Upload
others
View
2
Download
0
Embed Size (px)
Citation preview
University of Alberta
Nonlinear State-Space Control Design for Displacement-Based Real-Time Testing of Structural Systems
by
Seyed Abdol Hadi Moosavi Nanehkaran
A thesis submitted to the Faculty of Graduate Studies and Research in partial fulfillment of the requirements for the degree of
Master of Science
in
Structural Engineering
Department of Civil and Environmental Engineering
©Seyed Abdol Hadi Moosavi Nanehkaran
Spring 2011 Edmonton, Alberta
Permission is hereby granted to the University of Alberta Libraries to reproduce single copies of this thesis and to lend or sell such copies for private, scholarly or scientific research purposes only. Where the thesis is
converted to, or otherwise made available in digital form, the University of Alberta will advise potential users of the thesis of these terms.
The author reserves all other publication and other rights in association with the copyright in the thesis and, except as herein before provided, neither the thesis nor any substantial portion thereof may be printed or
otherwise reproduced in any material form whatsoever without the author's prior written permission.
Examining Committee Dr. Oya Mercan, Department of Civil Engineering and Environmental Dr. Alan Lynch, Department of Electrical and Computer Engineering Dr. Samer Adeeb, Department of Civil Engineering and Environmental
Abstract
This study presents the nonlinear design of a state space controller to
controlhydraulicactuatorsunderdisplacementcontrol,specificallyforreal‐
time pseudo‐dynamic testing applications. The proposed control design
processusesthenonlinearstatespacemodelofthedynamicsofthesystem
to be controlled; and utilizes state feedback linearization through a
transformationof the state variables.Comparisonsofnumerical simulation
resultsforlinearstate‐spaceandnonlinearstate‐spacecontrollersaregiven.
Alsorobustnessofthecontroldesignwithrespecttoidentifiedparametersis
investigated.Itisshownthatacontrollerwithimprovedperformancecanbe
designed using nonlinear state space control design techniques, provided
thatarepresentativemodelofthesystemisavailable.
Preface
Thereareseveraldynamictestingmethodsthathavebeenintroducedand
usedtoexamineand/orverifythedynamicperformanceofconventionaland
new structural systems. Some of thesemethods themselves are still under
research in an attempt to make themmore efficient and accurate. Among
these,pseudo‐dynamic(PSD)testingmethodandasitsextensionhybridPSD
testingmethod offer economical and practicalways to assess the dynamic
behaviourofstructuralsystems.Ifexecutedatfastrates(ideallyinreal‐time)
these testingmethods can handle load‐rate dependent structures (such as
theonesthathavedampersinstalledforseismichazardmitigationpurposes)
appropriately.
In aPSD test the equationofmotionof the test structure is solvedby a
direct step‐by step integration algorithm where the inertial and damping
forcecharacteristicsarekeptanalytical.Thesedisplacementsareimposedon
theteststructurebyhydraulicactuators;theresultingrestoringforcesfrom
the deformed structure are measured and fed back to the integration
algorithm for the computation of next step displacements. The method is
called hybrid PSDwhen thewhole structure is split into experimental and
analyticalsubstructurestoavoidfabricatingabigstructureinthelaboratory.
During a hybrid PSD test the command displacements generated by the
integration algorithm are imposed on the experimental and analytical
substructuresandtheresultsfrombotharecombinedandfedback.
Oneof themainchallenges inusingPSDmethod is thesensitivityof the
resultstotheexperimentalerrors.Thisisbecauseoftheclosedloopnature
of themethod; i.e. in each time step of the test procedure, the integration
algorithm uses themeasured information (whichmay be contaminated by
errors) from the previous step to generate the new displacements to be
applied.This fact implies that theerrors ineach stepareaffected from the
errorsofthepreviousstepswhichmaybecumulativelyaddedtogetheruntil
theendof thetest.AsaresultPSDmethodsuffers frompropagationof the
errorwhichatbestcausesaccuracyproblemsoratworstrendersthewhole
test unstable (i.e., the test needs to be aborted as the displacements grow
unboundedly).
From a stability (and also accuracy) point of view, the delay in the
measuredsignals (asopposedto lead,oramplitudeerrors)hasbeenfound
criticalinreal‐timetesting(MercanandRicles2008).Thedelaymainlyarises
due to the time it takes theactuators toreach thecommanddisplacements
issuedbytheintegrationalgorithm.
IngeneralaPIDcontroller(oramodifiedversion)isemployedtocontrol
the actuator displacements which may not provide acceptably accurate
tracking especially when large displacements need to be imposed at fast
ratesunderconsiderableload.Thismaybeduetothefactthatunderthese
typesofconditions,theservo‐hydraulicsystemnonlinearitiesareinvokedor
the test structure presents highly nonlinear behaviour. Hence, when
nonlinearities in the servo‐hydraulic system are invoked, theremight be a
needtousenonlinearstate‐spacecontrollertoaccountforthem.
Due to its matrix form, a state‐space control is preferable to a PID
controller in cases when physically coupled multiple degrees of freedom
needtobecontrolled.Themaineffortoftheresearchpresentedistodevelop
acontrolalgorithmbasedonadvancedcontroltheories(e.g.nonlinearstate‐
space) and assess its performance in comparison with a PID controller.
Simulations are done using a model of the servo‐hydraulic system that
accountsfornonlinearities.Thisstudyisintendedtoleadtheapplicationof
nonlinearstate‐spacecontrolstrategyinPSDtestingmethod.
Acknowledgement
Foremost, IwouldliketoexpressmysinceregratitudetomyadvisorDr.
OyaMercan for thecontinuoussupportofmyM.Sc.studyandresearch, for
herpersistence,interest,andhelp.Herguidancehelpedmeinallthetimeof
researchandwritingofthisthesis.Besidesmyadvisor,Iwouldliketothank
the rest of my thesis committee: Dr. Alan Lynch and Dr. Samer Adeeb for
theirinsightfulcomments,andhardquestions.
Last butnot the least Iwould like to thankmy family for their life‐time
support.
ContentsListofTables...........................................................................................................................1
ListofFigures.........................................................................................................................2
Acknowledgement................................................................................................................7
Preface.......................................................................................................................................4
1 Introduction.....................................................................................................................5
1.1 General......................................................................................................................5
1.2 SeismicTestingMethodsofStructures.......................................................5
1.3 PSDandHybridPSDtestingmethod...........................................................7
1.4 Researchgoalsandthesisorganisation....................................................13
2 ModellingoftheServo‐SystemandIdentificationofsystemparameters
(systemID)..................................................................................................................................15
2.1 ComponentsofaServo‐HydraulicSystem..............................................15
2.1.1 FlowControlServo‐Valve......................................................................17
2.1.2 LinearHydraulicActuator.....................................................................17
2.1.3 DisplacementTransducer......................................................................17
2.1.4 ServoController.........................................................................................17
2.2 DynamicsofaServo‐HydraulicSystem....................................................18
2.2.1 Servo‐valvedynamics..............................................................................18
2.2.2 ActuatorChamberPressureDynamics............................................25
2.2.3 PistonDynamics.........................................................................................26
2.2.4 LinearApproximationoftheDynamics...........................................26
2.3 SystemIdentification........................................................................................28
3 ControlTheory.............................................................................................................30
General................................................................................................................................30
3.1 FeedbackControlofDynamicSystems.....................................................30
3.2 LinearControlDesign(Basics).....................................................................33
3.2.1 LaplaceTransformandTransferFunction.....................................33
3.2.2 TheBlockDiagram....................................................................................37
3.2.3 S‐plane,PolesandZeros.........................................................................38
3.3 PIDControllerDesign.......................................................................................40
3.4 State‐SpaceControllerDesign......................................................................41
3.5 NonlinearStateSpaceControllerDesign.................................................44
4 ImplementationofControlMethods...................................................................50
General................................................................................................................................50
4.1 DynamicModelofaServo‐HydraulicSystem........................................50
LinearModel...............................................................................................................51
NonlinearModel........................................................................................................53
4.2 PIDControllerDesign.......................................................................................55
4.2.1 ControllerTuning......................................................................................57
4.3 LinearState‐SpaceControllerDesign........................................................58
4.3.1 State‐variableformofequations........................................................58
4.3.2 PolePlacement...........................................................................................60
4.4 NonlinearState‐SpaceControllerDesign................................................61
4.4.1 State‐variableformofequations........................................................61
5 SimulationsResults....................................................................................................70
General................................................................................................................................70
5.1 NumericalValuesforServo‐HydraulicSystemandTestStructure
inLinearandNonlinearModels....................................................................................71
5.2 ComparisonofControllersWithandWithoutSaturationLimits..72
5.3 ImplementingtheNonlinearState‐SpacecontrollerinaPSDTest
Simulation...............................................................................................................................77
5.4 RobustnessoftheControlDesign...............................................................85
5.5 Conclusion.............................................................................................................86
6 Testsetupdesign.........................................................................................................88
AppendixA.............................................................................................................................94
AppendixB SomeonDifferentialGeometry(Lynch2009).........................96
1 ChangesofCoordinatesorDiffeomorphisms........................................96
2 VectorFields.........................................................................................................97
3 DifferentialGeometryFunctionsUsedintheText...............................97
Liebrackets..................................................................................................................97
LieDerivative..............................................................................................................98
4 Distributions.........................................................................................................99
AppendixC MatlabCodesandSimulinkModels...........................................100
AppendixD ConferencePaper...............................................................................108
1
ListofTables
Table2‐1Nonlineardynamicsofaservo‐hydraulicsystemanditslinear
approximation
Table4‐1Linearizeddynamicsofaservo‐hydraulicsystem
Table4‐2Nonlineardynamicsofaservo‐hydraulic
Table5‐1Values for Parameters for Linear and Nonlinear Servo-Hydraulic
System Models
Table5‐2Approximatevaluesfortheelastomericdamperparameters
2
ListofFigures
Figure1‐1Structuresubjecttogroundaccelerationinarealearthquake
Figure1‐2PSDtestingmethod
Figure1‐3HybridPSDtestingmethod
Figure2‐1BlockdiagramofinnerloopinPSDtestmethod
Figure2‐2Crosssectionofatwo‐stageflowcontrolvalve
Figure2‐3Crosssectionofatwo‐stageflowcontrolvalveatoperation
Figure2‐4Turbulentflowthroughanorifice
Figure3‐1 anexampleforaunitstepresponseofasystem
Figure3‐2 Blockdiagramofaservohydraulicmodelincluding
disturbanceandnoise
Figure3‐3 amass‐damper‐springsystem
Figure3‐4 Threeexamplesofelementaryblockdiagrams
Figure3‐5 Timefunctionassociatedwithpointsinthe ‐plane(Franklin
etal.2010)
Figure3‐6 BlockdiagramofaPIDcontroller
Figure4‐1 Simulinkmodelforaservo‐hydraulicsystemwithlinearized
dynamics
3
Figure4‐2 Simulinkmodelforaservo‐hydraulicsystemwithnonlinear
dynamics
Figure4‐3 Simulinkmodelforaservo‐hydraulicsystemwithaPID
controller
Figure4‐4 Approximationoffunction
Figure5‐1 Stepresponsesforasystemwithoutsaturation
Figure5‐2 Stepresponsesforasystemwithsaturation
Figure5‐3 Responsetosinusoidinput
Figure5‐4 loadpressurevariationwithoutservo‐valvesaturation
Figure5‐5 loadpressurevariationwithservo‐valvesaturation
Figure5‐6 Servo‐valveopeningvariationwithoutservo‐valvesaturation
Figure5‐7 Servo‐valveopeningvariationwithservo‐valvesaturation
Figure5‐8 HybridPSDtestmethod(Mercan2007)
Figure5‐9 Simulinkmodelforreal‐timehybridPSDtestingwithMDOF
analyticalsubstructureandanelastomericdamperastheexperimental
substructure
Figure5‐10 Simulinkmodelforaservo‐hydraulicsubsystemwithaPID
controller
Figure5‐11 Simulinkmodelforaservo‐hydraulicsubsystemwithaNL
state‐spacecontroller
Figure5‐12 comparisonbetweencommandandmeasureddisplacement
(a)PIDcontroller(b)NLcontroller
Figure5‐13 comparisonbetweencommandandmeasureddisplacement
4
(a)PIDcontroller(b)NLcontroller
Figure5‐14 comparisonbetweenservo‐valvespoolopening(a)PID
controllerwithavelocityfeedforward(b)NLcontroller
Figure5‐15 effectof10%errorin ontheresponse
Figure6‐1ExperimentalsetupforPSDtest
Figure6‐2Expectedbaserotationatbaseusedindesign
5
1 Introduction
1.1 General
This chapter contains background information about seismic testing
methods.Themotivation andobjectiveof the research andorganizationof
thedissertationarealsodescribed.
1.2 SeismicTestingMethodsofStructures
Oneofthemaingoalsinseismicperformancetestingistoimposeloading
conditions on a test specimen that are representative of those that might
happen during a real earthquake. To achieve this goal, various forms of
earthquakeandstructuraldynamictestingmethodshavebeenthesubjectof
research.
Four experimental laboratory techniques are typically used in seismic
performancetestingofstructures:quasi‐statictesting,shakingtabletesting,
effectiveforcetesting(EFT)andpseudo‐dynamic(PSD)method
Inaquasi‐statictest,apredefinedcyclicdisplacementhistoryisappliedto
the structure or structural component under study and the behaviour is
observedandanalyzed.Thepredefineddisplacementhistory,ifnotselected
6
from some typically used displacement histories, is based on a response
computedfromadynamictimehistoryanalysispriortotesting.Thismethod
is commonly used and economical. However, it is limited in terms of
delivering the true earthquake response. This is because the model with
whichthepre‐testdynamicanalysisisperformedmaynotaccuratelypredict
thebehaviourand in turn, theresultingdisplacement timehistorymaynot
correspondtotherealearthquakeresponse.
Placing a structure on a shaking table and exerting a properly scaled
ground motion may be the most realistic method. In spite of this, due to
payloadrestrictionsofshaketables,shakingtabletestsareimplementedon
small‐scaleteststructures.Thisimpliesthatthegroundaccelerationneedsto
be scaled (compressed) accordingly. As a result the available time for
observing the behaviour will be very little during the test. Generally
speaking,despitethefactthattheshakingtablemayberepresentativeofthe
actual seismicbehaviour, thecombinedeffectsof theneed toconstruct the
complete structure, small scale test specimens, short observation time and
finallythecostlimittheuseofshakingtable.
Effective force testing (EFT) is a real‐time testingmethod. Thismethod
usesaforcecontrolapproachandcanbeemployedinreal‐timeearthquake
simulationoflargescalestructures.Knowingthestructuralmassandground
accelerationhistory,thecompleteforcehistorythatshouldbeappliedtothe
structure is calculated beforehand. Despite the conceptual simplicity, the
7
implementationof thismethodhasbeenobserved tobechallengingdue to
actuator‐structureinteraction(Dimigetal.1999).Toovercomethisproblem,
Zhao (2003) proposed a nonlinear velocity compensation scheme and
verified it through simulations and experimental studies under limited
conditions.
In the late 1970s and early 1980s PSD method was initiated as an
experimentaltechniqueinwhichthedisplacementresponseofstructuretoa
given ground acceleration is numerically calculated and quasi‐statically
imposedon the structure (Takanashi et al. 1975;Okada et al. 1980;Mahin
andWilliams1981;ShingandandMahin1983;MahinandShing1985).This
computed response is based on analytically predefined inertia and viscous
damping aswell as the experimentallymeasured structural resisting force.
