Copyright © IFAC Identification and System Parameter Estimation, Beijing, PRC 1988
ROBUST ADAPTIVE CONTROL FOR SYSTEMS WITH UNKNOWN TIME DELAY Ma Runjin* and Zhang Yongguang** *Division of Engin eeri ng, Northern China L'niversit)' of Technology, Beij'ing, PRC **/ nstitute of Systems Science, Academia Sinica, Beijing, PRC
Abstract.
Usualy a adaptive controler most in use is based on a lower - order model, due
to the simpleness o f model the unmad e led part of the true system often makes the co ntrol - 'C s be not robust. This paper considers a robust class of models that r.(s)=Ke /( 5-5 ) " , 1 (s - sn)' where
~
is time delay. For th is c lass of model the unmodeled part of the system
can be converted into a part of unknown time delay. Thus, the sampled system of G(s) has a fractional time delay time delay
p
p,
Basing on combination of esti mating the fractional
O ~p< l.
and co n struct ing a self - tun ing
controler, a rob ust adaptive co ntr ol
scheme is given in the paper . Th e simulations show that this adaptive control algorithm has indeed robustness.
Keywords .
Sampled data .ystem; Fractional time delay; minimum phase;
Self-tuK~ng
r egu -
lat or .
I NTRODUCTI ON
contrast with it,
in this paper we reduced order of
model further by using rat:onal aoproximation to For the systems having unknown time delay
~
t here
unknown time dela y and give a adaptive control by
are several adaptive control algorithms in the lit-
conbinating the estimation of fractional time delay
eratures, for example, Kurz and Godecke (1981).
and a self -tuntng
Wong & Bayoumi (1982), Mosca & Zappa (19 83 ) and De
pear good robustness also.
controler. The simulations ap-
Keys er & Van Cauwenberghe (19 85 ) . However, most of them are based on a assumption that time dela y
~
is a integer times of the sampling interval To.
A EXAMPLE WITHOUT ROBUSTNESS
Obviously, this ass umpti on is not natural. Usualy ,
1: =(d. El )To, where d is a integer and 0,. p < 1, this
P
is called fractional time delay . Zrang
&
Example 1 . Ma
Consider a third order system
G(s) = 1/(10s-1)(2.5s-1)(2s-1),
To=5s
(1 )
(19 88 ) gave much insight about fractional time delay, the analysis sh owed that the zeros of sam-
it has one main pole that 5 = - 0 . 1. According to 1
pled system are strongly effected by fractional
classical procedure it can be reduced ord er as
time de l ay
p .
Therefore, most adaptive controler
first order s ystem G' (s) =1/(105-1). The corre -
which depends on minimum phase condition would not
sponding impulse transfer function has following
be robust due to without considering fractional
form:
time delay
p H( z )
Recently, Souza, Goodwin, Mayne & Palaniswami
tor for sampled s ystem ( 2 ) . The simulation result
struct adapti ve control by using rational approxi ~.
(2)
lie designed a mini..mum vari..ance self-tuning regula -
(19 86 ) has given a interesting new thought to con -
mation to unknown time delay
b z- l / ( I.az - l )
.ppoared in Fig.l shows that the control is failed
Simulations given
to divergent. However, if we use a control model
in the paper show the adaptive control has good
having unknown time delay term
robustness. However, the cost is that the order of
€-~s as doing in
this paper, the .daptive cont rol by using the algo -
model will be increased one or two of order . In
rithm with considering fractional time delay given
287
288
Ma Runjin and Zhang Yongguang
later in this paper makes the closed loop s ystem be controled and have much better performa nce . The
Especially, for the case of n =2 zi ( p ) is inside the unit disc and z2(f )-OO ,s
p _ l . and the
result is showen in examp le 2 and Fi g . 2. relatio n hetween p and zi ( P ) for first order sys_ tern approx imately is SAMPLED SYSTEM WITH FRACTIONAL TIME DeLAY
Cousider a continuous system :
(5 ) and for second order system is ( 6)
(3)
where
q=4 or 5 .
Sampl in g the continuous system G(s) under time in terval To, thus, and 0
~p<
I,
Pis
~
=( d+ P )To, whe re d is a integer
ca lled fractional time delay.
