05/10/17

Time-domain output-error identification

Marco Lovera
Dipartimento di Scienze e Tecnologie Aerospaziali, Politecnico di Milano
The output-error method

Model class: } only uncertainty in measurement

there Is me
randomness in the dynamics
µ

can be non .
linear function

Assumptions: Xcok × . initial state

| mat of the tune in not Known no It Is also estimated

• y is a scalar measurement

• u(t) piece-wise constant with period Ts

• or equivalently true system belongs to the net

I to keep things mriple )

Marco Lovera
The output-error method

Measurement model:

• Measurements are discrete easier

really r
not necessary
but makes things

• Sampling is uniform and defined by

• Measurement equation:
From
deterministic value Is affected
( only
the expected value by error )

=
-

true value measurements

noise
of attnut

•
mean variance

• at different time
ewicorrelodeol
instant

one
random number Gowmoin
Joan be a function of K

Marco Lovera
The output-error method

• Under the previous assumptions the samples of the

measured output are such that by it Is htro mean assumption
q

A
Y ym

:
Ym
,

as y(k) is a deterministic sequence. immune

.

• In terms of variance we have

• Therefore

this must
xxx
be proven

Marco Lovera
The output-error method

• We have to check independence of the measurements,

which, under Gaussianity assumptions, reduces to
checking incorrelation:

• Expanding the product:

• And recalling that samples of y are deterministic:

htho mean there mean

/ ¥
www.t.mudemt from Vol
It

• Therefore =o?
}
eevysjiaf
H
they are uncorrelated

Marco Lovera
The output-error method

• Block diagram:

• Hence the name output-error for this set-up:

• Only measurement noise is considered
• No disturbances acting on the plant are included in the
model.
• In other words, the map from u to y is deterministic.

Marco Lovera
The output-error method

The joint density of the data can then be written as

so the logarithm of the joint density is

depends on the parameters
.

Marco Lovera
The output-error method

• Therefore the log-likelihood can be obtained by plugging

the measurements in place of the running variables, to get

• Note now that maximizing the log-likelihood is equivalent

to the minimisation of

Kent linear
if it in put of non system

compute min
it is not possible to
need to do it
in closed form ,
we

numerically
Marco Lovera
The output-error method

• Note that defining

we have that

• Therefore the cost function is equal to the sum of the

squares of the deviations between the measured outputs
and the model outputs.

• This is a particular case of ML estimation known as Least

Squares (LS) estimation.
Marco Lovera
The output-error method

• Note further that under the assumption

we have that

& t
measurement between true and model
error
discrepancy system

• Therefore if then
and the cost converges to

• The optimal cost is zero only in the noise-free case.

Marco Lovera
An optimization scheme for the OE method

initially
Consider a starting value for the parameter and a
perturbation

Taking a second order approximation imposing stationarity

we get

vector to
matrix & to
vector

Marco Lovera
An optimization scheme for the OE method

Solving for the increment in the parameter we get

? ¥olo=o
.

Therefore if the cost is truly quadratic then starting from the

initial guess we find the minimum in one iteration.

Marco Lovera
An optimization scheme for the OE method

If the cost is not quadratic we can use this result to set up an

iterative optimisation scheme:

Iteration is repeated until convergence of the cost function

and/or convergence of the parameter

is reached.
Marco Lovera
An optimization scheme for the OE method

This simple iterative scheme is known as the Newton-

Raphson method.

Other possible convergence criteria include:

• Relative rather than absolute changes in cost and/or

parameters.

• Gradient of the cost sufficiently close to zero.

Marco Lovera
An optimization scheme for the OE method

How do we compute the gradient and the hessian of the

cost?

Recall that

and therefore
problem to
compute

0
Marco Lovera
An optimization scheme for the OE method

This is a vector with components given by

As for the second derivative, element-wise we get ( element by element )

. to

free if we have already

must be computed
compute the gradient

Marco Lovera
An optimization scheme for the OE method

For the sake of simplicity in the expression of the second

derivative

the second term is often neglected (this avoids the need to

compute the second derivative of the model output), so that

Note that the approximate hessian is still symmetric.

