You are on page 1of 8

Aut omat i ca. Vol. 16, pp.

519-526
Pergamon Press Ltd. 1980. Prinled in Great Britain
International Federation of Automatic Control
000~ 1098/80/0901-0519 $02.00,0
Ms. 195
Step Response and Frequency
Response Methods
H. RAKE,
Key words--l dent i fi cat i on; Fouri er analysis; frequency response; step response.
Al mt raet --In this t ut ori al very simple met hods will be pre-
sented by which models for dynami c processes can be ob-
tained. These met hods employ step responses of t he process
or responses t o ot her nonperi odi c testsignals or process
responses to periodic signals.
I NTRODUCTI ON
ONLY methods giving linear models will be trea-
ted in spite of the fact that most processes of
interest will not be linear. Thus a linear model is
fitted somehow to the process.
The discussion' will be restricted further to
single input-single output (SISO) models. Many
processes with more than one input and/or more
than one output may be represented by a number
of SISO-models connecting each input to each
output. These models can often be determined by
investigating all signal paths separately from each
other.
Figure 1 shows the general setup of process
and model and will be referred to when discuss-
ing different methods. The process is exited by
the input (test) signal u provided by a signal
generator. The measurable process response y
will in general consist of the uncorrupted process
output plus additional process and/or measure-
ment noise n. Thus all noise is thought to be
condensed into this noise added to the process
output. It will be shown that different identifi-
cation .methods are more or less sensitive to
noise. It will also become apparent that in almost
any identification a large signal to noise ratio is
desirable. Thus in the design of experiments and
*Received 19 March, 1980. The original version of this
paper was presented at t he 5th I FAC Symposium on
Identification and System Paramet er Est i mat i on which was
held in Darmst adt , Federal Republic of Ger many duri ng
September 1979.
The published Proceedings of this IFAC Meeting may be
ordered from: Pergamon Press Limited, Headi ngt on Hill
Hall, Oxford OX3 0BW, England.
Thi s paper was recommended for publ i cat i on in revised
form by the Aut omat i c Editorial Board.
f l nst , f. Regelungstechnik, RWTH Aachen, Aachen,
Germany.
519
measurements all means should be employed to
diminish noise and to enlarge usable signals.
Test-signal amplitudes in general are subject to
restrictions. With regard to safety, product qua-
lity etc. large deviations from setpoints cannot be
tolerated in most real processes. In addition to
this, nonlinearities in many processes prevent the
use of large signals for determining linear models.
The identification methods to be presented will
be classified according to the test signals em-
ployed. This leads to two groups of methods one
of which uses nonperiodic test signals whereas
the other works with periodic signals. Both
groups will be subdivided into methods giving
nonparametric models and those giving para-
metric models respectively.
1 o
I
I
I
I
!
MOdel
Yu
Y
r -
I
I
I
I
FIG. 1. General setup of process and model.
Description of processes and models
As implied by Fig. 1 process models will be
chosen and rated on the basis of input and
output signals. Thus only the transfer characteris-
tics of process and model have to be described.
Models are graded according to their ability to
generate an output signal YM which is very
similar to the signal y generated by the process
when model and process are exited by the same
input signal u.
Models very often are described by their
weighting function h(t) or their frequency re-
sponse function H(je)). When no analytic ex-
pression for these functions is known a nonpara-
520 H. RAKE
metric model is obt ai ned, which is represent ed by
a pl ot or a number of sampl e values. Best known
member of this set of model s pr obabl y is t he
frequency response represent ed by a Bode plot.
Par amet r i c models are descri bed by the number
and the values of par amet er s of differential equa-
tions, frequency response funct i ons or transfer
functions. Fr om these anal yt i c represent at i ons
plots or values of interest of frequency response
or weighting funct i ons can be comput ed in ge-
neral wi t hout much difficulty whereas the inverse
process namel y deriving par amet er s from non-
par amet r i c model descri pt i ons is much mor e
difficult.
A fairly general par amet r i c model can be de-
scribed by t he differential equat i on
a, y~t ) ( t ) + . . . + al)Y'M(t) + a o y M( t ) = b o u ( t - r a)
+ b l t i ( t - - re) +. . . + b , , u ( m ) ( t - r a) (1)
or by the cor r espondi ng frequency response
bo + b l j o + . . . + b , , ( j o ) "
H ( j e ) ) =
ao + a l j og + . . . + a , ( j o ) n
exp(-jOgra).
