See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/23596822

An Eigensystem Realization Algorithm for Modal Parameter Identification

and Model Reduction

Article in Journal of Guidance, Control, and Dynamics · November 1985

DOI: 10.2514/3.20031 · Source: NTRS

CITATIONS READS
2,847 8,836

2 authors, including:

Richard S. Pappa
National Aeronautics and Space Administration
102 PUBLICATIONS 4,830 CITATIONS

SEE PROFILE

All content following this page was uploaded by Richard S. Pappa on 19 March 2015.

The user has requested enhancement of the downloaded file.

620 J. GUIDANCE VOL. 8, NO. 5

An Eigensystem Realization Algorithm

for Modal Parameter Identification and Model Reduction
Jer-Nan Juang* and Richard S. Pappa*
NASA Langley Research Center, Hampton, Virginia

A method called the eigensystem realization algorithm is developed for modal parameter identification and
model reduction of dynamic systems from test data. A new approach is introduced in conjunction with the
singular-value decomposition technique to derive the basic formulation of minimum order realization which is
an extended version of the Ho-Kalman algorithm. The basic formulation is then transformed into modal space
for modal parameter identification. Two accuracy indicators are developed to quantitatively identify the system
and noise modes. For illustration of the algorithm, an example is shown using experimental data from the
Galileo spacecraft.

Introduction modal parameter identification for flexible structures has not

been addressed.. . .
T HE state space model has received considerable attention
for system analyses and design in recent control and
systems research programs. One of these areas, in particular, is
In the structures field, the finite element technique is used
almost exclusively for constructing analytical models. This ap-
control of large space structures. In order to design controls for proach is well established and normally provides a model ac-
a dynamic system it is necessary to have a mathematical model curate enough for structural design purposes. Once the struc-
that will adequately describe the system's motion. The process ture is built, static and dynamic tests are performed. These test
of constructing a state space representation from experimental results are used to refine the finite element model, which is
data is called system realization. then used for final analyses. This traditional approach to
During the past two decades, numerous algorithms for the analytical model development may not be accurate enough for
construction of state space representations of linear systems use in designing a vibration control system for flexible struc-
have appeared in the controls literature. Among the first were tures. Another approach is to realize a model directly from the
the works of Gilbert1 and Kalman,2 introducing the important experimental results. This requires the construction of a
principles of realization theory in terms of the concepts of minimum-order model from the test data that characterizes
controllability and observability. Both techniques use the the dynamics of the system at the selected control and
transfer function matrix to solve the realization problem. Ho measurement positions. The present state-of-the-art in struc-
and Kalman3 approached this problem from a new viewpoint. tural modal testing and data analysis is one of controversy
They showed that the minimum realization problem is about the best technique to use. Classical test techniques,
equivalent to a representation problem involving a sequence of which may provide only good frequency and moderate mode
real matrices known as Markov parameters (pulse response shape accuracy, are often considered adequate for finite ele-
functions). By minimum realization is meant a model with the ment model verification purposes. On the other hand, ad-
smallest state space dimension among systems realized that vanced data analysis techniques that offer significant reduc-
has the same input-output relations within a specified degree tions in test time and improved accuracy have been
of accuracy. Questions regarding the minimum realization available,12'16 but are not yet fully accepted. For example,
from various types of input-output data and the generation of Void and Russell16 presented a method using frequency-
a minimum partial realization are studied by Tether,4 Silver- response functions and time-domain analysis for direct iden-
man,5 and Rossen and Lapidus6 using Markov parameters. tification of modal parameters including repeated eigenvalues.
Rossen and Lapidus7 successfully applied Ho-Kalman3 and A comparison of contemporary methods using data from the
Tether4 methods to chemical engineering systems. A common Galileo spacecraft test is provided by Chen.17
weakness of the preceding schemes is that effects of noise on Although structural dynamics techniques are generally suc-
the data analysis were not evaluated. Zeiger and McEwen8 cessful for ground data, further incorporation with work from
proposed a combination of the Ho-Kalman algorithm3 with the controls discipline is needed to solve modal parameter
the singular-value decomposition technique for the treatment identification/control problems. For example, it is known
of noisy data. However, no theoretical or numerical studies from control theory18 that a system with repeated eigenvalues
were reported by Zeiger and McEwen. Among follow-up and independent mode shapes is not identifiable by single in-
developments along similar lines, Kung9 presented another put and single output. Methods which allow only one initial
algorithm in conjunction with the singular-value decomposi- condition (input) at a time13 will miss repeated eigenvalues.
tion technique to incorporate the presence of the noise. Note Also, if the realized system is not of minimum order and
that the singular-value decomposition technique10'11 has been matrix inversion is used for constructing an oversized state
widely recognized as being very effective and numerically matrix, numerical errors may become dominant.
stable. Although several techniques of minimum realization Under the interaction of structure and control disciplines,
are available in the literature, formal direct application to the objective of this paper is to introduce an eigensystem
realization algorithm (ERA) for modal parameter identifica-
tion and model reduction for dynamical systems from test
Received July 3,1984; revision received Nov. 29,1984. This paper is data. The algorithm consists of two major parts, namely,
declared a work of the U.S. Government and therefore is in the public basic formulation of the minimum-order realization and
domain. modal parameter identification. In the section of the basic for-
*Aerospace Engineer, Structural Dynamics Branch. Member mulation, the Hankel matrix, which represents the data struc-
AIAA. ture for the Ho-Kalman algorithm, is generalized to allow ran-
SEPT.-OCT. 1985 MODAL PARAMETER IDENTIFICATION AND MODEL REDUCTION 621