The resulting approximate algorithm is called Gauss-Newton.
calculating the derivative of the output instead of cost function
we have only shift the problem in

Marco Lovera
An optimization scheme for the OE method

To complete the definition of the method we need a scheme

to compute the sensitivities of the model output with respect
to the parameters.

This can be done either numerically or analytically.

The numerical approach is unavoidable whenever nonlinear

models are considered.

Sensitivities can be computed

• Using forward differences

• Using central differences.
Marco Lovera
An optimization scheme for the OE method

£9
'

Using forward differences we get =µ0§g|← jth component

( r ÷ 5 %)

the perturbation should be small – general guideline: 1% of

the current value of the parameter component.

Clearly the computation of the vector of sensitivities requires

} It If mmilaton of the model is net long
⇒ is
useful
otherwise It Keeps too long time → available for not
simulations of the response of the model to the sampled to much complex
model
input.

Marco Lovera
An optimization scheme for the OE method

Using central differences instead we get

In this case the computation of the vector of sensitivities

requires

simulations of the response of the model to the sampled

input, but the computed sensitivities are significantly more
accurate.

Marco Lovera
An optimization scheme for the OE method

linear model
to more useful for
The analytical approach, on the other hand, starts from the
model equations:

Differentiating with respect to a component of the parameter

vector we get for the state equation we can modify this

ni
no condition if we want

include I. C.
to
.
parameters
into the
estimated
to be

is function
and note that
.

given
-

that doesn't depend en parameter

Marco Lovera
An optimization scheme for the OE method

Interchanging derivatives on the left had side we get:

and similarly for the output equation

Therefore, it is possible to compute the required sensitivities

by simulating the state space models defined by the above
state and output equations.

Marco Lovera
Local and global minima

• The developed optimisation scheme is local, in the sense

that at each iteration only point-wise information on the
derivatives of the cost are used.

• This means that in general the algorithm may converge to

a different solution depending on the initial guess for the

an
parameters.
←hfi•V
o¥'i→V←
0*0
' =/ GLOBAL MINIMUM LOCAL MINIMA
point
'

! ,
,

i minimum of local all minima stationary

i
minima
.

.
Oo

• This is a key issue with the OE method:

• a reliable initial guess for the parameters is necessary
• careful inspection of the computed estimates is also
necessary, to ensure they are physically meaningful.

Marco Lovera
 10/10/17

Asymptotic variance of the OE estimates

ML estimators are efficient, so we expect that

where
Crowner Rae bound

How can we evaluate M?

We reason as follows:

• The estimate has been chosen so as to maximise the likelihood

of the data;
• Therefore maximal probability (corresponding to the expected
value) should be attained at the optimal likelihood. no we can calculate M as

to
function of logl
for Gaumoin

Marco Lovera
Asymptotic variance of the OE estimates

As a consequence we can make the following

approximation: I can compute this because I have

, data and theta

Note that for the OE problem

.
MY
doll approximation
|g=g*

this quantity
•
in
given for free from our IAMTWE SCHEME ALGORITHM

•
Jloie Joy +
Yoel a. so +sotffIy*so
do

conditioning !!

:
Marco Lovera M
Asymptotic variance of the OE estimates

Therefore we can evaluate M as

Finally, note that the noise variance is also needed.

If it is not known, it can be estimated using the sample

variance, as in the preliminary examples on ML estimation:

Marco Lovera
Confidence intervals for the estimates

• The theory of ML estimation ensures that, asymptotically,

estimates are unbiased and achieve the C-R variance
bound.

• Therefore, asymptotically
*

• As a consequence, having obtained an estimate from a

given dataset, we can define confidence intervals using
properties of Gaussian densities.

