IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 54, NO.

1, FEBRUARY 2005 297

A Gray System Modeling Approach to the Prediction

of Calibration Intervals
Kuo-Huang Lin, Member, IEEE, and Bin-Da Liu, Senior Member, IEEE

Abstract—This paper discusses a class of data-preprocessed sta- stochastic model that can be tailored to the given measuring in-
tistical models for evaluating the optimal calibration interval of a strument.
measuring instrument. These models are based on the assumption Useful stochastic methods for data management and time se-
that the calibration status of a measuring instrument can be pre-
dicted using the instrument’s historical calibration data. On the ries analysis exist [3], [4]. In 1991, Stuckman et al. [5] pro-
basis of the gray threshold value prediction method, a series of his- posed a Wiener process model for the calibration drift of an
torical calibration data are preprocessed so that a monotone-in- electrical meter. In 1997, Bobbio et al. [6] proposed a class of
creasing series of data points will be created. Then, the first-order stochastic methods for modeling the drift of one observable pa-
gray model, exponential regression, linear regression, and general rameter of an instrument. However, these statistical approaches
polynomial regression are applied to fit the series of preprocessed
data points to predict the time at which the measured value of the require large amounts of data and need a typical procedure such
instrument will be outside of the allowable tolerance range. The as Gauss distribution or Markov procedure so that percentiles
effectiveness of each developed model was evaluated through the can be used.
actual data collected in a calibration laboratory. Results demon- In system theories, incomplete information for a given system
strate that the gray threshold value prediction based on second- is often called “the black box.” “Black” means that the informa-
order polynomial model, a modified autoregressive model, is the
best method for forecasting the calibration interval of a measuring tion is not entirely available. Conversely, “white” means that the
instrument. required information is complete. The meaning of “gray” can be
expressed as the characteristic between black and white. In other
Index Terms—Calibration interval, gray system theory, gray
threshold value prediction method, statistical model. words, the information is poor or uncertain. The purpose of the
“gray” system and its applications is to bridge the gap between
“black” and “white” (i.e., a “gray” system can be called a par-
I. INTRODUCTION tially known and partially unknown system). In the real word, it
is difficult to describe a proper mathematical expression, which
T HE characteristics of a measuring instrument are specified
immediately after calibration. Owing to factors such as
aging, mechanical wear, and long-term environmental changes,
becomes the model to express the dynamic behavior of a com-
plex system. The focal point of the “gray” system theory in-
the characteristics of a measuring instrument change over a pe- volves using a finite amount of available information to build
riod of time. A reliable estimation approach for determining the a “gray” model (GM) in order to approximate the dynamic be-
optimized calibration intervals (the time between two consec- havior of a system [7]. The “gray” system theory was pioneered
utive calibrations made by a measuring instrument) is a very by Deng in 1982 [8]. The modeling techniques are characterized
crucial problem in any quality assurance program. However, by by the following.
international standards, there is a lack of well-established and • The original data series does not require a probability dis-
recommended methods of determining the calibration intervals tribution.
for a measuring instrument [1]. The methods surveyed in [1] are • A minimum of three data points are required for modeling.
mostly of a statistical nature and can be applied to the same cate- • Little computational effort is needed to constitute the
gory of measuring instruments. In an alternative approach based model.
upon calibration drift, measurement nonlinearity, and measure- • The accumulating generation is provided to adapt system
ment uncertainty, the manufacturers of electrical instruments dynamic behavior.
specify some recommended calibration intervals. In effect, this The gray prediction, the most important section in the gray
mode of specification tends to be quite conservative and cannot system theory, is based on some treatments of the original data
be considered a real operation environment for the measuring and the establishments of the gray model to discover and to con-
system [2]. Therefore, it is practical to resort to a satisfactory trol the development laws of the system. The gray prediction
method has been applied broadly in academic fields, and all of
the resulting research has reached good results [9]–[14]. Five
Manuscript received October 30, 2003; revised May 15, 2004. kinds of gray prediction methods have been proposed for their
K.-H. Lin was with the Department of Electrical Engineering, National Cheng respective application [15], from which the gray threshold value
Kung University, Tainan, Taiwan 701, R.O.C. He is now with the Department of
Electrical Engineering, National Huwei University of Science and Technology, prediction was adopted for our prediction problem.
Huwei, Taiwan 632, R.O.C. As mentioned in [5] and [6], this paper also addresses the
B.-D. Liu is with the Department of Electrical Engineering, National problem of modeling calibration drift in a measuring instru-
Cheng Kung University, Tainan, Taiwan 701, R.O.C. (e-mail: [email protected].
ncku.edu.tw). ment. A class of data-preprocessed statistical models will be
Digital Object Identifier 10.1109/TIM.2004.840234 presented to predict the time when the measuring instrument
0018-9456/$20.00 © 2005 IEEE

