Anda di halaman 1dari 108

Methods for Flight Vehicle System Identification

Maneuver
Input

Optimized
Input

Actual
Response

FlightVehicle

Measurements
DataCollection
&Compatibility

Methods
APrioriValues,
lower/upper
bounds

Estimation
Algorithm /
Optimization

Models
Model
Structure

Mathematical
Model/
Simulation

Identification
Criteria

Response
Error

Parameter
Adjustments
ModelResponse

Identification Phase
Model
Complementary
Flight Data
Validation

Dr. Ravindra Jategaonkar

Validation Phase

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/1

This page is left intentionally blank.

Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/2

Model Equations
State equations:
x
u
w
F

x0

z (t k ) = y (t k ) + G v(t k ),

k = 1,..., N

measurement noise, N(0,I)

Unknown parameters:
x
y

y (t ) = g[ x(t ), u (t ), x , y ]

output variables

Discrete measurements:
v

x(t 0 ) = x0

state vector
input vector
additive process noise, N(0,I)
process noise distribution matrix

Observation equations:
y

x& (t ) = f [ x(t ), u (t ), x ] + F ( ) w(t );

T = ( xT ;

Ty ;

T ; x0T )

system parameters of the state eqs.


system parameters of the observation eqs.
elements of noise distribution matrix
initial conditions

Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/3

Model Extensions (I)


Multi run evaluation (concatenating several flight maneuvers)
Bias errors in measurements of input and output variables
Nonlinear systems
x&(t ) = f [ x(t ), u(t ) u(bu,l ), x ],

x(t0,l ) = x0 (bx,l )

z(tk ) = y(tk ) + G v(tk )

k =1, ... , N

y(t ) = g[ x(t ), u(t ) u(bu,l ), x , y ] + z(by,l ) , l =1, ... , nz

nz
N
m
n
p

number of time slices /experiments analyzed simultaneously


total number of data points analyzed
number of output variables
number of state variables
number of input variables

Unknown parameters:

{ }T

= Tx ; Ty ; bTx,1 , . . , bTx,nz ; bTy,1 , . . , bTy,nz ; bTu,1 , . . , bTu,nz

n initial conditions and m + p bias parameters.

Not all of them can be estimated (linear dependence, high correlation).


Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/4

Model Extensions (II)


Linear Systems:
- State variable transformation x:= x-x0 leads to equivalent model:
x& (t ) = f [ x(t ), u (t ), x ] + bx,l
y (t ) = g[ x(t ), u (t ), x , y ] + b y ,l
z (t k ) = y (t k ) + G v(t k )

= A x(t ) + B u (t ) + bx,l
= C x(t ) + D u (t ) + b y ,l

x(t 0,l ) = 0
l = 1,... , nz
k = 1,... , N

- bias terms bx = A x0 - B u

by = C x0 - D u + z

exactly n + m bias parameters


all can be estimated

Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/5

Parameter Estimation Methods


Complications:
1) Presence of noise --> no longer possible to exactly identify the values of unknown
parameters; instead , the values must be estimated by some criterion (to average out
the noise effects).
2) Modeling errors --> are deterministic, but are treated simply as noise (rigorously
not justifiable, but probably the best approach).
Equation error

- Linear in parameters and independent variables (one-shot)


- Nonlinear least squares (iterative solution)
- Accounts for process noise

Output error

- Accounts for measurement noise


- Nonlinear optimization problem: iterative solution

Filter error

- Accounts for both process and measurement noise


- state and parameter estimation

EKF/UKF

- Filtering approach to parameter estimation


Generalized
approach

Special case

Equation Error

Filter Error
No
Measurement
noise

Dr. Ravindra Jategaonkar

Special case

Output Error
No
Process
noise

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/6

Regression Analysis (1)


Model postulate

yi (tk ) = i1 x1(tk ) + i 2 x2 (tk ) + ... + ir xr (tk ) + ei (tk )


yi
x1,...,xr
i1,...,ir
ei

k = 1,... , N

dependent variable (response variable)


independent variables (regressors, explanatory variables)
unknown parameters
equation error

Assumptions:
- Measurements of y corrupted by noise (zero mean)
- Measurements of x are exact and noise free
Equation error:

(k ) = y (k ) x T (k )

LS cost function:

1
J ( ) =
2

2 (k) = 2 T

k =1

Directly in terms of error.


Not based on probability theory.

LS method applicable to linear as well as nonlinear models:


- As for the cost function, it is immaterial whether the errors result from
a linear or nonlinear model.
- The minimization procedure is model type dependent.
Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/7

Regression Analysis (2)


Linear least squares:
One shot procedure to solve Normal equation:

( X T X ) T = X T Y

x1 (1) x 2 (1) ...


x (2) x (2) ...
2
1
X =

x1 ( N ) x 2 ( N ) ...

x nq (1)
x nq (2)

x nq ( N )

Y = [ y (1) y ( 2) . . .

y ( N )]T

Nonlinear least squares problems can be solved only iteratively.


Statistical properties:
- Unbiased estimates under given assumptions
(noise zero mean; independent variables exact)
- In the presence of noise and systematic (bias, scale factor) errors in x,
LS estimates are biased and inefficient.
- Leads to two step procedure of aerodynamic model identification:
a) FPR to eliminate systematic instrument errors
b) estimate aerodynamic parameters using LS method
Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/8

Regression Analysis (3)


Total least squares (TLS)
Orthogonal Distance Regression (ODR)
Error-in-Variable (EIV) Modeling

y = ax

yOLS
(x2, y2)

yTLS

TLS accounts for noise in


the independent variables.

(x1, y1)

(x3, y3)

TLS solution based on SVD


and rank reduction.
Simple linear solution not possible.
Equivalently, TLS = ( X T X n2+1 I ) 1 X T Y : Includes a correction term
nq +1 smallest singular value.
TLS estimates are unbiased, provided noise is zero mean.
No general procedure within the framework of regression techniques
to account for systematic errors.
Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/9

Regression Analysis (4)


Data Partitioning
80
deg

angle of
55
attack

50

20

100

time

55
50

possible because LS does not rely on the


temporal relation between data points
Improved information contents leads to
reduced correlation
improved parameter estimates
Dr. Ravindra Jategaonkar

55

50
0

time

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

10
Methods/10

Output Error Method (1)


Input

Measurement
noise

Dynamic System

Measured
response

+
+

Mathematical Model
Integration of State Eq.
Observation Equations

y
Sensitivities

Parameter update
By optimization of
Likelihood Function

Response
Error

(z-y)

Advantages: Computationally (relatively) simple; Readily extendable to


nonlinear systems, widely applied (vast expertise available)
Limitations: Accounts for Measurement noise only;
biased estimates in the presence of atmospheric turbulence
Arguments: Flight tests to be performed in steady atmosphere
(hypothetical assumption; rarely feasible due to tight time
schedules and cost factors)
Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/11

Output Error Method (2)


Assumptions
1) Mathematical model: State space
2) Input sequence {u(k), k=1, 2, ... , N} is exogenous;
i.e., generated independent of the system output
==> special treatment for unstable aircraft with FCS
3) Control inputs {u(k)} are sufficiently and adequately (i.e., in magnitude
and frequency) varied to excite various modes of the system
4) System is corrupted by measurement noise only;
i.e., state equations represent a deterministic system
==> special treatment for process noise case (Stochastic system)
5) Measurement errors v(k) at different discrete time points are statistically
independent and Gaussian distributed with zero mean and covariance
matrix R;
i.e., E{v(t k )} = 0 ;
E{v(t k ) v T (t l )} = R kl
Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/12

Likelihood Function (1)


Definition of Likelihood function:
Conditional probability density function of measurements z(k)
for given and R.
For Gaussian distribution (assumption 5), the conditional probability density
p(z(k)| , R) at a particular discrete time point k is given by:

p ( z (t k ) , R ) = (2 ) m R

1 / 2

exp 12 {z (t k ) y (t k )}T R 1 {z (t k ) y (t k )}

Eq. (1)

Mathematically strictly speaking, it would be more appropriate to write the


likelihood function for the given data as p(z(k)| , R, u).
However, the argument u is dropped without loss of generality:
What is the reason?
Ref. for Eq. (1):
Davenport, W. B. and Rot, W. L., Random Signals and Noise, McGraw-Hill NY, 1958.
Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/13

Likelihood Function (2)


The reasons for dropping the argument u:
- Identical control inputs are used to excite the system
as well as the postulated model
- Control inputs are exogenous, noise free and known a-priori

Measurement errors at various discrete points are assumed independent:


=> the likelihood function can be expressed as:
N

p ( z (1), z (2), ... , z ( N ) , R ) = p ( z (k ) , R )

Eq. (2)

k =1

From Eqs. 1 and 2, the likelihood function follows as:

p ( z N , R) = (2 )

N / 2

1 N

exp 2 {z (t k ) y (t k )}T R 1 {z (t k ) y (t k )} Eq. (3)


k =1

m: number of output variables, i. e., the dimension of the vectors z and y.


