You are on page 1of 26

Nonlinear

identification

Luiz
Góes
Nonlinear identification using polynomial
Introduction

Nonlinear
NARMAX model and a stability analysis of an
aeroelastic
model aeroelastic system
Nonlinear
identification
using
polynomial
NARMAX Luiz Góes
model
Model structure
selection
Raphaela Barbosa, Airton Nabarrete, José Balthazar, David Zúñiga
Identification of the
system with noise Technological Institute of Aeronautics - ITA
A stability
analysis
Influence of the pitch June 15, 2019
angle

Conclusion
Nonlinear
identification

Luiz 1 Introduction
Góes
Introduction 2 Nonlinear aeroelastic model
Nonlinear
aeroelastic
model 3 Nonlinear identification using polynomial NARMAX model
Nonlinear Model structure selection
identification
using Identification of the system with noise
polynomial
NARMAX
model
Model structure
selection
4 A stability analysis
Identification of the
system with noise
Influence of the pitch angle
A stability
analysis
Influence of the pitch
5 Conclusion
angle

Conclusion
Introduction

Nonlinear
identification

Luiz Choosing a model... what type of model can I use?


Góes
white box model.
Introduction

Nonlinear
black box model.
aeroelastic
model gray box model.
Nonlinear
identification
Mathematic representations:
using
polynomial transfer function, difference equation, time series models
NARMAX
model (ARX, ARMA, ARMAX,...) and state-space equation.
Model structure
selection and the choice is based on...
Identification of the
system with noise
control system design, facility in mathematical
A stability
analysis manipulation, parametrization for MIMO systems.
Influence of the pitch
angle

Conclusion
Introduction

Nonlinear
identification System identification consists in choosing the best model
Luiz from a given model set based on the system input-output
Góes data.
Models obtained by identification process are valid under
Introduction

Nonlinear
some conditions and the parameters of model have no
aeroelastic
model
physical meaning.
Nonlinear
It’s difficult to obtain a universal mathematical model of
identification
using
the system by performing the system identification only at
polynomial
NARMAX
a single excitation level.
model
Model structure
selection
Identification of the
system with noise

A stability
analysis
Influence of the pitch
angle

Conclusion
Figure: A system representation: uk is the signal input and yk is the
signal output. The system has disturbance that is modeled as noise.
Nonlinear aeroelastic model

Nonlinear In this work the identification procedure presented will be suitable


identification
to apply to the identification of a typical section apparatus.
Luiz
Góes It’s a typical section with pitch and plunge degrees of freedom.
The rigid wing has one flight control surface and is supported by
Introduction
an elastic structure with low damping.
Nonlinear
aeroelastic
model

Nonlinear
identification
using
polynomial
NARMAX
model
Model structure
selection
Identification of the
system with noise

A stability
analysis (a) (b)
Influence of the pitch
angle

Conclusion Figure: Airfoil apparatus available at the Laboratory.


Nonlinear aeroelastic model

Nonlinear
identification
The mathematical model of the aeroelastic system is presented
as
Luiz
Góes
Introduction

Nonlinear
aeroelastic
model

Nonlinear
identification
using
polynomial
NARMAX Figure: Typical section model of an airfoil (Li, 2010).
model
Model structure
selection
Identification of the
          
mT mT xα b ḧ c 0 ḣ k 0 h −L
system with noise + h + h = (1)
mT xα b Iα α̈ 0 cα α̇ 0 kα (α) α M
A stability
analysis
Influence of the pitch
with static nonlinearity in the form
angle
2 3 4
Conclusion kα (α, t) = kα0 + kα1 α + kα2 α + kα3 α + kα4 α (2)
Nonlinear aeroelastic model

Nonlinear
identification
Rearranging the terms from Equation 1 the stiffness and damping
matrices are obtained including the aeroelastic influence under
Luiz the free stream velocity as
Góes
   
