Chung-Wen Chen

Research Associate,
Mars Mission Research Center,
North Carolina State University,
Frequency Domain State-Space
Raleigh, NC 27695-7910
System Identification
Jer-Nan Juang
Principal Scientist,
This paper presents an algorithm for identifying state-space models of linear systems
Spacecraft Dynamics Branch, from frequency response data. A matrix-fraction description of the transfer function
NASA Langley Research Center, is employed to curve-fit the frequency response data, using the least-squares method.
Hampton, VA 23665 The parameters of the matrix-fraction representation are then used to construct the
Mem. ASME Markov parameters of the system. Finally, state-space models are obtained through
the Eigensystem Realization Algorithm using the Markov parameters. The main
advantage of this approach is that the curve-fitting and the Markov-parameter-
G. Lee construction are linear problems which avoid the difficulties of non-linear optimi-
Professor, zation of other approaches. Another advantage is that it avoids windowing distor-
Mars Mission Research Center, tions associated with other frequency domain methods.
North Carolina State University,
Raleigh, NC 27695-7921

Introduction
State-space models of dynamic systems are usually required Realization Algorithm (ERA) (Juang and Pappa, 1985) or its
for many current control design methods as these control ap- variant ERA/DC (Juang etal., 1988) is used. The disadvantage
proaches are developed based upon some state-space repre- of this method is that the curve-fitting problem must be solved
sentation of the system. Recently, it has been found that state- either by nonlinear optimization techniques or by linear ap-
space models can be effectively identified through the Ob- proximate algorithms requiring several iterations (Bayard,
server/Kalman Filter System Identification method (OKID) 1992).
using time domain input-output data (Phan et al., 1991, 1992; This paper proposes a simple yet effective way of curve-
Juang et al., 1991; Chen et al., 1992). However, there are cases fitting the FRF data and of constructing the Markov param-
in which frequency response data rather than time histories eters. Instead of using a rational matrix function, this method
are available. This is often the case with the advent of so- uses a matrix-fraction for the curve-fitting. Thus the curve-
phisticated spectrum analyzers and associated automatic test fitting is reformulated as a linear problem which can be solved
equipment. Therefore, the technique of obtaining state-space by the ordinary least-squares method in one step; that is, no
models from frequency response data is of practical interest. iteration is required. The method can match the frequency
Classically, the Inverse Discrete Fourier Transform method response data perfectly if the FRF is accurate in an ideal case,
(IDFT) is used to transform frequency response data to time and will seek an optimal match if noise and/or distortion are
domain data, that is, to transform the frequency response involved in the data. This new approach retains all the ad-
function (FRF) of the system to its pulse response. The pulse vantages associated with the SSFD while avoiding the iterative,
response of discrete-time systems is also known as the Markov approximate curve-fitting procedures.
parameters. The disadvantage of this approach is that the Section 2 gives some background and the notation used for
Markov parameter sequence thus obtained is distorted by time- this problem. Section 3 discusses the curve-fitting method while
aliasing effects (Oppenheim and Schafer, 1975). Section 4 describes a method to compute the Markov param-
Recently, a method called the State-Space Frequency Do- eters from the parameters obtained from curve-fitting. The
main (SSFD) identification algorithm (Bayard, 1992) has been process of going from the Markov parameters to a state-space
developed. This method can estimate Markov parameters from model is discussed in Section 5. Finally, simulated data from
the FRF without windowing distortion and an arbitrary fre- a model of the Mini-Mast structure and experimental data from
quency weighting can be introduced to shape the estimation • a NASA testbed are used in Section 6 as illustrative examples.
error. The method uses a rational matrix description (the ratio The simulated data discuss an ideal FRF case (without dis-
of a matrix polynomial and a monic scalar polynomial de- tortion and noise) whereas the experimental data present a
nominator) to curve-fit the frequency data and obtains the practical case. The illustrative examples show that the method
Markov parameters from this equation. In obtaining the state- is effective in both cases.
space models from the Markov parameters, the Eigensystem

