Professional Documents
Culture Documents
A practical real-time process identification scheme must be relatively simple yet have a
reasonable basis in statistical estimation theory. The Instrumental Variable EquationError Method represents one attempt to satisfy these requirements.
Summary--The problem of identifying a dynamic process
from its normal operating data has received considerable
attention in recent years. The various techniques developed
range from largely deterministic procedures to sophisticated
statistical methods based on the results of optimal estimation
theory. The instrumental variable (IV) technique outlined
in this paper is intended as a compromise between these two
extremes; it has a basis in classical statistical estimation
theory, but does not require a priori information on the
signal and noise statistics.
The paper describes an IV approach to the problem of
identifying a linear process described by a differential
equation model and outlines the development of a simple
digital recursive estimation algorithm. It also discusses
briefly how the choice of input signal and the form of the
mathematical model can affect the identifiability of a process.
Finally, a number of representative experimental results are
included both to demonstrate the practical feasibility of this
particular approach to process identification, and to show
that it can be used to estimate either time invariant or
slowly variable process parameters.
these two extremes; it has a basis in classical statistical estimation theory, but does not require a priori
information on the signal and noise statistics.
The IV a p p r o a c h to least squares parameter
estimation has its foundations in classical statistical
estimation theory, where it represents one a p p r o a c h
to the problem of estimating structural model parameters [4, 5]. Probably the first application o f the
IV m e t h o d in the field o f process identification
was by P. JOSEPH, J. LEWIS and J. Tot; [6]. Although
they did not refer to it by name, Joseph et al.
suggested an IV procedure for identifying the
parameters of a process described by a difference
equation model. Subsequently, R. E. ANDEEN and
P. P. SHIPLEY [7] proposed an improved version of
the Joseph et al. method and applied it to the
theoretical design o f a digital adaptive flight control
system for an aero-space vehicle. Shortly afterwards
V. S. LEVADI [8] described the mechanisation o f a
purely analog IV method for identifying a linear
dynamic process described by a differential equation
model.
The identification scheme described in the present
paper is mechanised in a hybrid (analog-digital)
form. Here, the measured signals obtained f r o m
an u n k n o w n single-input, single-output process are
passed through analog "state variable filters"* and
then sampled to provide the input data to a simple
digital estimation algorithm. This special recursive
least squares algorithm was first suggested in an
unpublished research note [10]. The overall
estimation procedure can be considered as a
INTRODUCTION
THE PROBLEMo f identifying a dynamic process f r o m
its normal operating data has received considerable
attention in recent years [1]. The various techniques
developed range from largely deterministic procedures [2] to elegant statistical methods based on
the results o f optimal estimation theory [3]. The
instrumental variable (IV) technique outlined in
this paper is intended as a compromise between
272
P . C . YOUNG
d"x
~t
dmu
n~=oan-~:U"~m~=lbm'~
(1)
M
anXn=UO-k Z bmum
n=O
m=l
(2)
(3)
where
x =Exoxl..,
uM]
. . . --bM].
H=[100...0]
x~a=uo
(4)
(5)
v = Uo+ 11.
(6)
Here, H is a j x ( M + N + I )
observation matrix
relating the observed ./ vector,
Y=[YoYJ-..35"-1] T, with the augmented state
(7)
(j<M+N+I)
Di(s) = P~(s)/Qi(s)
(8)
- I
273
PROCESS
-I
M+N+
= Uo
V_. = H x A +
I
i
V = Uo+
(a)
I
U0
_]
STATE
VARIABLE
FILTERS
PROCESS
I
=
VARIABLE
FILTERS
z = XAD + ~D
Ill
uo
--
+ no
D i Is)
I
(b)
FIG. 1. Signal topology of (a) a general single input process; (b) single-input, single-output process with
state variable filters.
a.D,{x.(t)} = a.[(x.)]o,
= bm[(Um)]D ,
(9)
a.[(x.)]o,=[(Uo)]o,+
~ bm[(U..)]o,.
m=l
(I0)
274
P . C . YOUNG
[(Xo)]D,.
[(u m)] o ,
(12)
m= [
n=O
and
zTa = w
n=O
where
zT=
[[(Yo)]o,o[(Yo)]o,
. . . [(Yo)]o,N[(v)]o,,
...
[(O]o,M]
a.[(Xo)]D,.=[(Uo)]D,o+
(13)
[(Uo)]D,.,
(11)
m=l
w=[(v)]o,o.
Note that by the principle of superposition, the
following relationships are also true [see Fig. l(b)]:
z T = xSo + ~ = [ [ ( X o ) ] o , o
+ [[(~o)],,o.
[(Uo)]o,~,]
[(n)],,,]
w = u~ + n~ = [(Uo) ]~,o + [ ( n ) ] o , o
(14)
ei=azf~- wi.
(15)
J2 = ~ [-ZT~--Wi]2.