APPROXIMATION BETWEEN TIME DELAY AND UNHODELLED PART
Th e
discussio~
abAve shows a source of frac tional
time delay due to sampling, t here are, also, other
Let
30urces for fractional time delay that: a)
System existing somewhat
D(s)=e - t s
nonlin~arlty
where
discribed as linear system .
b) The lag due t o
c~ lc u lation
(7)
is
of control signal
after samp l ing.
r
is a positive integer, th e degree of ap -
proximation . Obviously, \ D(j W )\ =1 and -w"t j
also,
c) The assumed order of system is not correct,
Dr(j <...>)\
L
D(j W )=
( 1.w2 'C. 2 /r 2;'/'~
=
and
l _r tg- ( t..lt /r) . The comparision of
a nd so on.
both L D(j <...l ) and L
Dr(j LJ ) is shown in Fig.3, it
can be seen that L. Dr( j "" ) is v e ry closed to
All of these will introduce some modelling errors those make the integrated time delay be unknown and
L
D( j w ) and difference between them is less than
7% while w~ r4:.(mofe detail ca n be found in
result in robust less of adaptive co ntr01.
Sou za, Goodwi n, et ai, (1986». Elementary c lassiFor the ca s e of O.:.p< I, the impulse tran3fer func tion Can be obtained by using the modified z- trans formation (see Jury (1964». We denote the sompled
cal contro l arguments to ld one that it will be vir tually impossible to r obustly stabili ze the plant if the phase shift arising from the delay and approximation
system with zero-order hold as follows:
exceeds something like 100
0
over
the c l osed loop bandwidth. Thus, if the objective z
-d
is to design a contro l system, one is on ly inter-
este d (4)
i~
accurately modell ing the system u p to
phase shift of about 100
0
.
For the ca SE p =O, let H(z;O)=H( z ) which is obtained by orig i nal z-tr ansformation . Let zi ( p ), . .. ,
z~ ( p )
From the anal y sis above, if system ( 3) has main
be the zeros of H( z;p ) and z l' ... , zn_ l be the
poles sI ' s2' . . . , srn' the term l /( s-sm_l)
zores of H( z ). Much more insight wa s given in Zhang
( s-sn ) can be approximated as K· e
& Ma ( 1988 ) , here some main results are listed
~
- 'f1's
,where t
•
k " are adequate positive numbers. Thus,
wi t hout pr oof . (cf . Th e Appendix ) (s-s )
(i) H( z;p ) has n zeros, but H( z ) has n - l zer os. H( z;p ) h as n poles which are the same as those of
(i1) As z.i. ( ~ )
p_ O,
H(z;P) _
bandwidth. Obvious 1~1, G(s) and G' ( s ) are in the
H( z ;O)=H( z ) , O:1e of the
tends to 0, f o r example z i (0) =0. So
zeros
zi ( ~ )
-1
H(z ) , one o f the z~ (1) =oo.
So
the sample d s y s t e m i s non-minimum pha se while the
P is
main
p~leo
nOt more than 3 . In fact, control er de-
Signed based on ( 8 ) ha s good r obus tness than that
tends to 00, for example
fra ction al time delay
same class of mode I, out G' ( s ) has lower orde r. In
practice, one consider u s ualy the system having
there i s a c an ce ~la tion between 0 zero and 0 pole . (Hi) As p--. l, H( z ; P ) _ z
(8)
is an approximate of GCs) over the clo sed l oo p
H( z ), but there is one more pole that Pn_l=O.
zer os
:n
s ufficently close to 1 .
without consid2ring time delay i f we can give a
estimate of fractional time de l ay p .
Control for Systems with Cnknown Time Delay F~ACTIONAL
ESTIMATION OF
289
STR was presented in ~strom, Wittenmark (1973) and
TIME DELAY
.
Iserman (1977) . A modified STR to de al with nonIt is very difficult to f i nd an analytical expres-
p from the z e r o
sion of calculating
n<4. Znang & Ma (988) sh oued tions between
p
f )
z~(
and
z~(
P)
unless
a ppro xi mate rela-
minimum phase sampled system was given in
As ~ ["i5m
&.
Wittenmark (1985). Her e , we try to present ano ther mo difi ~ d
STR t o deal with th e system ha v ing unknown
t ime delay.
tha t
Co nsid e r the st a ble system t o be contra led as folfo r
Z;-I
l ow s:
n =1
G(s) = k e
-1:5
/(5 +5
(9)
and
1
(s+ s )
)
(13 )
n
which has one o r two main pol es. We try to model it by th e foll oW i ng mo de l: Based On th e s e relations, we c an estimat e a ppro-
C; (s )
ximal e ly f or n = 1, 2 as f o l lo\.;s :
I ) fit f o ll owi n g
mo d ~ l:
~
k'e
- t.'
s
I(s+a )
(14 )
(a:> 0)
if G(s) hriS on l y one main po l e . or
'" y( k-I ) . "a y( k- ZI -Ab ; ui k-1 )- 1'\ y( k)=a biu (k - Z) j 2
:b 3U( k-3 1.