ρUbClα ρUb 2 Clα 1 −a
       
mT mT xα b ḧ c 0 ḣ 2 ḣ
Introduction + h +  (3)
mT xα b Iα α̈ 0 cα α̇ ρUb 2 Cmα −ρUb 3 Cmα 1 −a α̇
2
Nonlinear
aeroelastic
model
ρU 2 bClα
       
kh 0 h 0 h −ρbClβ 2
Nonlinear + + = U β
0 kα (α) α 0 −ρU 2 b 2 Cmα α ρb 2 Cmα
identification
using
polynomial
NARMAX
model
Model structure
Table: Continuous model parameters.
selection
Identification of the Parameter Value Parameter Value Parameter Value
system with noise mT [kg ] 8 ξh 0.022 kα0 [N m/rad] 3
A stability Ia [kg m2 ] 0.0505 ξα 0.03 kα1 [N m/rad 2 ] -30
analysis xa [m] 0.084 ch [N s/m] 6.6373 kα2 [N m/rad 3 ] 6600
Influence of the pitch b [m] 0.1064 cα [N s m] 0.2939 kα3 [N m/rad 4 ] -21000
angle
a -0.8 kh [N/m] 2844.4 kα4 [N m/rad 5 ] 48000
Conclusion
Nonlinear aeroelastic model

Nonlinear
identification

Luiz
Góes
Introduction

Nonlinear
aeroelastic
model

Nonlinear
identification
using
polynomial
NARMAX (a) (b)
model
Model structure
selection
Identification of the Figure: Time history for initial conditions h(0) = 0.1 m and
system with noise
α(0) = 0.01 rad: (a) vertical displacement and velocity; (b) angular
A stability
analysis displacement and velocity.
Influence of the pitch
angle

Conclusion
Nonlinear identification using polynomial NARMAX
model
Nonlinear
identification

Luiz The NARMAX model was introduced in 1981 as a new


Góes representation for a wide class of nonlinear systems
Introduction
(Billings ans Leontaritis, 1981).
Nonlinear In practice, many types of model structures are available
aeroelastic
model to approximate the unknown mappinf F[.]. Some examples
Nonlinear are: polynomial NARMAX models, rational NARMAX
identification
using models and the extended model set representations.
polynomial
NARMAX
model
The nonlinear identification using polynomial NARMAX model
Model structure consists in the representation of the system output as a polyno-
selection
Identification of the mial function of y (n), u(n) and e(n) with nonlinearity order l.
system with noise
It follows
A stability
analysis l  
y (n) = F y (n − 1), · · · , y (n − ny ), u(n − nu − d), · · · , u(n − nu ), e(n − 1), · · · , e(n − ne ) + e(n) (
Influence of the pitch
angle

Conclusion
Nonlinear identification using polynomial NARMAX
model
Nonlinear
identification

Luiz
The model structure to represent the airfoil have two outputs,
Góes
ḣ(n) and α̇(n), with delays ny = 1, nu = 1, ne = 1, dead time
Introduction
d = 0, nonlinearity order l = 5 corresponding to polynomial of
Nonlinear
aeroelastic fourth order to represent the torsional stiffness kα (α).
model

Nonlinear The nonlinear identification deals with a typical least


identification
using
square problem. As so, the nonlinearities appear in the
polynomial
NARMAX
regression matrix and the problem is linear in the
model
Model structure
parameters. It becomes
selection
Identification of the
system with noise

A stability
y (n) = ψ T (n − 1)θ̂ + ξ(n) (5)
analysis
Influence of the pitch
angle

Conclusion
Model structure selection

Nonlinear
identification

Luiz
Góes The structure of the continuous-time model has the form
Introduction
ẍ + M−1 Cẋ + M−1 Kx = M−1 Q (6)
Nonlinear
aeroelastic
model
   
ḧ −L
Nonlinear where, ẍ = and Q = .
identification α̈ M
using
polynomial Rearranging, the continuous-time model is represented as
NARMAX
model    
Model structure ḧ −k1 h − [k2 µ + pkα (α)] − c1 ḣ − c2 α̇
selection = (7)
Identification of the
system with noise
α̈ −k3 h − [k4 µ + qkα (α)] − c3 ḣ − c4 α̇
A stability
analysis
Influence of the pitch
angle