Detailsfortheprocedurearegivenbelow.
1.3 PSDandHybridPSDtestingmethod
InaPSDtest, the teststructure is first idealizedasadiscreteparameter
system.Therebyforastructuresubjectedtogroundacceleration(Figure1‐1)
thegoverningequationsofmotioncanbeexpressedasa systemof second
orderordinarydifferentialequationswithrespecttotime.
, , Eq.1‐1
re
re
ex
th
a
F
1
(s
b
Fig
In the ab
espectively;
espectively;
xternal (eff
hegivengro
reassociate
orastep‐by
isused
Ateachst
seeFigure1
y‐stepinteg
gure1‐1Struc
bove equati
; and
; isthe
fective) load
oundaccele
edwiththe
y‐stepsimu
tepofacon
1‐2)theab
grationalgo
cturesubjectt
ion and
d are
resistingfo
d vector th
eration.All
degreesof
lationorte
nventional(
oveequatio
orithmtoge
,
ogroundacce
are the
e the accel
orcevector
at is obtain
ofthematr
freedom(D
sting(e.g.P
(slowandin
onofmotio
eneratethec
,
elerationinar
mass and
leration an
inthestruc
ned using th
ricesandve
DOFs)ofthe
PSD)adiscr
nextended
n issolved
commandd
realearthquak
damping
nd velocity
ctureand
hemassma
ectorsdefin
eidealizeds
etizedform
timescale)
usingadir
displacemen
8
ke
matrices,
y vectors
isthe
atrix and
edabove
structure.
mofEq.1‐
Eq.1‐2
PSDtest
rectstep‐
nts.
fe
a
g
d
st
d
in
is
m
th
n
a
These tar
eedbackcon
re measure
enerate tar
epending o
tructure, th
ependent),
nthemeasu
spaidtomin
Themain
measuredex
he integrat
onlinear ch
ccountedfo
integratio
rget displac
ntrolquasi‐s
ed and fed
rget displac
on the expe
he measur
dampingo
uredrestori
nimizethem
n feature of
xperimental
ion algorith
haracteristic
or.
onalgorithm
Figure1‐
ements are
statically.T
d back to t
cements for
eriment rate
ed restorin
rinertialfo
ngforceisn
movingmas
f PSD testin
llyfromthe
hm in com
cs associate
command
measured
‐2PSDtesting
e imposed b
Thentheres
the integra
r the next
e and dyna
ng force m
orces.Captu
notdesirab
ssintheexp
ngmethod i
edeformed
mmand gene
ed with the
ddisplacemen
d restoringfor
actuato
gmethod
by hydrauli
sistingforce
ation algori
step. It sh
amic charac
may includ
uringinertia
ble,thatisw
perimentalt
is that the
teststructu
eration. Th
e resisting
nts
rces
or
ic actuators
esfromthes
thm to be
hould be no
cteristics of
de stiffness
alforcecon
whyspecial
testset‐up.
resisting fo
ureandare
is assures
forces are
l
9
s under a
structure
used to
oted that
f the test
(strain‐
ntribution
attention
orces are
eusedby
that any
properly
loadcell
ex
fo
is
m
h
d
su
co
q
w
n
Hybrid P
xtensionof
orwhichar
s referred t
modeledina
ybrid PSD
isplacemen
ubstructure
ombinedan
In a conv
uasi‐statica
withmateria
otbeaccur
integratioanalyticalmodel
SD test me
f PSDtesting
reliableanal
to as the ex
acomputer
test the tw
nts generate
essimultane
ndfedback
ventional PS
ally (i.e. in a
al propertie
atelycaptur
onalgorithm&lsubstructure
Figure1‐3Hy
ethod (Figu
g.Inthism
lyticalmode
xperimental
(andreferr
wo substruct
ed by the in
eously,and
totheinteg
SD test, sin
an extended
es or comp
red.Exampl
command
measured&
ybridPSDtest
ure 1‐3) is
methodonly
elmaynote
l substructu
redtoasth
tures are co
ntegration a
therestorin
grationalgor
nce the targ
d time scale
onents that
lesofthese
displacements
restoringforc
tingmethod
s an econo
aportiono
exististeste
ure), and th
heanalytica
oupled toge
algorithm a
ngforcesco
rithm.
get displace
e), the beha
t are load‐r
load‐rated
ts
ces
analyt
experime
mical and
ofthetests
edphysicall
he remainin
alsubstructu
ether; the c
are imposed
omingfrom
ements are
aviour of st
rate depend
dependents
ticalsubstructu
entalsubstruc
10
practical
structure,
ly(which
ngpart is
ure).Ina
command
d to both
mbothare
imposed
tructures
dentmay
structural
ure
cture
11
components and devices are metallic, friction, visco‐elastic, viscous fluid,
tuned mass, tuned liquid, elastomeric dampers and lead rubber bearings
whichare introduced intostructures forvibrationmitigationpurposes.For
assessing the capabilities of structures equipped with these devices, it is
necessarytoperformthePSDtestdynamicallyatratesapproachingtoreal‐
time.
In PSD testing method the results are sensitive to experimental errors.
Thisisbecausethemethodhasaclosedloopalgorithm;i.e.ineachtimestep
of the test procedure, the integration algorithm uses the measured
informationfromthepreviousstep(whichmaybecontaminatedbyerrors)
togeneratethenewdisplacementstobeapplied.Thismeansthattheerrors
in each step are affected from the errors of the previous steps which are
cumulativelyaddedtogetheruntil theendofthetest.HencePSDmethod is
pronetopropagationoftheerror,whichatbestcausesaccuracyproblemsor
atworstrendersthewholetestunstable(i.e.,thetestneedstobeabortedas
thedisplacementsgrowunboundedly).Mercan(2007)showedthatactuator
delay in following the command displacements in experimental setupmay
impairthedynamicstabilityofthetestsetupinareal‐timePSDtest.Thisis
alsotrueforreal‐timehybridPSDmethod.
Oneofthemainsourcesoftimedelayisthetimeittakestheactuatorto
reach the command displacement calculated by the integration algorithm
which cannot be done instantaneously. This is because it takes time to
12
convert energies (electrical to mechanical) in an experimental setup.
Therefore it is critical touse theavailableequipmentefficientlyso that the
corresponding time delay is minimum. This is done by using the control
theorytodesignservo‐hydrauliccontrollers.Thecommonlyusedcontroller
in PSD tests is the well‐known PID (proportional‐Integral‐Derivative)
controller.PIDcontrollersarepopularbecausetheycanbeadjusted(tuned)
tocontrol complexsystemswithout theneed for completeknowledgeover
thedominatingdynamics.
Conducting a fast (ideally real‐time)PSD test requires accurate actuator
control by means of sophisticated servo‐hydraulic strategies and reliable
computationschemethroughefficientintegrationalgorithms.
Thedynamicsofanelectrohydraulicservosystemishighlynonlinearand
involvessignandsquarerootfunctions.Howeverinmostindustrialcontexts
linearcontroltheoryisused.Althoughlinearizationaboutanoperatingpoint
decreases design effort, it degrades the performance at regions off the
operating point. There have been several works on controlling
electrohydraulicservosystemsusingadvancedcontrolmethods.
Lim(2002)applied linearizationandpoleplacement.YanadaandFuruta
(2007) combined linear theory with an adaptive approach. Kwon et al.
(2007) applied full‐state feedback linearization. Seo et al. (2007) used
feedback linearization to design controllers for displacement, velocity and
differentialpressurecontrolofarotationalhydraulicdriveandAyalewand
13
Jablokow(2007)usedpartial‐inputfeedbacklinearizationforthecontrolof
electrohydraulic servo systems. Mintsa et al. (2009) used feedback
linearization to design an adaptive control for electrohydraulic position
servo system with the objective of enhancing robustness with respect to
variationsofsupplypressure.
1.4 Researchgoalsandthesisorganisation
Inordertoimprovethetrackingcapability,andinturntheaccuracyofthe
overall real‐time PSD test results, this thesis investigates the design and
implementationofadvancedcontrolalgorithms.
As stated before, currently the majority of the controllers used in PSD
testing arePIDbased controllers. They are tuned according to a linearized
modelofthesystemdynamics(forsingleinputsingleoutputsystems),orby
ad‐hoc tuning; where the accurate window of operation is limited to the
linearrangeofthesystemdynamics.Inthecaseofamulti‐degreeoffreedom
test structure, state‐space design offers considerable reduction in control
designeffort(Mercanetal.2006).Theabovementionedmethods,astheyare
designedbasedonlinearmodelsofthesystem,maynotprovideacceptably
accurate tracking when the servo‐hydraulic system nonlinearities are
invokedortheteststructurepresentshighlynonlinearbehaviour.
Inthisstudythedesignofanonlinearstate‐spacecontrollerforrealtime
pseudo‐dynamic testing of structural systems is presented based on
advancedcontroltheoriesusinganonlinearmodelofthesystem.
14
After the introductioninthischapter,Chapter2considersthegoverning
dynamicsandequationsofaservo‐hydraulicsystem.Chapter3beginswith
basicsofcontrol theoryandafterwards linearandnonlinearcontroldesign
methodsareintroducedanddiscussed.Chapter4appliesthecontroldesign
methods introduced in Chapter 3 on the dynamics discussed in Chapter 2.
Finally,Chapter5 illustrates the implementationof thecontrolmethodsby
computersimulationandcomparesthebehaviourofdifferentcontrollers.
Chapter6discussesthedesignofasingledegreeoffreedomtestsetupto
investigate different aspects of a PSD test method (e.g. applying different
controllers).Aconferencepapertitled“ModificationofIntegrationAlgorithm
toAccountforLoadDiscontinuityinPseudo‐DynamicTesting”thatwasdone
alongwiththisresearchispresentedasanappendix.
15
2 Modelling of the Servo‐System and Identification of
systemparameters(systemID)
Toworkoutanewcontrolapproachwithanimprovedperformancefora
servo‐hydraulic system, the behaviour of the system has to be well
understoodthroughsimulationofrepresentativemodels.Thesemodelsare
based on the important dynamics that govern the behaviour and include
physicalparameterssomeofwhichmayalreadybeknown(e.g.pistonarea)
while others need to be identified through system identification. While
decidingforthelevelofaccuracy(complexity)ofthemodel,oneneedsalso
toconsideriftheparametersthatappearinthemodelareeasilyobtainable
throughsystemidentificationornot. Inthenextsectiondifferentpartsofa
servohydraulicsystemandthegoverningdynamicsareelaborated.
2.1 ComponentsofaServo‐HydraulicSystem
Apositioncontrolledservo‐hydraulicsystemused inaPSDtesttypically
consists of a hydraulic power supply, a flow control servo‐valve, a linear
actuator, a displacement transducer, and an electronic servo‐controller.
Figure2‐1showsablockdiagramthatrepresentshowtheabovementioned
parts are interconnected. The controller compares the command
displacement with the measured displacement coming from the
displacementtransducer(e.g.anLVDT)todeterminethepositionerrorand
16
thensendsoutacommandsignaltodrivetheflowcontrolservo‐valve.Infact
the command signal introduces a spool displacement in the servo‐valve to
adjust the flow of pressurized oil from the hydraulic power supply to the
linearactuatorchambersinordertomovetheactuatorpistontothedesired
position.
Figure2‐1BlockdiagramofinnerloopinPSDtestmethod
ThefollowinggivesaconciseexplanationofthepartsshowninFigure2‐1.
HydraulicPowersupply
A hydraulic power supply provides the pressurized fluid (oil) for the
hydraulic system. The level of oil pressure in the power supply is selected
considering several factors. Low pressure systems have less leakage but
physicallylargercomponentsareneededtoprovideaspecifiedforce.Onthe
otherhand,highpressuresystemsexperiencemoreleakagebuthavebetter
ServoController
LinearActuator
LoadFlowControlServo‐Valve
DisplacementTransducer
HydraulicPowerSupply
CommandDisplacement
MeasuredDisplacement
AppliedDisplacement
17
dynamicperformanceandhavesmaller(lighter)components. Inmanyhigh
performancesystems3000psi(210bar)isselectedforthesystempressure.
2.1.1 FlowControlServo‐Valve
An electro hydraulic flow control servo valve is a servo valve which is
designedtoproducehydraulicflowoutputproportionaltoelectricalcurrent
input.Themechanismanddynamicsarediscussedlater.
2.1.2 LinearHydraulicActuator
A hydraulic actuator converts hydraulic energy to mechanical force or
motion.Theyareimplementedwherelargeactuationforcesandfastmotion
arerequired.Governingdynamicsaregivenlater.
2.1.3 DisplacementTransducer
They generally come built‐in with actuators and are often attached
directlytothepistonrod.Therearevarioustypesoffeedbacktransducersin
useincludinginductivelinearvariabledifferentialtransformer(LVDT).Itisa
common practice to include external displacement transducers in the test
setuptocheckthemeasurementsoftheinternaltransducersandtoexclude
theactuatorsupportmotion.
2.1.4 ServoController
A controller continuously compares the command displacement against
the actuator position that is measured by a displacement transducer. The
resultofthiscomparisonisdisplacementerrorwhichisthenmanipulatedby
18
acontrol law inorder togenerateandsendacommandsignal to theservo
valve.
2.2 DynamicsofaServo‐HydraulicSystem
2.2.1 Servo‐valvedynamics
Servo‐valvesareusedinservo‐hydraulicsystemstoconverttheelectrical
command signal, coming from the controller, to a spool displacement. This
displacement alongwith the differential pressure between the servo‐valve
ports results in oil flow through valve control ports into and out of the
actuatorchamberenablingthemotionofthehydraulicpiston.
Amongseveral typesofservo‐valvesarethe flowcontrolservo‐valves in
which the control flow at constant load pressure is proportional to the
electricalinputcurrent(Thayer1958,revisedin1965).
Figure 2‐2 shows a two stage flow control servo valve. It is called two‐
stage as it has two portions containing a hydraulic amplifier. A hydraulic
amplifierisafluidvalvingdevicewhichactsasapoweramplifier,suchasa
slidingspoolandanozzleflapperandwillbeelaboratedinthischapter.
to
d
re
d
In a two‐
orque moto
eflection o
estrictsflui
isplacethe
arm(torqu
spool
Figure2‐2
stage servo
or coils crea
of armature
dflowthro
spool.
matureuemotor)
acch
topowersupply
Crosssection
o‐valve the
ates a mag
e‐flapper a
oughoneof
nozzle
to1stctuatorhamber
nofatwo‐stag
electrical c
netic force.
ssembly w
thetwono
toreturntank
geflowcontro
command s
. This magn
where the
ozzlesandr
a ppe
m
to2ndactuatorchamber
lvalve
ignal applie
netic force
resulting d
redirectsth
pairofermanent
magnets
flapper
19
ed to the
causes a
deflection
heflowto
co
o
sp
fl
b
a
st
le
th
Figu
Spool dis
ontrol port
ther contro
pringshow
lapper asse
ecomesequ
ssemblymo
tate of equ
evel(Moog2
In summa
he input c
ure2‐3Cross
placement
s and at th
ol port. Th
ninFigure
embly. Spo
ual to the t
ovesbackto
ilibrium un
2010).
ary, aspoin
current. Als
sectionofatw
opens the
he same tim
e displaced
2‐3thatcr
ol moveme
orque from
otheneutr
ntil the elec
ntedoutbef
so with co
wo‐stageflow
supply pre
me opens th
d spool app
reatesarest
ent continu
m themagne
alposition
ctrical comm
fore, the sp
onstant pre
a
controlvalve
essure port
he return ta
plies a forc
toringtorqu
ues until t
etic forcea
andthespo
mand signa
poolpositio
essure drop
deflectedarmature‐flap
assembly
deflectedcantileversp
e atoperation
t ( ) to on
ank port (
ce to the c
ueonthea
the restori
nd that isw
oolisheldo
al changes t
n ispropor
p across t
pper
dring
20
ne of the
) to the
cantilever
armature‐
ing force
when the
openina
to a new
rtional to
he valve
21
(constant loadpressure)theloadflowisproportionaltothespoolposition.