Authorized licensed use limited to: National Cheng Kung University. Downloaded on October 16, 2008 at 02:43 from IEEE Xplore. Restrictions apply.
298 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 54, NO. 1, FEBRUARY 2005

needs to be recalibrated. The modeling process starts with the solution for the first-order differential equation is also in the ex-
gray threshold value prediction method adopted to preprocess ponential form, thus, many workers in system research are in-
the series of historical measuring instrument calibration records. terested in differential equations with the belief that differential
Four statistical methods, i.e., the first-order gray model, expo- equations deeply describe the essence of system development.
nential regression, linear regression, and polynomial regression The concept of gray derivatives is introduced so that we can
are then applied to fit the series of preprocessed data. Each fitting establish models similar to differential equations for sequences
equation is used to predict the calibration interval for the mea- of discrete data. Thus, an ordinary first-order differential equa-
suring instrument. By using the actual data from St. John’s and tion can be taken into consideration to approximate the behavior
St. Mary’s Institute of Technology (SJSMIT) Calibration Lab- of . According to GM(1, 1), we can form the following
oratory, three examples are given to demonstrate the validity of first-order gray differential equation:
fitting error and forecasting error for each modeling technique.
(5)
II. GRAY THRESHOLD VALUE PREDICTION BASED ON
STATISTICAL MODELS where the parameters and in the GM(1, 1) are, respectively,
In this section, we will discuss the gray modeling technique, called the development coefficient and gray action quantity. By
gray threshold value prediction method, and curve-fitting anal- solving the differential equation, we can get the prediction func-
ysis applied in this study. tion for the gray system. The sequence given by

A. Gray Modeling Technique

In general, the gray model is based on GM , where is (6)
the order of the differential equation, and is the number of vari- is said to be the time response sequence of GM(1, 1), where
ables. In a real control environment, GM(1, 1) is the most com- denotes the prediction of , and the optimal and are given
monly used model for studies and applications, where GM(1, 1) by
denotes a kind of single-variable sequence and first-order linear
gray model. In this section, some detailed original gray model (7)
operations that are needed for building a GM(1, 1) model are
described [16]. where
Assume that the original series of data with samples is ex-
pressed as (8)
and
(1)

in which the superscription (0) of represents that it is the .. .. (9)

original data series. After the first-order accumulated generating . .
operation (AGO) of , we obtain the sequence as

(2) Hence, is given as

(10)
where
Summarizing the above descriptions, Appendix A shows an ex-
(3) ample of the description for steps of building the GM(1, 1)
model.
After the process of accumulating generation was applied, the B. Gray Threshold Value Prediction Method
inverse accumulated generating operation (IAGO) plays the role
of returning the data to the original condition. So, AGO, and The gray threshold value prediction is appropriate for the pre-
IAGO are a pair of inverse sequence operators. The operation diction of a “disaster” that may occur in a certain period of time.
of IAGO for the first-order series is defined as follows: The task of the disaster prediction is to pinpoint the time mo-
ment(s) for one or several threshold values to occur so that rel-
and evant parties can have enough time to do preparations for disas-
for ters to come [15]. Assume that a time series data expressed as
(4) (11)
Obviously, for a non-negative original data series , the is a sequence of raw data. For a given upper threshold value ,
exhibits the increased data series and becomes more reg- the subsequence of
ular than . Suppose that we treat the accumulated result
as an exponentially behaved increasing energy, the tendency of
can be approximated using an exponential function. The (12)

Authorized licensed use limited to: National Cheng Kung University. Downloaded on October 16, 2008 at 02:43 from IEEE Xplore. Restrictions apply.
LIN AND LIU: A GRAY SYSTEM MODELING APPROACH TO THE PREDICTION OF CALIBRATION INTERVALS 299

is called an upper threshold value sequence. Relatively, for a

given lower threshold value , the subsequence of

(13)

is called a lower threshold value sequence.

Since different threshold value sequences require different
approaches to handle related details, we will not distinguish be-
tween the upper and lower threshold value sequences in the fol-
lowing discussions. Then

(14)

is called a threshold data sequence.

By using the relevant threshold data sequence to establish the
GM(1, 1) model, the date of disaster occurrences can be pre- Fig. 1. Actual values of HP3478A and gray threshold value prediction method.
dicted.

C. Curve-fitting
Curve-fitting is a technique for extracting a set of curve pa-
rameters from the data series to obtain a functional description
of the data series. The algorithm that fits a curve to a particular
data series is known as the least squared method. The curve-fit-
ting error is defined as

(15)

where is the functional description of the data series,

is the observed data series, and is the vector of parameters
which best describes the curve.
The least squared algorithm finds by solving a system of
Fig. 2. GM(1, 1) prediction model in experiment one.
algebraic equations

(16) given by the standard value multiplied by (1-uncertainty).

Each threshold value yields a straight line in the graph. In
We can then obtain an estimation of the observed data series for Fig. 1, the upper threshold value is 10.00100 V, and the lower
any value of using the functional description . threshold value is 9.99900 V. The monthly points index cor-
In LabVIEW, the curve-fitting virtual instrument (VI) is auto- responding to a calibration value greater than 10.00100 V are
matically set up, solves the Jacobian system, and returns the set . Otherwise, there is an empty set for the
of coefficients that best describes the data series. The following calibration value less than 9.99900 V.
curve-fitting VI’s are available: 1) exponential fit; 2) linear fit; Among the seven data points , as given in
and 3) general polynomial fit. Fig. 1, the front six points are used to develop
the statistical model, while the last point (12) is used for testing
III. NUMERICAL TESTS AND DISCUSSIONS the forecasting error. The mean squared error (MSE) is applied
to estimate the fitting error and forecasting error for each sta-
In this section, three practical experiments are given with tistical method. The MSE is a relative measure of the residuals
comparisons to demonstrate the validity of the fitting and fore- between the curve fitted data series and the actual observed data
casting based on the gray threshold value prediction method. series using the formula
The historical calibration records are collected in SJSMIT Cal-
ibration Laboratory from June 1999 to May 2000.
MSE (17)
A. Experiment 1: HP3478A
To investigate the practicability of the gray threshold value where represents values of the function that approximates
prediction approach, we collected the historical calibration data, represents the observed data, and is the number of
records from a HP3478A multimeter in SJSMIT Calibration sample points observed.
Laboratory. The actual calibration values and the corresponding Each of the above-mentioned statistical methods was simu-
piecewise linear curve is shown in Fig. 1. Let be the upper lated and the results were plotted in Figs. 2 to 7. The fitting and
threshold value given by the standard value multiplied by forecasting error caused by each model are both listed in Table I
uncertainty , and let be the lower threshold value for comparison. It was observed that the fitting MSE for each

Authorized licensed use limited to: National Cheng Kung University. Downloaded on October 16, 2008 at 02:43 from IEEE Xplore. Restrictions apply.
300 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 54, NO. 1, FEBRUARY 2005

Fig. 3. Exponential model in experiment one. Fig. 6. Third-order polynomial model in experiment one.

Fig. 4. Linear fit model in experiment one. Fig. 7. Fourth-order polynomial model in experiment one.