Background and Notation

Contributed by the Technical Committee on Vibration and Sound for pub-
lication in the JOURNAL OF VIBRATION AND ACOUSTICS. Manuscript received Aug. The objective of frequency domain state-space system iden-
1992; revised June 1993. Associate Technical Editor: J. Mottershead. tification is to identify state-space models from the given fre-

Journal of Vibration and Acoustics OCTOBER 1994, V o l . 116 / 523

Copyright © 1994 by ASME
Downloaded 25 Sep 2010 to 140.116.16.160. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
 quency response data—the frequency response functions (FRF). G(co,) =£> + CBe-^>T+ CABe"^7+ CA2Be'j3"iT
The state-space representation of a linear discrete-time system CO

is Y e Jka T
+ •••=!] «~ ' <6>
x(k+l)=Ax(k)+Bu(k) (1)
y(k) = Cx(k)+D»(k) (2) (Y0 = D, Yk=CAk~'B)
where x{k) £R"*[ is the state vector, u(k) e i ? r x l the input Substitution of Eq. (6) into Eq. (5) yields
vector, y(k) 67?"' xl the output vector; A, B, C, D are the
system matrix, input matrix, output matrix and direct-influ-
ence matrix, respectively. Matrices A, B, C, D are referred to
min E^(«/)«G(«/)-2*'*e"y*M'71 (7)
as state-space parameters or the state-space model. The relation ( * = ! , • • • , < » )

between the state-space parameters and the FRF G(w,) is The parameters Yk = CAK~lB (k=\, • • •, oo) are the Markov
G(^)^C(eiu>TI,,-AylB +D (3) parameters. However, the problem associated with this ap-
proach is that theoretically the number of Markov parameters
where T is the sampling time of the discrete-time system in is infinite. Though overall the Markov sequence is a decreasing
seconds and co; are the frequencies in rad/sec. Given G(o);)( sequence assuming the system is stable, it may take a large
the problem of frequency state-space system identification is number of terms to make CAk~1B = 0 (for all k>kt) for some
to find a set of state-space parameters, denoted by [A, B, arbitrarily large k\, especially when the system is lightly damped.
C, D] (hereafter " " denotes an estimated value), such that A large number of Markov parameters will make the opti-
the estimated FRF mization in Eq. (7) computationally too intensive and im-
G(o>i) = C(eJO,i'xIn~A)~lB +D (4) practical for many applications.
To avoid the problem of excessive number of parameters in
matches G(w,) optimally under some optimality criterion. Note
the optimization, an intermediate step should be taken. That
that G(o)j) is a matrix of dimension mxr. If the L2-norm of
is, curve-fit the FRF data using a finite-ordered matrix-fraction
the error is to be minimized, then an appropriate error criterion
first and then construct the Markov parameters from this re-
is
sult. This approach is detailed in the next section.
min 2vv 2 (a),)IIG(a>,)-G(a>,)ll 2 2 (5)
A,B,C,D ;=1

where w(w,-) is a specified weighting function of frequency Linear Curve-fitting

and I is the total number of the frequency data. The transfer function matrix of the system described by Eqs.
Optimizing Eq. (5) with respect to the state-space parameters (1) and (2) is
directly is a nonlinear problem, which may be difficult to solve. G(z'l)^C(zI„-A)~lB +D
To avoid the difficulties associated with the nonlinear opti-
mization, one possible alternative is to optimize first with re- Cadj(zI„-A)B + a(z)D
spect to the Markov parameters and then convert the Markov (8)
parameters to a state-space model, as optimizing Eq. (5) with a{z)
respect to the Markov parameters is a linear problem. To where a(z) is the characteristic polynomial of A, and adj (.)
formulate this alternative mathematically, it begins with ex- denotes the adjoint of a square matrix. Polynomial a(z), in
panding Eq. (3) general, is monic. The FRF is simply the transfer function