The governing dynamics of the servo valvewill be discussed in two parts;
valve spool dynamics which includes the relationship between the input
currentandthespooldisplacementandvalveflowdynamicswhichexplains
howthespooldisplacementrelatestotheflowfromthevalvetotheactuator
chambers.
2.2.1.1 ValveSpoolDynamics
Servo‐valves are complicated devices. Experience has shown that their
nonlinearandnon‐idealcharacteristicsmakeithardtotheoreticallyanalyze
servo‐valvedynamics in systemsdesign. Instead, it ismore convenientbut
alsoacceptablyaccuratetoapproximateservo‐valvedynamicswithsuitable
empirical transfer functions by using measured servo‐valve response
(Thayer 1958, revised in 1965). Depending on the frequency range of
interest the servo‐valve dynamics can be represented by a first order
transferfunction.
1
Eq.2‐1
Where , , and are servo‐valve spool opening (see Figure 2‐3),
differentialcurrentinputtoservo‐valve,servo‐valvestaticflowgainatzero
load pressure drop and apparent servo‐valve time constant. It should be
noted thaton the lefthand sideofEq. 2‐1 theLaplace transformsof spool
openingandinputcurrentareused.
2
se
T
fl
fl
n
2.2.1.2 Valv
Ascanbe
ervo‐valvea
Theseflows
lowthrough
2
, , a
lowingliqui
Rewriting
owthat
veFlowDyn
seeninthe
andtheactu
Figu
areclassifie
hanorifice(
2
and
idanddiffer
g the above
isthesame
2
2
namics
Figure2‐3,
uatorports.
ure2‐4Turbul
edasturbul
(seeFigure
aredischa
rentialpres
e turbulent
forallorifi
,fourflows
lentflowthrou
lentflows.E
2‐4)
argecoeffici
surebetwe
orifice flow
cesgives
canbereco
ughanorifice
Eq.2‐2repr
ient,orifice
enpoint a
w for Figur
ognizedbet
e
resentsthet
earea,dens
and respe
re 2‐3 assu
22
tweenthe
turbulent
Eq.2‐2
ityofthe
ectively.
uming for
Eq.2‐3
Eq.2‐4
23
2
Eq.2‐5
2
Eq.2‐6
Intheaboveequations, , istheorificeareaassociatedwith
flow , , and is thesupplypressure.Also and are the
pressuresateachoneoftheactuatorchambers.
The return pressure ( ) is assumed to be zero as it is usually much
smaller than the other pressures involved. If the return pressure is not
negligible, thesupplypressure in theaboveexpressionscanbe interpreted
assupplypressureminusreturnpressure.
Theareasoftheorificesarefunctionsofthespoolopening .Becausethe
valveorificesarematched,
and Eq.2‐7
Andbecausetheyaresymmetrical,
and Eq.2‐8
Itcanbeshownthatwhentheorificeareasarematchedandsymmetrical
and (Merritt1967) Eq.2‐9
24
CombiningEq.2‐3,Eq.2‐4,Eq.2‐7andEq.2‐9pressuresattwochambers
canberelatedtosupplypressureby,
Eq.2‐10
Bydefinitiontheloadpressureisthepressuredifferencebetweenthetwo
actuatorchambersandisexpressedas
Eq.2‐11
UsingEq.2‐10andEq.2‐11, and maybewrittenas
2
Eq.2‐12
2 Eq.2‐13
Load flow, which is the flow from the valve to one of the actuator
chambers,canbeexpressedas
Eq.2‐14
Finallyusingthederivedequationsforamatchedandsymmetricalvalve,
theloadflowmaybewrittenas
1 1
Eq.2‐15
25
For an ideal critical center valvewithmatched and symmetrical orifices
the leakage flows ( and when the spool displacement is positive) are
zerobecausethevalvegeometry isassumedideal.Ontheotherhandwhen
the spool displacement is negative and will be the leakage flows.
Considering the fact that the orifice areas are linear functions of the spool
openingastheproductof thespoolopeningandfullperipheryofthespool
( ),loadflowcanbeexpressedas(Merritt1967)
Eq.2‐16
where isdefinedas
1
Eq.2‐17
Eq 2‐16 is not differentiable due to sign function. Sign functionmay be
approximatedbysomesmoothfunctionandthenusedinthecontroldesign
(Chapter4)
2.2.2 ActuatorChamberPressureDynamics
The oil flow between valve and actuator chambers causes the actuators
piston tomove. Using conservation ofmass principle on both sides of the
actuatorchambers,theactuatorpressuredynamicscanbeexpressedas
4
Eq.2‐18
26
Inthisequation istheactuatorpistoncross‐sectionalarea; ispiston
velocityorthetimederivativeofpistondisplacement( ); istheleakage
coefficientofpiston; and areloadpressureasdefinedinEq.2‐11and
rateofloadpressurerespectively; isactuatorchambervolumeand isoil
modulus.
2.2.3 PistonDynamics
Writingtheforceequationofmotionandconsideringastaticfrictionand
theexternalforcefromatestspecimengives
Eq.2‐19
, and are piston mass, viscous damping coefficient of actuator
piston and static friction respectively and represents the effect of
stiffness,dampingandinertialforcesfromatestspecimen.
2.2.4 LinearApproximationoftheDynamics
Forcontrollerdesignpurposesalinearized,simplifiedmodelisnecessary
especiallywhenthecontrollerdesignwillbebasedonlinearcontroltheory.
Note that the valve spool dynamics (Eq. 2‐1) and actuator chamber
pressure dynamics (Eq. 2‐18) are already linear. Using Taylor series
expansion, Eq. 2‐15 (that represents flow‐pressure relationship) can be
linearizedaboutanoperatingpoint( 0)whileassumingzero
leakageflowandidealgeometry(Merritt1967)tobe
27
Eq.2‐20
and are called the flow gain coefficient and the flow pressure
coefficient respectively. And for the pistondynamics, in case the friction is
negligibleEq.2‐19canberewrittenas
Eq.2‐21
Table 2‐1 gives a summary of both nonlinear and linear equations
discussedsofar.
Table2‐1Nonlineardynamicsofaservo‐hydraulicsystemanditslinearapproximation
Dynamics Nonlinear LinearValvespool
1
(Eq.2‐1)
1
(Eq.2‐1)
Valveflow
(Eq.2‐16)
(Eq.2‐20)
Actuator
chamber
pressure
4
(Eq.2‐18)
4
(Eq.2‐18)
Actuator
motion
(Eq.2‐19)
(Eq.2‐21)
28
2.3 SystemIdentification
It is a good idea to identify the system starting from the simplified
(linearized) model and finalize it by adding relevant nonlinearities and
additionaldynamicsasrequiredbythemeasuredresponse.
Grey‐boxmodeling option of the system ID toolbox ofMATLAB is ideal
when the model to be identified is known (e.g. derived from the first
principles) and the numerical values of the parameters that appear in this
modelneedtobeestimatedfrommeasureddata.
In identifying the servo‐hydraulic systemmentioned above the transfer
functionrelatingtheinputcurrentandthespoolopeningisknowntobe
1
Eq.2‐22
For a free standing actuator that have equal to zero, a transfer
functioncanbeestablishedusingEq.2‐18,Eq.2‐20,andEq.2‐21which in
turnwheniscombinedwithEq.2‐22willgiveadirecttransferfunctionfrom
inputcurrenttoactuatorpistondisplacement.
4
4 4 Eq.2‐23
where and 1 ⁄ .
MATLAB identification toolbox has a general built‐in systemmodel that
canbeadjustedtohavefourpolesandnozeroesas
29
1 2 1
Eq.2‐24
Theterm isconsideredtotakecareofanytimedelaythatmayexist
inthesystem.
Theidentificationprocedurebasicallystartswithexcitingthesystemwith
somepredefinedinputsandloggingtheresponse.Thereisnorestrictionfor
selecting the inputs but usually step and sinusoid inputs with different
frequencies are used. Then the logged data is analyzed by MATLAB
identificationtoolboxandvaluesfortransferfunctionparametersarefound.
ComparingEq.2‐23andEq.2‐24andconsideringthefactthat
hasa
verysmallvalueandcanbeneglected(Zhangetal.2005),itcanbeshown
4
Eq.2‐25
4
Eq.2‐26
1 Eq.2‐27
Eq.2‐28
30
3 ControlTheory
General
In a servo‐hydraulic system the controller calculates the displacement
errorandusesitasaninputtoacontrollaw.Theoutputofthecontrollawis
the command signal to the servo valve. In this chapter some principles of
feedback control theory is given. Then classic control design, linear state‐
space control design and nonlinear state‐space control design are
introduced.
3.1 FeedbackControlofDynamicSystems
In thecontextof thisstudyacontroller isdesignedtogivethe following
characteristics to the system; (1) the ability to follow command
displacements(tracking);(2)theabilitytomaintainthesystemstability;(3)
the ability to reduce the sensitivity of the system to external disturbances
(disturbancerejection).Somecriterianeedtobedefined inordertohavea
basisof comparisonbetween theperformancesofdifferent controllers.For
this purpose specific input signals like step functions and sinusoidal
functionsareused.Figure3‐1showsatypicalresponseofasystemtoaunit
st
a
re
e
ch
p
d
th
a
ca
w
tepinput.A
n input it
eachedafte
For the ca
ssential fa
haracteristi
erformance
elaytaking
he resisting
mplitudeer
auses inacc
willoccur.
Aphysicalsy
cannot inst
rexhibiting
Figure3‐1
ase of aPS
ctor for th
ics and ste
ecriteria.M
thesystem
g force) has
rror.Phase
curacyand
transienrespons
ystemstore
tantly follo
gatransient
anexample
D test, othe
he system
eady‐state
Mercan(200
mtoapplyth
smuchmor
errorintrod
if it goesbe
nte
esenergyth
w it. Instea
tresponse.
eforaunitste
er than stab
to be op
error are
07)showed
hecommand
re of a detr
ducesaneq
eyonda cri
steares
hereforewh
ad a steady
epresponseo
bility requir
perational,
also consid
dthatphase
ddisplacem
rimental eff
quivalentne
itical value,
adystatesponse
steady stateerror
henitissub
y state con
fasystem
rement,wh
transient
dered as c
eerror(i.e.
mentandto
fect compar
egativedam
dynamic in
e
31
bjectedto
ndition is
hich is an
response
controller
the time
measure
red to an
mpingthat
nstability
32
Figure 3‐2 gives a block diagram of the system including the possible
disturbancesandnoises.Itshouldbenoticedthatexternaldisturbancesmay
alsobepresentinindividualcomponents(e.g.,theservovalve)butinorder
tobeconciseinFigure3‐2,onlytheresultantofthedisturbancesisshown.
Figure3‐2 Blockdiagramofaservohydraulicmodelincludingdisturbanceandnoise
Thedynamicsofasystemcanbedefinedbyaset(system)ofdifferential
equations which are obtained from principles of physics. For a quick
approximateanalysisandforcontrollerdesignpurposesusinglinearcontrol
theory, linear approximation of the differential equations (whenever they
entailnonlinearterms) isused.Ontheotherhand, torepresent thesystem
dynamics more realistically computer simulations including system
nonlinearities can be performed. Moreover, in the event that the linear
controllersthatarebasedonlinearmodelsarenotefficient,nonlineardesign
offeedbackcontrolmaybeused.
ServoController
LinearActuator
LoadFlowControlServo‐Valve
DisplacementTransducer
HydraulicPowerSupply
CommandDisplacement
MeasuredDisplacement
AppliedDisplacement
Disturbance
Noise
33
3.2 LinearControlDesign(Basics)
Oneoftheattributesofalineartime‐invariantsystemisthatitobeysthe
principle of superposition. This principle states that if the system has an
input that can be expressed as a sum of signals then the response of the
system can be expressed as the sum of the individual responses to each
signal. In control engineering the dynamics of the system are typically
studied using root locus (in s‐plane), frequency response or state‐space
methods. The first two methods are based on Laplace transform and are
mainly used in classical control analysis or design. State‐space based
methodsareusedinmoderncontroldesign.
3.2.1 LaplaceTransformandTransferFunction
The Laplace transform is the mathematical tool that transforms
differentialequationsintoanalgebraicformwhichareeasiertomanipulate.
Compared to the Fourier transformwhich is informative about the steady
state response, the Laplace transform yields complete response
characteristics (both transient and steady state response) (Franklin et al.
2010).
Theunilateral (onesided)Laplace transform fora timedomain function
like is
Eq.3‐1
a
fu
in
1
p
ex
2
A
tr
Where i
nd is th
unctionsin
nAppendix
.
The inte
erforming
xemplarwh
).
Theequat
Assumingze
ransformof
is a comple
he imagina
timedomai
A.therefor
rpretation
Laplace t
hereamass
Fig
tionofmoti
eroinitialco
fbothsides
exvariable
ry part of
in,theirLap
ethereisn
of the d
transformat
s‐damper‐s
gure3‐3 am
onforthea
onditions(
ofEq.3‐2
of the form
the compl
placetransfo
oneedtop
dynamic be
tion) is pr
pringsyste
mass‐damper
aboveoscilla
0 0,
m
lex variabl
formsareta
erformthe
ehaviour in
rovided us
em isconsid
r‐springsystem
atorintime
0 0)an
. is the
e. For ele
abulatedas
integration
n s‐domai
sing the f
dered(seeF
m
edomainis
dtakingthe
34
realpart
ementary
provided
ninEq.3‐
n (upon
following
Figure5‐
Eq.3‐2
eLaplace
35
⇒
Eq.3‐3
FromEq.3‐3 canbesolvedas
Eq.3‐4
If,forexample, isagivenforcetimefunctionlikeaunit‐stepfunction
(1 ),whichisdefinedas
10, 01, 0.
