TP rate, Precision, and F-measure but low | | | | | | | | | | S495 > 1.3638: Predict -1
increase in FP rate of fault detection models from applying the | | | | | S51 > 646.9073: Predict -1
boosting technique | | S16 > 423.3311: Predict -1

Fig. 10 Decision tree model for fault-detection in the

semiconductor process control

ISBN: 978-988-18210-3-4 IMECS 2011

ISSN: 2078-0958 (Print); ISSN: 2078-0966 (Online)
Proceedings of the International MultiConference of Engineers and Computer Scientists 2011 Vol I,
IMECS 2011, March 16 - 18, 2011, Hong Kong

V. CONCLUSION [12] S.J. Qin, G. Cherry, R. Good, J. Wang, and C.A. Harrison,
“Semiconductor manufacturing process control and monitoring: a fab-
In semiconductor manufacturing process control and wide framework,” Journal of Process Control, vol.16, pp.179-191,
monitoring, hundreds of metrology data are available for 2006.
[13] E. Tafazzoli and M. Saif, “Application of combined support vector
process engineers to analyze for the purpose of maintaining machines in process fault diagnosis,” in Proc. American Control
efficient operations and getting optimum yield of high Conf., St. Louis, USA, 10-12 June 2009, pp.3429-3433.
quality products. For such a large volume of measurement [14] G. Verdier and A. Ferreira, “Fault detection with an adaptive distance
data, automatic fault detection technique is essential. We for the k-nearest neighbor rule,” in Proc. Int. Conf. Computer &
Industrial Engineering, Troyes, France, 6-9 July 2009, pp.1273-1278.
thus investigate the application of data mining techniques [15] G. Verdier and A. Ferreira. (2010). Adaptive Mahalanobis distance
such as decision tree induction, naïve Bayes analysis, and k-nearest neighbor rule for fault detection in semiconductor
logistic regression, and k-nearest neighbor classification for manufacturing. IEEE Trans. Semiconductor Manufacturing.
Available: doi:10.1109/TSM.2010.2065531
creating an accurate model for fault case detection in the [16] WEKA. (2010, October 23). Available: http://www.cs.waikato.ac.nz/
wafer fabrication process of semiconductor industries. ml/weka/
From a series of experimentation, we found that naïve
Bayes model built from a subset of features selected by a
gain ration criteria can detect the fault cases at the very high
rate of 90%. But the false alarm rate, or false positive, is
also as high as 80%. The decision tree method built from
our MeanDiff feature selection method generates a more
comprehensible form of fault detection model with false
alarm rate at only 4.5%. But the precision and true positive
rate, or recall, of the tree model are still low at 20.5% and
16%, respectively.
We thus devise a boosting technique to improve the
precision of tree-based model for fault detection by
pumping the number of rare cases, or fault test, to the equal
number of majority cases, or pass test. The outcome is
surprising that the true positive rate of the tree-based model
can increase up to 100%, whereas the false alarm rate is still
low at the 16%. We plan to investigate this boosting
technique to other domains that show imbalance among data
classes in our future research.

REFERENCES
[1] E. Epistola. (2010, December 4). Semiconductor manufacturing.
Available: http://www.siliconfareast.com/manufacturing.htm
[2] Z. Ge and Z. Song, “Semiconductor manufacturing process
monitoring based on adaptive substatistical PCA,” IEEE Trans.
Semiconductor Manufacturing, vol.23, no.1, pp.99-108, Feb. 2010.
[3] B.E. Goodlin, D.S. Boning, H.H. Sawin, and B.M. Wise,
“Simultaneous fault detection and classification for semiconductor
manufacturing tools,” Journal of The Electrochemical Society,
vol.150, no.12, pp.G778-G784, 2003.
[4] I. Guyon and A. Elisseeff, “An introduction to variable and feature
selection,” Journal of Machine Learning Research, vol.3, pp.1157-
1182, 2003.
[5] Q.P. He and J. Wang, “Fault detection using the k-nearest neighbor
rule for semiconductor manufacturing processes,” IEEE Trans.
Semiconductor Manufacturing, vol.20, no.4, pp.345-354, Nov. 2007.
[6] A.M. Ison, W. Li, and C.J. Spanos, “Fault diagnosis of plasma etch
equipment,” in Proc. IEEE Int. Symp. Semiconductor Manufacturing,
San Francisco, 6-8 October 1997, pp.B-49-B-52.
[7] A.M. Ison and C. Spanos, “Robust fault detection and fault
classification of semiconductor manufacturing equipment,” in Proc.
Int. Symp. Semiconductor Manufacturing, Tokyo, 2-4 October 1996.
[8] G.S. May and C.J. Spanos, Fundamentals of Semiconductor
Manufacturing and Process Control. John Wiley & Sons, 2006.
[9] M. McCann, Y. Li, L. Maguire, and A. Johnston, “Causality
challenge: benchmarking relevant signal components for effective
monitoring and process control,” in Proc. JMLR Workshop, Canada,
12 December 2008, pp.277-288.
[10] SEmi COnductor Manufacturing. (2010, November 25). Available:
http://www.causality.inf.ethz.ch/repository.php
[11] G. Spitzlsperger, C. Schmidt, G. Ernst, H. Strasser, and M. Speil,
“Fault detection for a via etch process using adaptive multivariate
methods,” IEEE Trans. Semiconductor Manufacturing, vol.18, no.4,
pp.528-533, Nov. 2005.

ISBN: 978-988-18210-3-4 IMECS 2011

ISSN: 2078-0958 (Print); ISSN: 2078-0966 (Online)