Acronyms adj(.) adjoint of the square

CEM = CSI Evolutionary matrix inside the paren-
Model thesis
CSI = Controls-Structures B = input matrix R[,. . . , Rq = parameters of R (z ~')
Interaction C output matrix R «xl = real vector of dimen-
=
ERA = Eigensystem Realization D = direct-influence matrix sion n
Algorithm e ( '' = exponential function S(z ) = numerator matrix poly-
ERA/DC = Eigensystem Realization G(z~[) = transfer function matrix nomial of the LMFD
Algorithm using Data G(« f ) = frequency response S()> • • • > Sq = parameters of S(z~l)
Correlation function T = sampling time (sec)
FRF = Frequency Response h = identity matrix of order u(k) = input vector
Function n W(o>i) = frequency weighting
IDFT = Inverse Discrete Fourier J = \P-1 *(k) = state vector
Transform imag (.) = imaginary part of the Yk = CAk~xB Markov pa-
LMFD = Left Matrix-Fraction variable in the paren- rameters
Description thesis y(k) = output vector
OKID = Observer/Kalman filter m = number of outputs * = data matrix of the least-
system IDentification n = number of states squares
SSFD = State Space Frequency <? = order of the LMFD = parameter matrix of the
Domain r = number of inputs
e least-squares
real (.) = real part of the variable y = objective matrix of the
Symbols
in the parenthesis least-squares
A = system matrix Riz'1) = denominator matrix z = z domain variable
a(z) = characteristic polyno- polynomial of the «/ = frequency (rad/sec)
mial of A LMFD = estimated value