Conclusion
Model structure selection

Nonlinear The NARMAX structure follows the discrete-time model repre-


identification sented as
Luiz 2 3
ḣ(n) = θ1 ḣ(n−1)+θ2 h(n−1)+θ3 α(n−1)+θ4 α(n−1) +θ5 α(n−1) +θ6 α(n−1) +θ7 α(n−1)
4 5
(8)
Góes
+θ8 α̇(n − 1) + θ9 u(n − 1) + θ10 eḣ (n − 1) + eḣ (n)
Introduction
2 3 4 5
α̇(n) = γ1 α̇(n − 1) + γ2 h(n − 1) + γ3 α(n − 1) + γ4 α(n − 1) + γ5 α(n − 1) + γ6 α(n − 1) + γ7 α(n − 1)
Nonlinear
aeroelastic +γ8 ḣ(n − 1) + γ9 u(n − 1) + γ10 eα̇ (n − 1) + eα̇ (n)
model

Nonlinear
identification
using
polynomial Table: Discrete-time model parameters.
NARMAX
model Parameter Value Identified Parameter Value Identified
Model structure θ1 0.9986 0.9986 γ1 0.9938 0.9938
selection θ2 -0.3601 -0.3601 γ2 0.5099 0.5105
Identification of the θ3 -0.0041 -0.0041 γ3 -0.0737 -0.0737
system with noise
θ4 0.0018 0.0019 γ4 0.2006 0.2024
A stability θ5 -0.3944 -0.4060 γ5 -44.1227 -44.1026
analysis θ6 1.2548 0.9508 γ6 140.3904 135.9770
Influence of the pitch
θ7 -2.8680 21.3053 γ7 -320.8924 -238.6472
angle θ8 -0.0000 -0.0000 γ8 -0.0011 -0.0011
θ9 -0.0019 -0.1110 γ9 -0.0035 -0.2021
Conclusion θ10 -0.9986 -0.9529 γ10 -0.9938 -0.9515
Model structure selection

Nonlinear
identification The Error Reduction Ratio - ERR criterior
Luiz Assume that the residual signal e is uncorrelated with the
Góes past outputs signals, then the total energy can be
Introduction expressed as
Nonlinear
aeroelastic
model M
1 T 1 X 2 T 1
Nonlinear Y Y= gi wi wi + eT e (9)
identification N N N
using i=1
polynomial
NARMAX
model
Model structure and the error reduction ratio can be defined as (Billings,
selection
Identification of the
system with noise
2013)
A stability
analysis
< y, wi >2
Influence of the pitch
angle
ERRi = × 100%, i = 1, 2, ..., M (10)
< y, y >< wi , wi >
Conclusion
Model structure selection

Nonlinear
identification The ERR criterior provides a very simple but effective
Luiz means of determining a subset of significant regressors.
Góes In this work was applied the Forward Regression OLS (FROLS)
Introduction
algorithm to ordering the model terms (Billings, 2013).
Nonlinear
aeroelastic
model Table: Results of the FROLS algorithm applied to the identified
Nonlinear model. Obs: A large ERR means significant model term (with more
identification
using
energy).
polynomial
NARMAX Index Model term Parameter ERR(%) Index Model term Parameter ERR(%)
model 1 ḣ(n − 1) 0.9986 98.1860 1 α̇(n − 1) 0.9938 95.2994
Model structure
2 h(n − 1) -0.3601 0.0023 2 h(n − 1) 0.5105 0.0030
selection 3 α(n − 1) -0.0041 0.0001 3 α(n − 1) -0.0737 0.0078
Identification of the 4 α(n − 1)2 0.0019 0.0000 4 α(n − 1)2 0.2024 0.0000
system with noise
5 α(n − 1)3 -0.4060 0.0000 5 α(n − 1)3 -44.1026 0.0004
A stability 6 α(n − 1)4 0.9508 0.0000 6 α(n − 1)4 135.9770 0.0004
analysis 7 α(n − 1)5 21.3053 0.0000 7 α(n − 1)5 -238.8924 0.0008
Influence of the pitch 8 α̇(n − 1) -0.0000 0.0048 8 ḣ(n − 1) -0.0011 0.0049
angle 9 u(n − 1) -0.1110 0.0277 9 u(n − 1) -0.2021 0.0007
10 eḣ (n − 1) -0.9529 0.8450 10 eα̇ (n − 1) -0.9515 2.2135
Conclusion P P
=99.1860 =96.2994
Identification of the system with noise

