An Adaptive Nonlinear PID Controller

J.P. Segovia and D. Sbarbaro

P.O.Box 53-C. Department of Electrical Engineering, University of Conception
Concepcion, Chile.

Abstracf - This paper presents a nonlinear PCD controller which 2. THE GAIN SCHEDULING FRAMEWORK
combiiaes an adaptive algorithm to cope with the problem of tuning
the PID controller, and a assoCiative menlory to store the We consider a plant described by
parametas of the ConuoUer according to different operating regions,
which are defined by some exogenous variables The special
characteristic of associative memories and the properlies of their
learning algorithms are &fly exploited To illustrate the performance
of the controller several examples are given
where X ( I ) E R",u(r) E R,y(r) ER, and

1 INTRODUCTION f : R " x R-+ R", f ( O , O ) = O

gR" x R + R , g(O,O)=O
As a consequence of the nonlinearities of a real system, its dynamical
behavior and the parameters of the Linearized model change with the both continuously differentiable. We assume that ( I ) has a
opesating conditions. Thus, a linear adaptive controller will adapt its parametrized family of mnstant operating points [4]
parameters every time the process changes operating points, but this
adaptive process takes some time to converge and in some situations
can lead to an undesirable transient behavior. On the other hand, a
gain scheduling controller can cope better with a nonlinear process,
h s e it changes its parameters according to the operating where the parameter set t' c K is an open neighborhood of a fixed
conditions, which are specify by the values of one or more exogenous a, E R.Typically, the constant operating point family is parametrized
Variables. However, the design of the program of the scheduling by constant values of the input, the output or ones of the underlying
variable requires a deep knowledge of the controlled process state of the system That is, for a E r
To address these problems, the nonlinear PID controller presented in
this paper combines two ideas, i.e. an adaptive algorithm to cope
with the problem of tuning, and a gain scheduling implemented using
associative memories to store the parameters of the controller The linearization family corresponding to ( I ) is
according to different operating regions, which are defined by the
exogenous variables.

Associative memories are defined by a set of nonlinear base functions,

which are a suitable solution for the storing problem presented in this
case, because their simple structures lead to simple leaming
dgorithms. which are amenable to a parallel implementation

The paper is organized as follows Section 2 describes the general

h e w o r k for the gain scheduling controller, and the motivation to
use M adaptive algonthm to find the relationship between the
scheddig vanable and the parameters of the controller Section 3
describes a pattern recognition approach used to identify the main
f m r e s of the system responses Section 4 describes the structure of
the associative memory and the learning algorithm to update the Let the parametrized transfer function of the linearizing family for the
parameters of its parameters Section 5 shows the performance of the given nonlinear system be
controller trying to control a broad range of processes under a wide
range of opwational conditions Finally. in section 6 some conclusions H , =-
N(a,s)
are given d(a,s)'
and the parametrized transfer function for the PID controller

c, (s) = s ' K , ( a ) + K S, ( a ) s + K , ( a )

403
 then the parametrized transfer h c t i o n for the linearized closed loop
systcmiswritta!as
El
I I I
Ma.s)[s'K,(a) + K,(a)s+ K , ( a ) ]
T,W = sd( 9s)+ s'K,( a)+ K, ( a ) s + K,( a )

Assuming that the closed loop system is stable, it can be

demonstrated that for a slowly vllrying scheduling variable, the state
and tbe output values do not deviate fir fiom their instantaneous trim I
I
WIJW x ( a ( O ) . ~ ( u t r ) )respectively 111. Setpoint L)lsturbance Load Dtsturbance

For each operating point, the values of K , ( a ( f ) ) , K , ( a ( f ) ) , and Figure 2. Underdamped responses
K , ( a ( f ) ) mbe calculated, such that the closed loop response Figure 2 shows some typical patterns for a step change in the
satis@ a design criteria, for example a step response similar to a reference and in the disturbance. The main features for this type of
second order system. which cll~lbe specified by an overshoot and a response are detined as
damping ooefficient. As tbe system is assumed unknown, it is not
possible to set the dominant closed loop poles into the desired
position. In this paper, a pattern recognition approach is implemented
to find the relationship b e e n the parameters of the controller and
the scheduling variable a such that the desired response for the
closed loop system is obtained.
3. THE PATTERN RECOGNlTlON APPROACH
pa = I, - I ,
The adjustment of the parameters in each operating point is done by a
pattern recognition approach. Typically, with this approach, the If only one peak is detected in a given period of time, then the
closed loop is perturbed, then the patterns are observed, and the response is classified as an overdamped response. Figure 3 shows the
actual panem is then compared with the one that is desired. A responses of the overdamped system, where wmax is a maximum wait
laming algorithm enable the system to adjust the parameters of an time, which defines the maximum time that the algorithm waits for a
associative memory accordingly. The general scheme is shown in second peak. Usually this parameter is easily provided because a
Figure I . reasonable idea of the dominant process time constant can always be
Pattem Recognitionand obtained.
Scheduling module
Enor