5 2 4 / V o l . 116, OCTOBER 1994 Transactions of the ASME

Downloaded 25 Sep 2010 to 140.116.16.160. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
 matrix G(z~') calculated along the unit circle in the z-plane. Remark 1: Using a matrix polynomial as the left divisor
Note that G(z~l) does not change if the numerator and de- in Eq. (9) instead of using a scalar polynomial denominator
nominator of Eq. (8) are multiplied simultaneously by an ar- as in Eq. (8) has a remarkable advantage of making the pa-
bitrary monic scalar polynomial. This implies that the rameter estimation become a linear problem. With a scalar
expression of G(z~l) as a rational matrix polynomial is not polynomial denominator, Eq. (8) can also be formulated to
unique; therefore, one can over-specify the orders of the pol- have a form similar to Eq. (13); however, in this case some
ynomials. In the work of Bayard (1992) the SSFD uses this parameters have to be diagonal matrices with identical diagonal
rational matrix polynomial to curve-fit frequency data. By elements. With this constraint, the ordinary least-squares
doing so, however, the estimation of the parameters becomes method can no longer be applied.
a nonlinear problem. The SSFD therefore has to use an ap-
proximate, iterative method to solve the problem, which results Remark 2: In an ideal case where the FRF is accurate, if
in a value not generally optimal in any sense in the presence the order of R(z~l) and S(z~l) is over-specified, the rank of
of noise and/or incorrect model order. real ($) will be less than the row number of 9 . In other words,
On the other hand, it is also known that the transfer function the number of unknowns is more than the number of equations;
matrix can be expressed by a left matrix-fraction description therefore, the answer is not unique. However, a minimum
(LMFD) (Kailath, 1980) as norm solution still can be found using the least-squares method
(Haykin, 1986). In this case, the match is exact; i.e.,
G(z-l)=R-[(z~[)S(z~[) (9) real (3>)G is exactly equal to real (ty). The cases of using the
l
where both R(z~ ) and B(z~') are matrix polynomials: imaginary part or of using combined real and imaginary parts
are similar.
R(z-l)=Im +^z-[+- +RPZ~P (R,arxm) (10)
l l
S(z~ )=S0 + Slz~ + •+Spz~" (SiiR'"xr) Remark 3: In a sense, R(z~[) and S(z~l) in Eq. (9) can
(ID
be interpreted as the observer Markov parameters (Phan et al.,
This factorization is also not unique. For convenience one 1992; Phan et al., 1991; Juang et al., 1991). The order/? of
can choose the orders of both polynomials to be equal ( =p). Riz'1) and S(z~') can be arbitrary as long as it is set equal
Pre-multiplying Eq. (9) by R(z~l) one has to or greater than n/m (Juang et al., 1991) where n is the
R(Z-l)G(z-])=S(z'1) (12) system dimension and m is the number of outputs. Therefore,
if more outputs are available, a smaller order can be assigned
which can be rearranged to become to the matrix polynomials.
l l
G(z-')=-R1G(z~ )z~ - -RPG(z~ ' ) « - " +Sb
Remark 4: The frequency weighting M>(O),) is set to unity
+ Slz-i+---+SpZ-p. (13) in the above derivation. If this is not done, a weighted least-
l x
Note that with G(z~ ) and z~ known, Eq. (13) is a linear squares algorithm should be used.
equation. Because G(z~l) is known at z-e^'T, (/'= 1, • • •, I),
Remark 5: It is known that the least-squares method yields
there are a total of I equations available. Denoting eia'T by z,
unbiased estimates if the noise in data is independent with zero
and stacking up the ( equations, one has
mean (white noise in time domain), and yields biased estimates
•R if the noise is correlated with data. In frequency domain, noise
T T [ p could come from different sources such as measurement error,
G (zV)zV G (z{ )z[ I r zi%- zr% insufficient number of data, leakage, etc., depending upon
r p l P
T l
G (Z2 )Z2' G (z 2 ^)z 2 " / , Z2 Ir- Z2- Ir -K how the FRF is obtained. Usually the quality of the frequency
data can be improved by carefully arranging the experiments
T l l
si and by averaging a sufficient number of results. The results
G (zf )zi T l
G (zf )Z[ P
IrZT%"- P
Zi Ir of this method do rely on the quality of the given frequency
response data; random error in data is acceptable, but biased
distortion will result in spurious modes.
GT(z[>)
GT(Z2~') Estimation of Markov Parameters
(14)
After obtaining a solution to Eq. (14), it is now necessary
to construct the system Markov parameters. Equation (12) can
GT(zf') be written as
or in short,
*G = ^ (15) ' £ * * - ' ) (£y,z-')=£s,-z-'. (19)
\i=o I \/=o / /=o
where
From this relation, the following equations are derived by
eT=[-«i Rn ••• sp] (16) equating terms of like powers and recalling Eqs. (10) and (11):
r X 1 (Juang et al., 1991)
* =[G(zr') G(Z2 )
G(zi )} (17)
Y0 = S0 (20)
and $ is the large data matrix. Equation (15) is a normal
k
equation, where a least-squares solution of 9 can be found.
The least-squares solution of G could be complex numbers. Yk = S„-^]RjYk_j f o r * = l , •••,/> (21)
To avoid this, one can force G to be real by solving either the y=i
real part or the imaginary part of Eq.(15):
Yk= -J]RjY^j for k=p + 1, • • -, oc (22)
real ($)Q = real (¥) (18a)
y=i
imag ( * ) 9 = imag ( ^ ) (18&) Using the estimated Rj and S,- instead of the true R/ and S,- in
Or one can combine and solve both equations as: Eqs. (20) through (22), thus yields a set of estimated Markov
real ($) real(-f) parameters with as many terms as desired.
9 = (18c) Note that there are only p independent Markov parameters
imag ( $ ) imag {¥) excluding Y0; all the other parameters are linear combinations

Journal of Vibration and Acoustics OCTOBER 1994, Vol. 116 / 525

Downloaded 25 Sep 2010 to 140.116.16.160. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
 Frequency Response 0.9859 0.1500 0.9859 0.1501
A = diag
-0.1500 0.9859 -0.1501 0.9859
0.6736 0.7257
-0.7257 0.6736
0.4033 0.9025 0.3943 0.9064
12 15 0.9025 0.4033
Frequency (Hz)
-0.9064 0.3943