Nonlinear
identification
A polynomial NARMAX model is used to identify the system with
noise. The identification procedure consists on the application
Luiz
Góes of an Extended Least Square Estimator (ELS) as indicated by
Aguirre (2007), following the sequential steps below:
Introduction

Nonlinear
1 Mount the matricial equation y = ψθ + e from the measured
aeroelastic
model
data to determine θ̂LS .
Nonlinear 2 Calculate the residue ξ1 = y − ŷ.
identification
using 3 Take k = 2.
polynomial
NARMAX
model
4 Mount the regression matrix ψk∗ using the extended regressors

Model structure from the estimative θELS,k .
selection
Identification of the
system with noise
5 Calculate the residue ξk = y − ŷ∗ .
A stability
analysis
6 Take k = k + 1 and return back to Step 4. Repeat until it
Influence of the pitch
angle
converges. The criterion for looping stop can be performed

Conclusion
using the residue variance σξ2k or the parameters vector θELS,k .
Typically, occurs in the range 3 ≤ k ≥ 10.
Identification of the system with noise

Nonlinear
identification
The process is excited by the angular displacement of the control
Luiz surface β, with an aleatory signal having constant levels of 12 s
Góes
each as depicted in figure below. The ELS procedure is repeated
Introduction until iteration k = 10.
Nonlinear
aeroelastic
model

Nonlinear
identification
using
polynomial
NARMAX
model
Model structure
selection
Identification of the
system with noise

A stability
analysis
Influence of the pitch
Figure: Excitation signal for system identification.
angle

Conclusion
Identification of the system with noise

Nonlinear
identification

Luiz
Góes
Introduction

Nonlinear
aeroelastic
model

Nonlinear
identification
using
polynomial
NARMAX
model (a) (b)
Model structure
selection
Identification of the
system with noise Figure: (a) Output prediction for the linear velocity, ḣ(n) and (b)
A stability enlarged time for linear velocity.
analysis
Influence of the pitch
angle

Conclusion
Identification of the system with noise

Nonlinear
identification

Luiz
Góes
Introduction

Nonlinear
aeroelastic
model

Nonlinear
identification
using
polynomial
NARMAX
model (a) (b)
Model structure
selection
Identification of the
system with noise
Figure: (c) output prediction for the angular velocity, α̇(n) and (d)
A stability enlarged time for angular velocity.
analysis
Influence of the pitch
angle

Conclusion
Identification of the system with noise

Nonlinear The identified stiffness parameters are present in the follow ta-
identification ble.
Luiz
Table: Parameters of the nonlinear stiffness.
Góes
Parameter Value Identified
Introduction kα0 3 2.99
kα1 -30 -30.24
Nonlinear kα2 6600 6592.35
aeroelastic kα3 -21000 -20325.51
model kα4 48000 35672.42
Nonlinear
identification
using
polynomial
NARMAX
model
Model structure
selection
Identification of the
system with noise

A stability
analysis
Influence of the pitch
angle

Conclusion
Figure: A representation of the nonlinear behavior: moment as
function of the angular displacement.
A stability analysis

Nonlinear
identification A continuous state-space model may be expressed as follows
Luiz
Góes ẋ(t) = fµ [x(t)] + g [x(t)]µβ (11)
Introduction

Nonlinear   
0 0 1 0
   
x˙ (t) x (t) 0
aeroelastic  1   1 

 

x˙2 (t)  0 0 0 1  x2 (t) +  0  µβ
   
model = 
−k1 −[k2 µ + pkx2 (x2 (t))] −c1 −c2   (12)
x˙3 (t) x (t)
 3  g3
  

Nonlinear
x˙4 (t) −k3 −[k4 µ + qkx2 (x2 (t))] −c3 −c4 x4 (t) g4
identification
using
polynomial It can be observed that the term kx2 (x2 ) is kα (α) and due the
NARMAX
model nonlinearity it is necessary to calculate the Jacobian matrix for
Model structure
selection
evaluating the eigenvalues. Therefore the Jacobian matrix is
Identification of the given by
system with noise

A stability
 
0 0 1 0
analysis  0 0 0 1 
J= (13)
−[k2 µ + p(kα0 + kα1 α(t) + kα2 α(t)2 ) + kα3 α(t)3 ]