dom distribution of Markov parameters generated by free where y / ( / = !,...,/*—1) and */(/= l,...,s — 1) are arbitrary in-
decay responses. A unique approach based on this generalized tegers. For the system with initial state response
Hankel matrix is developed to extend the Ho-Kalman measurements, simply replace Hrs(k-\) by Hrs(k). Now
algorithm in combination with the singular-value decomposi- observe that
tion technique.10'11 Through the use of the generalized Hankel
matrix, a linear model is realized for a dynamical system C
matching the input and output relationship. The realized CA>*
system model is then transformed into modal space for modal
parameter identification. As part of ERA, two accuracy in-
dicators, namely, the modal amplitude coherence and the
modal phase collinearity, are developed to quantify the system CAjr-l
and noise modes. The degree of modal excitation and observa- and
tion is evaluated. The ERA method thus forms the basis for a
rational choice of model size determined by the singular values Ws=[B,A'iB,...,A's-iB] (5)
and accuracy indicators.
Experimental results for a complex structure—the Galileo where Vr and Ws are the observability and controllability
spacecraft—are given to illustrate the ERA method. The matrices, respectively. Note that Vr and Ws are rectangular
Galileo spacecraft is an interplanetary vehicle to be launched matrices with dimensions rp x n and n x ms, respectively.
in 1986 for a detailed investigation of the planet Jupiter and its Assume that there exists a matrix If satisfying the relation
moons.
(6)
Basic Formulations
A finite dimensional, discrete-time, linear, time-invariant where In is an identity matrix of order n. It will be shown that
dynamical system has the state-variable equations the matrix IP plays a major role in deriving the ERA. What is
//*? Observe that, from Eqs. (5) and (6),
x(k+l)=Ax(k)+Bu(k) (1)
(7)
y(k)=Cx(k) (2)
The matrix If is thus the pseudoinverse of the matrix Hrs(0) in
where x is an ^-dimensional state vector, u an m-dimensional
a general sense. For a single input, there exists a case [see Eq.
control input, and y a p-dimensional output or measurement (5)] where the rank of Hrs(0) equals the column number of
vector. The integer k is the sample indicator. The transition
matrix A characterizes the dynamics of the system. For flexi- Hn(0), then
ble structures, it is a representation of mass, stiffness, and
damping properties. The problem of system realization is as H«=[[Hrs(0)]THrs(0)]-J[Hrs(0)]T (8)
follows: Given the measurement functions y(k), construct
constant matrices [A, B, C] such that the functions y are On the other hand, there exists a case for a single output [see
reproduced by the state-variable equations. Any system has an Eq. (5)] where the rank equals the row number, then
infinite number of realizations that will predict the identical
response for any particular input. Let T be any nonsingular H#=[Hrs(0)]T(Hrs(0)[Hrs(0)]T]-1 (9)
square matrix. The triple [TAT'1, TB, CT'1] will also be a
realization.6 However, the eigenvalues of the matrix A are The matrix Hrs(l)H* has been used in the structural dynamics
preserved. field to identify system modes and frequencies.13 This is a
For the system Eqs. (1) and (2) with free pulse response, the special case representing a single input which cannot realize a
time domain description is given by the function known as the system that has repeated eigenvalues, or a noise-free system
Markov parameter unless the system order is known a priori. More discussion can
be found in the Appendix.
Y(k)=CAk-*B (3) A general solution for If is given below. For an «th-order
system, find the nonsingular matrices P and Q such that 1(U1
or in the case of initial state response
Hrs(0)=PDQT (10)
Y(k)=CAk[x1(0),x2(0),...,xm(0)]
where the rpxn matrix P and the msxn matrix Q are
where x/(0) represents the rth set of initial conditions and k is isometric matrices (all of the columns are orthonormal), leav-
an integer. Note that B is an n x m matrix and C is a p x n ing the singular values of Hrs(Q) in the diagonal matrix D with
matrix. The problem of system realization is: Given the func- positive elements [d1,d2,...,dn]. The rank n of Hrs(G) is deter-
tions Y(k), construct constant matrices [A,B,C] in terms of mined by testing the singular values for zero (relative to
Y(k) such that the identities of Eq. (3) hold. The algorithm desired accuracy)11 which will be described in the next section.
begins by forming the r x s block matrix (generalized Hankel Define
matrix)
Hn(0) =PDQT= [PD] [QT] = (11)
Each of the four matrices [Pd,QT, Ws, V?] has rank and row
Y(k) number n. By Eq. (5) with fc = 0, ,

... Y(j1 + k+ts_1) VrWs=Hrs(0)=PdQT (12)

T
Multiplying on the left by Pj and solving for Q yields
_ Y(Jr-i+V YUr-i
(4) TWs=(PTdPd)~1PTdVrWs = (13)
622 J.-N. JUANG AND R.S. PAPPA J. GUIDANCE