TABLE I
FITTING MSE AND FORECASTING MSE OF HP3478A MULTIMETER FOR
DIFFERENT MODELS

lower threshold value is 9.99800 V. The corresponding piece-

Fig. 5. Second-order polynomial model in experiment one. wise linear curve is shown in Fig. 8. The monthly points index
corresponding to a calibration value less than 9.99800 V are
, but there is an empty set for the calibration
model was distributed in the same order of . However, the
value greater than 10.00200 V.
difference in forecasting error is obvious, and the second-order
Among the six data points , the front five
polynomial model is the best one for forecasting. The practical
points are used to develop the statistical model
application is shown in Fig. 8. It demonstrates that the next pre-
while the last point (12) is used for testing the forecasting error.
diction point is 13.14, which means that the instrument needs to
Each of the above-mentioned statistical methods was simulated
be recalibrated in June 2000.
and the results plotted in Fig. 10. The fitting and forecasting
error caused by each model are both listed in Table II. It was
B. Experiment 2: Fluke 45 observed that, except the fourth-order polynomial model, the fit-
The actual calibration values from a Fluke 45 multimeter are ting MSE for other models were distributed between the order of
listed in Fig. 9. According to the same definitions as in the ex- and . However, the second-order polynomial model
periment 1, the upper threshold value is 10.00200 V, and the has excellent accuracy for forecasting.

Authorized licensed use limited to: National Cheng Kung University. Downloaded on October 16, 2008 at 02:43 from IEEE Xplore. Restrictions apply.
LIN AND LIU: A GRAY SYSTEM MODELING APPROACH TO THE PREDICTION OF CALIBRATION INTERVALS 301

TABLE II
FITTING MSE AND FORECASTING MSE OF FLUKE 45 MULTIMETER FOR
DIFFERENT MODELS

Fig. 8. gray threshold value prediction based on second-order polynomial

model in experiment one.

Fig. 11. Actual values of MT800 and gray threshold value prediction method.

Fig. 9. Actual values of Fluke 45 and gray threshold value prediction method.

Fig. 12. Fitting and forecasting curve for different prediction models in
experiment three.

corresponding to a calibration value greater than 10.0060 V

are , but there is an empty set for the calibration
value less than 9.9940 V.
Among the five data points , the front four
points are used to develop the statistical model
while the last point (11) is used for testing the forecasting error.
Each of the above-mentioned statistical methods was simulated
Fig. 10. Fitting and forecasting curve for different prediction models in and the results plotted in Fig. 12. The fitting and forecasting
experiment two.
error caused by each model are both listed in Table III. It was
observed that, except the fourth-order polynomial model which
C. Experiment 3: MT800 was not a number (NaN) and the third-order polynomial model
The actual calibration values and the corresponding piece- was fitted completely, the fitting MSE for other models were
wise linear curve from a MT800 multimeter is shown in distributed between the order of and . However, the
Fig. 11, where the upper threshold value is 10.0060 V, and the second-order polynomial model still has the best accuracy for
lower threshold value is 9.9940 V. The monthly points index forecasting.

Authorized licensed use limited to: National Cheng Kung University. Downloaded on October 16, 2008 at 02:43 from IEEE Xplore. Restrictions apply.
302 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 54, NO. 1, FEBRUARY 2005

TABLE III TABLE IV

FITTING MSE AND FORECASTING MSE OF MT800 MULTIMETER FOR SUITABILITY OF SECOND-ORDER POLYNOMIAL MODEL SELECTED IN
DIFFERENT MODELS EXPERIMENTS ONE, TWO, AND THREE