(15)
<- l lcl
(10) if G( s) has two main p ol es, wh e re aiO, bJO, a and b
If f i t first orde r modpj . ta kp 2) Solve [ he z e r os
f\. z;l
k)
~,~O and ~i ~ O '
., no
/\. " Z2(K,
of
may b e diffe rent p os iti ve - ea! ( aJb . a > O, b>O), the same po si tive re al (a=b. a >O, b>O) or a pair
liD ),
of co mp lex conju g ate (a= b). that i s the r oo ts of
A _ I' 71' I' R ' ( Z) = b i z -. b;z.b~
sampling i nterva ll th 0r~
n:::l)and tak e \'ii\<
on ly one ze r o zi(k) [ or
i s
\'i2\ .
in eve ry
usu .1 ly zi Rnd
z2
p
2 t he sampled syst ems given in
n~l,
(4) e i the r ar e minimum phas e O r exist on l y one zero
a re
out sid e unit disc, so wc can co nveni e ntly t a k e in to
negative r ea l nu mb er. 3) E stimate
For th e ca se of
p
a ccount the estimation o f
2 in e\'('r~' s:l l"!lpl ing i nt.E:' !"'v al by
acco rding to t he loca-
tio n of z e ro ou ts i de uni t di sc . Thus. anadaptive
co ntrol a lgorithm is s hown as fo llow s :
using the f o rmula that
A
St ep 1. Se l ec t d(O), 9 (0) and Pt O) , usu a ll y
(1 1 )
6 a CO)=O a nd P( 0)= 10 1.
Step 2 . Recursl ve l y es tima t e I'
Note that
P 2 1k )
mi ght be grea ter than I or less
e
( k) b y RLS o r
Et
o ther ide ntifi cati on a l g orithm, wher e
=(a , a , l 2
than 0 if z2 >4- or 0 < z2 <4- re spe ctinl \". I n fact, fo, 2A St e p 3. Calculate the ro ots of "B/ (z) =b;z .b z
this is po ssibl e i f dt0' is llnccrrectlv se l ected . Simulations sh ow th a t d... (K i "':",i g!lt bl? grea te r i f d (O)
is l ess th ar th e t :- 'Je
be less t han 0 i f
d~ O)
is
\ 'il ;~O
g~~A:~r
~ha n
a T'c d. ( k ) "'ight
than
: r ue
~he
valu e (cf . Fi g . :. an d 51. sh 0ul~
So the pro cedu r e
Recur~ively
4)
"3
" f\. when n=2 (o r root of 1\. B , ( z )=b ;z -bi wh en n= I) .
tet z2 has r.:.aximllm ab so lute va l ue r e lative to zi ''''hen n =2. St e p 4. Es ti mate
~ 0n:~ n~~
adj~ s~
2
-b
~j k'
3S
as
f01 !o~s:
f011 0~~n g
f ( k) and d(k) bv using the
p rocedu r e gi ven in s ection S .
role:
St e p 5. Compute the mul ti - st ep pr e dictions v ( k-i \ k) , i= l, 2, ... , d(k) as fo l lo ws:
d t k-I ' -I
A
Thus. "t.. ( k ' = CC(k 1-
p : ',{\ < c:
;:
C' !SP ~· 't< : <:·
d ( k -! I A
E' \k ·"~c.
5 ) k=k-l. then back t c
Cas e I: If ~ i n i ",um
I
z
i \< I ,
t h u 5 the
5 a mr
led s y 5 tern i
5
phase, the o rig i nal STR can s t ill be used .