Eq.3‐5
AgainlookingatthetableinAppendixA,itisknownthat
11
Eq.3‐6
Therefore the response of the system to a unit step input in Laplace
domain
1
Eq.3‐7
Togettheresponseintimedomain,inverseLaplacetransformcanbeused
anditisdefinedas
36
12
Eq.3‐8
Where isachosenvaluetotherightofallthesingularitiesof inthe
‐plane. ‐planeistheplaneonwhichthe ‐typevariables( )canbe
shown.In fact,Eq.3‐8 isseldomused. Instead,complexLaplacetransforms
arebrokendownintosimplerexpressionsthatarelistedinthetablesalong
with their corresponding time responses (AppendixA). For example, if the
numerical values of the physical properties are such that Eq. 3‐7 can be
writtenas
1 3 2
Eq.3‐9
Using the partial‐fraction expansion technique Eq. 3‐10 can be broken
downintosimplerexpressions
12 1
1
122
Eq.3‐10
Using thematching time functions from Appendix A, the corresponding
time functionof eachcomponent canbe foundand the total time response
for willbethesumofthesetimefunctions.Hence
12
12
0Eq.3‐11
37
3.2.2 TheBlockDiagram
AtransferfunctionisdefinedastheratiooftheLaplacetransformofthe
output to the Laplace transform of the input. Inmany control systems the
dynamicequationscanbewrittenso that theircomponentsdonot interact
exceptbyhavingtheinputofonetransferfunctionastheoutputofanother
one. The dynamics of a system having multiple components are easier to
representinablockdiagramformwhereeachblockrepresentsthetransfer
function of one component (e.g. G1(s), G2(s) in Figure 3‐4) and the input‐
outputrelationshipsbetweentheblocksareshownbylinesandarrows.The
resulting transfer function for thewhole canbeobtainedbyblockdiagram
algebra. This method is often easier and more informative than algebraic
manipulation. Some examples for block diagrams and their equivalent
algebraicinput‐outputrelationshipsareshowninFigure3‐4.
Figure3‐4 Threeexamplesofelementaryblockdiagrams
R(s) Y(s)
R(s) Y(s)+
+
R(s)+ Y(s)
1
(a) (b)
(c)
38
Figure3‐4(c) illustratesanegative feedbackarrangement that isusedto
compare the output of a system with the command input to perform a
tracking task. This is also referred to as closed‐loop control as opposed to
open‐loop control. Open‐loop control is generally simpler and does not
introducestabilityproblems.Althoughfeedbackcontrolismorecomplicated
andmayhavestability issues, ithas thepotential toachieveamuchbetter
performance.Moreover, if the process is naturally (in open‐loop) unstable,
feedbackcontrol is theonlypossibility toattainastablesystemthatmeets
anyperformancecriteria(Franklinetal.2010).
3.2.3 S‐plane,PolesandZeros
Inthedesignandanalysisofacontrolsystem,thetransferfunctionofthe
system gives useful information about the system characteristics including
itsfrequencyresponse.
Therootsofthenumeratorofthesystemtransferfunctionarecalledzeros
of the systemwhich correspond to the locations in the ‐planewhere the
transferfunctioniszero.Therootsofthedenominatorofthesystemtransfer
function are called the poles of the system. Apparently, the poles are the
locations in the ‐plane where the magnitude of the transfer function
becomesinfinite.Forexample,inthepreviousexamplethetransferfunction
hadnozeroandthreepoles.Thezerosandpolesmaybecomplexquantities
andtheirlocationcanbedisplayedinacomplexplane,whichisreferredtoas
the ‐plane.
n
re
d
g
a
p
w
ex
m
The poles
atural or u
esponseisd
omain asso
eneral rule
ssociatedw
olescloser
withpositive
xponential
Figure3‐5
A feedba
modifyingth
s of the sys
unforced be
determined
ociated wit
poles farth
withnatural
to the imag
erealvalue
functionsw
Timefunct
ack control
hesystem’s
STABLE
stem determ
ehaviour of
dbythepole
th poles at
her to the le
lsignals tha
ginaryaxis.
s(inrighth
whichareun
tionassociate
ller improv
poleslocati
mine its sta
f the syste
es.Figure3
different l
eft in the
atdecay fas
.Alsoas in
half‐plane,i.
nstable.
edwithpoints
ves the d
ons.
ability prop
m. Basicall
‐5showsth
locations in
‐plane (LHP
sterthanth
dicated in t
.e.RHP)cor
inthe ‐plane
dynamic re
UNSTABL
perties and
ly the shap
heresponse
n the ‐pla
P inFigure
hoseassocia
the figure, t
rrespondto
e (Franklinet
esponse by
LE
39
also the
pe of the
esintime
ane. As a
3‐5)are
atedwith
thepoles
ogrowing
al.2010).
y mainly
40
3.3 PIDControllerDesign
ThefactthatPIDcontrollersareabletocontrolcomplexsystemswithoutthe
need forprecise identificationof theirdynamicshasmade thempopular in
control applications. A PID controller controls the dynamic error based on
themagnitude, history and rate of the calculated error. The corresponding
transferfunctionofaPIDcontrollerhasthreeterms.
Eq.3‐12
Intheaboveequation istheerrorsignaland iscontrolleroutput.
, and arethegains(coefficients)correspondingtoeachoftheterms.
InfactdesigningaPIDcontrolleristodecideonacombinationofthesethree
gains to get the desired system behaviour. Figure 3‐6 shows the block
diagramofaPIDcontroller.
Figure3‐6 BlockdiagramofaPIDcontroller
Therefore the transfer function for a PID controller is as below, which
introducesanextrapoletothesystem.
E(s) + U(s)
+
+
41
Eq.3‐13
APIDcontrollermaybedesignedusingtheroot locusmethod, frequency
responsemethodorcanbetunedexperimentallyviaanin‐situapproachsuch
asusingZieglerNicholstuningrules(Franklinetal.2010).
Rootlocusmethodstudiestheeffectofanyoneparameterthatentersthe
equation linearly to modify the location of system’s poles in the ‐plane.
TypicallythatoneparameterischosenfromoneofthePIDgains(wherethe
othersareexpressedinrelationtothatone).
Theuseoffrequencyresponsemethodsismorecommoninthedesignof
feedback control systems for industrial applications. One of the reasons is
that with frequency response method it is easy to use experimental
information fordesignpurposes. Frequency responsemethodsutilizeBode
plotsthatportraythesteadystateresponseofasystemtosinusoidalinput.
Whenarepresentativeanalyticalmodelforthesystemisnotavailable,a
PIDcontrollercanstillbeusedbyusingexperimentaltuningapproacheslike
ZieglerNichols(Franklinetal.2010).
3.4 State‐SpaceControllerDesign
Studying thesystemdynamic instatespace formhas the followingmain
advantages(Franklinetal.2010):
42
Having the differential equations in state‐variable form gives a
compact standard form where multi‐input multi‐output systems
canbestudiedeasilyeveninthepresenceofnonlinearities.
Incontrasttotransferfunctionwhichrelatesonlytheinputtothe
outputanddoesnotshowtheinternalbehaviourofthesystem,the
state form connects the internal variables to external inputs and
outputs. This keeps the internal information at hand, which at
timesisimportant.
The state‐variable representation of a continuous linear time‐invariant
open‐loopdynamicsystemcanbeexpressedas
Eq.3‐14
where the 1 column vector is called the state (vector) of the
system, is the 1 inputvector, is the 1outputvector, is
the system matrix, is the input matrix, is the the
output matrix, is the direct transmittance matrix; and, as can be
understood fromabove, , , and are thedimensionsof state, input and
outputvectors,respectively.
As a result of the freedom in choosing the state vector, state space
representationofasystemisnotunique.However, foragivensystemthey
areequivalentintermsoftheinput‐outputrelationship.
43
The eigenvalues ofmatrix are the roots of the characteristic equation
(i.e., the roots of the denominator polynomial (poles) of the open‐loop
transferfunctionforasingle‐inputsingle‐outputsystem).
In state‐spacemethodmoving the closed‐loop pole locations to desired
locationsisaccomplishedthroughafullstatefeedback.Forthestate‐variable
systemdescribedabove,withfullstatefeedback,theinputvectorbecomes
Eq.3‐15
Where is the 1 reference input vector and is an gain
matrix. For example if the reference input is zero (sucha controller is
calledaregulator)fortheclosedloopsystemdynamics becomes
Eq.3‐16
Inthiscase,theeigenvaluesofmatrix (rootsof 0)are
the closed‐loop poles. It can be shown that the closed‐loop poles of the
systemcanbeplacedanywhereinthecomplexplaneaslongasallthestates
arecontrollable(How2007).Thisiscalledpole‐placementmethod.Thereis
a function called place in the commercially available software package
MATLAB which calculates matrix to have the closed loop poles of the
systemmovetothedesiredlocations.
Inthecaseoftrackingareferenceinput( 0),thenonzeroreference
input needs to be introducedproperly for a goodperformance in tracking.
44
This is doneby scaling the reference input and then combining itwith full
statefeedbacktogettheproperinputvector(How2007).
Eq.3‐17
where
Eq.3‐18
Eq.3‐17ensuresthat forastepinputtherewillbenosteadystateerror
aftertransientbehaviour.
Onewaytoselectthelocationofclosedlooppolesistoconsidertreating
thesystemasasecondordersystembyselectingapairofdominantpoles,
with the remaining poles having a real part corresponding to sufficiently
dampedmodes.Thiswillresultinasystemwhichissimilartoasecondorder
system(How2007).
3.5 NonlinearStateSpaceControllerDesign
Inalloftheaforementionedmethodsthedynamicsystemtobecontrolled
was assumed to be linear. Most dynamic systems have some sort of
nonlinearity. In some cases the nonlinearities may safely be ignored or
linearized about anoperatingpoint.However there are systemswhere the
nonlinearities cannot be ignored or the range of operation is beyond the
limitswherelinearapproximationsarevalid.Inordertodesignacontroller
forsuchsystemsnonlineartechniquesneedtobeused.
45
Instead of using linear approximations of the dynamics as done in
Jacobian linearization, feedback linearization is a nonlinear control design
approach which algebraically transforms nonlinear system dynamics into
linear ones and permits the subsequent application of linear control
techniques.
Thesenonlineartechniquesmakeuseofdifferentialgeometryconcepts.In
thefollowingsectionswhereveranewdifferentialgeometricconceptisused
it is defined briefly, and also a more detailed summary of differential
geometryisprovidedinAppendixB.
A single inputnonlinear system in theneighbourhoodof an equilibrium
point, , corresponding to 0 i.e. . can be expressed in state‐
variableformas
Eq.3‐19
orsimply
Eq.3‐20
InEq.3‐20 and areassumedtobesmoothvectorfieldsand .
Avectorfieldisamapthatassignseach avector ofthesamesize.So
forexample,if isastatevectorofsize 1,
46
⋮ Eq.3‐21
Aswill be elaborated later there arenecessary and sufficient conditions
underwhichthesystemdefinedbyequationEq.3‐20istransformableintoa
linear controllable system by nonlinear feedback and coordinate
transformation. This problem is called feedback linearization. Feedback
linearization is viewed as a generalization of pole placement for linear
systems.
The nonlinear single input system in Eq. 3‐20 is said to be locally state
feedbacklinearizableifitislocallyfeedbackequivalenttoalinearsystemin
Brunovskycontrollerform(MarinoandTomei1995)whichis
Eq.3‐22
where the state vector and input in the new coordinate are and
respectivelyand
0 1 0 ⋯ 00 0 1 ⋯ 0⋮ ⋮ ⋮ ⋱ ⋮0 0 0 ⋯ 10 0 0 ⋯ 0
Eq.3‐23
47
00⋮01
Eq.3‐24
where the state vector and input in the new coordinate are and
respectively.
InordertobeabletocastasetofequationsinBrunovskycontrollerform,
the theorem of feedback linearization needs to be satisfied. This theorem
indicatesthatthesingleinputsysteminEq.3‐20with statesislocallystate
feedbacklinearizableifandonlyifinaneighbourhoodoforigin:
(i) thedistributionspan ,… , isofrank ,and
(ii) the distribution span ,… , is involutive and of
constantrank 1.
Expressions like are iterative forms of Lie bracket which is a
functionindifferentialgeometryactingontwovectorfieldslike and .Lie
bracket function is illustrated in Appendix B. Also an exact mathematical
explanationofdistributionanditsrankisgiveninAppendixB.Howeverasa
simpleexplanation,condition(i)requiresthatthespacegeneratedfromthe
indicated vector fields has a dimension of whichmeans the vector fields
have to be linearly independent. Condition (ii) requires that the space
generated from the indicated vector fields has a dimension of 1.
Moreoveradistributionis involutiveif,givenanytwovectorfields and
48
belonging to that distribution, their Lie bracket, , , also belongs to the
distribution.
Theabovetwoconditionsguaranteetheexistenceofafunction : →
such that in the neighbourhood of origin, the following conditions are
satisfied.
⟨ , ⟩ 0
⟨ , ⟩ 0, 0 2
Eq.3‐25
Intheaboveexpressions iscalledgradientof andisdefinedas
, , … , Eq.3‐26
andinnerproductisdefinedas
⟨ , ⟩ . Eq.3‐27
HavingsolvedtheconditionsinEq.3‐25for ,thetransformationfrom
to willbe
, , … , , , … , Eq.3‐28
where the expression is called the Lie derivative of function
alongthevectorfield .ThisoperatorisalsodefinedinAppendixB
49
HencethedynamicsysteminEq.3‐20i.e.
Eq.3‐29
transformsinto
, 1 1
Eq.3‐30
Eq.3‐31
The system expressed above is dynamically equivalent to the original
system which means they have identical poles. is the input of the
transformedsystemand (input for theoriginalsystem)canbecalculated
fromitusingequationEq.3‐31.Sothestatefeedbackis
Eq.3‐32
50
4 ImplementationofControlMethods
General
InthischapterthecontroltechniquesintroducedinChapter3areutilized.
As discussed before an actuator delay resulting in a time delay in
experimentalsubstructurecanimpairthedynamicstabilityandaccuracyof
the system. Control theory is used to design servo‐hydraulic controller to
minimizeactuatordelay.
In thisstudya linearizedmodelof thesystemdynamicswasused in the
designof the controllerusing linear controldesign techniques.Anonlinear
model of the system whose parameters were obtained through system
identification (Mercan2007)wasused in thenonlinear state space control
designandalso in thenumerical simulations.This isbecause thenonlinear
model is accepted toprovideamore realistic representationof the system
dynamics.
4.1 DynamicModelofaServo‐HydraulicSystem
The dynamic equations that govern a servo‐hydraulic system were
explainedinChapter2andtwoversionsofthemodel(linearandnonlinear)
51
were introduced. These dynamics are modelled in commercial software
packageMATLAB/SimulinktosimulatetheinnerloopofaPSDtestsetup.
LinearModel
Equations for the linearmodel presented in Chapter 2 are summarized
here.
Table4‐1Linearized dynamicsofaservo‐hydraulicsystem
Valvespool(in
domain) 1
(Eq.2‐1)
Valveflow(Eq.2‐
20)
Actuatorchamber
pressure 4
(Eq.2‐
18)
Actuatormotion (Eq.2‐21)
Figure 4‐1 shows the Simulinkmodel for a servo‐hydraulic systemwith
the above dynamics that is connected to a linear single‐degree of freedom
teststructure.Thisstructurehasamassof ,dampingof andstiffnessof .
Asitisshowninthefigurethisaddsaloaddynamicstotheaboveequations
as
Eq.4‐1
Figure4‐1
Simulinkmodelforaservo‐hydraulicsystemwithlinearizeddynamics
52
53
NonlinearModel
Equationsforthenonlinearmodelare
Table4‐2Nonlineardynamicsofaservo‐hydraulic
Valvespool(in
domain)1
(Eq.2‐1)
Valveflow (Eq.2‐
16)
Actuatorchamberpressure
4
(Eq.2‐
18)
Actuatormotion
(Eq.2‐
19)
Figure4‐2showstheSimulinkmodelforaservo‐hydraulicsystemofthe
abovedynamics that is connected to a linear single‐degreeof freedom test
structure.