Frequency Response (Phase) -0.0407 -0.0454"

200 -0.5384 -0.6001
100 0.0746 -0.0669
0.9867 -0.8850
Angle 0
0.0164 -0.0373
(Dag) B=
0.0860 X10~
-100 1 0.0376
- 0.0460 -0.0421
-200 -0.0711 -0.0650
6 9 12 15
Frequency (Hz) 0.0653 -0.0655
0.0997 -0.1000
Fig. 1 The least-squares matching of the frequency response function
(the (1,1) element) of Mini-Mast (the estimated data coincides with the
true data). 0.8570 1.8701"
-0.0000 0.0000
1.5700 -1.2390
(with matrix coefficients) of this independent set. Therefore, -0.0000 -0.0000
the Markov parameters represent a signal containing a max- 1.4030 1.4254
imum of mxp states where m is the number of outputs. C 0.0000 0.0000
0.9016 1.7852
-0.0000 0.0002
Identification of State-Space Models 1.3509 -1.4748
Once the Markov parameters are constructed, the Eigen- - 0.0000 0.0000
system Realization Algorithm (ERA) or the Eigensystem Re- 0.0 0.0
alization Algorithm using Data Correlation (ERA/DC) can be D=
used to obtain a state-space model. The ERA (or ERA/DC) 0.0 0.0
has been proved valuable for modal parameter identification
from the pulse response. The ERA uses singular value decom- With these parameters, one can calculate the frequency re-
position to decompose a data matrix (referred to as the general sponse function (FRF) according to Eq. (3). Two hundred
Hankel matrix) and to compute a state-space model from the frequency data points equally spaced in a frequency range from
decomposed matrices. The system order is determined by ex- 0 to 16.67 Hz are calculated. Assuming the orders of i?(z _ 1 )
amining the magnitudes of the singular values, where the small and S(z~[) are ten (i.e., p=l0) and solving Eq. (14), one
values are ascribed to the effects of noise and are truncated. obtains an estimated 9 . This set of parameters matches the
The number of the retained singular values is taken as the FRF through Eq. (13) exactly, as shown in Fig. 1. The (1,1)
order of the system. elements of the true FRF and of the estimated FRF (calculated
Because the Markov parameters contain at most m xp states, using 9) are plotted in the same figure. Note the two curves
the order assigned to the system cannot exceed the number match exactly and thus can not be distinguished. The error
mxp. Real systems, in theory, have an infinite dimension; between the two curves is within the numerical precision of
therefore, the larger the number p is chosen, the better the the computer. Comparisons of the other FRF elements are
results will be. However, a large/) causes intensive calculation. similar.
Thus there is a trade-off between accuracy and computation. Figure 2 compares the true Markov parameters and the es-
In practice, by examining the peaks in the FRF, an approximate timated Markov parameters, calculated based on Q. Only the
number of dominant modes can be estimated to assist the (1,1) elements are shown as the other elements have similar
selection of a proper value for p. results. Four hundred points of the Markov parameters are
calculated. The two curves also coincide within a difference
of order 10" 16 , which can be regarded as numerical error. In
Illustrative Examples this case the ERA (or ERA/DC) is capable of recovering the
true state space model (under some equivalence transforma-
Two examples are given—one simulated example and one tion), and the reconstructed FRF also matches the true FRF
experimental example. The simulated example uses a model perfectly.
of a 60-ft-long cantilevered truss structure, whereas the ex-
perimental example is a 55-ft-long truss with a 15-ft diameter Figure 3 shows a comparison between the true Markov pa-
antenna. rameters and those parameters obtained using the inverse dis-
crete Fourier transform (IDFT) method. There are 201
Example 1. The first example uses an analytical model of frequency samples available, including data at O J T = 0 and
the Mini-Mast structure as described and reported by Pappa oiT= 7r; therefore, 400 points of the Markov parameters can
and Schenk (1990). The model has five fundamental modes. be calculated. Due to the finite number of frequency samples
The first two modes are closely spaced, representing the first and long-duration pulse response of the system, the Markov
bending modes in two orthogonal planes. The third mode is parameters thus obtained involve distortion as shown in Fig.
the torsion mode while the last two modes are also closely 3, even though the FRF data are perfect. Additional frequency
spaced, representing the second bending modes. The discrete- response data (by decreasing the frequency increment to in-
time state-space parameters of this system using a sampling crease the resolution) will reduce the distortion. However, to-
rate of 30 Hz are listed as follows: tally eliminating the distortion requires a large number of FRF

