Proadlnga of the 29th Conference

on Daclalon M d Control
Honolulu. Hawall December 1990 WP-14-2 = 5:40
AUTOMATIC TRANSFER FUNCTION SYNTHESIS FROM A BODE PLOT

P. Lopes dos Santos and J. L. Martins de Carvalho

Faculdade de Engenharia da Universidade do Porto - D.E.E.C.

I?.dos Bragas, 4099 Porto Codex, Portugal

ABSTRACT auoroaches have been reported, namely by Hauer [21, i n

connection w i t h Electrical Power System Identification,
in this paper a novel algorithm that automatically where such difficulties are overcome by a trial-and-
identifies continuous-time transfer functions from Bode error procedure, based on the analyst's experience and
plots i s announcea The identification i s carried out i n intuition. The present authors have been addressing this
two stages, In the f i r s t one, the model order and 'good' proalem ana recently [7,9] they announced a successful
guesses of the poles and zeros are obtained; i n the second algorithm for the minimization of ( 1 1 in the case of all-
stage, estimates are refined by means of a modified poi? (or all-zero) plants. Unfortunatelly, the
Newton-Raphson algorithm. Because poles and zeros slmultaneous occurrence of poles and zeros creates a
estimation only requires the magnitude curve, transport much more difficult situation.
delays, i f any, can be easily estimated by means of In this paper an algorithm i s announced that
additional information supplied by the phase curve. The identifies the poles and zeros of a T.F, through the
major and novel contribuition of the proposed method minimization of (I).Basically i t consists of a modified
resides i n i t s f i r s t stage, where qualitative notions Newton-Rapnson (N.R 1 method, initialized w i t h estimates
currently 'hldden' in the intuition of the designer are resulting from a previous asymptotic approximation to
explici ty represented, yelding a simple optimization the amplitude Hone diagram. Before this approximation,
procedure that i s not impared by the presence of saddle the rnodel structure i s automatically determined;
points or local minima, and that converges very fast to simultaneously pole-zero pairs are also cancelled when
the vicinity of the true solution. A detailed example i s they have no significant bearing upon the F.R.. It i s also
also provided to illustrate the value of the method. shown that the estimation of pure transport delays
becomes a trivial Droblem once the poles and zeros are
estimated.
1. INTRODUCTION Gwen that the estimation of DC gains, and poles and
The problem of estimating continuous-time transfer zeros at th? origin are well known problems, we consider
functions (T.F.? from frequency response (F.R.) orliy i n the sequel zero type plants w i t h unit gain.
measurements, nas always aeen a topic of researcn
interesti2,A- 131. 11. MODEL STRUCTURE IDENTIFICATION
When the purpose of the iaentification i s the design Given the T.F
of a controller, the estimated model must be as simple as
possible, describing only the relevant dynamics of the
Dlant. In this context, each measurement must be
weighted taking into account both i t s accuracy and the (2)
importance of the associated frequency i n the system
description. Given that it i s relatively easy t o achieve 1
i=
measurements w i t h equal relative errors and, on the p j€B,I = 1 ,,..,rip z@, i= 1 ,.,, ,nz
other hand, closed-loop stability can be achieved Dy
limiting those errors within the bandwidth of interest
[ l ] , we have good reasons for the use of algorithms based
on the minimization of relative errors. Bearing i n mind
that, for small values, relative and logarithmic errors
are almost equal, we have that algorithms minimizing the
cr iter 1on
N

(1)
k= 1
proviae a gooa alternative Here (Y(jokl), k = l , ,N Defirtirig
denotes the set of F R measurements and EVA the plant
model (4)
Unfortunately, ( Ij i s a non-convex function, w i t h
several local minima and saddle points which render i t s
minimization a non-trivial problem In Dractice. several we have the well known asymptotic approximation of the