D. Discussions
Generally, the distribution of a calibration data series for a
measuring instrument always fluctuates within a certain range.
As the predicted result will follow the trajectory of the pre-
vious data, the fluctuating sequence would make the prediction
unacceptable at the inflection point of the data series. Thus, if
we can remove the fluctuation from the original curve, the pre-
For ease of exposition, the following three basic factors are em-
dicted results would become more accurate [17]. This is why the
ployed to assess the suitability of the second-order polynomial
study began with the gray threshold value prediction method to
models selected in the above-mentioned Experiments One, Two,
convert the fluctuating sequence into a monotone increasing se-
and Three: 1) coefficient of multiple determination; 2) F test;
quence. The focal point then selects a well-fitted model appro-
and 3) confidence intervals (CIs) [20]. The results are listed in
priate for the monotone data points.
Table IV. It is observed that the CIs of the model’s parameters
In general, a higher order polynomial will reduce the devia-
for each experiment are not all in the significant range. The main
tion in the points from the curve until, where the polynomial
reason for this phenomenon is that the sample sizes in Experi-
order equals ( as the number of data points), the
ments One, Two, and Three are small. However, the F test for the
polynomial passes exactly through each point. However, solving
lack of fit listed in the third column indicates that there is a re-
a higher order polynomial equation has an addition difficulty
gression relation between the dependent variable (Index Value)
in undesirable property known as ill-conditioning [18]. Conse-
and the set of independent variables (Index). In addition, the co-
quently, round-off errors cause unusually large errors in the so-
efficients of multiple determination for Experiments One, Two,
lutions. The problem is not too great until or , but be-
and Three are all greater than 0.9. This seems to support the
yond this point, special methods are needed. Otherwise, an at-
suitability of the second-order association between Index Value
tempt to make a polynomial match the data points exactly will
and Index.
cause oscillations. These oscillations become larger as the poly-
nomial order increases. This phenomenon is called the polyno-
mial wiggle problem [19]. Therefore, the solution for a poor IV. CONCLUSION
polynomial fit is not to try higher-order polynomials. In statis-
tics, one increases the order of approximating a polynomial so On the basis of the gray threshold value prediction method,
long as there is a statistically significant decrease in the variance this paper presents a gray threshold value prediction based
, which is computed by on second-order polynomial model so that the date when an
instrument needs to be recalibrated can easily be estimated. By
testing and comparing the proposed method with the existing
(18) techniques that use actual data, the effectiveness and credible
usage of the proposed approach were shown. The possibility
of improving the calibration interval forecasting using gray
where is the deviation squared for the curve-fitting error. It is threshold value prediction based on the second-order polyno-
important to realize that the numerator of (18) should continu- mial prediction technique is, thus, proven.
ally decrease as the order of the polynomial is raised, whereas LabVIEW is a general-purpose programming language, but it
the denominator of (18) makes increase as we go above also includes function libraries and development tools designed
the optimum order. The criterion for (18) chooses the optimum specifically for data acquisition and instrument control. Most
order as two [18]. This encourages us to apply the gray threshold calibration laboratories use LabVIEW to create an automatic
value prediction based on second-order polynomial model to calibration system for measuring instruments. In this study, all
predict the calibration interval. programs were designed by LabVIEW. Therefore, it is easy to
Once a regression model is selected for an application, it is combine the calibration interval prediction function with an ex-
necessary to evaluate the appropriateness for the selected model. isting automatic calibration system.

Authorized licensed use limited to: National Cheng Kung University. Downloaded on October 16, 2008 at 02:43 from IEEE Xplore. Restrictions apply.
LIN AND LIU: A GRAY SYSTEM MODELING APPROACH TO THE PREDICTION OF CALIBRATION INTERVALS 303

APPENDIX A Step 5) Solve for the simulation value of .

EXAMPLE FOR DESCRIBING THE STEPS TO BUILD
A GM(1, 1) MODEL

Assume that (A.11)

Step 6) Restore to find the simulation value of . From

(A.1) (A.12)
is a given sequence of raw data, the following procedures we obtain the simulated sequence
demonstrate how to simulate with GM(1, 1) model.
Step 1) Applying 1-AGO on gives us
(A.13)

(A.2) Comparing the simulated sequence with the given sequence

of raw data, we have the sequence of errors
Step 2) Applying a consecutive neighbor generation to .
Let