526 / Vol. 116, OCTOBER 1994 Transactions of the ASME

Downloaded 25 Sep 2010 to 140.116.16.160. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
 Pulse Response Histories Frequency Response

A n A
Amplitude
(M)
o

VWVWvw
V "
Magnitude

Frequency (Hz)
Time (Sec)

Fig. 2 Comparison of the true and the estimated Markov parameters Frequency Response (Phase)
(pulse response histories) of the Mini-Mast.

Pulse Response Histories

1.5 1 0 ' 6 Angle
True IDFT (Deg)

mpiitude
(M) 0
Frequency (Hz)
Fig. 5 Comparison of the test FRF (solid line) and the reconstructed
FRF (dash line) obtained using the identified system matrices.
0 2 4 6 8 10 12
Time (Sec)
Fig. 3 Comparison of the true Markov parameters (pulse response Frequency Response
history) of the Mini-Mast and the IDFT-recovered Markov parameters
(pulse response history).

8 Proportional and
Bi-directional
Thrusters
Frequency (Hz)

Frequency Response (Phase)

Angle n
(deg)

Fig. 4 CSI Evolutionary Phase Zero Model

data points for this lightly damped system. The Markov pa- Frequency (Hz)
rameters obtained using the method proposed in this paper do
not have distortion, and the number of data required is small. Fig. 6 Comparison of the test FRF (solid line) and the reconstructed
FRF (dash line) obtained using the identified system matrices.
Example 2. The second example uses experimental data
taken from the CSI Evolutionary Model (CEM). The CEM
shown in Fig. 4 is a NASA testbed to study the controls and resenting a case of a strong signal, while Fig. 6 is the frequency
structures interaction problem (Belvin et al., 1991). The system response of output 2 with respect to input 1, representing a
has eight inputs and eight collocated outputs for control. The case of a weak signal. The signal is weak because sensor 2 is
inputs are forces generated by means of air thrusters and the orthogonal to input 1. Similar results are obtained for other
outputs are accelerations. The locations of the input-output input/output pairs which are not shown in this paper. The
pairs are depicted in Fig. 4. In this example, the structure was results show that the matching is better for the strong signal
excited using random input signals to four thrusters located cases. This is expected because the strong signal has a larger
at positions 1,2, 6,7. The input and output signals were filtered signal-to-noise ratio than the weaker signal. The results for
using low-pass digital filters with the range set to 78 percent other input-output pairs are similar and hence omitted.
of the Nyquist frequency (12.8Hz) to concentrate the energy
in the low frequency range below 10 Hz. A total of 2048 data Conclusion
points at a sampling rate of 25.6 Hz from each sensor are used A novel method of identifying state-space models from fre-
for identification. In this example, sixteen FRF's from four quency response data is developed. The method uses only a
input and output pairs located at positions 1, 2, 6, 7 are si- linear technique in the frequency domain without iteration to
multaneously used to identify a state space system model to obtain the Markov parameters from frequency response data.
represent the CEM. It is capable of avoiding windowing distortions inherent to
The order of the matrix polynomial is set to p = 25, which other frequency domain algorithms. In this method, the fre-
is sufficient to match as many as 50 modes (a system of di- quency response data given is assumed reliable. For accurate
mension 100). A state-space model is obtained using ERA/ frequency response data of an ideal linear system the method
DC with the system order assigned to 100. The reconstructed can identify the system perfectly; for experimental data of real
frequency response data (dash lines) are compared with the systems the method provides a least-squares fit to the frequency
experimental data (solid lines) in Figs. 5 and 6. Figure 5 is the response data. The method was tested on both numerical and
frequency response of output 1 with respect to input 1, rep- experimental data sets with successful results.