CH2917-3/90/0000-1093$1 .OO @ 1990 IEEE 1093

Authorized licensed use limited to: University of Leeds. Downloaded on August 24,2021 at 22:33:32 UTC from IEEE Xplore. Restrictions apply.
term Lmg!x,x,,tij. Computing the difference

ass,(x,x,,t,)-Lmy,(x,x,,t,)=
0 5t,ln( l+exp( -2(x-x,)))ex2x1
(5)

we see that it only depends of the distance (x-x,), the or

largest magnitude occurs when x=xI and it tends to zero
when ( x - x , ~_+a
When we have a set of observations [ Y ( ] q ) l ,
k=l, ,N, we can take advantage of the last mentloned
properties of ( 5 )to determine the model structure and t o
obtaln a f i r s t set of pole and zero estlmates xi,
Assuming xk- <xk, for all k, ( 1 1 ) and ( 12) allow us
i = l , ,np+nZ To that effect, let us define
t o obtain all the estimates xl,tl,X2,t2, ...,Xnptn,,tnp+nz
LmY(xk)=lnlY(j%)l, k= I , , ,N, xk=kI(Uk), and assume
recursively because X1=xk-OSln(exp(2L- 1 )) where
LmY!Yk!= FL,
In practice we only have discrete measurements and
therefore !I 1 ) only by accident would be satisfied.
However, this can be easily overcome by making a linear
interpolation between points xk and xk- 1 whenever we
detect

When the T F has the poles and zeros very far apart,
the t.erms Lrnyl(x), xl )xl
‘ ~ U ( J ( l 0 S l I i ~Ltny,tx)=Lmyl(x,xl,tl),
are negligible i n the neighbourhood of i1 We can
tnererore expect estimates w i t h great accuracy, in tnese
cases, the threshold L must be selected large enough to
prevent measurements errors giving rise to false pole or
zero estimates In the case of Doles and (or) zeros close
Defining
t o x1 we can no longer neglect the influence of the terms
Lmyl(x), pi, and consequently a loss of accuracy in the
rstirriates is unavoidanle Because our purpose at this
stage is the determination of the model structure, we
nlust assess In what extent this loss of accuracy might
lead t o incorrect models To this end, let us start by
c m i d e r i n g the unfavourable situation of a single pole p
of multiplicity np In the deterministic case we have

Because IiniLmY(x)=-np(x-R) we can ensure that at

X F
least ,n, poles w i l l be detected, provided that a sufficient
We are now i n a position to determine the model set of measurements, at high frequencies, i s available
structure by means of (5).To achieve that all we need i s The probleni here i s the likely determination of redundant
to define a threshold L ) 0. Then an estimate iCi, of a pole poles and zeros. leading t o an unnecessary extra
or zero. w i l l be obtained whenever computational burden i n the later stages of the
?Igorithm, dealing w i t h the retinement of the estimates
In order to avoid that, we need

1094

Authorized licensed use limited to: University of Leeds. Downloaded on August 24,2021 at 22:33:32 UTC from IEEE Xplore. Restrictions apply.
 111 ASYMPTOTIC APPROXIMATION TO THE BODE DlAGRAM-
Let's assume for the moment that we have a
contitwom set of measurements i n [xmln,xmaxl Then our
estimates can be refrned through the minimlzatlon of

where xLp is the npth frequency satisfying ( 1 1 ) which led

t o the estimation of inD. The l e f t hand slde of (15)
vanishes for x=xLp (by definition) and Is bounded above by

i=
1 1= 1
where n=np+nz. In the case of a l l pole or all zero T.F.3
With poles and zeros sufficiently far apart, it can be
To avoid the estimation of redundant Doles we must
have shown that the minimum of J(6) lies i n a region where V
i s convex, and which contains the minimum of V [81.
Clbviously this fact will simplify the optimization of V.
By setting the gradient of J(6) equal to zero we
arrive at the following iterative algorithm [81

"P
Fig I depicts &-np2 as d function of L for
I=1
different values of np We can see that, for each case,
there exists a lirnit beyond which (17) i s fulfilled
Because noise immunity increases w i t h L, it is
convenient t o use large thresholds in situations like this

i=2,...,n (20)

nP Notice that the algorithm is not disturbed when

F i q 1 , zki-npR as a function of L
1= 1

Let us analyze now a afferent type of situation

namely, a T F w i t h a pole p and a zero z

I I I prdctice where we have a finite set of N

BPcause ILmY(x)lilxp-x,l, the pole and zero can only nwasurenlents, LmYkxi in (20) can be defined by linear
be detected i f L~lxp-x,l However, i f pz)G, we gain very illterwldtiun With reaard t o computing time, the
little i n using a small 1, necause when xp=xz they alqorlthm i s rather fast because, i n each iteration, the
contribute very l i t t l e t o the system F R mo;t time consuming operations are 2*N floating point
The above discussion ShOWS that the ChOiCe ot L n u l t :pl :rat :5\ns
must be the resiult of a compromise between the ability Rejarding the alaorithm as a transformation
t o detect poles and zeros reasonably close and the Ftoi IR+R i t can De proved [81 that
lmmunity against redundant pole-zero detection induced
by rnultiple poles or measurement errors
1095