(16)
i=1
P~-~ = C~ = Z zizY
i=1
and
k
Bk =
E
i=1
ZiWi "
(17)
ak=ak_l--ek-lz~[zrPk-lzk+
{zrt~k- 1 --Wk}
(Ia)
or
ak = a k - 1 - P ~ { z k z r a k - 1 - zkw~}.
275
(18)
where the noise-induced term E[~/9~T] is identically zero only when there is no noise on the
observed data. It can be shown [13, 15] that the
presence of this noise-induced term introduces the
asymptotic bias on the parameter estimates.
One approach to the above problem that does
not require a priori knowledge of the noise statistics
has its foundation in statistical estimation theory
where it has been termed the "instrumental variable
method" [5]. The asymptotic bias is removed by
modifying the solution given by Eq. (17) in the
following manner
a=
PkBk
(19)
where
k
i=1
and
k
B k = Y, ~tiw i .
(20)
i=1
E[ zT] = E[ xlo] # 0.
Furthermore, it is clear that
E[~xro]~E[xaoxrao] as Jt~Xao.
(21)
276
P.C. YOUNG
([la)
v_l VARIABLE ~I
STATE
?
NOISE
INPUTS
PROCESS
T
x A a = uo
FILTERS
STATE
VARIABLE
FILTERS I
w J
v
z
-
I'~
A-D
U"
45
ll^
STATE
VARIABLE
FILTERS
A
x
-
*""
DIGITAL
COMPUTER
ESTIMATION
ALOGRITHM
oo6,I VARIABLE
s,,TE ]
q
FILTERS
II
II
FIG. 2. Instrumental variable method for identifying a process within a closed l o o p - - t h e auxiliary model
approach where F and K are k n o w n control system elements.
277
~(M+N+I)/2;
d=[(M+N+2)/2;
M+N+I
M+N+I
even
odd.
278
P . C . YOUNG
1"
(22)
(z~_,
- w~}
(H~a)
Pk/k- 1 = ~ k - 1 + D
( [ I lb)
z~P~/k-1.
( I i lc)
(23)
(24)
(25)
279
d0
0-955
1"055
0-964
1-104
dl
1.056
2-038
1"091
1-800
d2
0-677
1"600
0"748
1"212
d3
1"069
2"233
1-099
1'515
1.400
1.000
1.500
280
P.C. YOUN(;
[(%)1 c ~
l+s
l+s I
STATE V A R I A B L E FILTERS I
-I MOOELI
uo'
'
'NP~ I,
SIGNAL
~-I'q-I'+'iT-I'+'IT-I"II
I
GENERATOR J
---t
'
"x"'~' I-Zl_
'
{(xo)lL~
--
PAPER
I [
? <~
OFF-LINE
DIGITAL
COMPUTER
i-~ TAPE I
OUTPUT
--
PAPER
TAPE
PUNCH
PUNCH
ENCODER
DIE- T A L
V( LT
ME FER
FIG. 3. Equipment used in the experimental investigations. The method of multiple filters provides the
following type of relationship:
j!(n--j+l)[(yo)]L~-i...I"
0'5
8= ,897
o.
FO
(a)
-1,0
50
100
150
TIME, SEC
-0"I
50
100
150
TIME, SEC
FIG. 4. Typical signals used in the estimation experiments. Process output for: (a) pseudo random input
281
STEP RESPONSE
+ 1"0
2-0
~
ACTUAL
ESTIMATED
f"----
-~'0
10
1,0
o
15
10
TIME, sec
15
TIME, sec
+ 1"0
2.0
1.0
~3
o
<
r.-J
(3
5
z
1.0
10
0
15
10
TIME, sec
+I'O
2,0
1-0
co
-x0
0
10
TIME, sec
FIG. 5.
15
TIME, sec
0
15
....
I
10
15
TIME, sec
Comparison of actual and estimated time responses obtained when there is partial linear
dependence between the elements composing the augmented state vector.
282
P. C, YOUNG
2o[
0
0
--20
z
0
Q
o
--40
6--269
6 = .538
~-.3.59
~2 6 - .s97
ACTUAL
O C} ~ ~
40
ESTIMATED
(.)
--80
I-0
0-!
I0'0
FREQUENCY, r~/tec
-20
-40
I
t~ 6-.3s9 I
0
<
~t[:)
t~
6 = .269
-60
0 r'l A ~
ESTIMATED
-80
(b)
-100
I I 1 Ill
0.1
1 I I 111
1.0
I I I1
10-0
FREQUENCY, rm:l/141
FIG. 6. Comparison of actual and estimated frequency response obtained when there is partial linear
dependence between the elements composing the augmented state vector.
that Fig. 7 illustrates optimum performance in
the sense that the auxiliary model parameter was
varied in accordance with the actual process parameter variation. This approach proved necessary
because equipment delays prevented full hybrid
operation. Figure 8 is a similar example in which
both ao and ax were varied simultaneously. The
283
15
0
O0
.0
0
1.0 -
O0
co
o
~u
<
n-
0'5
~_T 200
= 0.359
ACTUAL VARIATION, a 0
O
0-0
ESTIMATED VARIATION, ~ 0
50
100
150
200
SAMPLES
o o -1.5];
10-4
10
P0 =
10
D=
10
0
0
284
P.C. YOUNG
1"5
0
1,0
=:LU
t,LU
~E
C
eL.