To compute th e multi - step pred ic tions by usi n g the
=, _"
cc ~t;~u e
the p:- o-
following model:
ced ur e. 1\
1-a (k) z l
l!~K~O\{:-;
TI:1E DELAY
Case 2: If
\
z2\~
-I A
- a (k) z 2
-2
1, fi rst fac to rize
(16 )
~' (z) =
" -I " -2 " -3 b; ( k)z +biCk)z -b ( k ) z into Band B-, wh e re
3
290
Ma Runjin and Zhang Yongguang wh ich has a pair of complex conjegate main poles
+ " + A+ -1" 1\ -1 B = b Ck)+b Ck)z = biCk) Cl-ziz ) 1 2 1\-
l+b Ck)z 1
-1
A
= l-z z
2
t hat (17)
-I
G(s)
e-
3s
2 /(9s +3s+1)Cl . 3s+1)CS+l)(0.5S+1) (24)
the n , to computing the multi-step predictions by The modelled part of (24) is that
using the following model : ,,_
All
b Ck)-b Ck)z 2 1
- 1
(25)
z- d(k) ( 18) The co ntrol result is shown in Fig.4.
Example 4. Co nsid er the first order system with
wher e
timevarying time delay that ( 19 )
(26)
t
The time delay is changed as follows: The purpose of the transformation is to make t he
0
static gain ke e p the same.
Step 6 . Determine the control u(k)
-C Ct)=
Case 1: If!z2 l
"'-80',11'0'
2.7 To
tion that 1\
uCk)
if 0~t<80To if
80To ~
t< 180To
if 180To ~ t<. 200To
1.3 To
if
0.2 To
if
200To~
t < 260To
t ? 260To
1\
[ Yr(k+d(k)+1 )+a (k)y(k+d( k) \ k) 1 1\ (k)y(k+d(k)-1 1\ \ k )-b " ( k)u(k-l) +a 2
2
(27) (20)
-'b3(k )u( k-2)] / bi Ck)
Assuming t:
1 z2 1 ~ 1, then us e the following
Case 2: If
(t )
is unkn own, the performance of the
adaptive co ntr ol by using th e algorithm propos ed in this paper is shown in Fig.5.
equation that " \ k) u(k) = [ Yr(k+d(k)+I)+a1\ (k)y(k.d(k) 1
+~2(k)Y(k + d(k)-1 \ k) -~( k )UC k- l)]
If we don't tak e accou nt into the unknown time de-
(21)
lay in the modelled parts of example 2, 3 and 4 the
Idl
r es ults will be divergent simil ar to exa mple 1. We
1
can conclude th at model c lass (13) is indeed a ro-
I b Ck)
bust modpl class and the adaptiv e co ntro l for the
where Yr(') is the set-point of output .
class by using first order or second order model Step 7. k=k +l. then back to step 2 to continue the procedure .
(14) and (15) with unknown time delay has good ro bustness .
Remark There are several mothods to compute the multistep predictions, for example, De Keyser &
REFERENCES
Cauwenberghe (1981) and Clarke, Hodgson & Tuffs (1983) .
Astrom, K. J. and Wittenmark, B. (1973) . On self tuning regulators, Automatica, 9,
185.
Astrem, K. J. and Wittenmark, B. (1985). The selfSIMULATIONS
tuning regulat ors re vi sited,
~th
IFAC
Symp . on Identification and Syst. Parameters Example 2. Consider system which is the sa me as ( 1) that
Estimation , York, U. K"
xx v.
~strom, K. J. Hagander, P. and Sternby, J. (1984), G(s) 1/( 105 +1)( 2.55-1 )(25 -1 )
(22 )
Zeros of sampled systems, Automatica, 20, 31 - 38. Souza, C. E., Goodwin, G.C. , Mayne, D.Q. and
But the modelled p3rt of (2 2 ) is changed as
Palaniswami, M. (1986) . An adaptive control (23 )
The resu l t of adaptive control introduced in this paper is presented in Fig.2. It is obviously seen that the performance is much better than that of example 1. Example 3 . Consider the fifth order system
algirithm f or linear systems having unknown time delay, Technical report EE8626, University of Newcastle, Australia.
Zhang, Y.G. and Ma, R. J. ( 1988 ). Zero< of
Control for Systems with Unknown T ime Delay
29 1
Compari ng H(z; P ) with H(z), H(z; p ) has n zeros, but
Acke rmann, J. (1972) . Abtastregelung, Springer-
H(z) has n-l zeros on l y. In addition, H(z; p ) has
Verlag.
poles which are the same as that of H(z ) , but H(z;
De Keyser, R.M.C. and Van Cauwenberghe, A.R.A. (1981). Self - -uning mul t istep predictor appli -
P)
has a additional pole: Pn+l=O .
c"tion, Automatica, 17, 167-174. Now we consider two limit processes
De Keyser, R.M.C. and Van Cauwenberghe , A.R.A. (1985). Ext en ded predi ct ion self-adaptive con trol, Pr oc . 7th IFAC Symp . on Identification and Syst. Parameters Estimation, York, U.K . ,
K~yser,
p~ o
Rnd
p_l. First, we investigate the case of
p~ O .