Authorized licensed use limited to: University of Leeds. Downloaded on August 24,2021 at 22:33:32 UTC from IEEE Xplore. Restrictions apply.
 - IldF/dx,ll~=eA 1 and xl-l or x,+lafe no longer
bpdatecl,
where 11.11 denotes Euclidean norm, lc, a necessary
- IICEI~~~JI~=+ ) I and i = 1 1
condition for the convergence t o the minimum eo
Notice m a t w i t h this rule, one ensures tnat tne most
Computing
critical estimates do not violate the necessary condition
Sf1 A,-LmY(iclf 1 for ccrnv~rgenre
t-
i2-21 2
IV F b
01AGRAM
The next step i n the identification process, i s the
minimization of ( 1 ) through a modified N.R. method,
initialized With tne estimates from the previous
asymptotic approximation. Reexpressing ( 1 ) as

Sfl
7 0 ,Jfl,Jfl-1,J*i+l (23)
SXi
3 suitable minimizing algorithm i s as follows:
and substituting Ai, i = I , ...,n for i t s value at the
equilibrium point we arrive at tne following necessary
condition for the convergence of ( 2 1 ) 181

ai=ass(ici,i)o,r)-LmY(R,) , I = I ,, .,n
I= n
I ,...

*here P i s a positive definite matrix Given that

most of :he times asymptotic approximation produces
estimates in the region o f convexity o f V containinq the
!n!n!n?vm [a!, a natura! cnoice for P i s a N R direction

When the poles of a T.F are sufficiently far apart,

ass(x&,r) w i l l almost overlap the asymptotic diagram
of LmY(x). As a consequence we have

in order to Drevent PL from being non positive

.Jet mitt. w e can set [ 8 ]
which guarantees that conditlon (24) is fulfilled pmilded - PX=ih+crIi-’, or=minlDI evervtime we find at least
xi+ l-ii,i=1 ,..,,n is sufficiently large. Let us now turn our dric rieqative elzment or D
attention to all-pole transfer functions. Despite the FL=(H+u11)- 1 or P=(H+~I+criIj-l,
unavailability of results ensuring convergence for an n,=rflax(AVdV11,0 1 ) when a pole estimate i s closed to a
arbitrarv number of poles, it is Possible to show that
multiple poles constitute the most unfavourable zerc estimate ( e g kl-i,kO I , ic, the pole and kl the zero)
situation and that for a T.F. p/(s+p)n, n=2, ...,
10 and 3nd the elements of D corresponding t o these estimates
are smr;lll !e g less tnan 0 1 )
;fmax=ln!1 0 0 ~ 1 lldF/d9ll~,i0~0.2c)S
, [8], - P = ( H +or~P?~ I =) I ~ , 001, wnenever
( H + Q I + Q ~u2=0
When the T I . has poles and zero5 and these are close cales (or zeros) estimates close t o each other ( e g
Lo the poles, 00 may nave pole-zero parrs cancelling each l \ , - ~ i lC’C01, xl and Y, are both poles or zeros) and the
other, preventing the convergence of the algorithm [81. elements of D corresponding t o these estimates are small
Given that for T.F. of reasonable order convergence ! e 3 , less than 00001 !
problems only occur i n the presence of pole-zero Edotlse that it i s advisable t o keeD the elements of
cancellations or for a too small xmax, we can overcome 1 1 , a ~ - r - w f i ioccI?t
n ~ t n faclllrate the detectlon of these
these limitations i f we stoD updating an estimate x
whenever we detect:
:xes ,hs an sdditicnal precaution. A@ must be divided by
1096