ThelinearmodelwillbeusedfordesigningaPIDandalinearstatespace
controllerandthenonlinearmodelwillbeusedtodesignanonlinearstate
spacecontroller.
Figure4‐2
Simulinkmodelforaservo‐hydraulicsystemwithnonlineardynamics
54
55
4.2 PIDControllerDesign
Figure4‐3showstheSimulinkmodel foraservo‐hydraulicsystemalong
with a PID controller. As it can be seen, the servo‐hydraulic system is
represented by a blockwhich has an input as a command signal,which is
issuedbythecontroller,andfouroutputsthatmaybemeasuredduringatest
andarenamelypistondisplacement,pistonvelocity,loadpressureandvalve
opening. In this study the nonlinear models introduced in Figure 4‐2 was
usedandembeddedtothisblock.
The following gives a summary of the roles of each PID gains
(Ahmadizadeh2007).
Proportional( )–Thisistohandlethepresentrequirements.Theerror
ismultiplied by .Hence, the greater the proportional gain, themore the
servo‐valve opens for a given error. There is a trade‐off for selecting an
appropriate .Although a largeproportional gainmaydecrease the error
resultinginaclosertrackingofreferencesignalandreducedresponsetime,
itdecreasesthestabilitymarginofthesystemandincreasesthefrequencyof
theoscillationinthetransientresponse.
Figure4‐3
Simulinkmodelforaservo‐hydraulicsystemwithaPIDcontroller
56
57
Integral ( )–Theerror is integrated(addedup)overaperiodof time,
multipliedbyaconstant andaddedtothecontrolsignal.Awell‐tunedPI
controller will converge to the reference signal (zero steady‐state error),
leadingtoareducederrorbetweencommandandfeedback.
Integral ( ) – This is to handle the future requirements. The first
derivativeoftheerrorovertimeiscalculated,multipliedby andaddedto
the control signal.Basically, this termcontrols the response to a change in
thesystem.Inpracticeduetonoisesthatenterthemeasuredsignalinareal
PSDtest,acontrollerwithoutaderivativetermmaybeused.
4.2.1 ControllerTuning
Tuning a controller is adjusting its parameters to get a desired system
response.Thegoalistohaveasystemthathasastableandfastresponse(i.e.,
ashortresponsetime)withasmallsteady‐stateerror.
One of the Ziegler‐Nichols tuning methods is called ultimate sensitivity
methodandisusedinthisstudy.Inthismethodthecriteriaforadjustingthe
parameters are based on evaluating the amplitude and frequency of the
oscillationsofthesystematthelimitofstabilityratherthantakingthestep
response. To use the method, the proportional gain is increased until the
system becomes marginally stable and continuous oscillations just begin,
withamplitudelimitedbythesaturationoftheactuator.Thecorresponding
gainisdefinedas (knownasultimategain)andtheperiodofoscillationis
(knownasultimateperiod). shouldbemeasuredwhentheamplitudeof
58
oscillation is as small as possible. Then the tuning parameters for a PI
controller are selected as 0.45 and 1.2. Experience has
shownthatthecontrollersettingsaccordingtoZiegler‐Nicholsrulesprovide
acceptable closed‐loop response for many systems. Ziegler‐Nichols gives a
goodstartingpoint for thecontrollerparametersandthe fine tuningof the
controller is still needed to achieve a desired behaviour. Details for this
methodcanbefoundin(Franklinetal.2010).
Tuning thenonlinearmodel inSimulinkresults in 80, 0.5and
0.Andclosed‐looppolesofthesystemcanbecalculatedasfollowing.
242, 141, , 21.9 31.2 , 0.00625
AsshowninEquation3‐13,thetransferfunctionofaPIDcontrolleraddsa
poleatorigintotheopenlooptransferfunctioninEquation2‐23.Here,inthe
closed loop transfer function this extra pole moves a bit off the origin
( 0.00625).
4.3 LinearState‐SpaceControllerDesign
4.3.1 State‐variableformofequations
As pointed out before, for a given system, depending on the states
selected,theremaybemultiplestatespacerepresentations.Usingthelinear
equationsintroducedinTable4‐1,fourstatesandoneinputcanbechosen.
59
, Eq.4‐2
Alternatively,iftheservo‐valveisassumedtohaveafastresponsetothe
inputsignalcomparedtotherestofthesystem;itsdynamicscanbeomitted
withoutcompromisingthetrackingcapability.
, Eq.4‐3
Considering these three states, the state‐variable representation
(Equation3‐14)ofthesystemwillbe
0 1 0
04 4
004
1 0 0
Eq.4‐4
Andifallfourstatesareconsidered
60
0 1 0 0
0
04 4 4
0 0 01
000
Eq.4‐5
Itshouldbenoticedthatthevalvespooldynamics(Equation2‐1)intime
domainis
1 Eq.4‐6
4.3.2 PolePlacement
Boththree‐stateandfour‐statepresentationofthesystemcanbeusedto
designthecontroller.Forthefour‐statecase,polesareselectedequaltothe
poles corresponding to thepreviouslydesigned systemwithPIDcontroller
ignoringthepoleassociatedtoPIDitself,i.e.
242, 141, , 21.9 31.2
61
Function place.m of commercial software packageMATLAB is used to find
matrix and thenusingequation3‐18matrix is calculated to introduce
thesignalinputas
Eq.4‐7
Appendix C contains the MATLAB code which is used to calculate these
matrices.
If the servo‐valve spool dynamics is ignored (assumed to be faster than
otherpartsof thesystem)consideringonly threestates, thecorresponding
poleneeds tobe ignored too.As illustratedbefore, thepoles far left in the
complexplanearerelatedtofastresponses.Thereforepole 242isthe
one associated to servo‐valve spool dynamics. Consequently, the selected
polestodesignathree‐statecontrollerare
141, , 21.9 31.2
4.4 NonlinearState‐SpaceControllerDesign
4.4.1 State‐variableformofequations
State‐variableformofanonlinearsystemneedstobewrittenintheform
of
Eq.4‐8
62
Itcanbeshownthattheonlywaytowritedownthenonlinearequations
intheaboveformistoconsiderallfourstates.Sohaving
, Eq.4‐9
state‐variableformofthedynamicsaccordingtoTable4‐2willbe
4 sign
000 Eq.4‐10
Tofollowthecalculationsthefollowingnotationwillbeusedinstead
4 sign
000 Eq.4‐11
ItwasstatedinChapter3that and mustbesmoothvectorfieldsbut
the sign function in the third term of does not allow differentiation
at 0.Figure4‐4showsthatforalargecoefficient signfunctioncanbe
approximated by using an Arc Tan. Mintsa et al. (2009) used an equation
involvingexponentialtermstoapproximatethesignfunction.
63
Figure4‐4 Approximationoffunction
HencethedynamicsinEq.4‐11canbeapproximatedas
000
Eq.4‐12
and inEq.4‐12aresmoothvectorfieldswhichcanbeusedtodesigna
nonlinearcontroller.
As it was stated in Chapter 3, for a system to be locally state feedback
linearizable two conditions must be satisfied. These conditions are now
checkedfortheabovenonlinearsystem.
-0.002 -0.001 0.001 0.002
-1.0
-0.5
0.5
1.0
2100
21000 2
106
64
Condition(i) thedistributionspan , … , isofrank .
000
000⋆
, Eq.4‐13
00
4
2
2
00⋆⋆
Eq.4‐14
Andinthesamemanner,
0⋆⋆⋆
Eq.4‐15
⋆⋆⋆⋆
Eq.4‐16
Intheaboveequations“⋆”representsanonzeroexpression.Since
span
000⋆
,
00⋆⋆
,
0⋆⋆⋆
,
⋆⋆⋆⋆
Eq.4‐17
65
hasarankequalto4,thefirstconditionissatisfied.
Condition(ii) thedistributionspan , … , is involutive
andofconstantrank 1.
Itisalreadyknownthat
span
000⋆
,
00⋆⋆
,
0⋆⋆⋆
Eq.4‐18
hasarankof3.Howeverthedistributionneedstobeinvolutive.
Itcanbeshownthat
,
00⋆0
Eq.4‐19
,
0000
Eq.4‐20
,
0⋆⋆0
Eq.4‐21
Allabovevector fieldsbelongtothedistribution inEquation4‐15which
meansthedistributionisinvolutiveandthesecondconditionisalsosatisfied.
66
Therefore the nonlinear system is locally state feedback linearizable and
thereisafunctionlike : → suchthat
⟨ , ⟩ 0
⟨ , ⟩ 0, 0 2
Eq.4‐22
Theaboveconditionsareexaminedinmoredetailinthefollowing.
1. ⟨ , ⟩ 0,whichgives
, , ,
000 0 ⇒ 0 ⇒ 0 Eq.4‐23
2. ⟨ , ⟩ 0,whichinthesamemannergives
0 Eq.4‐24
3. ⟨ , ⟩ 0,whichalsointhesamemannergives
0 Eq.4‐25
4. ⟨ , ⟩ 0,whichgives
0 Eq.4‐26
67
It should benoted that the function is notunique (Marino andTomei
1995).
Function maybeselectedtobe
Eq.4‐27
AswasintroducedinEquation3‐25,coordinatetransformationisdefined
as
, , , , , , Eq.4‐28
So
Eq.4‐29
1 0 0 0 Eq.4‐30
0 1 0 0 Eq.4‐31
Eq.4‐32
68
Thistransformationofcoordinatesletsthesystemdynamicstobewritten
inBrunovskycontrollerformas
Eq.4‐33
where
Eq.4‐34
andintheaboveequation
Eq.4‐35
and
69
4
12 ArcTan
2 ArcTan 4
Eq.4‐36
Thereforethestatefeedbackwillbecalculatedas
Eq.4‐37
Chapter 5 illustrates implementation of the above coordinate
transformationinaMatlab‐Simulinkmodelalongwithlinearcontrollers.
70
5 SimulationsResults
General
This chapter presents simulation results for the inner loop of PSD test
models i.e. the control of servo‐hydraulic‐test structure system. Simulink,
which isdevelopedbyMathWorks, is an interactive graphical environment
formodeling, simulatingandanalyzingdynamic systemsand itworkswith
MATLAB.TheSimulinkmodelspresented in this chapter include linearized
andnonlinearmodelofaservo‐hydraulicsystemdiscussedinChapter2and
the controllers discussed and implemented in Chapter 3 and Chapter 4
includingPID,linearstate‐spaceandnonlinearstate‐spacecontrollers.These
simulations only consider the servo‐hydraulic system connected to a test
structure.Tobetterunderstandtheproposedcontrolmethod,foracomplete
simulation of a PSD test, the outer‐loop needs to be introducedwhere the
integration algorithm and also the analytical substructuremodel (if it is a
hybridPSDtest)reside.
71
5.1 Numerical Values for Servo‐Hydraulic System and Test
StructureinLinearandNonlinearModels
The values of the parameters for linear and nonlinear formulation of
dynamics of a servo‐hydraulic system are given in Table 5‐1. These
numerical values are obtained through system identification for a servo‐
hydraulic system used in another PSD test research program (Zhang et al.
2005)andareassumedtoberepresentative.Forthestructureselection,the
important thing foranSDOFare thenaturalperiod(shouldbearound1or
1.3sec),damping(around2to5%). InaPSDtest, it isdesiredtokeepthe
inertialanddampingforcesanalytical.
Table5‐1Values for Parameters for Linear and Nonlinear Servo-Hydraulic System Models
parameter value mass of test structure 1025kg
viscous damping coefficient of test structure 500 N sec/m
stiffness test structure 5.00 N/m
mass of actuator piston 1025kg
viscous damping coefficient of actuator piston 356.18 N sec/m
actuator piston cross section area 0.0808m
4
actuator chamber volume 5.65 m /Pa
oil modulus supply pressure of the hydraulic system 207 Pa
Γ 1 ⟶orifice coeffiecient
Γ ⟶valve opening gradient ⟶hydraulic oil density
2.62 m /s/Pa .
servo-valve flow-pressure coefficient 1.5 m / Pa sec
servo-valve flow gain 0.035 m /sec
actuator leakage coefficient 3 E m / Pa sec
servo-valve gain 0.9613
servo-valve time constant 0.004sec
, natural frequency of the test structure 69.8 1 sec⁄
damping ratio of the test structure 0.00349
72
5.2 Comparison of Controllers With and Without Saturation
Limits
Numerical simulations were performed to compare the tracking
capabilities of the PID, linear state‐space and nonlinear state‐space
controllers. As explained in Chapter 4, all the controllersweredesigned to
have the same dynamic characteristics. This was done by figuring out the
polelocationsfromthetunedPIDcontrollerandusingthemfordesigningthe
linear and nonlinear state‐space controllers. Linear state space controller
uses the linearized state space representation (Table 4‐1) in the pole
placement procedure. On the other hand, the nonlinear controller design
accounts for the nonlinearity in servo‐valve flow‐pressure relationship
(Table4‐2). Itutilizesacoordinatetransformationtoexpressthesystemin
Brunovskyformwhichprovidesanequivalent linearsystemrepresentation
forthesystemdynamics.Itshouldbenotedthatthecalculatedinputtothis
equivalentsystem( )hastobetransformedbacktotheinputoftheoriginal
system ( ) using the corresponding expression in Equation 4‐34. While
performing thenumerical simulations, thedynamics of the servo‐hydraulic
test structure system was represented by the nonlinear model as it is
presumedtobemorerepresentativeoftherealbehaviour.Figure5‐1shows
step responses of the system with linear and nonlinear state‐space
controllers. Itwas observed that the PID responsewas almost identical to
linear state‐space controller which was expected as both were based on
linearizedrepresentationofthesystem.Thereforeonlythelinearstate‐space
re
o
co
o
d
st
w
w
p
st
esponse is
penasmuc
ontrolleris
f the respo
esign has s
teadystate
Figure 5‐2
withthediffe
when the s
erforming b
taterespon
shown. In
chasneede
observedto
onse. Moreo
some steady
error.