The matrix T is nonsingular because if Hankel matrix can be displayed quantitatively. The set of
singular values can be used to judge the distance of the matrix
with determined order to a lower-order one. Therefore, the
structure of the generalized Hankel matrix can be properly ex-
then TU=Iby Eq. (13). Since 7T/=/= LTfor nonsingular T ploited to solve the realization problem efficiently. These in-
and C/, then clude an excellent numerical performance, stability of the
realization, and flexibility in determining order-error tradeoff.
(14) Assume that, by Eq. (10),
Hence, by Eq. (14), (20)
= [Q] (15) with
The dimension of matrices Q and Pj with rank n are msxn (21)
and « x 77?, respectively. Define 0^ as a null matrix of order/?,
Ip an identity matrix, Ej= [/^O^...^], and £%=[!„, If the matrix Hrs(0) has rank n then all of the singular values
O w ,...,0 m ]. df (i = n + 1,..; ,N) should be zero. When singular values
With the aid of Eqs. (5), (6), and (15), a minimum order- dt (i = n +1,...,N) are not exactly zero but very small, then one
realization can be obtained from can easily recognize that the matrix Hrs(G) is not far away
from an w-rank matrix/However, there can be real difficulties
Y(k+l)=ETpHrs(k)Em=ETpVrAkWsEm in determining a gap between the computed last nonzero
singular value and what effectively should be considered zero,
= ETpVrWsH*VrAkWsH»VrWsEm when measurement noise is present. Possible sources of the
noise can be attributed to the measurement signal, computer
=ElHrs(0)QP»dVrAkWsQP»dHrs(0)Em roundoff, and instrument imperfections.
Look at the singular value dn of the matrix Hrs(G). Choose a
number d based on measurement errors incurred in estimating
the elements of Hrs(Q) and/or roundoff errors incurred in a
=EtPd[P»dHK(l)Q]*QTEm previous computation to obtain them. If d is chosen as a "zero
threshold" such that d<dn, then the matrix Hrs(Q) is con-
=ETpD'A[D-'APTHK(l)QD-'A]kDKQTEfft (16) sidered to have rank n. Unless information about the certainty
of the measurement data is given, the number d is defined as a
This is the basic formulation of realization for the ERA. The function of the precision limit in the computer. For example,
triple [D-1/2PTHrs(\)QD~1/2, D1/2QTEmJ ETpPD1/2 ] is a b-dn/dj cannot exceed the precision limit; further details are
minimum realization since the order n of P*dHrs(\)Q equals the found in Ref. 11.
dimension of the state vector x. The same solution, in a dif- After the test of singular values, assume that the matrix
ferent form, for the case wherey, = */ = / (/= l,...,r-1), can be [D-1/2PTHrs(k)QD~1/2} has rank n. Find the eigenvalues z
obtained by a completely different approach as shown in Refs. and eigenvectors ^ such that
3 and 19. The system [Eqs. (1) and (2)] with this realization is
written as (22)

x(k+l)=D-'/2PTHrs(l)QD-1/2x(k)+D'/2QTEmu (17) The modal damping rates and damped natural frequencies are
simply the real and imaginary parts of the eigenvalues, after
transformation from the z to the s plane using the relationship
y(k)=ETPD1/2x(k) (18)
where (23)

= WsQD~1/2x(k) (19) where AT is the data sampling interval andy is an integer. The
integer k is generally chosen as 1 for simplicity. Although z
Now, the case can be summarized as follows. and \// are complex numbers, computations of Eq. (22) can be
A finite dimensional, discrete-time, linear time-invariant performed using a real algorithm20 since the state matrix
dynamical system with multi-input and multi-output is realized for flexible structures has independent eigenvectors.
realizable in terms of the measurement functions if the system The triple [z, ^-lD1/2QTEm) ETpPD1/2^\ is obviously a
is controllable and observable (the ranks of matrices Vr and minimum order realization simply by observing Eq. (16). The
Ws are n). A simple exercise, such as replacing Y(k+\) by matrix E*PD1/2\I/ is called mode shapes and the matrix
Y(k) in Eq. (16), shows that the algorithm developed above is ^-1D1/2Q^Em initial modal amplitudes. To quantify the
also true for the realization of a system with initial state system and noise modes, two indicators are developed as
response. follows.
Note that no restrictions on system eigenvalues are given for
this case. In other words, this technique can realize a system Modal Amplitude Coherence y
with repeated eigenvalues. As byproducts of this approach, If the information about the uncertainties of the measure-
two alternative algorithms identified as (Al) and (A2) are ment is minimum, the rank thus determined by the SVD
derived in the Appendix. becomes larger than the number of excited and observed
system modes to represent the presence of noises in modal
Modal Parameter Identification space. In modal parameter identification, the indicator refer-
and Model Reduction red to as modal amplitude coherence is developed to quan-
The presence of almost unavoidable noise and structural titatively distinguish the system and noise modes. Based on the
nonlinearity introduces uncertainty about the rank of the accuracy parameter, the degree of modal excitation (con-
generalized Hankel matrix and, hence, about the dimension of trollability) is estimated.
the resulting realization. By employing the singular-value The modal amplitude coherence is defined as the coherence
decomposition (SVD) technique, the rank structure of the between each modal amplitude history and an ideal one
SEPT.-OCT. 1985 MODAL PARAMETER IDENTIFICATION AND MODEL REDUCTION 623

formed by extrapolating the initial value of the history to lat- is then defined as21
ter points using the identified eigenvalue. Let the control input
matrix (initial condition) be expressed as

(24) j = l,2,...,n (36)

; • } • ..;-

where the asterisk means transpose complex conjugate, and This indicator checks the deviation from 0-180 degree
the Ixm column vector 2?y corresponds to the system eigen- behavior for components of the y'th identified mode shape.
value $j (j = !,...,«)• Consider the sequence The parameter ^j can have only the values between 0 and 1.
fjij — 1 indicates that accuracy of the mode shape is high. On
(25) the other hand, if /*, is away from 1, the y'th mode is either a
noise mode or the mode is significantly complex.
which represents the ideal modal amplitude in the complex do-
main containing information of the magnitude and phase Model Reduction
angle with time step AT. Now, define vector q^ such that The dynamical system is composed of an interconnection of
1 1/2 T
all of the ERA-identified modes. The accuracy indicators
\I/~ D Q = [qlfq23...fqn]* (26) allow one to determine the degree of individual mode par-
ticipation. Model reduction then can be made by truncating all
The comrjlex vector </, represents the modal amplitude time of the modes with low accuracy indicators. The accuracy of
history from the real measurement data obtained by the the complete modal decomposition process can be examined
decomposition of the Hankel matrix. Let 7, be defined as the by comparing a reconstruction of Y(k) formed by Eq. (16)
coherence parameter for the y'th mode, satisfying the relation with the original free decay responses, using the reduced
model.
7;= (27)
Summary of ERA
where I I represent the absolute value. The parameter 7, can The computational steps are summarized as follows:
have only the values between 0 and 1. 7,-* 1 as q} ^qj indicates 1) Construct a block-Hankel matrix 7/r5(0) by arranging the
that the realized system eigenvalue Sj and the initial modal measurement data into the blocks with givdn r, s, tt (i =
amplitude bj are very close to the true values for the y'th mode 1,2,...,5-7) andy / (i=l,2,...,r-l), [Eq. (4)].
of the system. On the other hand, if 7, is far away from the 2) Decompose Hrs(G) using singular-value decomposition
value 1, the y'th mode is a noise mode. However, to make a [Eq. (10)].
clear distinction between the system and noise modes requires 3) Determine the order of the system by examining the
further study. Obviously, the parameter 7, quantifies the singular values of the Hankel matrix Hrs(Q) [Eq. (20)].
degree to which the modes were excited by a specific input, 4) Construct a minimum-order realization (A, B, C) using
i.e., the degree of controllability. a shifted block-Hankel matrix [Eq. (16)].
5) Find the eigensolution of the realized state matrix [Eq.
Modal Phase Collinearity /*
(22)] and compute the modal damping rates and frequencies
For lightly damped structures, normal mode behavior [Eq.(23)].
should be observed., An indicator referred to as the modal 6) Calculate the coherence parameter [Eq. (27)] and the
phase cbllinearity is developed to measure the strength of the collinearity parameter [Eq. (36)] to quantify the system and
linear functional relationship between the real and imaginary noise modes.
parts of the mode shape for each mode. Based on the accuracy 7) Determine the reduced system model based on the ac-
indicator, the degree of modal observation is estimated. curacy indicators, reconstruct function Y(k) [Eq. (16)], and
Define compare with the measurement data.
Note that the optimum determination of r, s, t; andy, in step
ETpPD1/^=[c1,c2,...,cn] (28) 1 above requires further development. This determination is
related to the choice of the measurement data to minimize the
where Cj (j = 1,2,...,«) is the mode shape corresponding to the size of the Hankel matrix Hrs (0) with the rank unchanged.
y'th realized mode. Let the column vector 7 of order p be

7 T =[7,7,...,7] (29)

in which p is the number of sensors. Now compute the follow-

ing quantities for the y'th mode shape.
(30)

(31)
RPM THRU STIR
cri=[Re(Cj-Cjl)]T[Im(Cj-Cjl)] (32) CLUSTER
SBA
400 N. ENGINE
(33)

e=(cn-crr)/2cri (34) SCAN

PLATFORM
(35) S/C ADAPTERS
SEISMIC
where Re( ) and Im( ), respectively, are the real and imag- BLOC
inary parts of the complex vector ( ), and sgn( ) is the sign of
the scalar ( ). The modal phase collinearity /*, for the y'th mode Fig. 1 Galileo spacecraft in launch configuration.
624 J.-N. JUANG AND R.S. PAPPA J. GUIDANCE

Example: Analysis of Galileo Test Data characteristics of the test, data acquisition, and data-reduction
the ERA method has been verified using multi-input and processes were exactly the same for both data $ets.
multi-output simulation data with or without noise for as- The random excitation signal, bandlimited to the interval
sumed structures with distinct and/or repeated eigenvalues. from 10 to 45 Hz, was generated digitally with the same test
The reader is directed to the original version of this paper22 for system which was also used to record the accelerometer
more information. Experimental results for the analysis of response signals. Approximately 5 s of data were recorded
Galileo test data are given in the following. following the end of the excitation signal. The responses were
The Galileo spacecraft is shown in Fig. 1. All appendages, digitized at a rate of 102.4 samples per second, resulting in
including the S-X-band antenna (SXA) at the top of the vehi- about 500 free-response points in each test.
cle, were locked in their stowed positions. The structure was
cantilevered from its base by bolting the bottom edge of the
conical spacecraft adapter ring to a massive seismic block. The
adapter ring is the interface between Galileo and a Centaur up- Identified Eigenvalues
per stage that will provide the interplanetary boost. Each ERA analysis was performed using a.single matrix of
For dynamic excitation, several electrodynamic shakers of data from all 162 response measurements and one initial con-
about 100 N capacity, hung on soft suspensions from dition (either x- or z-direction test) at a time. Each response
overhead cranes, were attached at various points. Response function Yas shown in Eq. (4) was thus a 162x 1 matrix. Us-
measurements were made with 162 accelerometers distributed ing jj = 1 and tf = i (i= 1,2,...,499) the Hankel matrix Hrs of
over the test article. 324 rows by 500 columns was formed to perform the analysis.
The finite element model of the structure in this configura- A summary of the identification results for the x direction
tion predicted 45 modes of vibration below 50 Hz, with the test is provided in Table 1, including identified frequencies,
lowest frequency at about 13 Hz. However, as discussed in damping factors, and accuracy indicators. The results for the
Ref. 23, many of the modes are of lesser importance based on z-direction test can be found in Ref. 24. These results closely
their predicted contribution to the dynamic launch loads. In agree with those obtained by other experimental techniques as
fact, only about 15 modes are major contributors. The shown in Refs. 17 and 25.
presence of the others, however, interspersed., in frequency
with the important ones, results in high modal density and
makes accurate parameter identification more difficult.
Table 1 ERA-identified results for x-direction
Complete details of the test configuration, analytical model
predictions, and the method used to estimate the relative im- Darnping Accuracy indicators
portance of the modes, can be found in Ref. 23. Mode Frequency, factor,
No. Hz Ta M* Max.c No.d
Test and Data Acquisition Procedures 1 13.613s 3.301 98.6s 99.0 0.104 6
All results to be presented were obtained from two sets of 2 14.153s 1.645 99.8s 94.2 0.063 5
free-response measurements recorded following single-point 3 18.369s 1.409 99.7s 98.7 0.270 59
random excitation of the structure. The first data set was ob- 4 18.680s 1.086 99.7s 79.8 0.068 17
tained using single-shaker, lateral excitation—in the global X 5 19.516s 0.659 96.5s 99.1 0.090 23
6 22.040 3-311 50.7 43.1 0.009 0
direction—and the second set with single shaker, vertical ex- 7 22.0476 0.777 98.9s 97.7 0.122 58
citation—in the global z direction. These tests are referred to 8 22.132s 0.356 97.5s 99.3 0.088 45
as simply the "x direction" and "z direction" tests. For both 9 22.640f 1.282 93. l f 86.6 0.018 4
tests, no special effort was made to select the position for the 10 25.126s 1.515 99.8s 89.8 0.239 27
shaker. Each position was chosen using only the knowledge 11 25.751g 5.569 80.2g 56.2 0.043 15
that many modes were excited from the location in previous 12 25.871s 0.806 98.6s 99.4 0.279 67
tests. Other than the point and direction of excitation, all 13 28.517s 0.835 97.7s 98.0 0.121 10
14 29.659f 1.723 91.7f 61.2 0.010 1
15 30.940s 3.003 98.2s 77.1 0.059 5
16 32.828 6.668 67.9 52.8 0.015 2
Frequency = 13.613 Hz Damping = 3.301% 17 33.288f 2.894 93. 4f 44.0 0.012 4
Modal Amplitude Coherence = 98.6% Modal Phase Colinearity = 99.0% 18 33.613s 1.326 98.1s 59.8 0.024 3
<u-90
M «»....*....».......,.. ... •*..».,.».'•..,—2u*M—,.~.> .• >.•*,*• .t ^ 19 34.402s 1.057 98.2s 98.4 0.047 15
o 90 20 36.890f 1.633 93.3f 73.1 0.020 2
0.270
100 21 37.631 5.957 50.0 76.2 0.041 2
38.019f 94. l f
| 50

f °
JL 22
23
24
39.551
41.021f
0.575
4.048
2.038
75.5
92.4f
44.9
99.2
89.8
0.008
0.240
0.123
0
5
17
| -50
-100 25 42.169 3.737 65.2 89.1 0.022 6
0 10 20 30 40 50 60 70 80 90 100 110 120 130 140 150 160 26 42.273f 1.124 92.9f 68.5 0.011 3
Measurement No.
(a) 13.613 Hz antenna bending mode from x-direction random test.
27 43.758f 1.614 94.2f 57.7 0.019 3
28 44.892s 0.394 99.4s 99.1 0.513 24
29 45:173 4.188 79.6 67.0 0.021 6
Frequency = 38.037 Hz Damping = .513%
Modal Amplitude Coherence = 99.8% Modal Phase Colinearity = 93.1%
30 45.174 0.835 78.5 78.7 0.041 3
31 47.191g 2.126 84.3g 50.8 0.008 0
32 48.311f 1.214 94.4f 78.3 0.034 5
33 48.414 3.300 59.2 57.0 0.014 3
34 50.3858 1.143 89.5 40.2 0.027 3
a
7%: Modal amplitude coherence.
b
c
fjL%: Modal phase collinearity.
Maximum modal participation in g's among the 162 components identified.
0 10 20 30 40 50 60 70 80 .90 100 110 120 130 140 150 160
Measurement No.
(The measurement noise floor was approximately 0.002 g.)
"Number of mode shape components of 162 with initial response amplitude
(b) 38.037 Hz "bounce" mode from z-direction random test.
>0.01 g. Indicator of local or global response.
Fig. 2 Example ERA-identified mode shapes illustrating various f
90<7o<7<95«7o.
degrees of phase angle scatter.
SEPT.-OCT. 1985 MODAL PARAMETER IDENTIFICATION AND MODEL REDUCTION 625