Marco Lovera
Confidence intervals for the estimates

• More precisely, letting

• We have, element-wise, that

*

• And in terms of probabilities:

m *

m *

an *

Marco Lovera
Multiple output case

• So far only the case of a scalar measurement has been

considered, for the sake of simplicity.

• In real problems, however, vectors of measurements must

be used, so the output equation is

• And the measurement model becomes

k
R matrix
pxp

Marco Lovera
Multiple output case

• Depending on the specific problem, the noise variance R can

range from
• A diagonal matrix, in the case of uncorrelated
niftier
.it
measurements (individual components of the output - Y1

provided by p different sensors) ↳

.- Yz

Yp

• A block-diagonal matrix, in the case of partially correlated

different sensors)
.int#FEHt*E.H
measurements (subvectors of the output provided by
→ a .
,wI
→
3Yn
Yz ( ) Q accelerometer

vector of seampeneuts
→
yq

• Full matrix, in the case of fully correlated measurements.

• The first two cases are the most common in practice.

Marco Lovera
Multiple output case

• Therefore, the density of the measurement noise is given

by

and following the same derivation as in the scalar output

case, the density of the measurements results

Marco Lovera
Multiple output case

• The likelihood is constructed as before, to get

and the cost function to be minimised becomes

Marco Lovera
Multiple output case

• Note that if R is diagonal the cost reduces to the sum of p

costs for each component of the output:

where

↳ Rii

• This highlights the importance of proper scaling of the

measurements in the formulation of the problem. ↳

J ÷i
÷1 e. ÷r → .

yn
→

-3
÷ to
Yz →
- ' 2

ez÷
-6
-6 → is
J3÷ro
÷ eo
y , he you

J=Jn+@
-
don't
we want to minimise care

he
of J3
N .B .
variance
Marco Lovera lmegbgiblejn )
becomes

532=0032
\ YJ =
CYs ⇐
10+6
{ t.tt#geuenPampaterghnY.oEa
no not .
this may lotnnfnskityd
llowlta
loss of
Multiple output case

• Finally, the gradient and the hessian of the cost ans the
Fischer information matrix can be generalised as

Marco Lovera
 MATLAB FILE : oe .
example .m

a= . a b= -2
S
ftp.axotxbm

S e Mlo )
,
O=[ F)
Ts =
goes nowupluig interval

v( K ) ~G( on ) measurement noise in discrete time → N .B a- 2=1 -3 < vlk ) E 3

⇒ o=n
.

with 99,7%
µH= senlt ) + sentst ) + senlst )

to the words

? We have to estimate a ,b

^)
Is a well posed problem ? -
structural uibntifrability Is
guaranteed because of its features
otherwise Gcsk
¥ →
identifiable

-
look at the input also

( use 2000-3000-4000
for deval samples nomore
)

TIME DOMAIN OUTPUT ERROR IDENT .

For LTI
SYSTEMS
bot tune domain
•
we can use
frequency -
domain and -

the model Is LINEAR I AH )× Ban linear

{
=
+
• now Alo )= oarsman

y=ao)x+Dlo)n force

• we obtain the name eat function

but now
you can
proceed analytically

Fade
DI
=H9s¥oj +
°ff0÷x + 3Bo÷Pn because Tfg=o ( u applied externally Admit )
(¥¥=° depend O on

mafn
guitars go that onbe amputee *
A
.# :] →
FF 't : :]
'

¥k¥t= noises 'x

YET Yolyn :#
nestle
+ +

H÷=aaEgts¥x+so÷u
Now these model we need to solve to compute wuntivitis Is linear

mmulaton

mutinies
nommalx
Mo + L
in
for LH

£ to get
tout

interval than the

• sampling → SHANNON -

NYRUST TIME Is 5÷eo tunes shorter

fattest dynamics of interest

Ts arbitrarily small is not a

good idea -
# data points to
-

tegainintouusef information onthenaam .

Introduction on PREFILTERWG