N: number of data points.
Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/14

Likelihood Function (3)


p ( z | , R ) is the probability of the measurements for a given parameter
vector and the noise covariance matrix R;

The Likelihood function p does not represent the probability distribution


of the unknown parameters; but of the measurements .
The parameters does not have any probability density, since they are not
random.
Maximum-Likelihood estimation means that the -vector is searched
which maximizes the function p ( z | , R )
Such a vector is the most plausible, because it gives the highest
probability to the measurements.
Equivalently, negative logarithm of the likelihood-function is minimized:
J = L( z , R ) =

Dr. Ravindra Jategaonkar

1
2

[ z (t k ) y(t k )]T

k =1

R 1 [ z (t k ) y (t k )] + N2 ln(det( R )) + 2 ln(2 )
Eq. (4)

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Nm

Methods/15

Optimization of Likelihood Function (1)


Cost function:
J = L( z , R ) =

1
2

[ z (t k ) y(t k )]T R 1 [ z (t k ) y(t k )] +

k =1

N
2

ln(det( R )) +

Nm
2

ln(2 )

The last term is a constant:


can be neglected without affecting the optimization.
Two cases are of interest:
Case 1: R Known:

J=

[ z (t k ) y(t k )]T

k =1

R 1 [z (t k ) y (t k )]

Apply any optimization procedure (Newton based methods)


Case 2: R Unknown: Brute-force approach to combine and R into single
vector and apply any standard optimization method:
Never practiced, because there is no closed form
solution to this minimization problem.
During the optimization, the estimates of depend
on the R and vice versa.
use relaxation strategy (two step optimization procedure)
Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/16

Optimization of Likelihood Function (2)


Equating J / R to zero gives maximum likelihood estimate of R:
1
R=
N

[ z (t k ) y (t k )] [ z (t k ) y (t k )]T

Eq. (5)

k =1

For any fixed value of , Eq. (6) maximizes the likelihood function w.r.t. R.

Estimation problem in this case reduces to minimization of J w.r.t. ;


subject to R given by Eq. (5).
Substituting Eq. (5) in Eq. (4) yields:
J1 () = 12 m N +

N
2

ln(det( R )) +

Nm
2

ln(2 )

Minimization of J1() is equivalent to minimization of:


T

1 N
J 2 () = det( R ) = det [ z (t k ) y (t k )] [ z (t k ) y (t k )]

k =1
Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Eq. (6)

Eq. (7)

Methods/17

Optimization of Likelihood Function (3)


Relaxation algorithm:
1) Choose suitable initial values for
2) Estimate R using Eq. (5)
3) Minimize J2(), Eq. (7), with respect to using any
suitable optimization method
4) Iterate on step 2 and check for convergence.
Proof of global convergence of relaxation procedures is difficult,
but they are convenient to use and mostly work well in practice.

Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/18

Optimization Algorithms (1)


Newton Raphson Method
Iterative method of finding a zero of a nonlinear function J() of
several variables.
i.e., zero of the gradient of the cost function J

J = 0

Necessary condition

Taylor series expansion about the k-th value:

(J

)k +1 = (J )k + 2 J 2 k k +1

with

k +1 = ( k +1 k )

2
J k is the second gradient of the cost function w.r.t. ,
Where
(Hessian matrix) at the k-th iteration

Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/19

Optimization Algorithms (2)


The change in on the k+1-th iteration
to make (J )k +1 approximately zero is:

(J

)k +1 = (J )k + 2 J 2 k k +1

= 0 ==>

2
2
The Newton-Raphson algorithm: k = J
k

(J

)k

Tangent line at k

J ()
Slope of tangent line = ( J () )

k+1
Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/20

Optimization Algorithms (3)


Simple Test Cases
1) Quadratic cost function
Let the cost function be: J = 8 2
The first gradient:

(J

) = 16

The second gradient:

2 J 2

= 16

The parameter update:

k = [ 16]1 16 * 4 = 4

The updated parameter:

= 0 + k = 4 4 = 0

Which happens to be the minimum.


Optimum in a single step: -- Quadratic cost function;
second gradient constant.
Minimum in a single step starting from any initial value.
Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

ExQuadFun.m

The starting value for unknown parameter: 0 = 4

Methods/21

Optimization Algorithms (4)


Simple Test Cases
2) Non quadratic cost function
Let the cost function be: J = 4 2 + 0.4 3
The first gradient:

(J

) = 8 + 1.2 2

The second gradient:

2 J 2 = 8 + 2.4

The starting value for unknown parameter: 0 = 4


The parameter update:
1
2
Iteration 1: 1 = [8 + 2.4 * 4] [8 * 4 + 1.2 * 4 ] = 2.9091

1 = 0 + 1 = 4 2.9091 = 1.0909
Iteration 2: 2 = [8 + 2.4 *1.0909]1 [8 *1.0909 + 1.2 *1.0909 2 ] = 0.9564

2 = 1 + 2 = 1.0909 0.9564 = 0.1345

Dr. Ravindra Jategaonkar

ExNonQuadFun.m

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/22

Optimization Algorithms (5)


Recall our ML cost function defined as:

J ( ) =

1
2

T
[
z
(
t
)

y
(
t
)]
W [ z (t k ) y (t k )]
k
k

k =1

(W = R 1 )

Partial differentiation of J w.r.t yields:

J =

1
N

[y

]T W [ z (t k ) y (t k )] = 0

and the partial differentiation of J w.r.t yields:


2

J =
2

1
N

[y

]T W [y ] +

1
N

[ 2 y

2 ]T W [ z y ]

Computation of the first gradient y is relatively straightforward;


computation of of the second gradient 2 y 2 very time consuming.

Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/23

Optimization Algorithms (6)


Modified Newton-Raphson (Gauss-Newton) method:
contribution due to 2 y 2 goes to zero as the process converges.
Residuals z(k)-y(k) summed over sufficiently large N are zero mean.

J =
2

1
N

[y

]T W [y ]

2
2
k = J
k

(J

)k

Quasi-Linearization method:
Recall our cost function defined as:

J ( ) =

1
2

T
[
z
(
t
)

y
(
t
)]
W [ z (t k ) y (t k )]
k
k

k =1

Gradient of J is given by:

J = [y ]T W [ z (t k ) y (t k )] = 0
Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/24

Optimization Algorithms (7)


Quasi-Linearization of y() about 0 yields:

y () y ( 0 ) + (y ) k
Substituting linearized y(Q) in the previous equation leads to:

J = [y ]T W [ z { y (y ) }] = 0
T
T
[

]
W
[
z

y
]
+
[

]
W [y ] = 0

[y

]T W [y ] = [y ]T W [ z y ]

Gauss-Newton and Quasi-linearination methods are equivalent.

Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/25

Optimization Algorithms (8)


Unconstrained Gauss-Newton Method (summary)
Iterative parameter update starting from initial guess
i +1 = i + with = F 1 G
Information Matrix: F =

1 y (tk )

R

2 k =1

2 J

N y (t )T
k

Gradient vector:

N y
J

= (tk ) R 1 [z (tk ) y (tk )]


G=
k =1

Iterative update requires:


computation of observation variables y
computation of response gradients y /
Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/26

Optimization Algorithms (9)


Sensitivity matrix
Partial differentiation of the observation equations y=g(x,u,) w.r.t. yields:

y g x g
=
+
x
Thus, to compute the response gradients, we need gradients of the states.
Partial differentiation of the state equations x& = f ( x , u, ) yields:

x& f x f
+

=
x
Analytical expressions
Or
Numerical approximations

Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/27

Optimization Algorithms (10)


Response gradient approximated by finite difference approximation:
Forward difference:

y (t ) y pi (tk ) yi (tk )
k ij
j

Central difference:

j
j
y (t ) y pi ( x,u, + j e ,tk ) yi ( x,u, j e ,tk )
k ij
2 j

y p ( j ) perturbed response variables

Numerical approximation offers flexibility to handle conveniently


different model structures without software changes

Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/28

Optimization Algorithms (11)


Numerical Aspects:
- Gauss-Newton method results in a full step within parameter space
- J() not a simple function of
- Approximation in computation of Hessian
- Numerical approximation of gradients
- Poor performance for starting values far from optimum
- intermediate local divergence or stalling
Solutions:
- Heuristic approach
on divergence, halving of parameter update
- Line Search ()
i +1 = i + with = F 1 G

Damping strategy widens the convergence region and overcomes


numerical problems
Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/29

Optimization Algorithms (12)


Bounded-Variable Gauss-Newton Method
Parameters often physically constrained to lie in a certain range,
e.g.,
- non negative time delays,
- Oswald factor (increase in drag due to nonelliptical lift
distribution) < 1

Linearly-constrained optimization problem


min J () subject to min max

Classical approaches:
Barrier function,
Langrangian,

Active set strategy

Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/30

Optimization Algorithms (13)


BVGN Method (continued): Active set strategy
Simple, direct and efficient approach
(retains the advantages of GN method)
update of the free parameters via

0.02
1/rad

Unconstrained
Gauss-Newton

1 G
free = F free
free

0.01

Clr

Requires sorting of gradient vector and


information matrix to compute

freeze parameters that hit the bounds

Bounded-variable
Gauss-Newton
0.0

Upper
bound

-0.005

re-activate parameter if the gradient


points no longer outside the feasible
region (Kuhn-Tucker optimality condition)
Gi < 0 for i = i _ max ;
Gi > 0

for

2
4
Number of iterations

i = i _ min

Optimality conditions guarantee that the gradients for the variables hitting the bounds
are such that they point outwards of the feasible region, implying that any further
minimization of the cost function would be possible only when the particular
parameters are not constrained within the specified limits.
Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/31

Optimization Algorithms (14)


Levenberg-Marquardt Method
i +1 = i + with = ( F + I )1 G

Levenberg-Marquardt parameter

controls whether the update direction is more like:


steepest descent ( ) or
Gauss-Newton direction ( 0)
Selection of :
- requires solution for two values of LM-parameter
- LM-overhead minor part, major part spent on
computation of sensitivity matrix
- adaptation of guarantees convergence to the optimum

Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/32

Optimization Algorithms (15)


Direct Search Methods

Simplex method of Nelder and Mead


Subspace searching method of Rowan
Powell's conjugate direction minimum search
Jacob's heuristic minimum search

Simplex method:
- reflection, expansion, contraction and shrinkage of q-dimensional
convex hull (a geometrical figure with q+1 vertices - Simplex)
- Robustness (w.r.t. initial values and discrete nonlinearities)
- Inefficient for large number of parameters
Subplex method:
- Generalization of Simplex method
- Decomposes higher dimensional problem into smaller-dimensional
subspaces in which Simplex method search is more efficient.
Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/33

Integration Methods

Euler integration
Runge-Kutta 2nd order
Runge-Kutta 3rd order
Runge-Kutta 4th order
Runge-Kutta-Fehlberg 4th and 5th order with step size control
- Computational load and accuracy increases with the order
- Step size control is useful for high bandwidth models

Gear's backward differentiation formula for stiff systems


- Special integration method for systems with large and small eigenvalues
- inefficient for non-stiff systems
- Not possible for retarded systems

Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/34

Solution of Linear Algebraic Equation


Calculation of the inverse of the information matrix via
Cholesky factorization
Singular value decomposition
Standard LAPACK subroutines for SVD.
Lower bound for singular values

Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/35

Hybridization of Optimization and Integration Methods


To generate better starting values using more robust (but
slower) direct search methods before applying the GaussNewton method
To overcome numerical problems near optimum due to finitedifference approximation used in the Gauss-Newton method
To reduce computational time by using a simpler integration
method for the first iterations

Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/36

A-priori Information
Cost function modified by additional term that penalizes deviations
from the a-priori values
N

J ( ) = [ z (tk ) y (tk )]T R 1 [z (tk ) y (tk )] + W ( *)T R21 ( *)


k =1

R2
W

a-priori values of the derivatives


error covariance matrix of the a-priori values
weighting factor
W 0: suppression of the a-priori values
W : suppression of the measurements

Modified system of equations for parameter increment


T

y (t k ) 1
y (t k )

y (t k )
1
1
1

R
W
R
R
[
z
(
t
)
y
(
t
)]


W R2 ( *)

2
k
k

k
k

Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/37

Statistical Accuracy of Parameter Estimates


Measure of accuracy
Clues into the effectiveness, or lack thereof, of model parameters
Fischer information matrix provides a good approximation to the parameter
error covariance matrix P:
N y (t )T 1 y (t )
k
k
P
R


k =1

Standard deviations (Cramer-Rao bounds):


i =

pii

Correlation coefficients:
i j =

pij
pi p j

Model fit and parameter accuracy: get back to this later (Model Validation).
Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/38

Implementation Aspects for Large Scale Systems (1)


Gauss-Newton, Bounded-Variable GN or LM methods require solving
the linear algebraic equation once per iteration.
Computational time:
Information matrix
Matrix Inversion
240 sec
0.1 sec
example: 10 states, 19 observations, 3772 data points, 59 parameters
Sensitivity matrix y / for each time point involves (m*q*N) entries.
m=60; q=1000; N=80000
Avoid storing huge matrix:
- simulations for different parameter variations are run in parallel
- sensitivity matrix is determined for the current data point
- summands of the information matrix and gradients are immediately
calculated and summed up.
Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/39

Implementation Aspects for Large Scale Systems (2)


Computation of sensitivity matrix

0
0
0

0
0
0

0
0
Block for
system
parameters

Dr. Ravindra Jategaonkar

0
0
0

0
0

1
2
3
Time segment
Block for initial conditions

0
0
0

0
0

For multiple experiment


evaluation, Sparse matrix
structure

0
0

+ Symmetric square
matrix (compute only
lower triangular part)

1
2
3
Time segment
Block for bias parameters
of observation equation

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/40

Different Approaches to Account for Turbulence (1)


Complications:
1) Presence of noise --> parameters must be estimated by some statistical criterion.
2) Modeling errors --> are deterministic, but are treated simply as noise:
(rigorously not justifiable, but probably the best approach).

3) Process noise --> Stochastic system: State estimator is necessary


Two possible approaches:
1) to measure the wind components, more appropriately said,
derive them from other measured variables such as
true airspeed inertial speed, attitude angles, and flow angles
2) to model (generically or explicitly) the turbulence mathematically
and estimate the corresponding parameters.

Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/41

Different Approaches to Account for Turbulence (2)


The first approach:
A data pre-processing step:
yields wind components along the three body-fixed coordinates,
those can be treated as known inputs and accounted for in the
estimation through minor modifications of the postulated models.
Advantage:
Fairly simple output error method can be applied directly.
Disadvantages:
Requires precise measurements of the said variables.
Any inaccuracies in the measurements, for example those resulting
from calibration errors or time delays in the recorded flow angles,
will affect the derived wind, and consequently the estimates of the
aerodynamic derivatives.

Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/42

Different Approaches to Account for Turbulence (3)


The second approach: Simultaneous modeling of turbulence
Two possible approaches to account for the atmospheric turbulence
in the estimation of parameters:
1) Explicit modeling of gust spectrum (Dryden)
2) Generic model for process noise
Differences in the two options:
- First option based on physical characteristics of gust
- State vector augmentation
- Leads to additional state equations
- Second option is more generic
- Treat turbulence as white noise
- No additional states
Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/43

Approach 1: Explicit modeling of gust spectrum (1)


x
V

Steady atmospheric conditions.


w
z

Consider the longitudinal motion:

u& X u
& Z / U
= u 0
q& 0
&
0
Dr. Ravindra Jategaonkar

X
Z / U 0

M
0

Mq
1

g u Xe
0 Z e / U 0
+
e
0 q Me

0 0

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/44

Approach 1: Explicit modeling of gust spectrum (2)


State vector:

x = [u , , q, ]T

Input vector:

u = [ e ]

Observation vector:

y = [ a x , a z , u , , q, ]

Unknown parameters: X (.) , Z (.) ,L M (.)


To account for the turbulence in the longitudinal and
vertical directions, we have to model the stochastic disturbances.
Assume stationary and homogenous gusts.

Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/45

Approach 1: Explicit modeling of gust spectrum (3)


Exact determination of the Autocorrelation function and power spectral
density of turbulence is not possible.
The PSD, S(), at low frequencies tends to have a constant value:

lim S() = const.

At higher frequencies the PSD, and in turn the energy, falls off:

lim S() ~ m

The different types of spectrum result depending upon the value of m.


Kolmogoroff, von Weizsaecker and Onsager determined a value of:

m = 5/3
Von Karman spectrum:

Dr. Ravindra Jategaonkar

S() = 2w

2 Lw

[1 + (1,339 L

2 5/ 6

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/46

Approach 1: Explicit modeling of gust spectrum (4)


Heisenberg: m = 7
Pritchard:

von Karman

1.2 < m < 2,2

Dryden

Dryden:

m = 2 (Pragmatic value)

S() = 2w

2 Lw
1 + L2w 2

Measured power spectral density*

Both von Karman and Dryden spectrum define PSD of turbulence through
just two parameters;
- the variance
- characteristics length
Advantage of Dryden spectrum:
- Simple structure and simple realization
- Amenable to state space representation through simple filter
(starting from white noise process).
Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/47

Approach 1: Explicit modeling of gust spectrum (5)


In order to incorporate such an explicit model in the parameter
estimation, we have to arrive at a model form compatible with
the state space model of our system equations
Consider a first order Gauss-Markov process of the form:

x& = a x + b
x is an arbitrary variable (either longitudinal or vertical gust component)
(t ) ~ N (0, q x ) is the white noise (zero mean and variance q x ).
For a stationary Gauss-Markov process, the power spectral density
is given by
a = E U 0 / L
b2 qx
1

S x ( ) =

a 2 1 + ( / a ) 2

b = U 0 2E / L

L Characteristic length (meter);


Corner frequency in rad/s;
qx mean square intensity of the gust
Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/48

Approach 1: Explicit modeling of gust spectrum (6)


For longitudinal and vertical gusts, typically,
E = 1 and 2.4 rad/s
they are characterized through the mean square intensities
and scale (characteristic) lengths:

(u g2 , Lu )

( wg2 , Lw )

Combining state models for longitudinal and vertical gusts


(two additional states ug and wg)
with model for the longitudinal motion
leads to an extended model representation of the form:
The unknown parameters are the intensities and scale lengths
of the gusts:

(u g2 , Lu )
Dr. Ravindra Jategaonkar

( wg2 , Lw )
AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/49

Approach 1: Explicit modeling of gust spectrum (7)


Xu
&
u
& Z u / U 0
0
q&
& = 0
0
u& g

w& g 0

X
Z / U 0
M
0

0
1
Mq
1

Xu

0
0
0

Zu / U 0

X e

Z / U

e 0

Me
+
e +
0

U 0
0

0
0
U0
Lu
0
0
0
0
0
2 / Lu
0

U0

X /U0
u
Z / U 02

M /U0
q
0


0
ug
EU 0 w
g
Lw

ug
0

0
wg

2 E / Lw

The inputs ug and wg are the white noise processes,


which are generated using the random number generator
Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/50

Approach 1: Explicit modeling of gust spectrum (8)


Parameter estimation based on such an extended model is by
no means simpler than the generic approach (Approach 2)
discussed next:
- still have to apply the complex filter error method
incorporating a suitable state estimator.
- Model dependent (longitudinal mode)
additional states for lateral-directional motion
- estimation of the scale lengths and gust intensities
posed convergence problems:
inconsistent estimates compared to the expected values
based on physical understanding of the atmosphere.

Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/51

Approach 2: Atmospheric Turbulence as white noise


State equations:

x& (t ) = f [ x(t ), u (t ), ] + F ( ) w(t )


x
u
w
F

x(t 0 ) = x0

state vector
input vector
additive process noise, N(0,I)
process noise distribution matrix

Observation equations:

y (t ) = g[ x(t ), u (t ), x ]

Discrete measurements:

z (t k ) = y (t k ) + v(t k ),

k = 1,..., N

measurement noise, N(0,I)

Unknown parameters:

T = ( xT ;

T ; x0T )

Filter Error method


Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/52

Filter Error Method (1)


Process
noise

Measurement
noise

Input

Dynamic System
Mathematical Model
State Estimator

Measured
response

+
+

z
~
y

Observation Equations
Sensitivities

Parameter update
By optimization of
Likelihood Function

Response
Error

z- ~y

Cost function: Likelihood function

1 N
~
~
T

N / 2

m
1

p( z | , R) = (2 ) det R
exp z(t ) y (t ) R z(t ) y (t )
k
k
k
k

2
k = 1

Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/53

Filter Error Method (2)


Linear system

State estimator

x& = A x + B u + F w

~
x (t k +1 ) = x (t k ) + B u (t k ) + bx

y = C x + Du

~
y (t k ) = C ~
x (t k ) + D u (t k ) + b y

z (t k ) = y (t k ) + v(t k );
k = 1,2, ... , N

x (t k ) = ~
x (t k ) + K [ z (t k ) ~
y (t k )]

where the state transition matrix and its integral is given by:
2
A t
2 ( t )
=e
= I + A t + A
+ ....
2!
t

= e

3
(t ) 2
2 ( t )
d = I t + A
+A
+ ....
2!
3!

Kalman Gain:

K = P C T R 1
Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/54

Filter Error Method (3)


Covariance matrix of the state prediction error P
Riccati Equation (First-order approximation):

AP + PAT

1
PC T R 1CP + FF T = 0
t

Solve Riccati equation by Potters method of Eigenvetor decomposition:


Define the Hamiltonian matrix as:

A
1 C T R 1C
t

T
FF

AT

Compute the eigenvalues and eigenvectors.


Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/55

Filter Error Method (4)


Partition the eigenvectors into four equal size matrices:

X
11

X
21

12

X
22

Where the eigenvectors corresponding to eigenvalues with positive


real parts are in the left partition.
For controllable and observable system it turns out that exactly
one-half of the eigenvalues will have +ve real parts.
Solution is then given by:

Dr. Ravindra Jategaonkar

1
P = X 11 X 21

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/56

Filter Error Method (5)


Optimization of cost function:
Relaxation algorithm
1st step:

1
R=
N

Estimate covariance matrix R

[ z (t k ) ~y (t k )][ z (t k ) ~y (t k )]T

k =1

2nd step:

Apply Gauss-Newton method

i +1 = i + ,

F = G

T
~
~
y (t k )
1 y (t k )
R
F =

k =1
N

T
~

y
(
t
)

1
k
G =
R
[ z (t k ) ~
y (t k )]

k =1
N

Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/57

Filter Error Method (6)


Two-Step procedure: (for both Output error and Filter error methods)
Step 1: Estimate R for specified/updated parameters
Step 2: for known R, estimate parameters
Iterate on step 1 and 2 till convergence
For filter error method: R = G GT + CPC T
For physical meaningful results, the indirectly obtained GGT must be
positive semidefinite.
Results from constraining the diagonal elements of KC < 1.
Optimization subject to nonlinear inequality constraints;
solved by quadratic programming method.

Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/58

Filter Error Method (7)


Gradients
x (t k +1 )
b
x (t k )
~

x (t k ) +
u (t k ) +
B u (t k ) + x +
bx
+
=

x (t k +1 )
b
~
x (t k )
A
B
x (t k ) +
u (t k ) + x ,
+

x (1)
~
=0

b y
~
~
y (t k )
x (t k ) C ~
D
=C
+
x (t k ) +
u (t k ) +

x (t k )
y (t k ) K
x (t k ) ~
~
=
K
+
[ z (t k ) ~
y (t k )]

Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/59

Filter Error Method (8)


Needs gradient of K (Kalman gain matrix):

K = P C T R 1

C T 1
K P T 1
=
C R +P
R

Partial differentiation of the Riccati equation,

1 C T 1
P A
AT P T 1 P T 1
A
P+P
A
C R CP P
R CP
+
+


t
t
1
1
F T F T
T 1 C
T 1 P
PC R
P PC R C
+F
+
F
t

= 0

P
Solve Lyapunov equations to get

Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/60

Filter Error Method (9)


Nonlinear Systems:
State Estimator:
Prediction step

Correction step:

tk
~
x (tk ) = x(tk 1) + f [ x(t ), u(tk ), ] dt
tk 1
~
y(tk ) = g[ x (tk ), u(tk ), ]
x(tk ) = ~
x (tk ) + K[ z(tk ) y(tk )]

Steady state gain Matrix: K = P CT R 1;

g[ x(t ), u (t ),
=
x

=t

Riccati equation for covariance matrix of state prediction error:


AP + PAT 1 PCT R 1CP + FF T = 0;
t

Dr. Ravindra Jategaonkar

f
[
x
(
t
),
u
(
t
),
]

where A =

x
t = t

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/61

Filter Error Method (10)


Response gradient approximated by finite difference approximation:
Forward difference or Central difference
Perturbed response variables y p ( j )
tk
~
x p (tk ) = x p (tk 1) + f [ x p (t ), u(tk ), + ] dt
tk 1
x p (tk ), u(tk ), + ]
y p (tk ) = g[~
x p (tk ) = ~
x p (tk ) + K p[ z(tk ) y p (tk )]
Computation of y p ( j ) by numerical integration of y p ( j ) straightforward.
Perturbed gain matrix: K p = Pp CTp R 1
(solution of Riccati equation with perturbed system matrices Ap and Cp ).

Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/62

Comparative performance of
Output Error and Filter Error methods
Time history plots of measured and estimated responses
Power spectral densities of residuals
comparison of parameter estimates
Example 1:
Estimation of lift, drag and pitching moment
coefficients
Example 2:
Estimation of lateral-directional derivatives

Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/63

Comparative performance (2)


Time histories
Output error method

Dr. Ravindra Jategaonkar

Filter error method

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/64

Comparative performance (3)


PSD of Residuals
Output error method

Dr. Ravindra Jategaonkar

Filter error method

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/65

Comparative performance (4)


Practical Utility
Weathercock stability (yaw stiffness)

Rudder effectiveness

Output Error Method


0.22
Cn

0.20

Output Error Method


-0.18

Flight
No.
209
112
214
219
221
216
222
223

-0.24

Cn

-0.30

0.18

0.16
flight in

Filter Error Method

C-160

-0.36

Filter Error Method


-0.18

0.22
flight
flightinin

Cn

turbulence

-0.24

0.20

Cn

-0.30

0.18

0.16
-3
Dr. Ravindra Jategaonkar

3
9
Angle of Attack, deg

15

-0.36
0.15

0.25

0.45
0.35
Mach Number

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

0.55
Methods/66

Comparative performance (5)


Practical Utility
Pros and Cons of Filter Error method:
- Computationally complex
- Extension to nonlinear systems possible
(Prediction step on nonlinear model; Correction step based on
first order approximation found adequate)
- Difficult for multiple experiment analysis
- Clues regarding modeling errors obliterated
- Capability to analyze data in turbulent atmospheric conditions
- Also, in seemingly steady atmospheric conditions yields better results
- Marked differences in derivatives w.r.t. states
- Comparable estimates of control derivatives
(because they are identified from the deterministic inputs, noise
level in control inputs is low).
- Improved convergence

Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/67

This page is left intentionally blank

Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/68

Identification of Unstable Systems (1)


Difficulties:
1) Open loop plant identification:
the basic / uncontrolled aircraft is unstable
(due to the aerodynamic design)
2) Aircraft states and controls are highly correlated
(due to the design of flight control laws)
3) Aircraft may be excited by process noise
(e.g., induced by forebody vortices)
Pilot
Commands

Dr. Ravindra Jategaonkar

Flight
Control

Surface
Deflections

Unstable
Aircraft

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Motion

Methods/69

Identification of Unstable Systems (2)


Difficulties

Consequences

Solution

Instability

Integration of Eqs. of
motion leads to divergence
(overflow problems)

Filter error method


EKF
Stabilized output error
Equation Decoupling
Regression
Frequency domain methods

Correlated
states & controls

Correlated estimates,
convergence problems

Reduced aerodynamic model


- fix some derivatives
- combined derivatives
- mixed estimation
Avoid correlation
- separate surface excitation
- Optimized input

Process noise

makes estimation more


difficult

Use appropriate method


- Filter error
- Regression
- EKF

Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/70

Identification of Unstable Systems (3)


In the case of highly unstable systems, special techniques and
modifications are necessary to prevent the growth of errors
introduced by poor initial guess values, round off or discretization
errors, and propagated by inherent instabilities of the system
equations. Such approaches are based either on limiting the
integration interval or making more efficient use of observed data.
-Least squares (LS)

- Total least squares (TLS)

- Filter error method (FEM)

- Output error method (OEM)

(LS, TLS, FEM, OEM already covered)


- Equation decoupling
- Eigen value transformation
- Extended Kalman filter / Unscented Kalman filter
- Stabilized output error
Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/71

Identification of Unstable Systems (4)


Artificial Stabilization
Output error method with artificial stabilization

x (t k ) = x(t k ) + S [ z (t k ) y (t k )]
S
x

stabilization matrix
stabilized state vector

0<S<I

S independent of system parameters


Stabilization matrix S corresponds to Kalman Filter gain matrix for the
filter error method

S = 0 =>
S=I

=>

output error method


equation error method

Unstable systems: numerical integration without stabilization diverges

Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/72

Identification of Unstable Systems (5)


Equation Decoupling
Reformulate state equations using measured states in such a way that
each differential equation can be integrated independently.

&x(t ) = A ( ) x(t ) + [B( ); A ( )] u(t )


D
OD
x (t )
m

AD - Diagonal matrix containing diagonal elements of A


AOD - Matrix containing off-diagonal elements of A
Decoupled system, implied by diagonal AD, mostly stable
- hence suitable for unstable systems.

Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/73

Identification of Unstable Systems (6)


Eigen value transformation:
Transformation though a real positive larger than the largest real
part of the most unstable eigenvalue
~
x (t ) = e T t x (t ) ,

~
y (t ) = e T t y (t ) ,

u~(t ) = e T t u(t )

~
z (t ) = e T t z (t ) ,

~(t ) = e T t w(t ) ,
w

v~(t ) = e T t v (t )

Original
Im axis

Transformed
Im axis

Leads to transformed stable system:


~
~ (t )
x& (t ) = ( A T I ) ~
x (t ) + B u~ (t ) + F w
~
y (t ) = C ~
x (t ) + D u~ (t )

~
z (t k ) = ~
y (t k ) + G v~(t k )

Apply classical output error method

Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/74

Identification of Unstable Systems (7)


EKF: Parameter Estimation by State Augmentation
Augmentation of state vector by unknown parameters:
xa = x

& =0
with

Convergence plots of estimates


-0.2

Extended model of the dynamic system:

x&a (t ) = f a[ xa (t ), u(t )] + Fa wa (t )
f [ x(t ), u(t ), ] F
0 w(t )

=
+

0 0
0

0
ya (t ) = g a[ xa (t ), u(t )]

z(tk ) = y(tk ) + G v(tk )

CNBET -0.4
[/rad]
-0.6
0.25

CLP*
[/rad]

0.15
0.05
0.5

CNR*
[/rad]

0
-1
-0.01

CLDA -0.03
[/rad]
-0.05

State estimation of augmented system by


extended Kalman filter.

10
time

15

sec

20

FFT and GGT to be specified a priori ==> Filter tuning.


Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/75

Identification of Unstable Systems (8)


EKF: Parameter Estimation by State Augmentation
Prediction step:

k
~
xa (t ) = xa (tk 1) + f a[ xa (t), u(tk )] dt
k
t k 1

~
y (tk ) = ga[~
xa (tk ), u(tk )]
~
Pa (tk ) = a (tk ) Pa (tk 1)Ta (tk ) + t Fa FaT
where the transition matrix is given by:
and the linearized state matrix by:

a = e Aa t

Aa (tk ) =

fa
xa x = x (t )
a
a k 1

obtained by numerical difference approximation.

Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/76

Identification of Unstable Systems (9)


EKF: Parameter Estimation by State Augmentation
Correction step (Update):

~
~
K a (tk ) = Pa (tk ) CaT (tk )[Ca (tk ) Pa (tk )CaT (tk ) + G GT ]1

xa (tk ) = ~
xa (tk ) + Ka (tk )[z(tk ) ~
y (tk )]
~
Pa (tk ) =[ I Ka (tk )Ca (tk )] Pa (tk )
~
= [ I Ka (tk )Ca (tk )] Pa (tk ) [ I Ka (tk )Ca (tk )]T + Ka (tk ) G GT KaT (tk )
The long form better conditioned for numerical computations and
Ensures that the covariance matrix P is positive definite.
The linearized observation matrix is given by: Ca (tk ) =

Dr. Ravindra Jategaonkar

ga
xa x = ~
a xa (t k 1)

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/77

Identification of Unstable Systems (10)


Example: Simulated data
The static stability parameter, Mw, is so adjusted as to result in an unstable system
with time-to-double of 1 s.
State equations:

System matrix

w& = Z w w + (U 0 + Z q ) q + Z e e
q& = M w w + M q q + M e e
Observation equations:

a z = Z w w + Z q q + Z e e
w=w
q=q

Dr. Ravindra Jategaonkar

U0 of 44.57 m/s
Eigenvalues: (0.6934, -5.8250)
Time to double: roughly 1 s
given by

A feedback proportional to the vertical velocity

e = p + K w

Zw U0 + Zq
A=

M
M
w
q

T2 =

ln 2

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/78

Identification of Unstable Systems (11)


Example: Simulated data
-- Estimated responses for starting parameter values

Output error method


Dr. Ravindra Jategaonkar

Stabilized output error method

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/79

Identification of Unstable Systems (12)


Example: Simulated data -- Estimated responses

Output error method


Dr. Ravindra Jategaonkar

Stabilized output error method

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/80

Identification of Unstable Systems (13)


Example: Simulated data -- Estimated parameters (Summary)

Zw

Zq
Z e
Mw

Mq
M e
Eigenvalues

Nominal
value

LS

TLS

-1.4249

-1.4249

-1.4249

-1.4768

-1.4768

-1.4768

-6.2632

-6.2632

-6.2632

0.2163

0.2148

0.2148

-3.7067

-3.6906

-12.784

-12.7056

0.6934
-5.8250

0.6887
-5.8042

LS/OEM

EqDecpl

EigTransf

FEM

EKF

UKF

OEM

SOEM

-1.4277
(0.01)

-1.4294
(0.01)

-1.42780
(0.01)

-1.4252
(0.0)

-1.4252
(0.0)

-1.3086
(2.48)

-1.4268
(0.01)

-1.47680*

-1.4768*

-1.4768*

-1.4768*

-1.4768*

-1.4768*

-1.4768*

-6.2632

-6.1681
(0.01)

-6.12825
(0.01)

-6.16658
(0.01)

-6.2622
(0.01)

-6.2653
(0.01)

-9.9477
(8.72)

-6.1711
(0.02)

0.21634

0.21616
(0.01)

0.21776
(0.01)

0.21717
(0.01)

0.2175
(0.04)

0.21713
(0.04)

0.05216
(5.85)

0.21709
(0.01)

-3.6907

-3.71215

-3.7086
(0.0)

-3.72645
(0.01)

-3.71967
(0.0)

-3.7373
(0.03)

-3.7343
(0.03)

-1.9277
(4.04)

-3.7201
(0.01)

-12.7059

-12.7687

-12.773
(0.01)

-12.8262
(0.01)

-12.8152
(0.01)

-12.8247
(0.02)

-12.8122
(0.02)

-7.7214
(1.87)

-12.815
(0.01)

0.6900
-5.8262

0.6936
-5.8495

0.6930
-5.8405

0.6913
-5.8538

0.6898
-5.8493

-0.0873
-3.1490

0.6925
-5.8395

-1.4249
-1.47680*

0.6887
-5.8043

0.6919
-5.8290

Program name

uAC_regLS

uAC_regTLS

ml_oem

ml_oem

ml_oem

ml_fem

mainRPE

mainRPE

ml_oem

ml_oem

test case

--

--

10

Function name for


state and observation
equations

--

--

xdot_TC06_
uAC_RegSt
obs_TC06_
uAC_ RegSt

xdot_TC07
_uAC_Eq
Decoup
obs_TC08_
uAC

xdot_TC10_
uAC_EigT
obs_TC08_
uAC

xdot_TC08
_uAC
obs_TC
08_uAC

xdot_TC08
_uAC
obs_TC
08_uAC

xdot_TC08
_uAC
obs_TC
08_uAC

xdot_TC08
_uAC
obs_TC
08_uAC

xdot_TC08
_uAC
obs_TC
08_uAC

Nx; Ny; Nu

0; 1; 3

0; 1; 3

2; 3; 3

2; 3; 3

2; 3; 1

2; 3; 1

2; 3; 1

2; 3; 1

2; 3; 1

2; 3; 1

NparSys

3; 3

3; 3

Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/81

This page is left intentionally blank

Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/82

Recursive Parameter Estimation (1)


Offline methods:
All data points are processed together at a time, yielding parameters
representing average system behavior.
Assume availability of data set over a fixed interval of time.
Implicitly assume system parameters constant over the period of observation.
Recursive methods:
Utilize the data point-by-point as they become available.
They are approximations of the more elaborative non-recursive methods.
By nature, they cater to systems with time-varying parameters.
Computer memory requirements are small, because storage of past data is
not required.
Ability to track time-varying parameters helps indirectly in aerodynamic modeling
leads to accounting for nonlinearities in system model, at least to some extent,
although the postulated model may be linear.
Likewise, it also helps to account for the changes in the flight conditions
Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/83

Recursive Parameter Estimation (2)


Limitations associated with recursive estimation methods:
Convergence of standard RPE methods is slow and may not be adequate for
real-time fault detection or to detect sudden changes in the dynamics.

Convergence can be improved by incorporating a forgetting factor to


discard older data and thereby relying more on recent data.
Leads to increased noise sensitivity. Shorter records are necessary for faster
adaptation, whereas longer records are necessary to distinguish noise.
A compromise between the rate of tracking parameter changes and noise
sensitivity is necessary.
Wrong choice of forgetting factor can result in estimates oscillating around
the true values.

Uninterruptedly running recursive estimation may face numerical


problems: lack of or limited information content pertaining to dynamic motion.
During parts of the flight phases, like steady level flight, the control and motion
variables could be below noise level.

Verification of data collinearity is not a part of RPE methods. Nearly


correlated states and controls affect estimation of stability and control derivatives.

Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/84

Recursive Parameter Estimation (3)


Basically, any iterative off-line estimation method can be transformed
into recursive form.
Two types of recursive methods:
Least Squares Based Recursive Methods
Recursive Least Squares (RLS)
Fourier Transform Regression (FTR)
Filtering approach
nonlinear state estimation by augmenting the basic system states
with system parameters:
Extended Kalman Filter (EKF)
Unscented Kalman filter (UKF)
Extended forgetting factor recursive least squares (EFRLS)
Examples
Short period motion;
Longitudinal motion; Flight test Data (with turbulence)
Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/85

Recursive Parameter Estimation (4)


Recall, from least squares estimation:
yi (tk ) = i1 x1(tk ) + i 2 x2 (tk ) + ... + ir xr (tk ) + ei (tk )

k = 1,... , N

= ( X T X ) 1 X T Y

Recursive least squares (RLS)


( k + 1) = (k ) + K ( k + 1) [ y ( k + 1) x T ( k + 1) ( k )]
K ( k + 1) =

P ( k + 1) =

Dr. Ravindra Jategaonkar

P ( k ) x ( k + 1)

+ x T ( k + 1) P ( k ) x ( k + 1)

{
P ( k ) K ( k + 1) x T ( k + 1) P ( k )}

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/86

Recursive Parameter Estimation (5)


RLS: Recursive Least Squares -- forgetting factor
Cost function : Standard RLS : J N = kN=1 2 (k )
Modified cost function : J N = sN=1 k s

2 ( s)

i.e., samples older than T0 = 1/(1-) with reduced weighting.


1

0.999
0.5

Weighting curves for


Different forgetting
factors

0.99

0.98 0.97
0
0
Dr. Ravindra Jategaonkar

500

1000

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/87

Recursive Parameter Estimation (6)


RLS: Recursive Least Squares forgetting factor
Standard RLS: Infinite memory,
(i.e., entire data is given equal weighting)

RLS with Forgetting factor:


T0: Memory index (Typical choice: 0.98 ~ 0.995)
- Improved convergence
- at the cost of increased sensitivity to noise ( << 1 -> oscillations)
- time varying forgetting factor or Kalman filter

- Thumb rule: balancing between process and measurement noise


- larger process noise variance -> small
- Process noise variance relatively smaller than measurement noise
variance - > large (i.e., use more data for averaging).

Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/88

Recursive Parameter Estimation (7)


FTR: Fourier Transform Regression
Methodology
The fourier transform of a signal x(t) is given by:
j t
jt
~
~
T
i
N

x() = 0 x(t) e

dt ; => x() = ti =1xie

= X() t

Simple Euler approximation to the finite Fourier transform


Apply Fourier transform to state space eq.

j ~
x () = A ~
x () + B u~()
y() = C ~
x () + Du~()
Equation error formulation for the k-th state equation:
~
~
~(n)|2
J = 1 m
|
j

x
(
n
)

A
x
(
n
)

B
u
k
k
k 2 n =1 n k

xk ( n) is the k-th
are the k-th row of matrices A and B, and ~
~
element of vector x for frequency n ; m the frequencies of interest.

Ak and Bk

Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/89

Recursive Parameter Estimation (8)


FTR: Fourier Transform Regression
Denoting vector of unknown parameter as , formulate problem as
LS Regression with complex data:
xk (1) u (1)
xk (1)
j1~

T
T
~
~
~

j 2 xk (2)
xk (2) u (2)

X =
where Y =

.
and
.

~
j m xk ( m)
~T
~
~T

Y = X +

~T

xk (m) u(m)

represents the complex equation error in the frequency domain.


Minimization yields:

] ( )

1
T

= Re(X X ) Re X T Y

On-Line Version

For a given frequency n , the discrete Fourier transform at the i-th time step
is related to that at the preceding step by:

X i ( ) = X i 1 ( ) + xi e jn i t ; e jit = e jt e j (i 1) t
Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/90

Recursive Parameter Estimation (9)


FTR: Fourier Transform Regression
Low computational effort:
at each time point, discrete Fourier transform can be calculated using one
addition and two multiplications.
The m frequencies over which the cost function is evaluated can be selected
as evenly spaced between min and max
Excluding the zero frequency removes trim values and measurement biases.
The most computational effort is the inversion of the matrix
performed by SVD.
- No tuning of any parameter
- Good convergence for low to moderate noise.
- Applicable to linear models
- Aerodynamic zero terms can not be estimated
it is necessary to remove the trim values from the measurements of the motion
variables as well as controls before using them in the algorithm.
Failure to do so leads to erratic behavior of the FTR algorithm, resulting from
disproportionately large component at zero frequency.
Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/91

Recursive Parameter Estimation (10)


EKF: Extended Kalman Filter
Augmentation of state vector by unknown parameters:
xa = x

& =0
with

Convergence plots of estimates


-0.2

Extended model of the dynamic system:

CNBET -0.4
[/rad]
-0.6

x&a (t ) = f a[ xa (t ), u(t )] + Fa wa (t )

f [ x(t ), u(t ), ]
0

F
+

0

0.25

0 w(t )

CLP*
[/rad]

0.05

ya (t ) = g a[ xa (t ), u(t )]

z(tk ) = y(tk ) + G v(tk )

0.15

0.5

CNR*
[/rad]

0
-1
-0.01

CLDA -0.03
[/rad]
-0.05

State estimation of augmented system by


extended Kalman filter.

10
time

15

sec

20

FFT and GGT to be specified a priori ==> Filter tuning.


Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/92

Recursive Parameter Estimation (11)


EKF: Extended Kalman Filter
1. Initialise parameter (state) vector, and
k
Covariance matrices, P , Q and R
k k
k
2. State extrapolation,
=
k +1 k +1 k
P T + Q
3. Covariance extrapolation, P
=
k +1 k +1 k k +1 k

P HT + R
4. Kalman Gain computation, K
= P H T H
k +1 k +1 k +1 k +1 k +1 k +1 k +1

5. State Update,
x
=
+K
H
z
k +1 k +1 k +1 k +1 k +1 k +1
=P
K
H
P
6. Covariance Update, P
k +1 k +1 k +1 k +1 k +1

Tuning of covariance matrices Q and R


Best when noise characteristics are known
Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/93

Recursive Parameter Estimation (12)


Unscented Kalman Filter UKF
The EKF performance is sensitive to nonlinearities in the system model.
Strong system nonlinearities, or in other words the higher-order terms neglected
in the propagation of states and error covariances, and wrong values of noise
statistics may result in biased estimates and in a worst case lead to divergence

Fundamental problem of not accounting for nonlinear transformation the


random variables undergo:
Sigma point filters (SPF) / Unscented Kalman filter (UKF)
- retains the standard Kalman filter form,
- involves no local iterations, and has a better performance

UKF: - Propagates a finite set of (sigma) points through nonlinear dynamics


- Approximates the distribution (mean and covariance) through a
weighted sum and outer (cross) product of the propagated points.
In contrast to the first order approximation used in the EKF for covariance
propagation, in the UKF nonlinear dynamics are used without approximations.

Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/94

Recursive Parameter Estimation (13)


Unscented Kalman Filter UKF
Actual (sampling)

Linearized (EKF)

Sigma-Point

sigma
points

covariance

mean
y=f(x)
= ( )
yfx

true
mean

true
covariance

linear
propagation
~
y y==f (f(x),
x ) P == P T

Yii = ff((X)
=

true
mean,
covariance

EKF
EKF
covariance mean

)
S-P
mean

S-P
covariance transformed
sigma points

Algorithmic steps: UKF Much more complex than EKF (See Ref. 1)
Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/95

Recursive Parameter Estimation (14)


EFRLS: Extended Forgetting factor Recursive Least Squares
An alternative approach based on an extension of the RLS with forgetting
factor incorporating dynamic system matrix.
x ( k + 1) = ( k ) x ( k ) + w( k )

y(k ) = C (k ) x(k )

z(k ) = y (k ) + v(k )

Algorithm
x ( k + 1) = ( k ) x ( k ) + ( k ) L( k + 1) [ z ( k + 1) C ( k + 1) ( k ) x ( k )]

L( k + 1) = P ( k ) T ( k ) C T ( k + 1) [I +
C ( k + 1) ( k ) P ( k ) T ( k ) C T ( k + 1)] 1
P ( k + 1) =

( k ) [ I L( k + 1) C ( k + 1) ( k )] P ( k ) T ( k )

Knowledge of noise covariances not required, but needs tuning of forgetting factor
Applicable to linear systems, with only states as observations

Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/96

Recursive Parameter Estimation (15)


Example 1: Short period dynamics:

& = Zo + Z + Zq' q + Z e
e
q& = M o + M + M q q + M e

Zq' = (1+ Zq )

Mq:totalpitchdamping;
X = [ q Zo Z ... LM o M ...]T
u = [ e]
y= [ q ]T

Flighttestdata:elevatormultistep inputs

Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/97

Recursive Parameter Estimation (16)


Example 1: Short period dynamics:
Parameter

Reference
values
(OEM)

Estimates
EFRLS

FTR

--

EKF

UKF

UKFaug

-0.0075
(15.2)

-0.0077
(14.6)

-0.0090
(2.78)

Z0

-0.009
(7.25)*

-0.0075

-0.483
(1.36)

-0.500

-0.477
(3.74)

-0.496
(2.28)

-0.496
(2.24)

-0.484
(0.50)

Zq

0.104
(5.28)

0.109

0.105
(1.29)

0.102
(9.61)

0.103
(9.34)

0.108
(1.88)

Ze

0.676
(1.76)

0.669

0.665
(5.47)

0.648
(3.57)

0.652
(3.49)

0.679
(0.67)

M0

0.475
(0.3)

0.465

0.467
(0.28)

0.468
(0.26)

0.474
(0.10)

-4.923
(0.33)

-4.846

-4.929
(1.52)

-4.871
(0.26)

-4.873
(0.25)

-4.914
(0.11)

Mq

-2.006
(0.57)

-1.972

-2.031
(2.96)

-1.965
(0.57)

-1.967
(0.54)

-2.012
(0.20)

Me

-7.208
(0.35)

-7.129

-7.254
(2.11)

-7.072
(0.37)

-7.074
(0.35)

-7.217
(0.12)

Dr. Ravindra Jategaonkar

--

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/98

Recursive Parameter Estimation (17)


Example 1: Short period dynamics:

Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/99

Recursive Parameter Estimation (18)


Example 2: Non-dimensional lift, drag, pitching moment derivatives:
State equations

F
qS
C D + g sin( ) + e cos( + T )
V& =
m
m
F
g
qS
& =
C L + q + cos( ) e sin( + T )
V
mV
mV
& = q
F
q Sc
C m + e ( l tx sin T + l tz cos T )
q& =
Iy
Iy

Observation variables:
Aerodynamic model:

V , , , q, q& , a x , a z
C D = C D 0 + C DV
C L = C L0 + C LV

V
+ C D
V0

V
+ C L
V0

Cm = Cm0 + CmV

V
qc
+ Cm + Cmq
+ Cme e
V0
2Vo

Unknown paramaters: = [C D 0 C DV C D C L 0 C LV C L C m 0 C mV C m C mq C me ]T
Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/100

Recursive Parameter Estimation (19)


Example 2: Non-dimensional lift, drag, pitching moment derivatives:
Parameter

Filter error method


(FEM)

RPE methods
EKF

UKF

UKFaug

C L0

-0.0929
(21.1)

-0.0853
(23.5)

-0.087
(22.9)

-0.099
(20.0)

C LV

0.149
(11.1)

0.144
(11.7)

0.147
(11.4)

0.157
(10.6)

C L

4.328
(1.08)

4.303
(1.14)

4.289
(1.14)

4.303
(1.13)

C m0

0.112
(3.27)

0.112
(4.28)

0.112
(4.29)

0.115
(3.40)

C mV

0.0039
(82.1)

0.0046
(90.5)

0.0045
(92.2)

0.0022
(154)

C m

-0.968
(1.12)

-0.971
(1.54)

-0.970
(1.54)

-0.983
(1.24)

C mq

-34.710
(2.27)

-34.937
(2.85)

-35.363
(2.82)

-35.098
(2.27)

Cm e

-1.529
(1.27)

-1.533
(1.65)

-1.539
(1.65)

-1.552
(1.31)

Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/101

Recursive Parameter Estimation (20)


Example 2: Non-dimensional lift, drag, pitching moment derivatives:

Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/102

Estimation of Time Delays (1)


Typical examples

Sensor calibration model


Scale factor and bias

p
= K p
+ p
d m
dyn nb
d

1: Recorded data
2. Multi-point aerodynamic model
- Downwash lag /
transit time effect

Time delay

pd m( t ) = K pdyn ( t q )nb( t ) + pd

Transit time
C mDLC = C L ( rH / l ) M DLC ( t )
where
CL

rh
DLC(t-t):
M

Dr. Ravindra Jategaonkar

stabilizer lift coefficient,


stabilizer angle
distance from wing NP to tail NP
time delayed DLC deflection
unknown downwash parameter

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/103

Estimation of Time Delays (2)


Time delays affect the estimates
Example: Estimation of pitching moment derivatives
0

Cmw

Cost
function

-2.5

-0.6

Dr. Ravindra Jategaonkar

sec
0
Time delay in e

0.4

-0.6

sec
0
Time delay in e

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

0.4

Methods/104

Estimation of Time Delays (3)


Different Approaches: Advantages and Limitations

Data pre-processing:
- Time shifting through a specified fixed value
- Simple, commonly used procedure
- Limited to measured parameters
- Only part to read flight data needs to be modified
Approximation by a first order lag
- Simple procedure, requiring no modifications to estimation program
- Additional first order differential eq. for each variable to be time shifted
- Some dynamic effects due to approximation, albeit small.
Delay matrix
- Accurate representation of time shift
- Possible for both positive and negative values
- Flexible: Can be applied to any parameter (directly measured as well as
those computed in the model)
- Requires estimation program for Nonlinear systems and a special
subroutine for delay matrix.
Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/105

Neural Networks for Aerodynamic Modeling


Neural Networks - Pros and Cons
Nonlinear Input-Output subspace mapping

ATTAS Lateral-Directional Motion


Training Phase
0.15

Sideforcecoefficient
0

x c Rn

y c Rm

-0.15
0.05

Yawingmomentcoefficient

Network properties:
- Number of hidden layers and nodes
- Weights estimated by Back-Propagation

-0.05
0

Black-box approach:
- No physical significance to the structure
or to the weights

Application spectrum:

20

time

40

60

Predictive Capability
0.15

Sideforcecoefficient
0.0

- Sensor failure detection


- Highly NL phenomenon: stall, separated flows
- Unstable systems

-0.1
0.025

Yawingmomentcoeff.
0

Less attractive for classical SysId applications


Incremental update or fine tuning of
sub models difficult
Dr. Ravindra Jategaonkar

-0.025
0

10

20

30

time

Flightdata

40

Neuralnetworkoutput

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/106

References (1)
Jategaonkar, R. V.,
Flight Vehicle System Identification: A Time Domain Methodology,
Volume 216, AIAA Progress in Astronautics and Aeronautics Series
Published by AIAA Reston, VA, Aug. 2006, ISBN: 1-56347-836-6
http://www.aiaa.org/content.cfm?pageid=360&id=1447

Chapman, G. T. and Yates, L. A., Nonlinear Aerodynamic Parameter Estimation and Model Structure
Identification, AIAA Paper 92-4502, Aug. 1992.
Chen, W., and Valasek, J., Observer/Kalman Filter Identification for On-Line System Identification of
Aircraft, AIAA Paper 99-4173, Aug. 1999.
Hamel, P. G. and Jategaonkar, R. V., Evolution of Flight Vehicle System Identification, Journal of Aircraft,
Vol. 33, No. 1, Jan.-Feb. 1996, pp. 9-28.
Jategaonkar, R. V. and Plaetschke, E., Algorithms for Aircraft Parameter Estimation Accounting for Process
and Measurement Noise, Journal of Aircraft, Vol. 26, No. 4, 1989, pp. 360-372.
Jategaonkar, R. V. and Thielecke, F., Evaluation of Parameter Estimation Methods for Unstable Aircraft,
Journal of Aircraft, Vol. 31, No. 2, May-June 1994, pp. 510-519.
Jategaonkar, R. V., Bounded-Variable Gauss-Newton Algorithm for Aircraft Parameter Estimation,
Journal of Aircraft, Vol. 37, No. 4, July-Aug. 2000, pp.742-744.
Julier, S., Uhlmann, J., and Durrant-Whyte, H. F., A New Method for the Nonlinear Transformation of Means
and Covariances in Filters and Estimators, IEEE Transactions on Automatic Control, Vol. 45, No. 3, 2000,
pp. 477-482.
Klein, V., Estimation of Aircraft Aerodynamic Parameters from Flight Data, Progress in Aerospace Sciences,
Vol. 26, Pergamon, Oxford, UK, 1989, pp. 1-77.
Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/107

References (2)
Maine, R. E. and Iliff, K. W., Identification of Dynamic Systems - Applications to Aircraft. Part 1: The Output
Error Approach, AGARD AG-300, Vol. 3, Pt. 1, Dec. 1986.
Morelli, E. A., Real-Time Parameter Estimation in the Frequency Domain, AIAA Paper 99-4043, Aug. 1999.
Murray-Smith, D. J., Methods for the External Validation of Continuous System Simulation Models:
A Review, Mathematical and Computer Modelling of Dynamical Systems, Vol. 4, No. 1, 1998, pp. 5-31.
Plaetschke, E., Ein FORTRAN-Program zur Maximum-Likelihood-Parameterschtzung in nichtlinearen
Systemen der Flugmechanik Benutzeranleitung, DFVLR-Mitt. 86-08, Feb. 1986.
Seanor, B., Song, Y., Napolitano, M. R., and Campa, G., Comparison of On-Line and Off-Line Parameter
Estimation Techniques Using the NASA F/A-18 HARV Flight Data, AIAA paper 2001-4261, Aug. 2001.
Wan., E. A., and van der Merwe, R., The Unscented Kalman Filter for Nonlinear Estimation, Proceedings of
the IEEE Symposium 2000 on Adaptive Systems for Signal Processing, Communication and Control ASSPCC, Lake Louise, Alberta, Canada, 1-4 Oct. 2000, pp. 153-158.

Dr. Ravindra Jategaonkar

AIAA Short Course: Flight Vehicle System Identification in Time Domain, Aug. 2006

Methods/108

Anda mungkin juga menyukai