The best single accuracy indicator now available is the The mode shown in Fig. 3, at approximately 38 Hz, is the
modal amplitude coherence. Its value is used in Table 1 to rate fundamental z-direction "bounce mode" of the structure. As
the identified modes at various degrees of accuracy. The rating would be expected, the modal amplitude plots confirm that
scale is noted in the key beneath the table. A brief descrip- the mode was more strongly excited in the ^-direction test than
tion of the information in each of the three columns on the far in the x-direction test. This conclusion is reflected in the modal
right of the table is also contained in the keys. The significance amplitude coherence shown in Fig. 3a compared with that in
of modal phase collinearity will be discussed in the next Fig. 3b.
subsection. The data in the last two columns are two addi-
tional indications of the strength of the modal response signals Data Reconstruction
relative to the instrumentation noise floor. These indicators The final aspect of the ERA method available for assessing
are computed using modal participation values in physical identification accuracy is the process of data reconstruction.
units, which are the products of the mode shapes and the in- This procedure consists of comparing the original free-
itial modal amplitudes. response, time histories (and their frequency spectra) with
thos^e calculated using the ERA-reduced model, which con-
Identified Mode Shapes tains 22 modes with high modal amplitude coherence. If the
Two typical ERA-identified mode shapes are shown in Fig. ERA modal decomposition process is performed accurately,
2. Based on these results and observations from other data not the reconstruction results will closely match the original data.
shown, the following conclusions can be drawn. Figure 4 shows a typical such comparison from the Galileo
1) The local behavior of many of the Galileo modes—ex- data analysis. The two time histories are compared at the top
emplified by the antenna mode shown in Fig. 2a in which only of the figure, and their fast Fourier transforms (FFT), in both
about five measurements show motion—makes it more dif- amplitude and phase, in the lower plots, the reconstruction
ficult to identify all 162 components of the mode shapes ac- result, the smoother of the two lines in the FFT plots, is seen
curately because of the low response levels. • *;* to closely follow the original data in both amplitude and
2) A good measure of the effects of noise on the mode phase. The increase in amplitude of the test data below 2 Hz is
shape accuracy is often indicated by the amount of scatter in due to residual motion of the shaker on its soft suspension.
the identified modal phase angles from the ideal 0-180 deg These response characteristics, well below the first flexural
normal mode behavior. Of course, true complex-mode
behavior needs to be differentiated from identification scatter
due to noise and nonlinearity. The best remedy is to compare
10°
the results for the same mode obtained in several different
tests. io-2
3) The parameter referred to as the modal phase collinear-
ity can be used to measure how closely the modal phase angle
results for each mode cluster near 0 and 180 deg. Calculated 2 .3 b) 2 3
using principal component analysis, it indicates the extent to Time, sec Time, sec
which the information in each complex-valued mode shape is
Fig. 3 Example modal amplitude results. This mode near 38 Hz is
representable as a real-valued vector. It ranges from a value of the fundamental z-direction "bounce mode" of the structure. Com-
zero for no collinearity to 100% for perfect collinearity. parison of a and b shows that the mode was more strongly excited in
4) Based on studies with simulated data, the accuracy of the z-directipn test, as ekpected. a) 38.037-Hz modal amplitude from
mode shapes showing clustering of the identified phase angles ^-direction test; modal amplitude coherence = 99.8%. b) 38.019-Hz
near 0 and 180 deg, such as in Fig. 2, can generally be accepted modal amplitude from ^-direction test; modal amplitude
with little questioning. However, those modes with coherence = 94.1%.
significantly more phase angle scatter should not be used
without further confirmation.
5) Most mode-shape components whose identified phase
angles are displaced from the 0- and 180-deg lines are those
LRecc
.06
with the smallest amplitudes. This characteristic is consistently AcceL, •^Reconstruction AcceL, Measurement
observed in the result shown in Fig. 2b. Small modal
amplitude results for these components, however, usually in-
9
'Jfc0 4 8 0 4 8
dicate accurately that the response amplitude is, in fact, very Time, sec Time, sec
small. This information is all that can be expected from a
measurement standpoint, and is all that is required in many 180i
instances. FFT J rRecOnstr.
Phase °1
Identified Modal Amplitudes -180
The ERA modal amplitude coherence indicates the purity of 10' •2..M,,.MM"M|.M.|MMJMM|.U.,,M.|.M. !,.?.,

the individual modal amplitude time histories. For each iden-

tified eigenvalue, a modal amplitude time sequence is obtained
for each initial condition. These data provide a direct indica- 10
tion of the strength with which the mode was identified in the
analysis. For strongly identified modes, the modal amplitude FFT 1Q-4
history is a pure, exponentially decaying sinusoid of the cor- Modulus10
responding frequency and damping, which decays smoothly
over the entire analysis interval. For weakly identified modes,
the modal amplitude history is distorted. In particular, the
history is a sequence of noise for any eigenvalue not cor-
responding to a structural mode.
Typical examples of modal amplitude results from the 10 20 30 40 50
Galileo analysis are shown in Fig. 3. These data were selected Frequency, Hz
to illustrate the variation in the purity of the modal amplitude
histories for a mode which was strongly excited is only one of Fig. 4 Example comparison of a measurement from jc-directioh test
the two ERA tests. with a 22-mode reconstruction.
626 J.-N. JUANG AND R.S. PAPPA J. GUIDANCE

mode of the spacecraft, do not affect the modal results and where x(k) = Vrtfx(k). Or, using Eq. (A2),
were not retained in the identified parameter set.
x(k+l)=fl»Hrs(l)x(k)+Emu