Journal of Vibration and Acoustics OCTOBER 1994, Vol. 116/527

Downloaded 25 Sep 2010 to 140.116.16.160. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
 Acknowledgment tification of Observer/Kalman Filter Markov Parameters: Theory and Experi-
ment," Proceedings of the AIAA Guidance, Navigation and Control Conference,
The experimental data in Example 2 were generated by Dr. New Orleans, LA, Aug. 12-14, pp. 1195-1207; NASA Technical Paper 3164,
Kenny B. Elliott of NASA Langley Research Center and Mr. June 1992 and also Journal of Optimization Theory and Application, to appear.
Juang, J.-N., and Pappa, R. S., 1985, " An Eigensystem Realization Algo-
Joseph W. Gregory of Lockheed Engineering and Sciences rithm for Modal Parameter Identification and Model Reduction," Journal of
Company. The frequency response functions (FRF) of the Guidance, Control, and Dynamics, Vol. 8, Sept.-Oct., pp. 620-627.
experimental data were provided by Mr. Gregory. Dr. Elliott Juang, J. N., Cooper, J. E., and Wright, J. R., 1988, "An Eigensystem
and Mr. Gregory are currently continuing with the analysis of Realization Algorithm Using Data Correlation (ERA/DC) for Modal Parameter
this data as well as other data sets associated with different Identification," Journal of Control Theory and Advanced Technology, Vol. 4,
No. 1, pp. 5-14.
experiments. Kailath, T., 1980, Linear System, Prentice-Hall Inc., Englewood Cliffs, NJ,
pp. 345-372.
Oppenheim, A. V., and Schafer, R. W., 1975, Digital Signal Processing,
References Prentice-Hall Inc., Englewood Cliffs, NJ, pp. 96-98.
Bayard, D. S., 1992, "An Algorithm for State Space Frequency Domain Pappa, R. S., and Schenk, A., 1990, "ERA Modal Identification Experiences
Identification without Windowing Distortions," Proceedings, IEEE Conference with Mini-Mast." 2nd USAF/NASA Workshop on System Identification and
on Decision and Control., Tucson, AZ, Dec. pp. 1707-1712. Health Monitoring of Precision Space Structures, Pasadena, CA.
Belvin, W. K., et. al, 1991, "Langley's CSI Evolutionary Model: Phase 0," Phan, M., Juang, J. N., Longman, R. W., 1992, "Identification of Linear
NASA Technical Memorandum 104165, September. Multivariate Systems from a Single Set of Data by Identification of Observers
Chen, C.-W., Huang, J.-K., Phan, M., and Juang, J.-N., 1992, "Integrated with Assigned Real Eigenvalues," Journal of Astronautical Sciences, Vol. 40,
System Identification and State Estimation for Control of Large Flexible Space No. 2, April-June, pp. 261-279.
Structures," Journal of Guidance, Control and Dynamics, Vol. 15, No. 1, Jan.- Phan, M., Horta, L. G., Juang, J.-N., and Longman, R. W., 1991, "Linear
Feb., pp. 88-95. System Identification Via an Asymptotically Stable Observer," Proceedings of
Haykin, S., 1986, Adaptive Filter Theory, Prentice-Hall Inc., Englewood the AIAA Guidance, Navigation and Control Conference, New Orleans, LA,
Cliffs, NJ, pp. 331-341. Aug. 12-14, pp. 1180-1194, and the Journal of Optimization Theory and Ap-
Juang, J.-N., Phan, M., Horta, L. G., and Longman, R. W., 1991, "Iden- plication, to appear.

528 / Vol. 116, OCTOBER 1994 Transactions of the ASME

Downloaded 25 Sep 2010 to 140.116.16.160. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm