You are on page 1of 12

Copyright ©1996, American Institute of Aeronautics and Astronautics, Inc.

AIAA Meeting Papers on Disc, 1996, pp. 1547-1557


A9626954, NAS1-18690, NAS1-19622, AIAA Paper 96-1486

Uncertainty modeling for structural control analysis and synthesis


Mark E. Cambpell
MIT, Cambridge, MA

Edward F. Crawley
MIT, Cambridge, MA

IN:AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics, and Materials


Conference and Exhibit, 37th, Salt Lake City, UT, Apr. 15-17, 1996, Technical Papers. Pt. 3
(A96-26801 06-39), Reston, VA, American Institute of Aeronautics and Astronautics, 1996,
p. 1547-1557

The development of an accurate model of uncertainties for robust control of structures that undergo a change in
operational environment is studied. The application used in this work is the development of an on-orbit uncertainty
model based on ground modeling and experimentation. A ground-based uncertainty model consisting of mean errors
and bounds on critical structural parameters is developed. These uncertainties are localized to specific degrees of
freedom in the form of mass and stiffness uncertainties using a first-order sensitivity method. The mass and stiffness
uncertainties of the ground-based system are projected onto the on-orbit system to create an analogous on-orbit
uncertainty model. The Middeck Active Control Experiment is introduced as experimental verification for the
localization and projection methods. (Author)

Page 1
96-1486

A96-26954 AIAA-96-1486-CP

Uncertainty Modeling for Structural Control Analysis and Synthesis


Mark E. Campbell* and Edward F. Crawleyt
Space Engineering Research Center
Massachusetts Institute of Technology
Cambridge, Massachusetts, 02139 USA

Abstract finite element model to create an input-output model.


Even higher accuracy measurement models are devel-
The development of an accurate model of uncertainties
oped to fit open loop data. By comparing measurement
for robust control of structures that undergo a change
models and the input-output model, a 1-g uncertainty
in operational environment is studied. The application
model consisting of mean errors and bounds on criti-
used in this work is the development of an on-orbit uncer-
cal parameters is developed. Control design in 1-g can
tainty model based on ground modeling and experimen-
be accomplished using the 1-g input-output model with
tation. A ground-based uncertainty model consisting of
parameters shifted by their mean errors, and an uncer-
mean errors and bounds on critical structural parameters
tainty model consisting of the parameter bounds.
is developed. These uncertainties are localized to specific
A 0-g input-output model is created by removing the
degrees of freedom in the form of mass and stiffness un-
gravity effects and suspension degrees of freedom from
certainties using a first order sensitivity method. The
the 1-g finite element model, and appending any ad-
mass and stiffness uncertainties of the ground-based sys-
ditional dynamics. A 0-g uncertainty model can then
tem are projected onto the on-orbit system to create an
be developed by projecting the mean errors and bounds
analogous on-orbit uncertainty model. The Middeck Ac-
from the 1-g uncertainty model into 0-g, using both the
tive Control Experiment is introduced as experimental
1-g and 0-g finite element models.
verification for the localization and projection methods.
Figure 1 shows that developing an accurate 0-g uncer-
Introduction tainty model is a complex blending of many disciplines
including finite element modeling and updating, system
Many engineering applications require testing in an en- identification, probability analysis, and finite element er-
vironment that is different from its eventual use: wings ror localization. Finite element error localization tech-
and rotorcraft are tested in a wind tunnel; space struc- niques attempt to localize errors to specific degrees of
tures are tested using a suspension system on the ground. freedom using measured frequencies and mode shapes,
Practical control design for space structures typically re- and a finite element model. A number of papers give
quires a very accurate model of the dynamics, in addi- excellent surveys of work in this area.[~3
tion to a robust control design procedure. Two common Sidhu and Ewins developed the most well known error
approaches exist: develop a model of the structure and localization method called the Error Matrix Method,4'5
controllers prior to flight; perform an on-orbit identifi- where a system is first reduced to the measured degrees
cation and design controllers during the flight. The ob- of freedom, and errors in the flexibility matrix are iden-
jective of this work is to develop accurate uncertainty tified and then inverted to get the measured stiffness
models for control analysis and synthesis of structures
in space, based solely on modeling and experimentation
on the ground.
To develop the most accurate model possible for con-
trol design on a space structure prior to flight, extensive
ground modeling and testing must be performed. Fig-
ure 1 shows this process schematically. A 1-g finite el- 1 -8 Uncertainty Model
ement model is developed and updated using open loop
data to obtain good accuracy. Additional dynamics for
sensors, actuators, time delays, etc. are appended to the
'Postdoctoral Associate, Dept. of Aeronautics and Astronautics,
Member AIAA, ASME, AAS f 0-g Uncertainty Model

t Professor of Aeronautics and Astronautics and Mac Vicar Faculty


Fellow, Fellow AIAA.

Copyright ©1996 by the American Institute of Aeronautics and Figure 1: Development of control design arid uncer-
Astronautics, Inc. All rights reserved. Paper No. 96-1486. tainty models for 0-g.
1547
errors. The sensitivity method6'7 expresses mass and Table 1: Structural uncertainties, and their capturabil-
stiffness error matrices as a function of design param- ity using mean/variance uncertainty.
eters (E, I, p, etc.) and calculates first order sensitivi- Capturable
ties. The Force Balance method8 examines the system Group Sub-group Type Easy Diff.
for large eigensystem residues using (expanded) exper- Linear Physical Material V
imental mode shapes and frequencies, thus indicating Parameters Geometric V
Modeling Discretization V
errors at specific degrees of freedom. There are also fre-
quency domain methods that build an error matrix as
Incorrect Physics V
Internal Boundary Conditions V
a function of frequency rather than corrections to the Coupling Joints/ Attachments V
mass or stiffness matrices.9'10 Act. /Sens. Dynamics V
Many of the finite element error localization methods
Time Delays V
require measurement of all degrees of freedom, reduc-
External Aerodynamic J
Coupling Gravity V
tion of the original system, or expansion of experimental Control System V
mode shapes to the size of the original system. In prac- Testing Sensor/Process Noise v/
tical applications for space structures, measurement of
Sensor Bias V
Nonlinear Weak V
all degrees of freedom is not possible. Reduction of the Strong V
original system leads to errors and error locations that
are suspect.5 While expansion of mode shapes has been
explored by many researchers, there is no general corre- model. This section examines possible modeling un-
lation between the actual mode shape and the accuracy certainties in a structural system, and how they man-
of the expanded mode shape.11 The only method which ifest themselves into the model. (Note that this may be
is capable of handling partial information, i.e. measure- quite difficult in some cases.) All uncertainties can be
ment of a subset of frequencies and degrees of freedom, divided into the general groups of Linear and Nonlin-
is the sensitivity method. ear. Within these groups, 5 sub-groups can be defined:
One of the most thorough examinations of the repeata- Physical, Modeling, Coupling, Testing, and Nonlinear-
bility of ground and space hardware is the Middeck 0-g ities. Table 1 gives a summary of these uncertainties,
Dynamics Experiment (MODE).12 On the ground and in and in addition, the capturability, or ability to localize
space, variation in modal parameters were examined as a given uncertainty to a physical mean or variance un-
functions of disturbance force amplitude, joint preload, certainty as either easy or difficult.
reassembly, shipset (different hardware), and suspension. Physical uncertainties are modeling parameters: ma-
Collins et al.13 assume statistical properties for model- terial uncertainties such as the modulus of elasticity,
ing parameters in order to iteratively localize uncertain- density, and Poission's ratio; and geometric uncertain-
ties. This is very similar to the sensitivity method, ex- ties such as area and inertia (or their primitives such as
cept the variances are chosen to reflect the user's confi- radius, width, etc.), and length. The physics of the sys-
dence in the parameter estimate. Hasselman et a/.14'15 tem has been captured, but the parameters have incor-
group together generically similar structures into three rect values. These uncertainties are traditionally treated
databases, an example of which is Research Models of with updating techniques and are therefore usually easy
Large (truss-type) Space Structures. The databases can to localize. Modeling uncertainties are those errors in
thtsn be used to create a probability model of a space which the physics of the hardware has not been repre-
structure, consisting of a mean model and distribution sented correctly, including discretization issues, selection
of modal errors. of correct element types, or unmodeled boundary condi-
In this paper, a thorough assessment of structural tions. Because the physics has not been captured cor-
modeling uncertainties is given. The development of a rectly, these uncertainties are quite difficult to localize.
1-g uncertainty model is then summarized. Next, the Coupling uncertainties are components that are usu-
sensitivity method is used to localize both mean and ally appended to the finite element model. It is assumed
variance uncertainties to specific degrees of freedom, and that they are modeled correctly, but the parameters need
project them into 0-g to develop the 0-g uncertainty adjustment. Internal couplings are components such as
model. Finally, the Middeck Active Control Experi- boundary conditions, joints and attachments, actuator
ment, a shuttle flight experiment which flew on STS-67 and Sensor dynamics, time delays, etc. External cou-
in March 1995, is introduced for experimental verifica- plings are components such as aerodynamic, gravity, and
tion of the methods presented. control systems (and many others). Most of these un-
certainties can be updated using a parameterized model
Assessment of Modeling Uncertainties that separately includes the modeled components, and a
parameter estimation technique to update these param-
A thorough understanding of all possible modeling un- eters. The coupling of gravity into a system produces
certainties in a structural system is very important to effects such as geometric stiffening of members, initial
both the development of a ground-based uncertainty deformation, and effects on actuators and sensors, and
model, and prediction of a space-based uncertainty are usually modeled as an addition to the stiffness nia-

1548
trix. Localization of aerodynamic uncertainties would be nN physical degrees of freedom. Assuming proportional
a great challenge because of the large uncertainties that damping, the dynamics can be described by solving the
exist in the homogeneous disturbance and aerodynamic nN degree of freedom, linear structural generalized eigen-
effects. value problem
Depending upon the type of testing, many uncertain-
ties can also enter during experimental testing. Issues N =0 (2)
such as noise, bias, and bandwidth in sensors and time
delays in computers can all can impede accurate mea- where AN € K"N X"N and $N 6 K"N X " N are the eigenval-
ues and eigenvectors given by
surements of the system. It is assumed that the noise un-
certainties, because they are small, will manifest them-
selves as larger variance uncertainties. Random bias er-
rors that are large, however, cannot be represented as AN = diag [ A N i (3)
a linear physical change to the system. Many of these
It is assumed that all nN eigenvalues and eigenvectors of
uncertainties can be avoided by using quality actuators
the model are available. This assumption can be easily
and sensors, and modeling them correctly.
relaxed, although as will be seen later, it is important
Nonlinearities act on every system in some form: ma-
to solve for as many modes as possible, even though the
terial properties, stiction, dead band stiffness in joints,
accuracy might be suspect. Note that the eigenvectors
and changing boundary conditions such as loss of tension
are assumed to be mass normalized such that
in a cabled suspension system. Strong nonlinearities only
degrade the uncertainty localization process, and usually
require a nonlinear model, in addition to the finite ele-
ment model. If they are weak, or can be made weak by
- AN (4)
the addition of another component to the system such The finite element model is assumed to be an accurate
as a controller, these types of nonlinearities are assumed representation of the hardware. The mean error and
to be identified using variance uncertainties. variance uncertainties which exist between the nominal
In order to successfully identify and localize uncertain- finite element model and hardware can then be defined
ties, steps must be taken to reduce the effects of those
using a first order perturbation to the nominal system.
uncertainties which are not capturable. These include:
• Refining the finite element mesh to prevent dis- MN = MN + AMN
cretization and joint and/or attachment errors.
• Understanding the physics of the hardware such
that the correct element types are used. AN = A N + AA N
• Completely modeling boundary conditions and ac- (5)
tuator and sensor dynamics.
• Completely understanding all coupling effects, and where (r) refers to the linear model description of the
their impact on the finite element model. hardware (i.e. data).
• Completely understanding all nonlinear effects, in- To ensure the practical application of these methods,
cluding modeling all moderate nonlinearities.
the nominal uncertainty model (and subsequent uncer-
Once these precautions have been taken, the uncertainty tainty localization and projection), is based on incom-
localization process has the best chance of capturing the plete information: the number of frequencies (eigenval-
most accurate mass and stiffness uncertainties. ues) that are measured (n N ) is less than (or equal to) the
number of modes in the system, or nN < nN; the number
Uncertainty Localization and Projection of distinct actuators and sensors (mode shapes) that are
The uncertainty localization and projection methods are measured (r N ) is less (or equal to) the number of original
developed in the most general form. Therefore, the 1-g degrees of freedom, or rN < n N -
system is referred to as the "Nominal" system and is de- Assuming the nominal eigenvectors $N are a sufficient
noted (-) N . The 0-g system is referred to as the "Modi- set of basis vectors describing the hardware dynamics,
fied" system and is denoted (-) M . This allows the meth- changes in the eigenvectors, A$ N , can be written as a
ods to be easily developed and applied to systems differ- linear combination of the original eigenvectors, .
ent from the 1-g to 0-g example described in this work.
Problem Formulation
It is this equation which requires as many of the eigen-
The dynamics for the discretized finite element model of vectors as possible to be retained in the finite element
the nominal system can be written as solution. The higher frequency modes, even though they
= 0 might not be accurate, create a sufficient basis for repre-
(1)
senting the eigenvalue perturbations. Note that an anal-
where MN € HT"*"", and KN 6 R"Nxn N are the mass ogous formulation of each of these equations can be de-
and stiffness matrices respectively and r;N is a vector of veloped for the modified system.

1549
Model Hardware Step 2: Uncertainty Localization and Mapping

AN,$N Uncertainty Model Once the nominal uncertainty model has been developed,
Nominal MN,KN A N ,|>N the next step is to localize the uncertainties to specific
AA N , A*N
degrees of freedom, and map them onto the modified sys-
Uncertainty Localization(AA/N, tem. A first order sensitivity method is used, whose roots
Mapping of DOF(AM M , A.K"M) can be traced back three decades.13'6 The method de-
scribed here, however, develops a complete uncertainty
AM,$M Projection localization strategy that not only uses first order sen-
Modified AM, I'M
MM, KM AA M , A$M sitivities, but also addresses problems such as systemat-
ically choosing update parameters and handling an in-
Figure 2: Uncertainty model development problem. sufficient number of modal measurements. In addition,
the technique is adapted for the localization of variance
Figure 2 shows the approach for the uncertainty lo- uncertainties, as well as mean errors.
calization and projection process. Given a finite ele- The linear structural eigenvalue problem for the true
ment model and data sets for the nominal system, an dynamics of the nominal hardware can be written as
uncertainty model is developed, as shown in the two
top blocks. This uncertainty model is in the form of — 0 (7)
mean and covariance uncertainties of the eigenvalues and
eigenvectors between the model and hardware. These Substituting Equations 5-6 into Equation 7, and premul-
uncertainties are localized to specific physical degrees of tiplying by <j%. and <^., yields
freedom of the finite element model, in the form of mean
error and variance uncertainties of the physical mass and N - A Ni AM N ] 4>Ni (8)
stiffness matrices. A mapping of degrees of freedom be- [ANi - ANJ- 0Ni (9)
tween the nominal and modified systems is used to cre-
ate the mass and stiffness uncertainties of the modified The mass and stiffness perturbations are assumed to
be in the form,
system. Finally, the modified finite element model is
used to project the mapped mass and stiffness uncer-
tainties, creating an uncertainty model for the modified
system in the form of mean and covariance uncertainties
of the eigenvalues and eigenvectors. This is shown in the AMN = (10)
bottom two blocks of Figure 2. Each of these steps is 3=1
examined in detail in the following sections.
where Kj and Mj are macroelements, aj and /3j are
Step 1: Development of Nominal Uncertainty Model macroelemental scale factors that scale the relative size
The first step in this compilation of methods is the devel- of each macroelement, and kN and mN refer to the total
number of macroelements for the nominal system. These
opment of the nominal uncertainty model. A set of tests
is needed to accurately measure each of the uncertain- macroelements can be in the form of a finite element, a
ties itemized in Table 1. This can be done by varying part of an element, or groups of elements. A detailed
test parameters such as using different actuators, sen- discussion of these macroelements is given subsequently.
Substituting Equation 10 into Equations 8-9 and re-
sors, disassembly/reassembly, time between tests, distur-
bance levels, and adding a non-structural change such as arranging, the eigenvalue and eigenvector errors are
a servo system. By testing each of these parameters, a
large database of the nominal system can be created. AA Ni - £ [faKjfa] a, - ANi fj^Ni]0j (11)
An identification/parameter estimation method is 3=1
then used to find parameter estimates for the measured
"N
frequencies (eigenvalues) and eigenvectors of the sys-
tem. This technique must parameterize the finite ele- E ••0MV K4 0Nfc
*'*——* T.Vr.
ANfc —Ajvji
a,- (12)
ment model and component models (sensor, actuator, Lfc^i
time delays, etc.) separately, and allow easy extraction
and combination of measured modal parameters. Once
the parameters are identified, the nominal uncertainty -£ £•
model is created by calculating mean errors and bounds
on critical parameters of the system, i.e. frequencies and
mode shape entries (and damping ratios). For this work, The above equations are linear in terms of the macroele-
uncertainties are not developed for parameters in the
mental scale factors a.j and /3j. Therefore, all of the
measured errors can be written in the form
component models (sensor, actuator, time delays, etc.),
but can easily be added. A complete explanation of this vec {AANi} 1 _
(13)
procedure is given in Ref. 16.
1550
The matrix SN is a called a sensitivity matrix. Note that Note that the sensitivity method is quite flexible in lo-
there are nN < nN measured eigenvalues and rN < nN calizing particular errors. Macroelements can be defined
measurement locations. using modeling parameters and gravity effects that su-
In order to calculate the macroelemental scale factors, perpose. When projecting into the modified system,
and subsequently the mass and stiffness errors, a pseu- the gravity macroelements are thrown out, and only the
doinverse is employed. If the choice of macroelements modeling parameter macroelements are used. This gives
and scale factors is made such that their number is less the correct physical results.
than or equal to the number of measured errors, or Therefore, the scale factors (and therefore macroele-
(14)
ments) for the modified system can be found by ordering
kN +mN < (rN + l)-n N
the nominal scale factors such that
then Equation 13 is overdetermined, rather than under-
determined, and the resulting solution is physically real- aN = (20)
izable, in addition to being mathematically optimal.
A weighted pseudoinverse is used to calculate the
macroelemental scale factors Note that this assumes the number of macroelements
vec {AA Ni } for the modified system is less that the number for the
(15) nominal system, or kM < kN and mM < mN, which is
Pfi vec A0Ni
generally the case.
where the [-]+ refers to the pseudoinverse and WN is a The next step is to find the eigenvalue and eigenvector
diagonal weighting matrix uncertainties in terms of the mass and stiffness uncer-
CAN tainties for the modified system. An analogous deriva-
WN = diag tion of the sensitivity equation, given in Equations 11,13
for the nominal system, can be made for the modified
and C(.) are confidence factors that are chosen is a some- system using the macroelements in Equation 19. This
what ad hoc manner. For instance, eigenvalue uncer- again gives a set of linear equations between the un-
tainties are usually easier to calculate than mode shape known eigenvalue and eigenvector errors and calculated
uncertainties, and are not as prone to measurement er- scale factors,
rors. Therefore, the confidence factors are chosen to be
CANJ = 1-0 and C0N, = 0.1. vec {AA Mi } QM
(21)
Because the macroelemental scale factors are linear vec A^ Mi
in terms of the known errors, the expected value and
covariance can easily be calculated where 5M is the sensitivity matrix for the modified sys-
tem. The expected mean and covariance are then calcu-
lated quite easily.
[ vec {AA M ,} 1 = F QM 1
[ vec A$Mi J L PM J
c
Step 3: Uncertainty Projection °* "^^ =s»c™ [ Z} * ™
L Tin* j ^ r* tvi j

The physical mass and stiffness uncertainties correspond The mean errors and bounds for the relevant param-
to particular degrees of freedom; therefore they can then eters of the modified system, i.e. the frequencies and
be mapped to the modified system. For instance, sup- eigenvectors, can then be computed. The eigenvector
pose there is a truss structure with a plate that has a 5% mean errors, (•) and bounds, (•) are computed directly,
mass error. If the plate is also in the modified structure,
the mass error of the corresponding degree(s) of freedom
is also 5%. Under this assumption, a mapping of degrees
= 3• (24)
of freedom is denoted TMN, or
T?M = TMNTJN (18) where var[-] denotes the variance operator, or the di-
agonal of the covariance matrix. The eigenvalue mean
Using this general mapping of degrees of freedom be-
errors are found similarly. More meaningful uncertainty
tween the two systems, the mass and stiffness errors for
bounds for control design, however, are bounds on fre-
the modified system are given by
quencies, rather than eigenvalues, of the modified sys-
tem. Assuming the eigenvalue error is small compared to
the eigenvalue, the frequency perturbation can be writ-
ten using a first order Taylor expansion.
AMM - (19)
(25)
2/M
1551
Using this equation, the mean errors and bounds for each structure that does not have sensors, making it difficult
sample frequency can be found. to localize errors to these degrees of freedom. Taking a
term from the control systems community, the element
_ E[AA errors may be unobservable. Mathematically, these un-
observable element errors manifest themselves as singu-
_ 3Vvar[AA Mi ] larities in the sensitivity matrix.
(26)
~~
To examine this problem, a singular value decomposi-
tion of the (weighted) nominal sensitivity matrix is per-
Macroelements formed
Macroelements are user defined matrices that attempt WNSN = U^Vf (27)
to span the space of all possible modeling errors. There-
where UN and VN are left and right unitary matrices, and
fore, many types of macroelements exist when using the
EN is a matrix of singular values. If there are element
sensitivity method. Factors which influence the choices
errors that are unobservable, they produce singularities,
include: the number and location of the measurements;
or small singular values in the decomposition. The cor-
the number and type of modes; the types of uncertain-
responding column of the VN matrix yields a null vector
ties that may exist; and the size of the finite element
that can aid in the grouping elements.
model. The most influential factor, however, is forcing
the number of macroelements to be less than or equal to Practical Algorithm
the number of measured errors, as shown in Equation 14,
to ensure an overdetermined pseudoinverse. As a result, Too many times have there been methods developed that
three types of macroelements are defined: sub-elements, work quite well for simple problems, but the practical
elements, and groups of elements. application of such methods is severely lacking. This
Sub-element based macroelements are those errors is exemplified in the error localization literature where
which lie within a particular finite element. Physical most of the algorithms assume the measurement of (or
uncertainties such as modulus of elasticity and density expansion to) all degrees of freedom. Although inter-
are easily represented as errors within the element. Er- esting, practical application of many of these methods
rors in couplings such as actuator/sensor dynamics that is not realistic. This section describes the application
are not in the finite element model are localized using of uncertainty localization and projection methods to a
the identification procedure. Gravity effects such as geo- practical experiment, specifically addressing issues such
metric stiffening and initial deformations are additions to as type of uncertainty localization technique, choices of
the stiffness matrix, and therefore macroelements can be macroelements, and effects of large modeling errors.
developed using the superposing stiffness adjustment.17 The sensitivity method is the best method avail-
Because testing uncertainties and nonlinearities cannot able to localize uncertainties because it can handle in-
be: expressed using linear macroelements, it is assumed complete information. The choice of macroelements
that they manifest as mean error and variance uncer- for this method is heavily influenced by the num-
tainties in the macroelements developed previously. ber of errors measured: the number of macroelements
Element based macroelements are those errors which must be smaller than the number of known uncer-
lie in different finite elements. A single macroelement tainties. To localize the most accurate and physi-
can be formed that is the summation of all sub-element cal mass and stiffness uncertainties, given uncertain-
errors corresponding to a finite element. This macroele- ties in the eigenvalues and eigenvectors, the localiza-
ment can then be used to localize errors to that element, tion must be performed at the sub-element level. Ac-
even if the sub-element errors are distinct. knowledging the practical case of a small number of
And finally, groups of elements based macroelements measurements, a three step procedure is proposed:
are those errors that lie in different areas of the struc- 1. - Localize with macroelements: groups of elements
ture, as defined by different groups of elements. When - Reduce number of macroelements
creating a finite element model, a finer and finer mesh of 2. - Localize with macroelements: elements
elements may be needed to increase the model accuracy. - Reduce number of macroelements
As a result, all degrees of freedom are usually not exper- 3. - Localize with macroelements: subelements
imentally measured. It is therefore occasionally useful - Reduce number of macroelements
to group finite element errors into one macroelement. The uncertainties are first localized to areas of the
The simplest grouping procedure is using insight from structure, using macroelements based on groups of el-
the modeling process. For instance, if an element is split ements. The number of macroelements is reduced by
into two equivalent elements to obtain a finer mesh, er- throwing out those macroelements that do not contain
rors in these two finite elements can be grouped together large modeling errors. The remaining macroelements are
into one macroelement for the localization process. then divided to create element based macroelements. A
It is also useful to obtain a grouping procedure that similar procedure is used in steps 2 and 3, resulting in
is a function of the problem, as opposed to using model- localized errors within each finite element in parameters
ing insight. For instance, there may be an section of the such as modulus of elasticity, density, and gravity ef-

1552
original finite element model of the nominal system. A
new system is then formed by combining the finite ele-
ment model and the localized mass and stiffness mean er-
rors. This new system is then taken as the finite element
model, and another iteration of mean error localization is
performed. An iterative procedure is not performed for
the variance uncertainties, however, because a solution
must be found for a nonlinear statistics problem.
Cut Off Point
Small Order Example
Figure 4 shows a small order example used to analyze
the techniques developed and add insight to the uncer-
tainty localization and projection methods. It is a planar
Number of Macroelements
system with 4 tubular cross-sectioned struts and three
concentrated mass/inertia collars. There are 8 beam el-
Figure 3: Example of the reduction of macroelements
using the weighted least squares cost ements and 8 nodal points with 3 degrees of freedom
for each (vertical, horizontal, and rotational) giving 24
degrees of freedom. The mass and stiffness matrices
fects. Note that the strategy could start with steps 2 or
for these elements are standard displacement based two-
3, depending on the number of measured errors.
dimensional beam elements. The mass matrix is chosen
To reduce the number of macroelements, the weighted
to be in consistent form. Note that the "L" shape cou-
least squares problem is examined. Defining an error
ples the vertical and horizontal displacements.
using Equation 13,
This example is also used to verify the projection tech-
vec {AANi} niques proposed in this work. The operational change
-5 N E (28) is a cantilevered configuration to free-free configuration,
vec A4>Ni 0*
shifting frequencies by 20-50%. (Note the damping is
a quadratic cost can be defined proportional and 1% for all modes.) The free-free con-
figuration has 9 nodal points and 27 degrees of freedom.
J = (29) There are 4 modes and 4 evenly spaced rotational de-
grees of freedom (ui,U3,U5, and ur) measured, giving
where WN is the weighting matrix in Equation 15. Mini- 4 + 4 x 4 = 20 measured errors. In addition, uncertain-
mization of this cost leads to the weighted pseudoinverse. ties (both deterministic and stochastic) in four areas
To reduce the number of macroelements, the pseudoin- of the system are introduced: modulus of element 2,
verse is calculated, and the weighted least squares cost £2 = JV(5%,2%); length of element 4, L4 = 7V(-5%,0);
is evaluated. Then, one macroelement is eliminated at outer radius of element 6, re = A^(—5%, 2%); and mass
a time, and the cost is evaluated again. The macroele- and inertia of the collar between elements 6 and 7,
ment that reduces the cost the least is then eliminated. mer = 7V(3%,0). The uncertainties are presented in the
This procedure is repeated until only one macroelement form of a normal variable with mean and standard devi-
is left. The reduction procedure also gives a rough or-
ation errors, or (•) = 7V(A(-),(jA(.))-
dering of the importance of the macroelements to both For this example, the stochastic uncertainties are sam-
the pseudoinverse uncertainty localization method. Fig- pled, the uncertainties are introduced into the model,
ure 3 shows a sample plot of the normalized cost at each and a large number of cases are simulated in a Monte
iteration of this procedure. The usual cut off point for Carlo type simulation. There are 10 simulated cases for
reducing the number of macroelements is when the cost both the nominal and modified systems, with the mean
begins to increase, or at approximately macroelement 6 and variance uncertainties in measured frequencies and
in Figure 3. This may change, however, if there are a lot mode shapes calculated for each.
of uncertainties. Note that a distinct, but similar pro-
A two-step localization procedure was used. Table 2
cedure can be developed for the reduction of macroele-
shows the actual and localized results for a few "entries
ments when localizing variance uncertainties.
The last aspect of the practical uncertainty localiza-
tion and projection strategy is addressing the use of a
first order linear perturbation to the structural eigen-
value problem, which is inherently nonlinear. If the per-
turbations are small, the strategy works quite well. How-
ever, when they become large, the assumptions made
during the derivation become invalid. To eliminate these
problems, an iterative process similar to model updat-
ing is used. The mean errors are localized using the Figure 4: Small order example: cantilevered "L" beam
1553
Table 2: Results of the small ordei example showing tainty model for the MACE flight,19 while the sensitivity
entries in the mean error matnces. method was developed after the flight to improve upon
stiffness mass inadequacies in the matrix method, i.e. the use of partial
entry Actual ] Local. entry Actual Local. information and improved localization of gravity errors.
1 -1 2.4 2.2 7-7 -2.6 -2.7 There were multiple configurations of the MACE test
2-2 2.4 2.2 8-8 -2.6 -2.7 article used during the flight. Figure 6(a) shows the
3-3 2.4 2.2 9-9 -7.7 -8.0 MACE test article in Configuration I in 1-g, while the re-
8-8 7.7 7.8 10- 10 -0.2 -0.2
9-9 2.6 2.6 11- 11 -0.2 -0.2 moval of the suspension system (and gravity) gives Con-
11 -11 7.7 7.9 12- 12 -0.1 -0.1 figuration I in 0-g. Figure 6(b) shows a second configu-
13-13 -2.6 -2.4 13-13 -2.6 0.0 ration of the test article used during the flight, termed
14-14 -7.7 -7.7 16-16 2.6 2.6 Configuration II. Two experiments are used to evaluate
16- 16 -4.9 -4.6 17-17 2.5 2.6 these methods: MACE 1-g system as the nominal to pre-
18-18 -7.7 -7.7 18- 18 2.9 2.8 dict the modified MACE 0-g uncertainty model; MACE
0-g Configuration I system as the nominal to predict the
modified MACE 0-g Configuration II uncertainty model.
Free-free FEM
Simulated free-free samples MACE Configuration I 1-g to 0-g
Projected free-free tneun
'rnjec-ted free-tree bound The first experiment examined is MACE Configuration
I
I 1-g, shown in Figure 6(a) to Configuration I 0-g (re-
a s.v.i I——•-——1°
moval of gravity and suspension from Figure 6(a)). For
MACE, this example was the basis for most controllers
developed prior to flight. In addition, because of the re-
R
moval of gravity and suspension degrees of freedom, this
prediction is quite a challenging task.
Details of this example are given in Table 3. The nom-
inal system is MACE Configuration I in 1-g. The 1-g
finite element model contains nN = 678 degrees of free-
0.97 0.99 1
Nonnali/*d to free-free FEM frequency dom, and 375 modes are retained. There are 16 data sets
Figure 5: Frequency results using the sensitivity method available for identification, producing rN = 17 measured
for the small order example. degrees of freedom (9 XY axes + 8 Z axis) and nN = 24
measured frequencies (14 XY axes + 10 Z axis). This
in the mass and stiffness matrices. In most cases, the gives 230 measured modal uncertainties in the MACE
physical errors are predicted quite well. An overall mass
mean error of 2.35% is reduced to 0.10% (for those en-
tries in error), and a stiffness mean error of 3.98% is
reduced to 0.15%. Figure 5 shows the FEM frequen-
cies, projected frequencies and bounds for four sample
frequencies, and 10 simulated cases for the free-free sys-
tem. The sensitivity method works quite well in predict-
ing both the mean errors and bounds, such that each
simulated case is within the predictions. This is a re-
sult of both the macroelements correctly spanning the
modeling error sources and using a pseudoinverse on an * Secondary Ptyload
overdetermined system. Note that because the correct
mass and stiffness uncertainties are predicted, the pro- (a) MACE Configuration I 1-g
jection method works quite well.

Experimental Verification
The Middeck Active Control Experiment (MACE) was
designed to be a reusable dynamics and control labo-
ratory for the investigation of issues associated with a
change in operational environment of a flexible space-
craft from ground to space based operations. An exten-
sive set of modeling and control experiments were per- Optical Encoder-
formed on the test article during 14 days of on-orbit
(b) MACE Configuration II 0-g
operations while on STS-67 in March, 1995.18 Note that
a matrix method was used to develop the 0-g uncer- Figure 6: The MACE test article.
1554
1-g uncertainty model (140 XY axes + 90 Z axis). Note o 0-B FEM
56.97
that this puts an upper limit on the number of macroele- x 0-Kduta
49.67 * Projected 0-g Meun
ments used in the uncertainty localization. I——I Projected 0-g Bound
The modified system is MACE Configuration I in 0-g. 49.18

The 0-g finite element model contains nM = 480 degrees 39.56

of freedom, and 280 modes are retained. There are 12 25.03


data sets available for identification, producing rN = 17 I7.K4
measured degrees of freedom (9 XY axes + 8 Z axis) and ,16.11
nN — 24 measured frequencies (14 XY axes + 10 Z axis).
9.46
The modified 0-g finite element model is developed by XX_1K X X_
3.73
removing the suspension degrees of freedom, and elimi- X >K X X
2.26
nating the nonlinear iterative technique to add the grav-
ity stiffening and initial deformation. The 1-g and 0-g O.K5 09 0.95 I 105 I.I
Nonnali/ed to 0-g FEM Frequency
finite element models are set up such that the mapping
between systems given by r/M = TMNr]N is Figure 7: MACE Configuration I 1-g to 0-g projection
using the sensitivity method.
rj-i __
-L MN —— *480 0'480-198 (30)
tensioning in the suspensions cables. These gravity ef-
There are 99 stiffness elements, 59 mass elements, 9 fects are also used in the development of macroelements
concentrated spring elements, and 39 concentrated in- at the sub-element level using the superposing adjust-
ertia elements, for a total of 206 elements in the nom- ments to the stiffness matrix.17
inal 1-g finite element model. Because there are 230 Although not pictured, the localized uncertainties are
measured nominal uncertainties, only two steps of un- primarily in both beam element and gravity parameters
certainty localization are needed: at the element and of the Lexan beam elements. Surprisingly there are few
sub-element levels. localized uncertainties in the suspension cables, where
Uncertainty localization at the sub-element level re- there are very large gravity effects modeled in the ten-
quires an examination of the elements. Because most of sioning of the cables. This may be a result of the lack of
the elements are beams with tubular cross-sections, there measured suspension modes and degrees of freedom.
are six physical uncertainties that can be used in develop- Figure 7 shows the frequency results of the projection
ing macroelements at the sub-element level: modulus of of both the 1-g mean error and variance uncertainties
elasticity (A.E), density (Ap), length (AL), radius (Ar), into 0-g for MACE using the sensitivity method. The
thickness (Af), and Poisson's ratio (Av). For this exam- prediction of mean errors and bounds is excellent for
ple, only.AS, Ap, and AL are used, giving two indepen- all but the 2.26 and 9.46 Hz modes. It is thought that
dent mass and two independent stiffness sub-elements. these modes are difficult to predict because of their cou-
The physical uncertainty At can be found by using a pling with the gravity effects. Note that in a few cases,
combination of A.E and Ap, and At/ is assumed to be there are one or two data points outside the predicted
negligible. The physical uncertainty Ar is very similar bounds. For these data points, the parameter estimate
to AE1 and Ap, with only small differences in the tor- had not converged in the identification. With more time
sional and rod degrees of freedom. Because the identified domain data, these estimates would be within the pre-
modes are bending and measured degrees of freedom are dicted ranges as well.
rotational, it is assumed that Ar can be reproduced us- To show the overall results of the uncertainty local-
ing a combination of AE and Ap sub-elements. ization and projection including mode shape uncertain-
In addition to the physical uncertainties, gravity ties, transfer functions are plotted. Figure 8 shows the
macroelements are also defined. The primary gravity transfer function from primary X gimbal to primary X
effects are gravity stiffening and initial sag in the Lexan rate gyro using the 0-g finite element model, the 0-g
beam elements; gravity stiffening in the gimbals; and finite element model shifted by the projected mean er-
rors, and three data sets taken on-orbit. Note that 0-g
Table 3: MACE testing summary. damping ratios are shifted by mean errors from -the 1-g
Property CI 1-g | CI 0-g / CII 0-g uncertainty model. Although similar, there are two im-
DOF (n ( .)) 678
portant areas of improvement of the shifted model. At
480
FEM modes 375 280 high frequency, the 56.97 Hz mode has been predicted
measured DOF (r(.)) 17(9XY,8Z) 17(9XY,8Z) correctly, while there is a 15% error in the finite element
measured freqs. (H(.)) 24(14XY,10Z) 24(14XY,10Z) model. Secondly, the residue of 17.84 Hz mode in the
data sets 16 12/4 shifted model matches that of the data, thus changing
stiffness elements 99 69 the sign of the finite element model residue.
mass elements 59 32 The process of localizing uncertainties for the MACE
cone, spring elem. 9 0 system was quite a difficult task for the MACE 1-g to
cone, mass elem. 39 33 0-g example. It is thought that the number and type of
1555
MACE 0-g Configuration I to Configuration II
The second experiment used to verify the uncertainty
-0-8 FEM
-0-g shifted model localization and projection methods is MACE 0-g Con-
0-g data
figuration I (removal of gravity and suspension from Fig-
ure 6(a)), to 0-g Configuration II, shown in Figure 6(b).
This example is much simpler than the previous 1-g to
0-g projection because there are no modeling of gravity
effects or suspension degrees of freedom.
Details this example are given in Table 3. The nominal
system is MACE Configuration I in 0-g. There are 12
data sets and 134 measured modal uncertainties in the
MACE 0-g uncertainty model (80 XY axes + 54.7, axis).
The modified system or MACE Configuration II is quite
I01 102
10"' 10°
Frequency (H/) similar, except for only 4 data sets.
Figure 8: MACE 0-g FEM, 0-g FEM shifted by pro- The Configuration II finite element model is developed
jected mean errors, and three sets of data. by rotating particular degrees of freedom of Configura-
tion I. The nominal to modified mapping of degrees of
freedom is therefore given by
sensors make it difficult for the sensitivity method to as-
certain the difference between a physical uncertainty in T — ^360 0360-120
an element, such as an error in the modulus, and a grav- -t MN — (31)
Ol20-360 ^~120
ity modeling uncertainty, such as an initial deformation.
In addition, the sensitivity method (and localization and where T^o is a rotation matrix.
update methods in general) work best for small errors in Figure 9 shows the frequency results of the MACE 0-g
Configuration I to Configuration II projection. Most of
the already defined parameters. Table 4 shows a sum-
mary of the mean changes in frequency of the data and the Configuration II data points lie within the mean er-
finite element model from 1-g to 0-g. For some modes rors and bounds predicted using the uncertainty local-
such as 4.60 Hz, the finite element model predicts the ization and projection methods. The frequencies which
changes quite well. However, for modes such as 17.82, do not (8.41 and 17.60 Hz) have only one point outside
the predicted ranges, and in each case, the parameter
24.89, and 56.16 Hz, the predicted change is incorrect
in sign and magnitude. This indicates a more thorough estimate again had not converged. Application of the
modeling of the gravity effects may be needed. uncertainty localization and projection methods to this
example was much simpler than the 1-g to 0-g example
In order to accurately model (and update) gravity ef- because of the elimination of gravity effects and suspen-
fects, a large number and different types of sensors must
sion degrees of freedom.
be used. Additional sensors on the test article nodes and Table 4 also shows the change between the Config-
suspension system would aid in the localization of errors
uration I and Configuration II frequencies of the finite
in the modeling of gravity effects. In addition, more user
element model and data. Notice that although the mag-
input into nonlinear iterative gravity stiffening procedure
nitude is not always correct, the general trends are very
is needed. Although it appears this procedure works well
good. This shows that the overall modeling of the MACE
for gravity stiffening in the cables and gimbals, it may
test article is very good, which makes it quite easy to lo-
need additional tuning for the gravity effects within the
calize and project the remaining small uncertainties.
test article that is not available.
Table 4: Data and finite element model frequency 49.87
o CII 0-g FEM

changes for both MACE examples. x CII 0-g dala


42.7(1 * Projected CII 0-g Mean-
I——[Projected CII 0-g Binjnil
1-g Freq. 1-g to 0-g (%) 0-g CI to CII (%) •£31.02
X
(Hz) FEM Data FEM Data £31.77

2.24 -1.1 -7.8 -4.6 -8.0 Sf25.97


u.
4.60 18.9 17.7 -53.0 -43.4 3 18.99
9.79 3.3 2.4 13.6 11.1 £17.60
16.19 0.4 4.1 -18.3 -17.8 G
8.41
17.82 -0.1 1.7 0.5 1.3
5.35
24.89 -0.5 2.9 -4.5 -3.7
2.45
39.41 -0.3 0.0 26.9 19.7
48.84 -0.6 4.2 8.8 13.1 0.95 1 1.05
rNurmaii/cvJ to
Nurmali/ed tti CII
*_u 0-g
u— g FEM Frequency
rc.p/i rrcqucncy
49.01 -1.3 -0.9 28.6 25.4
Figure 9: MACE 0-g Configuration I to Configuration
56.16 -1.4 7.8 0.5 12.4
II projection using the sensitivity method.
1556
Conclusions [5] Gysin, H., "Critical Application of the Error Matrix
Method for Localisation of Finite Element Model-
A set of methods to identify stochastic uncertainties be- ing Innacuracies," Proceedings, International Modal
tween a structural model and set of hardware in a nom- Analysis Conference, 1986, pp. 1339-1351.
inal environment, and predict an uncertainty model in a [6] Lallement, G., "Indirect Identification Methods I:
modified environment have been developed. A summary Adjustment of Methematical Models by the Results
of the many uncertainties that may exist in a structural of Vibrating Tests: Using Eigensolutions," Identifi-
system is given. The sensitivity method, because of its cation of Vibrating Structures, 1982, pp. 179-194.
ability to handle incomplete measured uncertainties, is [7] Piranda, J., Lallement, G., and Cogan, S., "Para-
used to localize uncertainties to specific degrees of free- metric Correction of Finite Element Models by Min-
imization of an Output Residual: Improvement of
dom. A practical algorithm is presented which limits the the Sensitivity Method," Proceedings, International
number of mass and stiffness uncertainties, and uses a Modal Analysis Conference, 1991, pp. 363-368.
layered approach to localize uncertainties in groups of [8] Fissette, E. Stavrinidis, C. and Ibrahim, S., "Er-
elements, then in elements, then in sub-elements. ror Location and Updating of Analytical Dynamic
Demonstration of the uncertainty localization and pro- Models Using a Force Balance Method," Pro-
jection methods on a sample problem showed the sensi- ceedings, International Modal Analysis Conference,
tivity method worked well in identifying both mean and 1989, pp. 1063-1070.
variance uncertainties in the mass and stiffness matri- [9] Gordis, J. H., "An Exact Formulation for Struc-
ces using partial information, i.e. a subset of measured tural Dynamic Model Error Localization," Pro-
ceedings, International Modal Analysis Conference,
modes and degrees of freedom. 1993, pp.159-167.
When applied to a 1-g to 0-g example on the Mid- [10] He, J., "Sensitivity Analysis and Error Matrix
deck Active Control Experiment, the methods again per- Method Using Measured Frequency Resonse Func-
formed quite well, but with small residual errors. This tion (FRF) Data," Proceedings, International Modal
is most likely caused by large errors in the modeling of Analysis Conference, 1993, pp. 1079-1082.
[11] Gysin, H., "Comparison of Expansion Methods for
gravity effects, and difficulties in localizing superposable
uncertainties such as physical and gravity modeling er- FE Modeling Error Localization," Proc,, Interna-
rors using the small number of sensors given. In another tional Modal Analysis Conf., 1990, pp. 195-204.
MACE example, the uncertainty model of a second "L" [12] Crawley, E. F., Barlow, M. S., van Schoor, M. C.,
configuration for MACE in 0-g was predicted, based on Masters, B. P., and Bicos, A. S., "Measurement of
the Modal Parameters of a Space Structure in Zero
the original straight configuration. The uncertainty lo- Gravity," AIAA J. of Guidance, Control, and Dy-
calization and projection worked very well for this exam- namics, Vol. 18, No. 3, May-June 1995, pp. 385-392.
ple, and did not have the residual errors as in the 1-g to
[13] Collins, J., Hart, G., Hasselman, T., Kennedy,
0-g example. The improved predictions are a result of B., "Statistical Identification of Structures," AIAA
the elimination of superposable gravity effects and sus- Journal, Vol. 12, No. 2, 1972, pp. 185-190.
pension degrees of freedom. [14] Hasselman, T. K., Chrostowski, J. D., and Ross,
T. J., "Interval Prediction in Structural Dynamic
Acknowledgments Analysis," Proceedings, AIAA Structures, Struc-
tural Dynamics, and Materials Conference, Dallas,
This work was supported by the NASA IN-STEP Pro- TX, Apr. 1992, pp. 1272-1284.
gram and NASA LaRC CSI Office with Mr. Gregory [15] Hasselman, T. K., Chrostowski, J. D., and Ross,
Stover and Dr. Jerry Newsom as Contract Monitors, T. J. C., "Propagation of Modeling Uncertainty
under contracts NAS1-18690 and NAS1-19622. through Structural Dynamic Models," Proceedings,
AIAA Structures, Structural Dynamics, and Mate-
rials Conference, Hilton Head, SC, Apr. 1994.
References [16] Campbell, M. E., "Identification and Parameter Es-
[1] Caesar, B., "Updating System Matrices Using timation for Control Design," Proceedings, IF AC
Model Test Data," Proceedings, International 13th World Congress, San Francisco, CA, July 1996.
Modal Analysis Conference, 1987. [17] Rey, D. A., Crawley, E. F., Alexander, H. L.,
[2] Link, M., "Identification and Correction of Errors Glaese, R. M., Gaudenzi, P., "Gravity and Suspen-
in Analytical Models Using Test Data - Theoretical sion Effects on the Dynamics of Controlled Struc-
and Practical Bounds," Proceedings, International tures," Proc., AIAA Structures, Structural Dynam-
Modal Analysis Conference, 1990, pp. 570-578. ics, and Materials Conf., La Jolla, CA, Apr. 1993.
[3] Maia, N., Reynier, M., and Ladeveze, P., "Er- [18] Miller, D. W., de Luis, J., Stover, G., and Crawley,
ror Localization for Updating Finite Element Mod- E. F., "MACE: Anatomy of a Modern Control Ex-
els Using Frequency Response Functions," Pro- periment," Proceedings, IFAC 13th World Congress,
ceedings, International Modal Analysis Conference, San Francisco, CA, July 1996.
1994, pp. 1299-1308. [19] Campbell, M. E., Grocott, S. C. O., How, J. P.,
[4] Sidhu, J. and Ewins, D. J., "Correlation of Fi- Miller, D. W., and Crawley, E. F., "Verification Pro-
nite Element and Modal Test Studies of a Practical cedure for On-orbit Controllers for the Middeck Ac-
Structure," Proceedings, International Modal Anal- tive Control Experiment," Proc., American Control
ysis Conference, 1984, pp. 756-762. Conference, Seattle, WA, June 1995, pp. 3600-3605.
1557

You might also like