Concluding Remarks y(k)=ETpHrs(0)x(k)

An eigensystem realization algorithm is developed for
modal parameter identification and model reduction for where x(k) = Wsx(k).
dynamical systems. Two developments are given in this paper. These realizations are not of minimum order, since the
First, a new approach is developed to derive the basic formula- dimension of x is the number of either columns or rows of the
tion of minimum realization for dynamical systems. As matrix Hrs(0) which is larger than the order n of the state
byproducts of this approach, two alternative less powerful matrix A for multi-input and multi-output cases. Examination
algorithms, identified as algorithms (Al) and (A2), are de- of Eqs. (Al), (A2), and (16) reveals that algorithms (Al) and
rived. A special case of (A 1) is shown to be equivalent to an (A2) are special cases of ERA. Algorithm (Al) is formulated
approach currently in use in the structural dynamics field * Sec- by inserting the identity matrix (6) on the right-hand side of
ond, accuracy indicators are developed to quantify the par- the state matrix A as shown in Eq. (Al). On the other hand,
ticipation of system modes and noise modes in the realized algorithm (A2) is obtained by inserting the identity matrix (6)
system model. In other words, the degree of controllability on the left-hand side of the state matrix A as shown in Eq.
and observability for each participating mode is determined. (A2). However, ERA is formed by inserting the identity
A model reduction then can be made for controller design. matrix (6) on both sides of the state matrix A as shown in Eq.
Based .on the results of other analyses of simulated data, the (16). Because of the different insertion, algorithms (Al) and
parameters referred to as modal amplitude coherence and (A2) do not minimize the order of the state transition matrix.
modal phase coliinearity are good indicators of the identifica- Mathematically, if the singular-value decomposition tech-
tion accuracy that is achieved. Using these indicators, approx- nique is not included in the computational procedures [see
imately 15 modes from the Galileo analysis are judged to be of Eqs. (8) and (9)], Eqs. (A 1) and (A2) cannot be numerically
high accuracy. Less than 10 min of CPU time on a CDC main- implemented, unless a certain degree of artificial noise and/or
frame computer 'were required for the Galileo analysis. system noise are present. Noises tend to make up the rank
However, additional research is needed to correctly assess the deficiency of the generalized Hankel matrix Hrs(0) for
degree of accuracy achieved by those modes showing signifi- algorithms (Al) and (A2). Since the degree of noise con-
cant identification scatter. The effects of structural tamination is generally unknown, algorithms (Al) and (A2)
nbnlinearities oil the results also need further attention. are not recommended.
Important features of the eigensystem realization algorithm
are summarized as follows. <f- .
1) From the computational standpoint, the algorithm is at- References
tractive since only simple numerical operations are needed. Gilbert, E.G., "Controllability and Observability in Multivariable
2) The computational procedure is numerically stable. Control Systems," SIAM Journal on Control, Vol. 1, No, 2, 1963,
3) The structural dynamics requirements for modal pp. 128-151.
2
parameter identification and the control design requirements Kalman, R.E;, "Mathematical Description bf Linear Dynamical
Systems," SIAM Journal on Control, Vol. 1, No. 2, 1963, pp.
for a reduced state space model are satisfied. 152-192.
4) Data from more, than one test can be used simulta- 3
Ho,B.L.,Kalman, R.E., "Effective Construction of Linear State-
neously to efficiently identify closely spaced eigenvalues. Variable Models from Input/Output Data," Proceedings of the 3rd
5) Computation requirements are moderate. Annual Allerton Conference on Circuit and System Theory, 1965, pp.
449-459; also, Regelungstechnik, Vol. 14, 1966, pp. 545-548.
4
Tether, A.J., "Construction of Minimal Linear State-Variable
Models from Finite Input-Output Data," IEEE Transactions on
Appendix Automatic Control, Vol. AC-15, No. 4, Aug. 1970, pp. 427-436.
5
In view of Eqs. (4) through (6), the measurement function Silverman, L.M., "Realization of Linear Dynamical Systems,"
Y(k+\) can be obtained through either of two other IEEE Transactions on Automatic Control, Vol. AC-16, No. 6, 1971,
algorithms [Eqs. (Al) and (A2)] . The algorithm (Al) is pp. 554-567.
6
Rosseh, R.H., and Lapidus, L., "Minimum Realizations and
System Modeling: I. Fundamental Theory and Algorithms," AIChE
Journal, Vol. 18, No. 4, July 1972, pp. 673-684.
7
Rossen, R.H.. and Lapidus, L., "Minimum Realizations and
System Modeling: II. Theoretical and Numerical Extensions," AIChE
Journal, Vol. 18, No. 5, Sept. 1972, pp. 881-892.
=E}[Hi,(l)If-]*»H(0)Em (Al) 8
Zeiger, H.P. and McEwen, A.J., "Approximate Linear Realiza-
tions of Given Dimension Via Ho's Algorithm," IEEE Transactions
on Automatic Control, Vol. AC-19, No. 2, April 1974, pp. 153.
and the algorithm (A2) is 9
Kung, S., "A New Identification and Model Reduction Algorithm
Via Singular Value Decomposition," 12th Asilomar Conference on
Circuits, Systems and Computers, Nov. 1978, pp. 705-714.
10
Golubi G.H. and Reinsch, C., "Singular Value Decomposition
and Least Squares Solutions," Numerical Mathematics, Vol. 14,
=ETpVrWs[H*VrAWs]kEm 1970, pp. 403-420.
u
Klema, V.C. and Laub, A.J., "The Singular Value Decomposi-
(A2) tion: Its Computation and Some Applications," IEEE Transactions
on Automatic Control, Vol. AC-25, No. 2, April 1980, pp. 164-176.
12
Hence, by Eq. (3), {Hn(\)If Hrs(Q)Em, ETp} or [H«Hrs(\), Stroud, R.C., Smith, S., and Hamma, G.A., "MODALAB—A
Ey, £jjF/rs(0)] is a realization. The system [Eqs. (1) and (2)] New System for Structural Dynamic Testing," Shock and Vibration
with realization (Al) will be transformed into the following Bulletin, No. 46, Pt. 5, Aug. 1976, pp. 153-175.
13
Ibrahim, S.R. and Mikulcik, B.C., "A Method for the Direct
equation: Identification of Vibration Parameters from the Free Response,"
Shock and Vibration Bulletin, No. 47, Pt. 4, Sept. 1977, pp. 183-198.
x(k+l)=Hrs(l)H»x(k)+Hrs(0)Emu 14
Brown, D.L., Allemang, R.J., Zimmerman, R. and Mergeay,
M., "Parameter Estimation Techniques for Modal Analysis," SAE
Paper 790221, Feb. 1979.
SEPT.-OCT. 1985 MODAL PARAMETER IDENTIFICATION AND MODEL REDUCTION 627