Figure5‐1
2 shows th
erencethat
servo‐valve
better than
sephases.
displacement(m)
Figure 5‐1
ed (i.e., the
obefastert
over, in the
y state erro
Steprespo
he step resp
ittakessat
opening i
n the linear
the servo‐v
systemdoe
thanthelin
e steady sta
or while th
onsesforasys
ponse for th
urationlim
is limited,
design dur
line
valve is ass
esnot satur
earoneint
ate part of
he nonlinea
stemwithouts
he same sy
itsintoacco
the nonlin
ring both tr
nonlinear desig
eardesign
sumed to b
rate).Then
thetransien
the respon
r design ha
saturation
ystems in Fi
ount.Ascan
near design
ransient an
gn
time(sec)
73
e able to
nonlinear
ntportion
nse linear
as a zero
igure 5‐1
nbeseen,
n is still
nd steady
sa
co
InFigure
aturation.N
Figures5
ontrollers l
Figure5‐
5‐3respon
Nonlinearst
F
‐4givesac
ookingat lo
displacement(m)
displacement(m)
‐2 Stepresp
nsetoasinu
tate‐spacec
Figure5‐3 R
comparison
oadpressur
no
lin
ponsesforasy
usoidinput
ontrollerca
Responsetosi
betweenn
revariation
onlinear
near design
linear
nonline
ystemwithsa
isshownfo
anbeseent
inusoidinput
onlinearan
n for theca
design
ear design
aturation
forasystem
tohaveless
ndlinearsta
sewhenth
time(sec)
time(sec)
74
mwithout
timelag.
ate‐space
here isno
sa
ca
aturationin
asewhensa
F
loadpressure(Pa)
loadpressure(Pa)
ntheservo‐
aturationha
Figure5‐4 lo
Figure5‐5
nonline
linear
linear d
‐valve.Figur
appensinse
oadpressurev
loadpressur
ear design
r design
nonlinear design
design
res5‐5give
ervo‐valve.
variationwith
evariationwi
esthesame
houtservo‐val
ithservo‐valv
ecompariso
lvesaturation
vesaturation
time(sec)
time(sec)
75
onforthe
n
st
ca
sa
b
a
Inadditio
tate‐space c
ase when t
amecompa
enoticedth
chievedby
Figu
Servo‐valvespoolopening(m)
on,Figures5
controllers
there is no
risonforth
hatthebett
asmallchan
ure5‐6 Serv
nonli
linear design
5‐6givesa
looking at
saturation
hecasewhe
terbehavio
ngeinservo
vo‐valveopeni
inear design
n
comparison
the servo‐v
in the serv
nsaturation
urofthepr
o‐valveopen
ing variationw
nbetweenn
valve openi
vo‐valve. Fi
nhappensi
roposedno
ning.
withoutservo
nonlineara
ing variatio
igures 5‐7 g
inservo‐val
nlinearcon
o‐valvesaturat
time(sec)
76
ndlinear
n for the
gives the
lve.Itcan
ntrolleris
tion
5
F
Fig
5.3 Imple
TestS
Theimple
igure5‐8.
Servo‐valvespoolopening(m)
gure5‐7 Ser
menting t
Simulation
ementation
rvo‐valveope
theNonlin
n
ofhybridP
nonline
linear design
ningvariation
nearState
PSDtestme
ar design
nwithservo‐v
e‐Spaceco
thodinaflo
valvesaturatio
ontroller in
owchartis
time(sec)
77
on
naPSD
shownin
w
S
si
v
in
u
Toinvesti
was conduct
imulink mo
imulations
elocity feed
nvestigatet
sed.
Figure
igatethebe
ted at Lehi
odel of the
duringthat
d forward.
thebehavio
5‐8 Hybrid
ehaviourof
igh Univers
e test simu
tstudy,whe
In this mo
urof thesy
dPSDtestmet
f anelastom
sity (Merca
ulation that
ereaPIDco
odel the sam
ystemwhen
thod(Mercan
mericdampe
n 2007). F
t was used
ontrollerwa
me Simulin
naNLstate
2007)
erahybrid
Figure 5‐9 g
d in the n
asusedalon
nk model is
e‐spacecon
78
PSDtest
gives the
numerical
ngwitha
s used to
ntroller is
sy
su
su
th
Figure5‐substr
As shown
ystem and
ubsystem2
ubsystem2
hesefigures
‐9 Simulinkructureandan
n in Figure
the experi
whenaPID
2whenaN
stheservo‐h
kmodelforreanelastomeric
e 5‐9 subsy
mental sub
Dcontroller
L state‐spa
hydraulicsy
al‐timehybriddamperasth
ystem 2 co
bstructure.
risusedan
ce controlle
ystemhastw
dPSDtestingwheexperiment
onsists of
Figure 5‐10
dFigure5‐
er isused.A
wosimilars
withMDOFanalsubstructur
the servo‐h
0 shows de
11showsd
As it canbe
servo‐valve
79
nalyticalre
hydraulic
etails for
detailsfor
e seen in
es.
Figure5‐10
Simulinkmodelforaservo‐hydraulicsubsystem
withaPIDcontroller
80
Figure5‐11
Simulinkmodelforaservo‐hydraulicsubsystem
withaNLstate‐spacecontroller
81
v
A
lo
si
ex
In
co
Sincethe
aluesofstif
Again,polep
ocations.Ca
imulations.
xceptforth
T
m v s
Figure5‐1
n each of
ommandre
Figure5‐12
displacement(m)
displacement(m)
behaviour
ffnessandd
placementw
anogaPark
Allthepara
heteststruc
Table 5-2 Appr
mass of test strviscous dampinstiffness test str
12istheres
the figures
esponsefrom
2 comparison
oftheelast
dampingwa
wasdones
earthquake
ametersuse
tureparam
oximate valuesparameter
ructure ng coefficient oructure
sultsimulat
s the meas
mthecontro
nbetweencom
tomericdam
asusedtode
uchthatbo
egroundmo
edindesign
etersthata
s for the elasto
of test structure
tionsforPID
sured displ
oller.
(a)
(b)mmandandmNLcontroller
mperisnon
esignthesta
othsystems
otionwasu
nwerethes
areasbelow
omeric damper
e 0.56
DandNLsa
lacement is
measureddisplr
nlinear,app
ate‐spaceco
shave ident
usedinthef
sameasinT
w.
parameters value 1335 kg
618 N sec/m3.92 N/m
ate‐spaceco
s given aga
lacement(a)P
82
proximate
ontroller.
ticalpole
following
Table5‐1
m
ontroller.
ainst the
PIDcontroller
time(sec)
time(sec
(b)
)
c)
o
g
ex
v
v
sh
sp
For both
verlapping
iveninFigu
Figure5‐1
It can be
xtent. This
elocity feed
elocity feed
hows that s
pacecontro
displacement(m)
displacement(m)
controllers
in this tim
ure5‐13.
13 compariso
seen that t
much of
d forward t
d forward
spool openi
ollerisused
, command
me scale. He
onbetweenco(b
the NL cont
improveme
to PID cont
may distor
ing has am
.
d andmeasu
ence a narr
(a)
(b)
ommandandmb)NLcontroll
troller is im
ent could a
troller but a
rt the spoo
much smoot
ured displa
ower view
measureddispler
mproving th
also be ach
as shown in
ol opening
ther behavi
measured
command
measured
comman
acements ar
of the resp
placement(a)
he tracking
hieved by
n 5‐14 (a)
response.
iourwhenN
e
d
83
re almost
ponses is
)PIDcontrolle
g to some
adding a
adding a
5‐14 (b)
NL state‐
time(sec)
time(sec
er
c)
Figure5‐1
Valveopening(m)
Valveopening(m)
14 comparisoonbetweensefeedfo
(a)
(b)
ervo‐valvesporward(b)NL
oolopening(acontroller
a)PIDcontrolle
84
erwithaveloc
time(sec)
time(sec)
city
5
p
w
ex
p
v
e
ro
p
p
m
S
a
5.4 Robus
There is
arameters
whatisreally
xamined w
arameters.
ariationsin
nhancedfee
obustness o
hysical par
ercent of o
model rema
upply pres
ccumulator
displacement(m)
stnessofth
always the
used in the
yinthesys
with respec
Mintsa et
nsupplypre
edback‐line
of the prop
rameters in
over/under
ained intact
ssure is co
rs.
Figu
heContro
e possibilit
e controller
tem.Thero
ct to pote
al. (2009)
essurebyde
earization‐b
posed contr
n the contr
restimation
t. , ,
nsidered a
ure5‐15 effe
No
lDesign
ty that the
r block are
obustnessof
ential error
addressed
esigningaL
asedcontro
rol design,
roller block
error whi
and ar
almost cons
ectof10%erro
10%erro
error
identified
inaccurate
fthecontro
rs in the
the robust
Lyapunovap
ollaw.Here
the values
k were con
le their va
re the para
stant due
orin onthe
t
or
values of
e and differ
oldesignne
identified
tness regar
pproachtod
eintoinvest
s of some i
taminated
alues in the
ameters con
to the pre
eresponse
time(sec)
85
physical
rent from
eedstobe
physical
rding the
derivean
tigatethe
identified
by some
e system
nsidered.
esence of
86
Figure5‐15comparesthecasewhenthevalue for in thecontroller is
overestimatedby10%withthecasewhenthereisnoerror.Itisreadilyseen
that 10% of error has made the system unstable. A similar graph was
conducted for 5% error (not shown in figure for sake of clarity) and the
control design was able to handle it without stability problems. Similar
comparisons were done for negative percent errors (underestimation)
wherein the accuracy did not change much and there were no instability
problems.
Conducting the same procedure for , and showed that the
proposed control design can easily handle errors within 20% of the real
value.
5.5 Conclusion
Thecontroller’sabilitytotrackthecommanddisplacementsaccuratelyis
crucialintheoverallstabilityandaccuracyofthereal‐timePSDtesting.For
thedisplacementcontrolloop,itwasshownthatacontrollerwithimproved
performance can be designed using nonlinear state space control design
techniques,providedthatarepresentativemodelofthesystemisavailable.
Simulationsweredone fora servo‐hydraulic systemattached toa linear
structure for cases with and without saturation in servo‐valve and it was
concludedthataNLcontrollercanimprovethetrackingcapabilities.
87
Moreover,theproposedNLcontrollerwascomparedtoPIDcontrollerby
implementing it in a real‐time PSD test simulation to show that it can
decreasethetimelagofthemeasureddisplacement.
Robustnessofthenonlinearcontroldesignwithrespecttosomeidentified
physicalparameterswasinvestigatedanditwasobservedthaterrorswithin
reasonablerangecanbeeasilyhandledintermsofinstabilitywithoutlossof
accuracy.
Future research is needed on application of the nonlinear controller to
multi‐input, multi‐output system control together with a comprehensive
numericalstudyandexperimentalvalidation.
88
6 Testsetupdesign
The design of the test specimen considers the displacement and force
limitationsof theexistingactuators.Theseactuatorshaveastrokerangeof
5"(12.7cm)andhaveaforcelimitof5.5Kip(24.5kN).
Theaimistodesignatestsetupthatexhibitsnonlinearbehaviourwithin
aboveactuatorslimitations.Figure3‐1showsasketchofthetestsetup.This
single‐degree‐of‐freedomsystemconsistsofashortcolumnwithwideflange
sectionthat ishingedtoasupportbymeansofaplateandaclevis.As it is
seen in the figure, replaceable steel coupons are used to providemoment
resistance for the support. Since the moment resistance capacity of the
coupons is smaller than the columns, the nonlinear behaviour (yielding)
initiates in the coupons and the column is expected to remain elastic.
Therefore, after each nonlinear PSD test, the sacrificial coupons can be
replacedeasilyatalowcost.
h
h
co
h
co
cr
Looking a
ave been u
ingecanbe
22
Twoof th
ompression
ingeandth
oupons. i
rosssection
Fi
at the defle
used, the int
workedou
hecoupons
n.Intheabo
hecoupons
is themodu
nalareaofo
igure6‐1Expe
cted shape
ternal force
utasthepro
willexperi
oveequatio
inthedirec
ulusof elast
onecoupon.
erimentalsetu
and assum
e in the cou
oductofthei
ence tensio
n istheh
ctionofapp
ticityof the
upforPSDtes
ming that tw
upons ( ) a
irdeflection
onandtheo
horizontald
pliedforce.
e couponm
st
wo pairs of
at either sid
nandaxials
other twow
istancebetw
istheleng
material and
coupo
o
89
f coupons
de of the
stiffness.
Eq.6‐1
willbe in
weenthe
gthofthe
d is the
n
y
th
n
ca
2
Also it is
ieldingforc
2
is th
hebeamand
As menti
onlinearity
alculatedby
2
s obvious t
ce,so
hemaximum
d isyield
oned befor
.Therotati
yequating
⇒
Figure6
that the fo
m force tha
ingstresso
re the cou
onoftheba
inEq.6‐1
‐2Expectedb
rce in the
at twocoup
ofthecoupo
upons are
aseatwhich
to inE
baserotationa
coupons c
ons canpro
onsteel.
intended t
hthecoupo
Eq.6‐2.
atbaseusedin
cannot exce
ovideaton
to experien
onsyield(
ndesign
90
eed their
Eq.6‐2
ne sideof
nce some
)canbe
Eq.6‐3
91
Inordertobeabletoobservenonlinearresponseintheexperimentaltest
set‐up,thedesignbaserotation willbesethigherthantheyieldrotation
.
1 Eq.6‐4
Ontheotherhand,becauseoftheactuatorstrokelimit,testsetupcannot
movemorethan5”(12.7cm)toeitherside.HenceaccordingtotheFigure3‐
1
sin 5" Eq.6‐5
orapproximately
5" ⇒
5" ⇒ 5" Eq.6‐6
Alsoforasingle‐degreeoffreedomsystemsuchastheoneinFigure3‐1,
ignoringtheself‐weightof thesystem,staticequilibriumofmomentsabout
point gives
⇒ Eq.6‐7
Andtheforceappliedbytheactuatorcannotexceed5.5Kip 24.5kN .So,
92
5.5Kip Eq.6‐8
whichaccordingtoEq.6‐2gives
2 5.5Kip ⇒
1
5.5 Kip2
Eq.6‐9
CombiningEq.6‐6withEq.6‐9showsthatthecolumnneedstobeinthe
followingrange.
2 5.5Kip
5"
Eq.6‐10
Itwasdecidedtouseawide‐flangesectionof 6 20forthecolumnand
for the coupons a diameter of 0.25" 0.635cm)was selected that gives an
area of 98.2 10 in 0.634cm . The clevises available in the
structurallaboratoryattheUniversityofAlbertadictatesacouponlengthof
8.13" 20.65cm .Thedistance between the coupons, as indicated in the
Figure3‐1,.waschosentobe 6" 15.24cm .
Assuming 34ksi 234 , 29000ksi 2 10 and 3,
Eq3‐10becomes
93
234ksi 98.2 10 in 6"
5.6kips5"
6" 29000 ksi3 8.13" 34 ksi
⇒ 7.16" 1000"
Eq.6‐11
Ascanbeobservedforheightofthecolumnthereisawiderangetoselect
from.Howeveroneshouldbeawarethattheupperlimitiscorrespondingto
3. That implies that any height less than 1000" corresponds to bigger
valuesof .Accordingtothespaceavailableinthestructurallabaheightof
50" 127cm waschosentobeused.
Buckling of the coupons also needs to be considered. Knowing that the
endsofthecouponsarefixed,bucklingloadofacouponis
29000ksi 40.25′′2
0.58.13′′3.32kips 14.8kN
whiletheyieldingforceis
340.254
1.67kips 7.43kN
Thismeansthatthecouponswillyieldlongbeforebuckling.
94
AppendixA
TableofLaplaceTransforms
Number , 0
1 1
2 1
1
3 1
4 2!
5 !
6 1
7 1
8 1
11
9 1
10 11
11
95
12 1
13
1 1
14
15 sin
16 cos
17 cos
18
sin
19 1 cos sin
96
AppendixB SomeonDifferentialGeometry(Lynch2009)
1 ChangesofCoordinatesorDiffeomorphisms
A nonlinear change of coordinates is a function defined on
⊆ andmappingto . It must have the properties
1. is an invertible function, i.e., there must exist an inverse
function suchthat
, ∀ ∈
2. Both and are mappings.
When ⊆ thechangeofcoordinatesiscalledglobal.Whenthisisnot
thecase, thechangeofcoordinates issaid local.Asufficientcondition fora
mapping tobe a local changeof coordinates is given by the Inverse Function
Theorem (Marino and Tomei 1995). Note that Condition 2 is required, i.e., that
the change of coordinate is , since these coordinates changes will be applied to
the state of a system, and the systems are required to be expressed in the new
coordinatesalsotobe orsmooth.Aninfinitedegreeofsmoothnessisnot
usually required, but it is assumed to avoid keeping track of the degree of
97
smoothness. The invertibility Condition 1 allows to uniquely recover the original
state coordinate from the new one. In addition to applying a change of coordinates
to a system’s state, transformation of a system’s output variable or time is also
possible.