Authorized licensed use limited to: University of Leeds. Downloaded on August 24,2021 at 22:33:32 UTC from IEEE Xplore. Restrictions apply.
 two, whenever V(X,@+ l,r))>V(X,ex,r). compute F(SPG(S)/Gr(S) and regard it as an estimate of
Despite the availability of methods that ensure exp(-sT). Because i t s phase i s a linear function of
(theoretically) error free F.R. measurements, these are frequency] T can be estimated by the least squares
unavoidable in practice Therefore an identification method. Although such estimate Is biased, i t s
met.hod based on F.R, measurements must be robust computation i s straightforward and good estimates are
against measurements errors, When the log. of the possible in most situations.
measurement errors are independent, identically
dlstributed gaussian random variables, it can be shown VI.
that the minimizat.ion of V(X,i,r) leads t.0 maximum ifthe purpose of our identification i s the synthesis
likelihood estimat.es 181. Since In(a)-ln(b)=(a-b)/b, i.e., of a controller, it i s important to know the degree of
relatlve errors are approximate equal t o the absolute robustness of our design. Recalling that if, in a pole-
error of the logs, we have assured immunity against placement design, the closed loop T.F. Gm is stable for
measurement errors i n a large number of cases. model G, then it remains stable for Go i f [11

v,
If the T.F. contains a pure transport delay T, we can
write

G(s)=G,Wexp(-sT) (29) (30)

where Gr(5) is a ratlonal function i n s. Since where Gff and GfD are the forward and feedback controller
IG(]w)l=IGr(]w)l, the algorithm just descriaea can be used T.F.3 respectively, then it i s natural to define a
robustness index as

(31)

In this perspective it is obvious that the greater R

the greater the robustness of the working model.

V l l . EXAMPLE
In this section, the performance of the algorithm
will be ilustrated by means of a numerical example.
- The plant t o be identified i s
,I 1 10 too 1000
o (rad/s)

Fig. 2A - Amplitude Bode Plots of GI and G1 With 100 undisturbed observations between
0.1 rad/s and 200 rad/s w i t h equally spaced logs and a
detection threshold for poles and zeros set t o 3dB, the
01 following model was identified,

1: w 5 554+1 1
G \(SI=- R=29.41 (33)
40 (s/O 874+ 1 )(s/2 1.341 1 )
+

Comparlng the Bode dlagrams of GI and G1 (fig. 2)

we find that, despite the order reduction, the F.R. are
almost indistinguishable.
In a comparative study of model order reductlon
methods (Lai and Hung 1987) GI was reduced to models
with the same structure as 6, above. The methods under
comparison were : Cauer f i r s t (CAUERI) and thlrd
1 10 100 1000 (CAUER3) forms, matching one time moment and three
o (rad/s) Markov parameters (T 1 M3), matching three time
Fig. 28 - Phase Bode Plots of G1 and G1 moments and one Markov parameter (T3M 1 ), Davidson's
method (DAW, partial realization method (PR), least-
to identify the size of the poles and zeros of G,(s). squares methoa (LS) and singular value decomposition
Knowing the right half-plane poles and zeros, we can method (SV). Table 1 compares the different values of R

1097

Authorized licensed use limited to: University of Leeds. Downloaded on August 24,2021 at 22:33:32 UTC from IEEE Xplore. Restrictions apply.
 TABLE I

CAUERI CAUER3 TIM3 T3M1 DAV RP MQ SV

3.75 15.87 14.29 3.77 0.35 8 8.4 13.51

obtained w i t h these methods, our method ( G ~ ( s ) ) and pxtensive simulations In the example provided an
Levy's method (LEVYlWith the exception of Davidson's n~itomatir, order reduction took place leadlng to an
method they a l l display a high degree of robustness. excellent robustness index ( 3 I ),
However, 6l(s)i s the one w i t h greatest R.
In order to assess the effect of measurement References
errors, a comparative study was made between our
method and Levy's. Several identification experiments [ 11 Astram, K. J, ( 1 9 8 0 l "Robustness of a Design Method
were carried out based on noise corrupted measurements, Based on A S S 1 y " t of Poles and Zeros". JFFF Trans,
and the results are summarized i n table 2 w i t h the help Automat,Cont r., vol AC-25, 588-59 1.
of 95% confidence intervals (when applicable). The table 121 Hauer, J . F. (1Y82)."Power System Identification by
was generated on the basis of 10 sets, each consisting of Fitting models to to Measured Frequency Response".
100 observations corrupted by independent, equally IEEE TratxPower Aoo,and Syst., vol PAS-101, 4,
distributed gaussian relative errors: 9 15-923.
[31 Lai, J. and Hung, J. C. (1987). "Pratical model order
LmY(x~=lnlG~c~o)(l+~/100)1 reduction methods". IEEE Trans. Ind. Elect., vol IE-34,
p-N(O,& (34) I , 70-77.
141 Lawrence, P. J. and Rogers (1979). "Sequential
TABLE 2 Transfer Functlon Synthesis from Measured
Frequency Response Data". Proc IFF, vol 126, 1, 104-
106.
MODEL p lpll lp21 lZIl
151 Levv. E L . iIY59). "Complex Curve Fitting". IRF Trans.
. . , I