aT
0.5
[ a0
a I
1"0 - 1,5]
[a" 0 a* 1 1 . 0 5 - 1 5 ]
-aT190
6 = o.359
=
ACTUAL VARIATION
O
ESTIMATED VARIATION
(a)
5O
100
150
200
SAMPL ES
2~0
/2
aT = [ a 0 a 1 1.0-1.5]
IV190
-
(J
~)f~.a~
( go ~1 1.os - 1.5)
.359
1"5
Lid
I'-tll
~E
O~
1.0
_2/0
OO
ACTUAL VARIATION
O
ESTIMATED VARIATION
(b)
0'5
100
150
SAMPLES
FIG. 8. Estimation results for two variable parameters,
200
rior = [ 0 0 0 - 1 . 5 ] ;
10-4
10
Po'=
10
D=
10-4
0
10
0
285
1"5
oooo
1.0
d"
</</,
6 - 0-369
[
<
a-
\
0'
o to I
~ ~12J ESTIMATEDVAR|ATION
~.zx
aa~t:baaaaaaaoaaa#aaaa:a~..~a~
I
50
I
100
I
150
SAMPLES
FIG. 9. Estimation results for two variable parameters, a0 and at; time invariant, approximate auxiliary model:
~or = [0 0 0 - 0-2"] ;
10-4
10
P0 =
10
D=
10
10-*
0
REFERENCES
[1] P. C. YOUNG: Process parameter estimation. Control
and Automation Progress 12, 931-937 (1968).
[2] P. M . LION: Rapid identification of linear and nonlinear systems. Proc. J A C C 605-615 (1966).
[3] R. E. KoPP and R. J. ORtORD: Linear regression
applied to system identification for adaptive control
systems. A I A A Journal 1, 2300-2306 (1963).
286
P . C . YOUNG
im Echt-Zeit-Verfahren dutch Auswertung von Betriebsdaten wurde in den letzten Jahren starke Aufmerksamkeit
geschenkt. Die verschiedenen cntwickeltcn Techniken
reichcn von deterministischen Prozeduren bis zu statistischen
Methodcn, die auf den Ergebnissen der Theorie der optimalen Sch/itzung beruhen.
Die Benutzung von Hilfsvariablen soll einen Kompromig
zwischcn den beiden Extrema darstcllcn. Sic griindet sich
auf die klassische statistische Sch~itzungstheorie, ohne eine a
priori-Information tiber das Signal und das Rauschen zu
benStigcn.
Die Arbeit beschreibt die Ausarbeitung eines einfachcn
durch tin Differcntialgleichungsmodell beschriebcnen digitalen rekursivcn Sch/itzungsalgorithmus zur LSsung des
Problems. Es wird diskuticrt, wie die Wahl des Eingangssignals und die Form des mathematischen Modells die
ldentifizierbarkeit eines Prozesses beeinflussen k6nnen.
SehlieBlich werden repr/isentative experimentelle Ergebnisse
angeftihrt, um die praktische Durchf'tihrbarkeit dieses
slaeziellen Verfahrens zur Kennwertermittlung zu veranschaulichen und seine Anwendung zur Sch~itzung von
zeitinvarianten oder langsam veriinderlichen ProzeBparametern zu zeigen.
Pe3mMe---Hpo6neMa OHO3HaBaHH~JIHHaMR~ecKoro 06~KTa
HCXO~IH3 e r o HHOOpMaIIHHB HOpManbHOt~pa60Te BI~3BaJIa
Maoro s m ~ a m ~
3a
nocne~n~M ro~.
Pa3Jm~tI~Ie
287
o r o pcKypch'BHOrO ~ 0 p p o B o r o a n r o p n T M a oI~eHI~. O H a
T a K ~ e x p a T K O o6cy~]laeT BHHKI-IH BI~16opa BblXO~HOrO
c a r H a n a H ~ O p M M MaTCMaTh~/CCKOfi MO~CJIH Ha CHOCOSHOCTb
K OHO3HaBaHHIO 0 6 I , CKTa. O H a BI~IOtlaCT, HaKOHCI.[ HCKOT o p o c ~.~ICJ-IOxapaKTepHbIX 3KCrlepHMeHTaJII~HbIX p e 3 y ~ TaTOB, C ~BOI~HOI~ I~CJIbIO IlOXa3aTb BO3MO~KHOCTb npaKTHICCKOrO ocylI~eCTBHCHH~I 3TOFO n o n x o ~ a K OH~)3HaBaHHIO
O~KTOB
H rioKa3aTb ~ITO O H MO)KCT ~I~ITb HCHOJIb3OBaH ~21g
OI~eHKH, JIH~O HCH3MeHHMX HO BpCMCHH~ n i l 6 0 Me~.rlCHHO H3MeHHIOmHXC~I IIapaMCTpOB O~bCKTa.