12)
Solving t he differential equat i on for t he i nput
signal
for, 0
u ( t ) = 6 ( t ) for t = 0
~ b ( t ) d t = l (3)
- : r j
with 6(t) being the unit impulse funct i on gives
t he so-called weighting funct i on
account namel y periodic and nonper i odi c ones.
In case of a nonper i odi c i nput (test) signal the
process being in equi l i bri um operat i on, corres-
pondi ng to zero initial condi t i ons, is di st urbed.
The process may t hen settle into a new equilib-
ri um which can be equal to the former one or
st art drifting out of the region of normal oper-
ation. Whi ch way the process will behave de-
pends on t he propert i es of the process and the
i nput signal.
Peri odi c signals in .general are empl oyed in
such a way t hat the process is r unni ng in a
st eady st at e with the out put oscillating with the
same frequency (or frequencies) as the i nput and
all transients having subsided.
In cont r ast to t he det ermi ni st i c i nput signal
well known in advance the noise n (Fig. 1) has to
be t reat ed as a st ochast i c signal about which onl y
its statistical propert i es if any at all might be
available. In many cases of pract i cal interest the
noise can be appr oxi mat ed by st ochast i c signals
with Gaussi an ampl i t ude di st ri but i on and a
spectral densi t y which is either i ndependent of
frequency (white noise) or a not t oo compl i cat ed
funct i on of frequency (col oured noise, e.g. low
pass-filtered noise). In most cases t he noise will
be consi dered as havi ng zero mean
E [ n ] = 0 ( 6 )
and its i nt ensi t y will be descri bed by its vari ance
a # = E[ n 2] ( 7 )
with a as so called st andar d devi at i on or r oot
mean square (r.m.s.) value.
y u ( t ) = h ( t ) (4)
as cor r espondi ng out put function. Thi s weighting
funct i on if represent ed by a pl ot or a number of
sampl e values can be regarded as a nonpar a-
metric model descri pt i on and be used in the
convol ut i on integral
y M ( t ) = S h(z).u(t-r)dz ( 5 )
0
METHODS FOR NONPERIODIC TEST SIGNALS
N o n p e r i o d t e s t s i g n a l s
Nonper i odi c test signals in common use are:
impulses of different shape and relatively short
dur at i on; step funct i ons; and r amp functions.
Tabl e I shows a col l ect i on of nonper i odi c
test signals.
TABLE l. NONPERIODIC TEST SIGNALS
to comput e t he out put YM(t) for ar bi t r ar y inputs
u ( t ) .
Of course t he differential equat i on can be
solved by Lapla'ce- or Four i er - Tr ansf or m techn-
niques empl oyi ng t ransfer funct i ons or frequency
response functions.
S i g n a l s
It has been ment i oned before t hat t wo different
types of i nput signals u (Fig. 1) will be t aken into
( a )
t m
( d)
t - - - ~
( b )
I
A
( e )
(c)
( f )
/
Step response and frequency response met hods 521
When choosi ng a test signal for a specific
application the properties of the signal, the possi-
bilities for generat i ng and appl yi ng the signal and
the ki nd of i nformat i on want ed from the test
have among others t o be t aken into account .
Thus universal rules are har d t o formul at e.
Nonparametric models f rom nonperiodic signals
As stated before nonpar amet r i c process models
of common interest are: the weighting funct i on
h(t); and the frequency response funct i on H(jog).
An obvi ous and fairly simple way for obt ai ni ng
the weighting funct i on h(t) woul d be t o use
u ( t ) = K . 6(t) (8)
as an i nput signal, giving an out put
y ( t ) =K " h(t) + n(t) (9)
ponent s in the out put signal y(t ) will be treated
in the same way as genuine process out put s are.
Four i er t ransforms of signals measured are
usually comput ed by digital comput ers via the
Discrete Four i er Transform (DFT), which can be
realized in a comput at i onal l y efficient manner by
al gori t hms using the so called Fast Fouri er trans-
form (FFT). Wi t hout going into t oo much detail,
it seems wort hwhi l e to bear in mi nd t hat the
Fouri er t ransform F(eo) of a function of time f i t )
is
F(co) = f ( t ) . e x p ( - f i o t ) d t (12)
and t hat it exists onl y if
I f ( t ) l dt < ~ . (13)
- - 05
which in case of negligible noise n(t) will give
h(t) wi t hout furt her effort. In practice the Di rac
impulse of equat i on (8) has t o be appr oxi mat ed
by ~mpulse functions wi t h short but finite du-
rat i on and limited ampl i t ude (Table l a, b). If
these impulse funct i ons are of a dur at i on t hat is
short in compari son to the dur at i on of the
weighting function, syst emat i c errors resulting
from use of these signals instead of a Dirac
impulse mi ght be tolerable. The (approxi mat ed)
weighting funct i on t hen can be comput ed as
y(t )
~( t ) =
u( t ) dt
0
(10)
In cases where impulse functions of non-
negligible dur at i on have to be used, inversion of
the convol ut i on integral equat i on (5) leads to a
(numerical) deconvol ut i on met hod for the de-
t er mi nat i on of the weighting function. Though
feasible in principle this met hod is not used very
often due t o the facts (among others) t hat :
weighting funct i on model s are not very common;
deconvol ut i on may suffer significantly from ill
condi t i oned i nput s and/ or non-negligible noise.
As model represent at i ons via frequency re-
sponse funct i ons are used very often, met hods for
deriving frequency responses from i nput - out put
signal pairs are of common interest. Frequency
response functions are calculated from Four i er
t ransforms of i nput and out put signals.
Y(co) ~ [ y ( t ) ] (11)
B( j m) = U(~o) = ~' [u(t )]"
Equat i on (11) will give satisfying results onl y if
the noise (Fig. 1) is negligible as noise c o r n -
Here signals u(t) and y( t ) are t o be t ransformed.
By suitable definitions one can ensure
y ( t ) =u ( t ) =O for t < 0 (14)
and t hus
O~3
Y((o) = S y ( t ) . e x p ( - j o g t ) d t . (15)
0
For numeri cal eval uat i on of equat i on (15) the
signal y(t ) is sampl ed at equi di st ant points n' T
in time and sample values y(n) are stored and
processed. For practical reasons functions of
time, in order to fulfill equat i on (13) and t hus be
Fouri er-t ransformabl e, have t o t end to zero for
large t > T N = N ' T . Such signals can be repre-
sented by a finite number of samples y(n), n
=0 . . . . . N- 1 and their DFT can be obt ai ned by
eval uat i ng the finite sum
with
N - 1 / ' 2 7 t ~ c n ~
Y(k. Aco)= . =o ~ Y ( n ) ' e x p k - j - W )
k = 0 . . . . . N- 1 (16)
2~
Ao = - - (17)
N -At
instead of the infinite integral equat i on (15).
Equat i on (16) indicates, t hat the DFT
Y(k-A~o) is defined for N discrete values of
frequency and t hus the associated frequency re-
sponse is defined for the same N discrete frequen-
cies. Evi dent l y the DFT of the input signal can
be obt ai ned in the same manner as described
above for the out put signal. Often st andar d input
signals will be used whose Fouri er t ransform is
")' 3
5 , , H. RAKE
known bef or ehand or can be cal cul at ed analyti-
cally or l ooked up in a suitable table. In such
cases the out put signal onl y has to be processed.
Fr om equat i on (13) is to be seen t hat step
functions and step response funct i ons t endi ng to
a finite nonzer o value for large time values are
not Four i er - t r ansf or mabl e. Thi s difficulty can be
ci r cumvent ed by not t r ansf or mi ng the signals
themselves but t hei r first derivatives. One can
pr ove t hat using simple differences
y' (n) = y(n + 1 ) - - y( n) (18)
u ' ( n ) =u ( n +l ) - u ( n )
in or der t o obt ai n
/-t ( k - A o g ) = , ~ - [ y ' ( n ) ] ( 1 9 )
~ ' [ u ' ( n ) ]
does not i nt r oduce syst emat i c er r or s as i nput and
out put signal are preprocessed in the same way.
As ment i oned before noise is a big pr obl em in
syst em identification with nonper i odi c signals. All
met hods using and processi ng such signals are
feasible onl y in cases of high signal t o noise ratio.
Thi s can easily be seen for t he det er mi nat i on of
frequency response functions. If t he process has a
frequency response H(~o) and t he noise n(t ) gives
a Four i er t r ansf or m
T~t
N(~o)= j' n ( t ) . e x p ( - j o g t ) d t (20)
0
t he measurabl e process out put signal will have a
Four i er t r ansf or m
Y( oo) =U( og) ' H( oJ) +N( ~n) (21)
and the model frequency response will be
N(o~)
~' ~o" Y(~o) = H( o g ) - + - - - . (22)
~= U(co) U(~o)
Thus t he addi t i onal noise leads t o
errors in the frequency response. By
several response funct i ons one mi ght
the effects of noise to a cert ai n degree.
addi t i onal
averagi ng
at t enuat e
Parametric models f rom aperiodic signals
Par amet r i c model s usually are descri bed by
their frequency response funct i on
H( ~ ) = bo + bljo9 +. . . b,, (jog)" " e xp( - j oyr )
a o + aljo . . . a, ( j o) "
(23)
with a j, b~, m, n, z as paramet ers.
Many aut hor s have pr oposed met hods for de-
t ermi ni ng par amet er values from time funct i ons
of the process out put pr ovi ded the process input
is a well det er mi ned signal. Of pr i mar y i mpor-
t ance are step funct i ons as input and step re-
sponses as out put signals. Her e the discussion
will be limited to this signal pair t hough t here
are ot her i nput signals (Tabl e 1) with interesting
propert i es.
Al most all met hods for the eval uat i on of step
response functions use a small number of charac-
teristic values of t he response function. One of
t he oldest and best known is t hat i nt r oduced by
KfipfmiJller giving a first or der lag model with
time del ay and a frequency response
K
/4(e9) = - - ' exp( - j ogz) . (24)
1 +j~or 1
The par amet er s K, r~, ~ can be deri ved from t he
step response as i ndi cat ed in Fig. 2 with u o being
I
FIG. 2. Evaluation of step responses according to Kiipfmfiller.
the ampl i t ude of the i nput step signal. Anot her
met hod has been devel oped by Strejc giving
par amet er values for
/:/(09)= K exp( - j ~oz) . (25)
(I + j ~ z I )"
Equat i on (25) describes a model of or der n with
equal time const ant s and a time delay. The gain
fact or K can be obt ai ned in the usual way by
dividing t h e final value y~ of t he step response
Fig. 3 by t he i nput step ampl i t ude u o. The
remai ni ng par amet er s r, ~l and n are f ound by
const r uct i on of the t angent with maxi mum slope
and measur ement of several i nt ermedi at e values
as i ndi cat ed in Fig. 3. Then with t he aid of Tabl e
2 the model or der is est i mat ed f r om Tu/T q and
Yl/Y~, Usual l y an integer value j ust below or
equal to t hat i ndi cat ed by these quot i ent s will be
chosen and Tu, r and T; adj ust ed if necessary.
Fr om T i, T~, Tg and T,, as measur ed accordi ng to
Fig. 3 and quot i ent s T~/T, T,/T, To/T and T, , / T in
Tabl e 2 cor r espondi ng with t he or der n chosen
before four values of T are obt ai ned. If these do
agree fairly well an average is to be comput ed
St ep response and frequency response met hods 523
, t
Y~o
Y~
Ti ~ Tm
-I
t ,~
Fit;. 3. Exaluation of step responses according to Strejc.
TABLE 2. CHARACTERISTIC VALUES OF STEP RESPONSES
n T, 7". 7 " . T i Y_2L T.,
T T ~ Y y = T T a
1 1 0 0 0 0 1 1
2 2.718 0.282 0.104 1 0.204 2.000 0.736
3 3.695 0.805 0.218 2 0.323 2.500 0.677
4 4.463 1.425 0.319 3 0.353 2.888 0.674
5 5.199 2.100 0.410 4 0.371 3.219 0.629
6 5.699 2.811 0.493 5 0.384 3.510 0.616
7 6.226 3.549 0.570 6 0.394 3.775 0.606
8 6.711 4.307 0.642 7 0.401 4.018 0.599
9 7.146 5.081 0.709 8 0.407 4.245 0.593
10 7.590 5.869 0.773 9 0.413 4.458 0.587
and t he t i me const ant is obt ai ned by
T
"r 1 =- - . (26)
n
o
In case t hese values for T di sagree a different
(usually hi gher) or der n and choi ce of anot her
(usually smaller) del ay t i me r mi ght alleviate this
di sagreement .
A still mor e det ai l ed met hod has been de-
vel oped by Schwarze. Wi t h t he aid of tables and
gr aphs t he user is put i nt o a posi t i on where he
can eval uat e step response , and r amp response
funct i ons and det er mi ne par amet er s for several
different model st ruct ures. For reasons of brevi t y
onl y one model st r uct ur e and one set of in-
t er medi at e values t o be ext r act ed f r om a step
response will be discussed.
The par amet er s K, T, b and n of t he frequency
response
K
(27)
t l ~oJ~ = (1 +jeJT)" (1 +jmbT)"-t
describing a model of or der n can be obt ai ned
via t he t i me values Tie, Tso, Tgo measur ed as
i ndi cat ed in Fig. 4. Fr om t hese e.g. t he quot i ent
T9o/Tso is comput ed and used t o det er mi ne a
sui t abl e model or der wi t h t he aid of Fig. 5. Fr om
this figure one mi ght find t hat T9o/T5o=2 in-
Yoo m D ~
0 . g y , o
! ,
O l S y = " = - -
- , . 4 " ; i
O.l Ym " - - I I
0 T , o T e o 1 " 9 o
t =
FIG. 4. Evaluation of step responses according to Schwarze.
T
3. 5
n = I
f
n : 2
3
2. 5
b(4) "-~ I ~ n ~ . ~
b ' ( 5 ) ' - " . ~ ~ I
b ( 6 ) h - " I S i 1 ~ I
2 3 4 5 6
n
FIG. 5. Determination of model order.
dicates a model or der of n = 3 with equal t i me
const ant s ( b = l ) or one of n > 3 with b < l .
Havi ng chosen n (usually as small as possible)
graphs similar t o Fig. 6 and Fig. 7 can be used t o
det er mi ne t he f act or b and a cor r espondi ng set of
quot i ent s T9o/T, Tso/T and Tlo/T f r om which
t hr ee values of t he t i me const ant T are t o be
obt ai ned. In case of good agr eement all para-
2
I 0 2
5
2
T v t J O f
T 5
2
I 0 o
5
T o t 3
~o
2
i O- I
4
9O
5O
IO
I I I i I I I I J I
I I ~ I I 2 5 I 0 15 2 0
2 0 15 I0 5 2
b
FIG. 6. Determination of time constant and factor b for n =3.
524 H. RAKE
2
tO 2
5
l 2
Tv i O I
T
5
2
;0 o
5
2
iO-~
4
t
rg.o 3
T~o
2
9 0
5 0
0
, , , , , , , , , ,
I__ i -I I 2 5 10 15 2 0
~0 15 ro 5 2
b "
FIG. 7. Determination of time constant and factor b for n=4.
met er s have been det er mi ned. In case of signi-
ficant di sagr eement a d;fferent or der n shoul d be
tried.
A qui t e different met hod uses limit t heor ems of
the Lapl ace t r ansf or m t o det er mi ne t he coef-
ficients of the model t ransfer funct i on
/ t (s ) = b o + b l s +. . . + b , j "
1 +al s+. . . +a. s"
(28)
f r om the uni t st ep r esponse r l ( t ) = y ( t ) / u o. As
one has
lim r l ( t ) = K o = l i m l C I ( s ) = b o ( 2 9 )
t ~3 s ~ O
b o = K o (30)
wi t h K o f r om Fig. 8 as i ndi cat ed there. Wi t h the
i nt egral
r/t ( t ) - S [Ko - r/(r)] dr (31)
0
I K ' = / " * [ K o - r~ ( ? ) ] d i "
K o - - ~ 7
O | J i
? ,"
FIG. 8. Evaluation of unit step response.
and cor r espondi ngl y
//~ ( s ) = - I [ K o - / l ( s ) ]
s
132)
the limit t heor em gives
lim r h ( t ) = Kl = l i m/ 4t ( s ) = K o "al - bt
t ~ ~ 0
Anot her st ep of i nt egrat i ng etc. leads to
(33)
bz + Kt "a I - K o "a 2 = K 2
(34)
and t he general f or mul a
( - 1)rbr + K , _ l a 1 - K , _ 2a2 -+-...
+ ( - 1 ) ' - t Ko a , = K , (35)
gi vi ng a syst em of l i near equat i ons.
I n case all bj in equat i on (28) but bo are zero
one has
K 0 = b o
K 1 = K o . a 1
K z = K I " a l - K o ' a 2 i36)
which can be sol ved recursi vel y wi t hout much
effort.
Of cour se det er mi nat i on of K~ for i>=2 can be
difficult and pr one to er r or s l i mi t i ng the ar ea of
useful appl i cat i on to low or der model s.
Equat i on (33) pr ovi des an i mpor t ant infor-
mat i on for model s wi t h frequency response
H( o ) ) = K (37)
f i ( 1 + j ( . o r i )
i = 1
As in this case
and
K = K o (38)
a~ = ~ ri (39)
i = 1
the area in Fig. 8 f r om which K t ii to be
obt ai ned is equi val ent to the sum of all t i me
const ant s mul t i pl i ed by the gai n K. Thi s pr oper t y
of t he st ep r esponse may be used for first appr o-
xi mat i ons as well as for checki ng purposes.
METHODS FOR PERI ODI C TEST SIGNAL
Pe r i o d i c t e s t s i gnal s
All measur ement s with peri odi c test signals are
conduct ed with the process in st eady state, i.e. all
t ransi ent s pr oduced by initial condi t i ons have
vanished.
Step response and frequency response met hods 525
Test signals in common use are simple si-
nusoi dal or bi nary monof r equent signals or multi-
frequency signals cont ai ni ng significant signal
ampl i t udes at more t han one frequency. The
latter signals can be of sinusoidal or bi nary t ype
(Table 3).
TABLE 3. PERIODIC TEST SIGNALS
Mo n o f r e q u e n t MultlfrN~t
D
1
Notrparametric models f r om periodic signals
Model s det ermi ned from periodic i nput and
out put signals are usually of the frequency re-
sponse type. This t ype of model is most easily
det ermi ned. In the simplest case of a mono-
frequent sinusoidal i nput and an uncor r upt ed
out put of a linear process, ampl i t ude and phase of
the frequency response can readily be comput ed
from a recordi ng of bot h signals by a two
channel recorder (Fig. 9) giving
t
I/-)(c o )l= Y , 4 ) ( c 0 ) = - - ~ . 3 6 0 . ( 4 0 )
. ! y r - '
~ l u ( t )
FIG. 9. E v a l u a t i o n of s i n u s o i d a l r e s pons e .
7
U .COS wt
s ~ ~ t
R
U ~lb
FIG. 10. De t e r mi n a t i o n of f r e q u e n c y r e s p o n s e by c o r r e l a t i o n .
The not-so-simple probl em of significant noise
and/ or process nonl i neari t y can be handl ed with
a setup accordi ng to Fig. 10 which is the prin-
ciple of a couple of commerci al l y available fre-
quency response analysers. By mul t i pl yi ng the
process out put by reference signals obt ai ned from
the test signal generat or and integrating the
product s over an integer number K of test signal
periods Tone gets
K ' T ' U r . r
R = RelSI(co)+ ~ n( t ) ' c os c ot dt
b
K. T
I = K " T ' U oS n( t ) ' si no~t dt .
(41)
As can be shown the average of the i nt egrat ed
noise in equat i on (41) is zero and the variance of
these integrals is bounded. Thus measurement s
t aken over a sufficient number of periods can be
made as accurat e as desired even in cases of low
signal to noise ratio.
If the power spectral density S. . of the noise is
known the st andar d deviation of the frequency
response as det ermi ned accordi ng to Fig. 10 and
equat i on (41)is
e . IH((o)[ K . T 3
(42)
which can be decreased by increasing measure-
ment time K T.
The met hod described above and illustrated by
Fig. 10 can easily be interpreted as det ermi nat i on
of the basic component of a Fouri er series expan-
sion of the process out put . In fact all moder n
transfer funct i on analysers are organized as
Fouri er analysers processing by digital har dwar e
signal sample values t aken at equi di st ant points
in time. Wi t h such equi pment or by processing
signals recorded in a suitable manner by general
purpose digital comput ers bi nary input signals
can be empl oyed. This can be advant ageous as
simple switching cont rol elements onl y are
needed to appl y this test signal to the process.
Anot her interesting possibility is the use of mul-
tifrequency sinusoidal or bi nary test signals
(Table 3). The process out put resulting from such
input signals will of course cont ai n signal com-
ponent s with different frequencies. These have to
be separat ed by Fouri er series expansion giving
several e.g. up to ten values of the frequency
response at predet ermi ned frequencies of interest.
Though t reat ed rat her briefly here frequency
response measurement s with periodic signals if
applicable have t o be regarded as the most
526 H. RAKE
powerful met hod at hand for det er mi ni ng process
model s wi t h hi gh accur acy even f r om noisy
measur ement s.
P a r a me t r i c mo d e l s f r o m p e r i o d s i gnal s
I f par amet r i c (frequency response) model s are
want ed f r om meas ur ement s wi t h peri odi c signals
the usual way is t o det er mi ne in a first st ep a
sufficient numbe r of frequency r esponse val ues i.e.
a nonpa r a me t r i c model . Thi s nonpa r a me t r i c mo-
del is used as a basis for furt her oper at i ons.
An i mpressi ve numbe r of met hods have been
devel oped which ai m at the det er mi nat i on of
val ues of coefficients in anal yt i c frequency re-
sponse functions. One r at her si mpl e but fairly
effective gr aphi cal met hod is based on the Bode
pl ot of f r equency responses. Fr o m the results of
t ransfer funct i on meas ur ement s a Bode pl ot is
dr awn and this in t urn is appr oxi mat ed by
el ement ar y frequency r esponse funct i ons. Thi s
trial and er r or met hod can be made mor e effec-
tive by use of pr ef abr i cat ed dr awi ng aids etc.
Comput e r or i ent ed met hods can be di vi ded
i nt o i nt er pol at i ng met hods and er r or mi ni mi zi ng
met hods. The f or mer i mpl y all measur ed val ues
t o be cor r ect and det er mi ne anal yt i c expressi ons
which satisfy each and every measur ement . In
cont r ast t he er r or mi ni mi zi ng met hods det er mi ne
par amet er s of an anal yt i c expressi on. The st ruc-
t ure of this expressi on usual l y has t o be specified
in advance. Par amet er s t hen are det er mi ned in
such a way t hat a measur e of t ot al er r or e.g. the
(weighted) sum of er r or s squar ed becomes
mi ni mal .
CONCLUSION
Ident i fi cat i on and process par amet er esti-
mat i on will al ways be different t hi ngs to different
peopl e. As t here are ma ny different reasons for
exper i ment al process i dent i fi cat i on t he pract i si ng
cost -consci ous scientist and engi neer needs to
know at least in pri nci pl e a vari et y of met hods at
hand f r om whi ch he has to select a sui t abl e one.
For reasons of si mpl i ci t y and economy, met hods
using det ermi ni st i c peri odi c and nonper i odi c test
signals deserve close at t ent i on.
REFERENCES
Eykhoff, P. (1974). System Identification. John Wiley, London.
Isermann, R. (1971). Experimentelle Analyse der Dynamik
industrieller Prozesse. Identifikation I. Bibliographisches
Institut, Mannheirn.
Schwarze, G. (1964). Algorithmische Bestimmung der
Ordnung und Zeitkonstanten bei P-, I- und D-Gliedern mit
zwei unterschiedlichen Zeitkonstanten und Verz6gerung bis
6. Ordnung. Messen, Steuern, Regeln 7, 10-19.
Str.ejc, V. (1959). N~iherungsverfahren Rir aperiodische
Ubergangscharakteristiken. Regelungstechnik 7, 124-128.
Strobel, H. (1975). Experimentelle Systemanalyse. Akademie-
Verlag, Berlin.
Wilfert, H.-H. (1969). Signal- und Frequenzganganalyse an
stark gest6rten Systemen. VEB Verlag Technik, Berlin.

You might also like