Influence of the pitch −k1 −c1 −c2 
angle
−k3 −[k4 µ + q(kα0 + kα1 α(t) + kα2 α(t)2 ) + kα3 α(t)3 ] −c3 −c4 |α=α
op
Conclusion
A stability analysis

Nonlinear According to equations of motion it is possible to verify that the non-


identification
linearity affects the stability when a negative damping factor occurs,
Luiz resulting in unstable behavior. The follow figure illustrates this behav-
Góes ior.
Introduction

Nonlinear
aeroelastic
model

Nonlinear
identification
using
polynomial
NARMAX
model
Model structure
selection
Identification of the
system with noise

A stability (a) (b)


analysis
Influence of the pitch
angle
Figure: (a) Root locus for the free stream velocity variation
Conclusion
considering the equilibrium point and (b) damping factor as functions
of the free stream velocity.
Influence of the pitch angle

Nonlinear The system becomes unstable due to the nonlinearity as shown before.
identification
In this case, the instability was obtained in the figure below varying
Luiz the system operation point (named xop ). In this case, the angular dis-
Góes placement sweep occurred because the torsional stiffness is a function
Introduction of the α values.
Nonlinear
aeroelastic
model

Nonlinear
identification
using
polynomial
NARMAX
model
Model structure
selection
Identification of the
system with noise

A stability
analysis
Influence of the pitch (a) (b)
angle

Conclusion
Figure: (a) Root locus for variation in nonlinear stiffness considering
six free stream velocities and (b) enlarged region.
Influence of the pitch angle

Nonlinear The root locus considering the free stream velocity of 8 m/s with the
identification
stiffness varying approximately from 0 Nm/rad to 27.07 Nm/rad is
Luiz shown in the figure (a). It is observed in the figure (b) the region of
Góes instability for stiffness varying from 6.11 Nm/rad to 8.56 Nm/rad.
Introduction

Nonlinear
aeroelastic
model

Nonlinear
identification
using
polynomial
NARMAX
model
Model structure
selection
Identification of the
system with noise

A stability
analysis (a) (b)
Influence of the pitch
angle

Conclusion
Figure: (a) Root locus for V∞ = 8 m/s varying α between 0 rad and
0.07 rad , (b) damping factor as function of a nonlinear stiffness.
Influence of the pitch angle

Nonlinear
identification
It is observed in the follow figure that the region of instability for
stiffness varying from 6.11 Nm/rad to 8.56 Nm/rad. Through
Luiz
Góes several simulations it is possible to note that the LCO frequency
increases as the stiffness rises, considering the stability region.
Introduction

Nonlinear
aeroelastic
model

Nonlinear
identification
using
polynomial
NARMAX
model
Model structure
selection
Identification of the
system with noise

A stability
analysis (a)
Influence of the pitch
angle

Conclusion
Figure: Stiffness kα versus angular displacement α.
Conclusion

Nonlinear
identification
The nonlinear identification results in a suitable nonlinear
Luiz torsional stiffness. The proposed analysis is validated to be
Góes
applied with the experimental data from the aeroelastic
Introduction system.
Nonlinear
aeroelastic The polynomial NARMAX model is obtained using the
model
extended least square estimator (ELS) to obtain unbiased
Nonlinear
identification estimation even with colored noise in measures. The
using
polynomial estimates converges quite quickly within ten iterations.
NARMAX
model The identified model serves to the stability analysis and
Model structure
selection the control system design with nonlinear approach.
Identification of the
system with noise
From the stability analysis, the mapping of the unstable
A stability
analysis regions is outlined based on the eigenvalues of the
Influence of the pitch
angle Jacobian matrix. It concludes that the nonlinearity can
Conclusion lead the system to be dynamically unstable.
Nonlinear
identification

Luiz
Góes
Thank you!
Introduction

Nonlinear
aeroelastic
goes@.ita.br
model machado@.ita.br
Nonlinear
identification
using
polynomial
NARMAX
model
Model structure
selection
Identification of the
system with noise Acknowledments: ITA and CAPES
A stability
analysis
Influence of the pitch
angle

Conclusion

You might also like