(A.3)
(A.14)
Then, it follows that
which can be used to check the feasibility and quality of
GM(1, 1) model.
(A.4)
REFERENCES
Therefore
[1] “Establishment and Adjustment of Calibration Intervals,” NCSL, Nat.
Conf. Standards Labs., Boulder, CO, Tech. Rep. RP1, 1989.
[2] A. S. Morris, Measurement and Calibration for Quality Assurance.
(A.5) Englewood Cliffs, NJ: Prentice-Hall, 1991.
[3] G. E. P. Box, G. M. Jenkins, and G. C. Reinsel, Time Series Analysis.
Englewood Cliffs, NJ: Prentice-Hall, 1994.
so [4] C. R. Rao, Linear Statistical Inference and Its Application, 2nd ed.
New York: Wiley, 1973.
[5] B. E. Stuckman, C. D. Perttunen, J. S. Usher, and B. A. Mclaughlin,
“Stochastic modeling of calibration drift in electrical meters,” in Proc.
Uncertainty Modeling Anal. Symp., 1990, pp. 530–536.
[6] A. Bobbio, P. Tavella, A. Montefusco, and S. Costamagna, “Monitoring
(A.6) the calibration status of a measuring instrument by a stochastic model,”
IEEE Trans. Instrum. Meas., vol. 46, no. 4, pp. 747–751, Aug. 1997.
[7] J. L. Deng, “Introduction to gray system theory,” J. gray Syst., vol. 1, no.
and 1, pp. 1–24, 1989.
[8] , “Control problems of gray systems,” Syst. Contr. Lett., vol. 5, pp.
(A.7) 228–294, 1982.
[9] , “Application of gray system theory in China,” in Proc. Uncertainty
Modeling Anal. Symp., 1990, pp. 285–291.
Step 3) Perform a least squared estimation for a parametric [10] X. G. Long, H. Y. Li, and F. G. Jia, “Analysis and envelope gray predic-
sequence . We can obtain that tion in fruits and forest yields,” J. gray Syst., vol. 5, no. 2, pp. 163–170,
1993.
[11] S. R. Wei and Y. F. Xie, “gray prediction for notifiable diseases,” J. gray
(A.8) Syst., vol. 6, no. 1, pp. 59–68, 1994.
[12] F. Y. Xu, “gray prediction and assessment for the geological structure in
coal mine,” J. gray Syst., vol. 6, no. 1, pp. 239–256, 1994.
Step 4) Determine the model. We have [13] Y. S. Zhao, Z. J. Luo, and R. M. Liu, “gray prediction of accident injury
and death rate,” J. gray Syst., vol. 9, no. 1, pp. 89–93, 1997.
[14] C. C. Wong, W. C. Liang, H. M. Feng, and D. A. Chiang, “gray predic-
(A.9) tion controller design,” J. gray Syst., vol. 3, no. 3, pp. 163–170, 1991.
[15] J. L. Deng, “Five kinds of gray forecasting,” Fuzzy Math., vol. 2, pp.
33–42, 1985.
and the time response sequence [16] S. F. Liu and Y. Lin, An Introduction to gray Systems.. Slippery Rock,
PA: IIGSS Academic, 1998.
[17] Y. P. Huang and T. M. Yu, “The hybrid gray-based models for temper-
ature prediction,” IEEE Trans. Syst., Man, Cybern., vol. 27, no. 2, pp.
284–292, Apr. 1997.
[18] C. F. Gerald and P. O. Wheatley, Applied Numerical Analysis.
Reading, MA: Addison-Wesley, 1989.
[19] M. J. Maron, Numerical Analysis. New York: Macmillan, 1982.
[20] J. Neter, M. H. Kutner, C. J. Nachtsheim, and W. Wasserman, Applied
(A.10) Linear Statistical Models. New York: McGraw-Hill, 1996.

Authorized licensed use limited to: National Cheng Kung University. Downloaded on October 16, 2008 at 02:43 from IEEE Xplore. Restrictions apply.
304 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 54, NO. 1, FEBRUARY 2005

Kuo-Huang Lin (M’04) received the B.S. and M.S.

degrees in industrial education from the National
Taiwan Normal University, Taipei, Taiwan, R.O.C.,
in 1977 and 1984, respectively, and the Ph.D. degree
in electrical engineering from the National Cheng
Kung University, Tainan, Taiwan, in 2004.
He joined the faculty of the National Huwei Uni-
versity of Science and Technology, Huwei, Taiwan,
in 1984. He is currently an Assistant Professor in
the Department of Electrical Engineering, National
Huwei University of Science and Technology. His
research interests include gray system theory, fuzzy-logic controller-chip
design, and SoC design.