5% 1.565i0.097 31.70i2.20 14.69i 1.88 2.87i0.24 Automat. Contr., vol AC-4, 37-43.
[61 Lilia, M. (1988). "Least Squares Fitting to a Rational
LEVY I % 0.9Oi 0.009 20.88iO. I 3 5.6 1 i0.08 19.61f 1.92 T r i n i f e r Function w i t h Time Delay". Control 8&
Oxford. U.K. - IEE Conference Publication, 285, 143-
OX 0.833 20,86 5.21 1 [7l Lopes dos Santos, P. and Martins de Carvalho, J.
( 1988). "A Robust Computer Algorithm for
5% 0.870i0.010 21,19iO,3; 5,s1 io. 10 26,3 1 f 2,08
Determining the T.F.of an Industrial System from i t s
((SI t w 0.a73~0.002 2 i . 3 1 ~ 0 . 0 5 5.54i0.01 2857i 0.82 Frequency Respponse". Anais 7QCongress0 R rasileirQ
be Automitica (IFAU, S.P., Brasil, vol I , 124-129
ow 0.874 21.341 5.554 29.4 1 !in PortugueseX
181 Lopes dos Santos,P and Martins Ue Carvalho, J.
I ! 9893. "Continuous-time System Identification
in the frequency interval u = O . l rad/s %=ZOO rad/s, from Bode Plots" Internal Reoort INFSC-N . I aruo de
An inspection of thls table shows the superiority of the Mornl3llher.Porto. PortuaaL
proposed method and i t s remarKable noise insent ivity [91 Looes dos Santos, P. and Martins de Carvalho,J.(1990).
property. " A Robust Algorithm for Identifying Transfer
Function Poles from a Frequency Response".
VIII. CONrl USIONS . 4 n d Model. Simul. , 7, Akademie-Verlag Berlin.
In this paper an algorithm for (reall poles and zeros [101 Rake, H. (1980). "Step Response and Frequency
Identification from a Bode plot i s announced. Response Methods". Automatica, vol 16, 5 19-526.
The identification i s carried out in t w o stages: In the [ I 11 Sanathanan, C. K. and Koerner, J. (1963). "Transfer
f i r s t one the model order and 'good' guesses of the poles Function Synthesis as a Ratio of two Complex
and zeros are obtained; then, i n the second stage, Polvnomials". IEFE Trans. Automat. Contr., AC-8, 56-
estimates are refined by means of a modified N.R. [ I ? ] Whitfield, A. H. (1987). "Asymptotic Behaviour of
algorithm. Because poles and zeros estimation only Transfer Function Synthesis Methods". Int. J. Contr,
requires F.R. magnitude curve, transport-delays, i f any, Vol 45,3, 1083-1092.
can be estimated i n a third stage by a simple least [131 Young, P and Patton, R. J. (1988). "Frequency Domain
squares procedure based i n the phase plot. I dent i f ication of Remotely-Piloted Helicopter
At the moment it i s s t i l l necessary a-priori Ciynam1i.s Using Frequency-Sweep and Schroeder-
knowledge of the number of the right half plane poles and Phased Test Signals", Proc. AIAA Flight M e c h a
zeros. &orifer-ence. Minneauolis, U.S.A., 1-9
The performance of the method has been tested by
extensive simulations. In the example provided an
automatic order reduction took place leadina to an
excellent robustness index (31)

1098

Authorized licensed use limited to: University of Leeds. Downloaded on August 24,2021 at 22:33:32 UTC from IEEE Xplore. Restrictions apply.