2 VectorFields
Avector field on ⊆ isa mappingfrom to . It is customary to
write vector fields using one of two notations. The first notation represents vector
fields as column vectors. The basis used to represent the vector field is implied
and this might lead to confusion.
⋮, , … ,
Alternately
where is the th unit tangent vector in the ‐coordinates. The latter
notation provides more information in that the basis used to express the
vectorfieldisstatedexplicitly.
3 DifferentialGeometryFunctionsUsedintheTextLiebracketsLie brackets between two vector fields and is another vector field
whichisdefinedinlocalcoordinatesas
98
,
IteratedLiebracketsaredefinedas
, , 1
LieDerivative
Liederivativeoffunction alongvectorfield istheinnerproductof
thegradientof definedas
, , … ,
andvectorfield ,therefore
⟨ , ⟩
andtheiterativeformofitisexplainedas
⟨ , ⟩, 1
SequentialformofLieDerivativeisdefinedas
⟨ , ⟩
99
4 Distributions
A ‐dimensionaldistribution∆definedon ⊆ isamapwhichassignsto
each ∈ ,a ‐dimensionalsubspaceof (ormoreprecisely the tangent
space ) such that for all ∈ there exists a neighbourhood ⊆
containing and smoothvectorfieldssuchthat
1. , … , arelinearlyindependent∀ ∈
2.∆ , … , forall ∈ .
NotethatinCondition1,“linearindependence”istheusualdefinitionon
fromlinearalgebra.InCondition2thespanontheRHSinvolvesreal linear
combinations of the constant vectors , … , in , that is, it is a
subspace of . In many cases we use so‐called generating vector fields
, 1 to define a distribution, however, the are only one
choiceofbasisforthesubspace∆ andotherscanbechosen.Forexample
if ∆ , then we also have∆
, .
Thedimensionof adistribution∆,denotedasdim ∆ is a functionof
and is equal to the dimension of the subspace ∆ . A distribution is said
nonsingularon ifithasconstantdimensionon .Innonlinearcontrol,we
oftenassumetherelevantdistributionsarenon‐singular.
Givena ‐dimensionaldistribution∆definedon anda vector field on
⊆ ,wesay belongsto∆if
∈ ∆ , ∀ ∈
Formoreinformationpleaserefertothereference.
100
AppendixC MatlabCodesandSimulinkModels
TypicalMatlab Code to define the parameters and design the controller
(poleplacement)usingthe“place”functionofMatlab.
clear all; clc; % Structure m=1000; c=500; k=5000000; % Both systems mp=1025; bp=356.18e3; %ap=0.0808; ap=0.1095; %v4b=5.65e-10; v4b=4e-11; cp=3e-11; tm=0.004; xvmax=2.73e-3; %xvmax=2.73e+3; % Linear kc=1.5e-11; kq=0.035; % Nonlinear ps=207e5; n=2.617e-3; %xvmax=2.73e-3; kv=1; wv=1/0.004; % PID kp=80; kd=0;
101
ki=0.5; %State-space design for linear model with 3 states a1=[0 1 0; -k/(m+mp) -(bp+c)/(m+mp) ap/(m+mp); 0 -ap/v4b -(kc+cp)/v4b]; b1=[0; 0; kq/v4b]; c1=[1 0 0]; p=[-123,-26.8+29.9*1i,-26.8-29.9*1i]; kcl1=place(a1,b1,p); nbar1=-inv(c1*((a1-b1*kcl1)\b1)); %State-space design for linear model with 4 a2=[0 1 0 0; -k/(m+mp) -(bp+c)/(m+mp) ap/(m+mp) 0; 0 -ap/v4b -(kc+cp)/v4b kq/v4b; 0 0 0 -wv]; b2=[0; 0; 0; kv*wv]; c2=[1 0 0 0]; p=[-242,-21.9+31.2*1i,-21.9-31.2*1i,-141]; kcl2=place(a2,b2,p); nbar2=-inv(c2*((a2-b2*kcl2)\b2)); % Nonlinear state-space design for non-linear model with 4 states a3=[0 1 0 0; 0 0 1 0; 0 0 0 1; 0 0 0 0]; b3=[0; 0; 0; 1]; c3=[1 0 0 0]; p=[-242,-21.9+31.2*1i,-21.9-31.2*1i,-141]; kcl3=place(a3,b3,p); nbar3=-inv(c3*((a3-b3*kcl3)\b3));
102
Functiondefinitionfornonlinearbehaviouroftheservo‐valve.
function flowB = fcn(xv,PL) xv;PL; %%% this fcn calculates the flow considering the leakage Cd=0.58; rho=700; %kg/m^3, mass density PS=207e5; %Pa , supply pressure d=0.038; %m, spool diameter % xvmax=2.73e-3; %m max spool opening xvmax=2.73e-3; %%%% leakage properties of the individual valve %xvlap=6/100; xvlap=0; %Aeffnull=2.056e-6; %m^2 Aeffnull=0; xvlapm=-1*xvlap; %% default A1 and A4 A1=0; A4=0; %%% define the orifice areas if xv <= xvlapm A1=0; A4=((pi*d*xvmax-2*Aeffnull)/(1-xvlap))*abs(xv)-(((pi*d*xvmax-2*Aeffnull)/(1-xvlap))*xvlap-2*Aeffnull); end if xv >= xvlap A4=0; A1=((pi*d*xvmax-2*Aeffnull)/(1-xvlap))*abs(xv)-(((pi*d*xvmax-2*Aeffnull)/(1-xvlap))*xvlap-2*Aeffnull); end if xv > xvlapm && xv <=0 A1= -(Aeffnull/xvlap)*abs(xv)+Aeffnull; A4= (Aeffnull/xvlap)*abs(xv)+Aeffnull; end if xv > 0 && xv < xvlap A1= (Aeffnull/xvlap)*abs(xv)+Aeffnull; A4= -(Aeffnull/xvlap)*abs(xv)+Aeffnull; end %%% calculate the flow flowB=(Cd*A1*((PS-PL)/rho)^0.5)-(Cd*A4*((PS+PL)/rho)^0.5);
FigureC‐1
SimulinkmodelfornonlineardynamicsofaServo‐hydraulic
103
FigureC1
SimulinkmodelfornonlineardynamicsofaServo
hydraulic
FigureC‐2
Servo‐hydraulicsystem
withaPIDcontroller
104
FigureC2
ServohydraulicsystemwithaPIDcontroller
FigureC‐3
Servo‐hydraulicsystem
withalinearstate‐spacecontroller
105
FigureC3
Servohydraulicsystemwithalinearstatespacecontroller
FigureC‐4
Servo‐hydraulicsystem
withanonlinearstate‐spacecontroller
106
FigureC4
Servohydraulicsystemwithanonlinearstatespacecontroller
FigureC‐5
nonlinearstate‐spacecontroller
107
FigureC5
nonlinearstatespacecontroller
108
AppendixD ConferencePaper
109
MODIFICATIONS OF INTEGRATION ALGORITHMS TO ACCOUNT FOR LOAD DISCONTINUITY IN PSEUDODYNAMIC TESTING
S. Hadi Moosavi1, Oya Mercan2
ABSTRACT When there is a sudden change in the loading (e.g., rectangular pulse), the discretized version of the load history will involve an artificial impulse, which manifests itself as an amplitude distortion in the structural response obtained by the numerical solution of the equation of motion. An approach to account for load discontinuity by modifying existing integration algorithms used in the solution of force equation of motion is introduced in this paper. Modified versions of four different algorithms, namely Central Difference, Newmark Explicit, α-method with a fixed number of iterations, and Rosenbrock-W integration algorithms are presented. The general approach in modifying an integration algorithm to account for load discontinuity is discussed and the improved accuracy of these modified algorithms is presented through numerical simulations.
Introduction
Pseudodynamic (PSD) test method is a displacement based experimental technique that can be used to determine the behavior of structural systems subjected to dynamic loading. In a PSD test, a direct step by step integration algorithm generates the command displacements by solving the force equation of motion. These displacements are imposed on the test structure by a servo-hydraulic system, and using the measured restoring force feedback from the deformed test structure, the integration algorithm computes the subsequent command displacements. For load-rate insensitive structures, PSD testing method can be applied in slow time (using an expanded time axis), or for structures that exhibit load-rate dependent vibration characteristics, it can be applied at fast rates (ideally in real-time). Both the slow-time and real-time PSD testing have been successfully applied for seismic loading (Mahin 1985, Nakashima 1999), but if the loading history has a sharp discontinuity as in the case of pulse loading (see Fig. 1-a), the numerical solution of the force equation of motion will have an amplitude distortion which may render the PSD test results inaccurate. This is due to the extra impulse (shaded area in Fig. 1-b) introduced during the discretization of the loading history.
To circumvent this problem, the use of step by step solution of the momentum equation of motion was suggested and the resulting improved accuracy was verified through numerical simulations (Chang, 2001, 2002, 2007a) and experiments (Chang, 1998). In this approach, the force equation of motion is replaced by its integral form, which is the momentum equation of motion. As a result of the time integration of the load that appears on the right hand side of the momentum equation, provided that the area under the load history is computed correctly, the discontinuity in the load history is eliminated. Although the success of the momentum approach
110
has been presented for Newmark explicit integration algorithm, replacing the force equilibrium equation with the momentum equation may not be a trivial task if one wishes to use an integration algorithm customized (e.g., unconditionally stable, implicit, real-time compatible) for particular testing needs. Other than the momentum approach, in an attempt to obtain an accurate solution in the presence of load discontinuity, Chang, (2007b) also proposed the use of a single time step immediately after the discontinuity that is much smaller than the discretization step size. This small step was recommended to be one-hundredth of the discretization step size or smaller. Especially within the context of real-time PSD testing, a variable time step is detrimental for it would result in inaccurate velocities as the displacement commands are typically imposed using a digital controller with a constant clock speed.
The study presented here introduces an approach, which is referred to as limit approach,
to account for the load discontinuity by modifying a given integration algorithm that solves the force equation of motion in its final form. Depending on the way a particular integration algorithm is formulated, these modifications generally involve updated force and/or acceleration values at the time of discontinuity. In the paper, the general approach (i.e., the limit approach) that introduces modifications to a given integration algorithm is discussed and then implemented to derive modified versions of the Central difference, Newmark explicit, α-method with a fixed number of iterations, and Rosenbrock-W integration algorithms. Both α-method and Rosenbrock-W algorithms are suitable for real-time testing, where the former is an implicit scheme. For each of the four integration algorithms considered in this paper, a summary of the original formulation is provided together with its modified version. The improved accuracy of the modified algorithms is presented through numerical simulations.
The Limit Approach
The limit approach starts by defining an intermediate step just after the load
discontinuity between times (where the discontinuity takes place) and 1 (see Fig. 1-c). In order to be able to set it apart from the discretization time step size of ∆ and thereby make the implementation of the limit approach for the modification of an algorithm easier to follow, the time step size associated with step is identified as ∆ . It should be noted the load value is the value of the load at the lower end of the discontinuity at step . From the original formulation of a given integration algorithm, the information for step (which may include the displacement ( , velocity ( ) and/ or acceleration ( ), or an intermediate quantity defined by the particular integration algorithm to march forward) can be obtained using the information from step (see Fig. 1-c) and considering and ∆ . In order to obtain the information at the lower end of the discontinuity ∗ and thereby account for discontinuity effects properly, next step involves taking the limit of the expressions that define step information where ∆ goes to zero. On Fig. 1-d the information (i.e., the expressions for the load, displacement etc.) associated with ∗are the results of this limit process. In programming the resulting modified algorithm, a flag needs to be set in order to identify the time step when discontinuity (i.e., step ) takes place. When that happens, the numerical values of for ∗ , ∗etc. need to be evaluated from the expressions obtained by the limit approach, and using these, the integration algorithm in its original form can march forward to compute information for step 1.
111
Implementation of the Limit Approach
Central Difference Method
The discretized form of the equation of motion for a single-degree-of-freedom (SDOF) system at time step is:
(1) Central difference method uses a finite difference approximation for velocity and acceleration (Chopra, 2007). With a constant time step size of ∆ , the velocity and acceleration at step are expressed as:
2∆ (2)
1∆ ∆ ∆
2∆
(3)
Substituting the velocity and acceleration from Eqs. (2) and (3) in Eq. (1), and solving the expression for :
∆ 2∆2∆
/∆ 2 ∆
(4)
Implementation of Limit Approach to Obtain ( ∗) Information
Considering Fig. 1-c and rewriting Eqs. (2) and (3) for step , velocity and acceleration at that step are:
∆ ∆ ′ ,1
12 ∆ ∆ ′ ∆ ′ ∆
(5), (6)
Also, using the same equations, velocity and acceleration at step can be expressed as:
Fig. 1. (a) Load history with discontinuity, (b) Discretized load history, (c) Load history with step , (d) Load history after performing the limit
∆ ∆ ∆
∆ ∆ ∆
, , ,
, 1, 1, 1
1
, , ,
, ∗
∆ ∆ ∆∆ ∆ ′ ∆
load
time
load
time
load
time
load
time
, , ,
11 1
, 1, 1, 1
∗, ∗, ∗, ∗
112
∆ ∆ ′ ,
112 ∆ ∆ ′ ∆ ∆ ′
(7), (8)
Eqs. (9) and (10) express the equation of motion at time steps and , respectively:
(9)
(10) Once Eq. (9) is solved for (after substituting for and from Eqs. (5) and (6)), the result can be used to replace in Eq.(10) which then can be solved for (after substituting for and from Eqs. (7) and (8)). Performing a limit where ∆ ′ → 0 gives an expression for in terms of and :
2∗
∆ 2 ∆2∆
∗
/∆ 2 ∆
(11)
Comparing Eq. (4) with Eq.(11), application of limit approach to Central difference
method reveals that to account for the load discontinuity, only the load value for ( ∗) needs to be redefined as the average value of the loads at each end of the discontinuity.
Explicit Newmark method
Newmark family integration methods (Newmark, 1959) can be customized by selecting
two parameters ( and ) which specify the variation of acceleration over a time step. These parameters also determine the accuracy and stability characteristics of the method. For 0.5
0 Newmark's method becomes explicit and conditionally stable:
Δ 0.5Δ (12)
0.5 Δ0.5 Δ
(13)
Δ 0.5 0.5 (14) Implementation of Limit Approach to Obtain ( ∗) Information Defining the information for step using Eqs. (12)-(14) as required by the limit approach:
′ 0.5 ′ (15)
113
0.5 ′0.5 ′
(16)
0.5Δ (17) Performing a limit on the above expressions where∆ ′ → 0, will yield
∗ , ∗ (18), (19)
∗∗
(20)
where ∗
From above equations, it turns out that, while the values for displacement and velocity remain the same, the acceleration value at the lower end of the discontinuity ∗ needs to be updated. By stepping forward to step 1 using the updated information from Eqs. (18)- (20) derived consistently with Newmark explicit method formulation, the load discontinuity will be taken care of properly.