Bin-Da Liu (S’79–M’82–SM’95) received the B.S.,

M.S., and Ph.D. degrees in electrical engineering
from the National Cheng Kung University, Tainan,
Taiwan, R.O.C., in 1973, 1975, and 1983, respec-
tively.
From 1975 to 1977, he served as Electrical Officer
in the Combined Service Forces of Taiwan. Since
1977, he has been on the faculty of the National
Cheng Kung University, where he is currently Dis-
tinguished Professor in the Department of Electrical
Engineering and Director of the SoC Research
Center. From 1983 to 1984, he was a Visiting Assistant Professor in the Depart-
ment of Computer Science, University of Illinois at Urbana-Champaign. From
1988 to 1992, he was the Director of Electrical Laboratories, National Cheng
Kung University. He was the Associate Chair of the Electrical Engineering
Department from 1996 to 1999 and the Chair from 1999 to 2002. From 1990
to 1993, he was a member of the Evaluation Committee for Junior Engineering
College, Ministry of Education. Since 1995, he has been a consultant of the
Chip Implementation Center, National Applied Research Laboratories. From
2002 to 2003, he was the Chair of the Digital IP Consortium, VLSI Educational
Program of Ministry of Education. He has been the Coordinator of the SoC
Design Promotion Program of National Science Council since 2002. He has
published more than 210 technical papers. He also contributed chapters in the
book Neural Networks and Systolic Array Design (D. Zhang Ed. Singapore:
World Scientific, 2002) and the book Accuracy Improvements in Linguistic
Fuzzy Modeling (J. Casillas, O. Cordón, F. Herrera, and L. Magdalena Eds.
Heidelberg, Germany: Springer-Verlag, 2003). His current research interests
include low-power circuit design, SoC system integration and verification,
and VLSI implementation for fuzzy-neural networks and audio/video signal
processors.
Dr. Liu is a Board member of Taiwan IC Design Society and a member of
Phi Tau Phi, Taiwan SOC Consortium, International Union of Radio Science,
Chinese Fuzzy Systems Association, and Chinese Institute of Electrical Engi-
neering. He has been the Chair of IEEE Circuits and Systems Society—Taipei
Chapter since 2002. He received the Dragon Distinguished Paper Award from
the Acer Foundation in 1991 and 1997, the Best Paper Award from the CIEE
in 1995 and 2002, the Golden Silicon Award from the Macronix Foundation in
2001, 2002, and 2003, the MPC Chip Design Award from the Chip Implemen-
tation Center in 2002, 2003, and 2004, the Low-Power Design Contest Award
form the ACM/IEEE in 2003, and the Research Award from the National Sci-
ence Council annually since 1988. He organized the Taiwan Student VLSI De-
sign Contest in 1998, 1999, and 2000. Since 1992, he has served as a Member
of the Steering Committee of VLSI Design/CAD Symposium and served as the
General Chair in 1994. He was a member of the Technical Program Committee
of R.O.C. National Conference on Fuzzy Theory and Its Applications in 1997
and 1998. He served as a member of the Technical Program Committee of the
1998 and the 1999 IEEE Workshop on VLSI Signal Processing Systems, the
1998 and the 2000 IEEE Asia Pacific Conference on Circuits and Systems, the
First, Second, and Third IEEE Asia Pacific Conference on ASICs, the 1997,
1999, 2001 and 2003 International Symposium on VLSI Technology, Systems,
and Applications, and the 2003 International Conference on Informatics, Cy-
bernetics, and Systems. He is a member of International Advisory Committee
of the 2003 IEEE International Conference on Neural Networks and Signal Pro-
cessing and has been a member of the International Steering Committee of the
IEEE Asia-Pacific Conference on Circuits and Systems since 2001. He was the
Technical Program Chair of the 2003 Workshop on Consumer Electronics and
the General Co-Chair of the First International Meeting on Microsensors and
Microsystems in 2003. He is the General Chair of the 2004 IEEE Asia-Pacific
Conference on Circuits and Systems. He is currently a CAS Associate Editor of
the IEEE Circuits and Devices Magazine and an Associate Editor of the IEEE
TRANSACTIONS ON CIRCUITS AND SYSTEMS–I.

Authorized licensed use limited to: National Cheng Kung University. Downloaded on October 16, 2008 at 02:43 from IEEE Xplore. Restrictions apply.