15 21
Coppolino, R.N., "A Simultaneous Frequency Domain Tech- Moore, B.C., "Principal Component Analysis in Linear Systems:
nique for Estimation of Modal Parameters from Measured Data," Controllability, Observability, and Model Reduction," IEEE Trans-
SAE Paper 811046, Oct. 1981. actions on Automatic Control, Vol. AC-26, No. 1, Feb. 1981, pp.
16 17-32.
Void, H. and Russell, R., "Advanced Analysis Methods Improve 22
Juang, J.N. and Pappa, R.S., "An Eigensystem Realization
Modal Test Results," Sound and Vibration, March 1983, pp. 36-40. Algorithm (ERA) for Modal Parameter Identification and Model
17 Reduction," NASA/JPL Workshop on Identification and Control of
Chen, J.C., "Evaluation of Modal Testing Methods," AIAA
Paper 84-1071, May 1984. Flexible Space Structures, San Diego, Calif., June 1984.
23
18
Chen, J.C. and Trubert, M., "Galileo Modal Test and Pre-Test
Chen, C.T., Introduction to Linear System Theory, Holt, Analysis," Proceedings of the 2nd International Modal Analysis Con-
Rienhart, and Winston, Inc., New York, 1970. ference, Orlando, Fla., Feb. 1984, pp. 796-802.
24
19
Kalman, R.E., and Englar, T.S., "Computation of a Minimal Pappa, R.S. and Juang, J.N., "Galileo Spacecraft Modal Iden-
Realization," A User Manual for the Automatic Synthesis Program, tification Using an Eigensystem Realization Algorithm," The Journal
NASA CR-475, June 1966, Chap XV. of the Astronautical Sciences, Vol. 33, No. 1, Jan.-March, 1985, pp.
15-33.
20 25
Gantmacher, F.R., The Theory of Matrices, Vol. 1, Chelsea Stroud, R.C., "The Modal Survey of the Galileo Spacecraft,"
Publishing Co., New York, New York, 1959. Sound and Vibration, April 1984, pp. 28-34.

From the AIAA Progress in Astronautics and Aeronautics Series

SPACE SYSTEMS AND THEIR INTERACTIONS WITH

EARTH'S SPACE ENVIRONMENT—v. 71
Edited by Henry B. Garrett and Charles P. Pike, Air Force Geophysics Laboratory

This volume presents a wide-ranging scientific examination of the many aspects of the interaction between space systems
and the space environment, a subject of growing importance in view of the ever more complicated missions to be performed
in space and in view of the ever growing intricacy of spacecraft systems. Among the many fascinating topics are such matters
as: the changes in the upper atmosphere, in the ionosphere, in the plasmasphere, and in the magnetosphere, due to vapor or
gas releases from large space vehicles; electrical charging of the spacecraft by action of solar radiation and by interaction
with the ionosphere, and the subsequent effects of such accumulation; the effects of microwave beams on the ionosphere,
including not only radiative heating but also electric breakdown of the surrounding gas; the creation of ionosphere "holes"
and wakes by rapidly moving spacecraft; the occurrence of arcs and the effects of such arcing in orbital spacecraft; the
effects on space systems of the radiation environment, etc. Included are discussions of the details of the space environment
itself, e.g., the characteristics of the upper atmosphere and of the outer atmosphere at great distances from the Earth; and
the diverse physical radiations prevalent in outer space, especially in Earth's magnetosphere. A subject as diverse as this
necessarily is an interdisciplinary one. It is therefore expected that this volume, based mainly on invited papers, will prove of
value.
Published in 1980,737pp., 6x9, illus., $35.00 Mem., $65.00 List

TO ORDER WRITE: Publications Order Dept., AIAA, 1633 Broadway, New York, N.Y. 10019

View publication stats