Rosenbrock
Considering an SDOF system the general implementation of Rosenbrock integration method proposed by Lamarche (2009) for real-time pseudo-dynamic testing is described in Fig. 2.
Implementation of Limit Approach to Obtain ( ∗) Information
Introducing step and taking the limit for the Rosenbrock algorithm, it can be easily shown that and approach to and , respectively (or equivalently ∗ and ∗ are the same as and , respectively). As a result, for Rosenbrock algorithm to handle the load
Fig. 2. Rosenbrock integration algorithm
1
Initial restoring force: Effective mass: Δ Δ
Δ12
12
,12
Δ
Δ ,
Δ
Mid-step between and 1
Where
Impose and to the structure and measure the restoring force
,
Δ2
Δ
Δ
Where
Impose and to the structure and measure the restoring force
2
3
4
5
1
114
discontinuity properly, the information at step 1 right after the discontinuity needs to be calculated using ∗, ∗ and ∗ which are the same as , , and , respectively.
Alpha method
Alpha method (Shing et al 2002) is an implicit method used in real-time pseudo-dynamic
testing which is based on Hilber -method (Hilber et al 1977). As can be seen from Fig. 3, Alpha method starts by computing a predictor displacement in stage 1 , which is followed by a fixed number of iteration substeps in stages 2 and 3 ; during the last iteration substep an equilibrium correction is performed. Through the equilibrium error correction (stages 4 and 5 ) the displacement and restoring force values are made available for the computation of the next step predictor displacement and, as a result, the actuator moves without interruption.
Implementation of Limit Approach to Obtain ( ∗) Information
Upon the application of the limit approach starting from stage 1 in Fig. 3, the predicted displacement at step approaches to after setting∆ ′ → 0 (or equivalently ∗ ) and the displacement term in stage 2 becomes constant (i.e., ∗ ). Considering that both parameters ∗ and approach to application of the limit to stages 4 and 5 gives:
Δ Δ 0.5Δ
1 1 1 Δ
Predicted displacement
Δ 1
Impose , measure the displacement and measure the restoring force
1Δ
Δ Δ 0.5 β
Δ 1
Δ 1
1
1
1
2
4
3
∗
∗
5
6
∗ Δ 1 1 Δ
Note that: 0 1,
Fig. 3. Alpha method algorithm
115
(21)
lim∆ ′→
∗ (22)
lim∆ ′→
∗ (23)
Eq. (23) is true for linear elastic systems. In stage 6 the expression for acceleration can be revised for step as
1
Δ ′Δ ′ Δ ′ 0.5 β
(24)
Upon substituting the expression of from stage 2 which has the expression for embedded from stage 1 , and as a result of the cancellations that take place, the zero over zero indeterminacy as ∆ ′ approaches to zero is eliminated; and the acceleration for ∗ is obtained:
∗ 1 ∗
(25)
where ∗ And the expression for velocity in stage 6 yields
∗ (26)
The above application of the limit approach in modifying Alpha method reveals that, to
compute the information at step 1 right after the discontinuity, information from ∗ needs to be used where the updated acceleration from Eq.(26) is used together with the value of the load at the lower end of the discontinuity.
Numerical Simulations
In order to verify the success of the proposed modifications in handling the load discontinuity, numerical simulation results for each integration algorithm are presented here. An undamped linear SDOF system with 0.2533 . / , and 10 / (i.e.,undamped natural period 1 .) subjected to the two loading cases shown in Fig.4 is considered. Fig. 4 (a) is a step pulse with an amplitude of 10 kips and duration of 0.1 sec.; whereas in Fig. 4 (b) the load value changes from +10 to -10 kips at the discontinuity and then increases to zero linearly over a duration of 0.1 sec.
W
the equatturn the amodifiedsimulatiothat of Chduration
differenc
amplitud2007).
F(unmodifversions set to ideintroducethe unmowith the t
-Dis
plac
emen
t (in
)
-Dis
plac
emen
t (in
)
When discrettion of motioamplitude did algorithms ons was selechang (2001)when the mo
ce and Newm
de decay and
ig. 5 comparfied) and moof each algo
entify the dised as derivedodified and ttheoretical r
0 0.2-1
-0.5
0
0.5
1
0 0.2-1
-0.5
0
0.5
1
izing the loaon, selectionistortion intrin handling cted as 0.1 s), where it womentum eq
mark explicit
period elon
Fig. 5. Sim
res the theorodified versioorithm were scontinuity, ad here. As catheoretical reresponse. Sim
0.4 0.6time (sec)
Central Differecenc
0.4 0.6time (sec)
Rosenbrock-W
(kips)
10
0.1
Fig. 4. Lo
ading historyn of a very smoduced in thlarge extra i
sec.; which awas shown thquation is use
t algorithms
gation introd
ulation result
retical responons of the inprogrammedand the modan be seen fresponses; whmilarly, for F
0.8
ce
modifiedtheoreticalunmodified
0.8
modifiedtheoreticalunmodified
(sec) 0.2
ad cases with
y for subsequmall time stehe response. impulses, thealso made theat time step ed. It needs
∆
=0.1 sets
duced by the
ts for the load
nse with thentegration algd in MATLA
dified expresrom Fig. 5, thereas the mFig. 6 which
1
0-1
-0.5
0
0.5
1
Dis
plac
emen
t (in
)
1
0-1
-0.5
0
0.5
1
Dis
plac
emen
t (in
)
h discontinuity
uent use in thep will minimTo be able te time step se results of tcan be selecto be pointe
the limit for
e algorithm m
ding case in F
results obtagorithms conAB. In the msions for loa
there is a sigmodified resuh considers n
0.2
0.2
(kips)
10
10
00.1
y
he step-by-smize the extrto check the size for the nthis study cocted as large d out that fo
r accuracy, b
may be signi
ig. 4(a)
ained from thnsidered in t
modified versad and/or accgnificant diffults are in gononzero load
0.4 0.6time (sec)
Newmark Explicit
0.4 0.6time (sec)
Alpha Method
(sec) 0.2
step solutionra impulse aability of th
numerical omparable w
as the impuor Central
beyond whic
ificant (Chop
he original this study. Bsion, a flag wceleration wference betwood agreemend values after
0.8 1
modifiedtheoreticalunmodified
0.8 1
modifiedtheoreticalunmodified
116
n of and in he
with ulse
ch the
pra,
Both was
were ween
nt r the
117
discontinuity, the proposed modifications provide considerable improvements in the accuracy of the numerical results. The unmodified responses from each algorithm between Fig. 5 and 6 are the same as expected, since with a time step size of 0.1 sec., the discretized versions of the load cases in Fig. 4 (a) and (b) are the same (see Fig. 1 (b)). Depending on the accuracy characteristics of the particular integration algorithm, the agreement between the theoretical response and the modified numerical solution can be improved by using a smaller time step.
Fig. 6. Simulation results for the loading case in Fig. 4(b).
Conclusion
Pseudodynamic testing method has been implemented successfully both in slow and real-
time for seismic loading of structures. However, when a sharp discontinuity exists in the loading history as in the case of pulse loading, the discretized version of the load will have an extra distortion which manifests itself as an amplitude distortion in the numerical response and may render the pseudodynamic test results inaccurate. Other than using very small time steps in discretizing the load, previous studies proposed the use of numerical solution of the momentum equation of motion that replaces the force equation of motion. Although the success of the momentum approach has been presented using Newmark explicit integration algorithm, replacing the force equilibrium with the momentum equation may not be a simple task if one wishes to use an integration algorithm customized for particular testing needs. The study presented here introduces a limit approach that considers the force equation of motion and modifies the integration algorithms in their final form to account for the load discontinuity. The general modification approach and its implementation to four integration algorithms are provided together with numerical simulation results that show the improved accuracy of the modified algorithms.
References
Chang, S. Y., Tsai, K. C., & Chen, K. C. (1998). Improved time integration for pseudodynamic
tests. Earthquake Engineering & Structural Dynamics, 27 (7), 711-730.
0 0.2 0.4 0.6 0.8 1-1
-0.5
0
0.5
1
time (sec)
disp
lace
men
t (in
)
Central Difference
modifiedtheoreticalunmodified
0 0.2 0.4 0.6 0.8 1-1
-0.5
0
0.5
1
time (sec)
disp
lace
men
t (in
)
Newmark Explicit
modifiedtheoreticalunmodified
0 0.2 0.4 0.6 0.8 1-1
-0.5
0
0.5
1
time (sec)
disp
lace
men
t (in
)
Rosenbrock
modifiedtheoreticalunmodified
0 0.2 0.4 0.6 0.8 1-1
-0.5
0
0.5
1
time (sec)
disp
lace
men
t (in
)
Alpha Method
modifiedtheoreticalunmodified
118
Chang, S. Y. (2001). Analytical study of the superiority of the momentum equations of motion for impulsive loads. Computers & Structures, 79 (15), 1377-1394.
Chang, S. Y. (2002). An improved on-line dynamic testing method. Engineering Structures , 24 (5), 587-596.
Chang, S. Y. (2007 a). A technique for overcoming load discontinuity in using Newmark method. Journal of Sound and Vibration , 304, 556-569.
Chang, S. Y. (2007 b). Approach for overcoming numerical inaccuracy caused by load discontinuity. Journal of Engineering Mechanics, 133 (5), 555-565.
Chopra, A. K. (2007). Dynamics of Structures. New Delhi: Prentice-hall Of India Pvt Ltd. Hilber, H. M., Hughes, T. J., & Taylor, R. L. (1977). Improved numerical dissipation for time
integration algorithms in structural dynamics. Earthquake Engineering & Structural Dynamics, 5, 283-292.
Lamarche, C. P., Bonelli, A., Bursi, O. S., & Tremblay, R. (2009). A Rosenbrock-W method for real-time dynamic substructuring and pseudo-dynamic testing. Earthquake Engineering & Structural Dynamics, 38, 1071–1092.
Mahin, S. A., & Shing, P.-s. B. (1985). Pseudodynamic method for seismic testing. Journal of Structural Engineering, 111 (7), 1482-1503.
MATLAB is a registered trademark of The MathWorks, Inc. http://www.mathworks.com/ Nakashima, M., & Masaoka, N. (1999). Real-time on-line test for MDOF systems. Earthquake
Engineering and Structural Dynamics, 28 (4), 393-420. Newmark, N. M. (1959). A method of compuation for structural dynamics. Journal of
Engineering Mechanics Division, 67-94. Shing, P. B., Spacone, E., & Stauffer, E. (2002). Conceptual design of fast hybrid test system at
the University of Colorado. 7th U.S. National Conference on Earthquake Engineering, (2389-2398). Boston, MA (USA).
119
References
Ahmadizadeh,M.2007.Real‐timeseismichybridsimulationprocedures for
reliable structural performance testing. Doctor of philosophy, State
UniversityofNewYorkatBuffalo,UnitedStates,NewYork.
Ayalew,B.andJablokow,K.2007.Partialfeedbacklinearisingforce‐tracking
control: Implementation and testing in electrohydraulic actuation, Control
Theory&Applications,IET,1:689‐698.
Dimig, J., Shield, C., French, C., Bailey, F., andClark,A. 1999.Effective force
testing: A method of seismic simulation for structural testing, Journal of
StructuralEngineering,125:1028‐1037.
Franklin, G.F., Powell, J.D., andEmami‐Naeini, A. 2010. Feedback control of
dynamicsystems.Pearson,UpperSaddleRiverN.J.
How, J. 2007. Course materials for 16.31 feedback control systems. MIT
OpenCourseWare,MassachusettsInstituteofTechnology.
Kwon,J.,Kim,T.,Jang,J.S.,andLee,I.2007.Feedbacklinearizationcontrolof
a hydraulic servo system. In SICE‐ICASE, 2006. International Joint
Conference,pp.455‐460.
Lim,T.2002.Poleplacementcontrolofanelectrohydraulicservomotor. In
PowerElectronicsandDriveSystems,1997.Proceedings.,1997International
Conferenceon,Vol.1,pp.350‐356.
120
Lynch, A. 2009. Differential geometry ‐ basic, notes for nonlinear control
design.UniversityofAlberta.
Mahin, S.A. and Williams, M.E. 1981. Computer controlled seismic
performance testing. In Second ASCE‐EMD Specialty Conference
onDynamicResponseofStructures,Atlanta,GA.
Mahin, S., A. and Shing, P., B. 1985. Pseudodynamic method for seismic
testing,JournalofStructuralEngineering,111:1482‐1503.
Marino,R.andTomei,P.1995.Nonlinearcontroldesign:Geometric,adaptive,
androbust.PrenticeHall,.
Mercan, O. and Ricles, J.M. 2008. Stability analysis for real‐time
pseudodynamicandhybridpseudodynamictestingwithmultiplesourcesof
delay,EarthquakeEngineering&StructuralDynamics,37:1269‐1293.
Mercan,O.,Zhang,X.P., andRicles, J.M.2006.State‐spacecontroldesign for
real‐timepseudodynamictestingapplications.InProceedingsofthe8thU.S.
National Conference on Earthquake Engineering, San Francisco, California,
USA.
Mercan, O. 2007. Analytical and experimental studies on large scale, real‐
time pseudodynamic testing. Ph.D., Lehigh University, United States ‐‐
Pennsylvania.
Merritt,H.E.1967.Hydrauliccontrolsystems.Wiley,NewYork.
121
Mintsa, H.A., Kenne, J.P., and Venugopal, R. 2009. Adaptive control of an
electrohydraulicpositionservosystem.InAFRICON,2009.AFRICON'09.,pp.
1‐6.
Moog,I.C.D.2010.G761seriesservo‐valves,productcatalog.EastAurora,NY.
Okada,T.,Seki,M.,andPark,Y.J.1980.Asimulationofearthquakeresponse
ofreinforcedconcretebuildingframestobi‐directionalgroundmotionbyIIS
computer‐actuator on‐line system. In Seventh World Conference on
EarthquakeEngineering,Istanbul,Turkey.
Seo, J., Venugopal, R., and Kenné, J.P. 2007. Feedback linearization based
control of a rotational hydraulic drive, Control Engineering Practice, 15:
1495‐1507.
Shing, P.B. and and Mahin, S.A. 1983. Experimental error propagation in
pseudo‐dynamic testing. UCB/EERC‐83/12, Earthquake Engineering
ResearchCenter,UniversityofCalifornia,Berkeley,California.
Takanashi, K., Udagawa, K., Seki, M., Okada, T., and Tanaka, H. 1975.
Nonlinear earthquake response analysis of structures by a computer‐
actuatoron‐linesystem.8,EarthquakeResistantStructureResearchCenter,
InstituteofIndustrialScience,UniversityofTokyo,Tokyo,Japan.
Thayer, W.J. 1958 (revised in 1965). Transfer functions for MOOG servo‐
valves,technicalbulletin103.MOOGInc.,ControlDivision,EastAurora,NY.
122
Yanada,H.andFuruta,K.2007.Adaptivecontrolofanelectrohydraulicservo
systemutilizing online estimate of its natural frequency,Mechatronics,17:
337‐343.
Zhang,X.,Ricles,J.M.,Mercan,O.,andChen,C.2005.Servo‐hydraulicsystem
identificationfortheNEESreal‐timemulti‐directionalearthquakesimulation
facilty.05‐14,ATLSS,LehighUniversity.
Zhao, J. 2003. Development of EFT for nonlinear SDOF systems. Ph.D.,
UniversityofMinnesota,UnitedStates‐‐Minnesota.