Based on
the co ntinuity of modified z-transformation it is
true that
1255-126e. De
3$
R.M.C. (1986). Adaptive dead-time e sti-
lim p- O
mation , Pro c . 2nd IFAC Wo rk shop on Ada p tive Systems in Co ntrol and Signal Processing , Lund,
lim
bie r )
lim
b~+I ( P )
p..... o
Jury, L1. (1964). Theor y and Application of the
p-a
Z-transformation, John Wi l ey , New York.
Verlag.
b'(O) =b., 1
( 198 3). Th e o ffset problem and k-incr emental pr e dictors in self - tuning control, Proe.
i=1,2, . .. ,n
1
(3 1 )
Let Zl' z2' ... , zn_l be the n-I zeros of H(z) in ( 28 ) and zi (f ) ' ... ,
Clarke, D.W., Hodg son. A. J .F . and Tuffs, P.S .
(30)
This means thdt
Sweden, 209 - 213 .
Is erma nn, R. (1977) . Digitale Regclsystem, Springer
H(z)
H( z ;0)
H(z; P )
P)
z~( f )
be the n zeros of H(z ;
in (29). According to polynomial theory we have
lEE,
(32 )
130 , PtD (5), 217-225. Kurz, K. and Goedec ke, W.
(1981) . Dig ital parame-
b~+l(O)=O,
Because
so there must be one zi(O) at
ter - adaptive control of pro cess with unknown
least that it equals O. Without los s of generility,
dead-time, Automatica, 17 , 245 - 252.
we can assume zi (O)=O .
Mosca, E. and Zappa, G. ( 1983 ) . A MV adaptive con troller for plants with time-varying 1/0 transport delay. Pro c . IFAC Works shop on Adaptive
According to (29) there is a 0 pol e Pn+l' so there would be cancellation between zi (O)=O and Pn+l =O as
p_O and
Systems in Con trol and Signal Processing, SAC - 5 lim
Wong, K.Y. and Bayoumi, M.M . (1982) . A self-tuning
zi+ l ( P )= zi+l (O)=zi'
i=I,2, .. . ,n-I
(33)
/4-0
algorithm for syste ms with unknown time-delay,
For the convenience of theoritically mentioning we
proc. 6th IFAC Symp. Identification and Syst .
assume that H( z; O) and H(z) ha ve a 0 zero and a 0
Parameters Estimation, 1064- 1069.
pole constantly. Next
consider the case of
'Ne
p_ l. Also based on
the con t inuity of modified z-transformation we have
APPENDIX
lim
H(z;
f'~
When E> =O the transfer function of the sampled sys -
lim
bi+l( r )=bi ~I(1 ) =bj'
lim
bi( P )= bi(!)=O
P- l -n
n
H(z)
'-a 1 z
- 1
-a z
= H(z;l) = z-I H( z )
z-d
z -a z
Because bi ( 1 )=0, thus
1
For the case of
n
(28)
- a
O
n
z~ (!)
(36)
, ~ ,
This implies that there must be at least one zi ( l )
to obtain the transfer svste~
function of the sampled
I
r~ 1
-n
n
n- l
i=1, 2, ... n-1 (35 )
1\
n
(34)
This means that
tem (with zer o-order hold) is as follows: - b z
p)
1
that equals infinite, without loss of generility we
the modified z-trans-
assume
z~ (1)= 00 .
and
z ~ (1 )
form method is used . we have ' ( D ) - ( n-l 1 b 1' ( P) z - 1 . . .. -b n-I I .z
H( z: P ) = 1-a z
-1
1
- . .. -a z
- n
n
n- l
- ... -an ) z
i=1,2, . . . ,n-1 (37)
z' Cl) n
n
bi ( f l zn- .. . -b~(r)z-b~_I(f) ( z Ta1z
z
1
-d
Z~ ( I)= oO
z-d
is equivalent to a backward shift operator
( 29) z- l, which is absorbed by
Ma Runjin an d Zha ng Yongguang
292
Input
A
Output
Esti .. "tion ---=-=:=~==------==- . - . ,--
. 4' ".
"
K
Esti .. "tion ~)
i.
" Fig. 4
Fig.