I
Wmax Wmax
'1 Time
Figure 3. Overdamped responses

Y
1
Set point t PID Non
+Q
E ) Controler -+ process
a r I" > The damping fkctor is calculated as

&=-- e
El - e '
Figure I . Adaptive nonlinear P D controller
The pattern recognition phase begins only when the variance of the
The pattern to be recognized is the error. The distinctive features are error exceeds a nominal threshold. All the real measurement have the
the absence or presence of peak heights, and the time between peaks. potential of being noisy, because the intrinsic nature of the
The pattern classified as a member ofone of the following classes: measurement technology or the measurement location, this noise does
overdamped not contain any useful information for control. The threshold prevents
criticallydamped the algorithm to extract information from the error signal under these
underdamped situations.
oscillatory
unstable The parameters of the associative memory remains fixed during the
pattern recognition phase, but the algorithm start watching the error
The features used to define a pattern are: damping (h),
overshoot magnitude until a pattern is detected and classified.
(oY).and an estimate ofthe ultimate period (pa).

404
 4 LEARNINGALGORITHM

Associative memories can be decomposed into a fixed nonlinear

mapping and adaptive linear mapping. The fixed nonlinear h c t i o n
are local, which mean that for a particular imput, only a small region where a(/)is the scheduling variable used to address the elements in
of the associative memory will involved in the calculation of the
the memory, Q represents the basic element of the memory, and
output.The mapping is constructed as a linear combination of basis
functions, which must cover the input space. The parameters of the A I , A I ,A, represent real constants less than 1.
controller are the output of the associative memory defined as follows
If pa is maintained constant for a given a, equations (2) and (3)
converge to

which are local tuning rules, l i e the ones suggested by the Ziegler-
Nichols method [SI.
where pJ are the base functions. At the beginning, the controller start working as a conventional
The base function gives a mean of interpolation between a set of adaptive PID controller. i e., k,, = K,,k, = K , , k , = KdrVlj,but
points. In many industrial applications these hctions are linear and through out the closed loop operation of the process, the proposed
can be represented by h c t i o n s defined as controller can learn the map between the scheduling variables and the
parameters of the controller In this way, the PID controller can
exploit any stored parameter to improve the process control.

5 TESTINGTHEALGORITHM

A gain scheduling controller must function in a wide variety of

figuh 4.0 shows a set of base hndions A nice property of these process control situations. Several systems are simulated trying to
hnctions is that they satisfy the interpolation property, i e cover extremes situations.

The first system has a gain characteristic which is affected by

the input amplirude. i.e., the parametrized transfer fUnclion is

T'
k(a)
H a =-
1s+l

Where the parameters are given by the foll,. Aing expression

Parameters

4.0 sec
Figure 4.0 Basis functions 1.osec

The base hnction may be a binary fbnction, radial base hnction

(gaussians). or piecewise potynomials (B splines). Any of rhex Tt,c initial parameters of the associative memory were set to the same
functions can be included, if a different interpolation schemes is values, such that the setting of the PID controller gave an unstable
needed. Tbc modelling capability of an associative memory is specified closed loop behavior The centers of the base hnctions were selected
by the number of functions and their characteristics. (0.0. IO, 20, 30,40, 50. 60,70. 80, 90,I 0 0 ] yo

to update the parameters of the associative memory The associative PID Parameters Value
memories are particularly suited, because the output is linearly Proportional Band 600 Yo
dependent upon the parameters of the memory and only those Integrative Time sosec
paramaen which contribute to the output are modified The learning ' Derivative Time oosec
algorithm is described by the following equations Noise Rand (nb) 0 5 Yo
Wmax 25 0 sec
Da 0 25
km (I + I ) = k,, ( I )+ A , p, (a(l))k,, (OD1
ov 0 20
 Setpoint and Process Variable
1O( I -r--T- I' 1 ---I I
I

8C .. . . I..
'!
I
, .
.. .
I
... .. . ... -,I .._

. ., , I . . ..
I

20

0
0 200 400 600 800 1000 1200 I400 1600 1800 2000
time [sec]

Proportional Gain Integrative Gain

I 0 2 7 -
I -----------

0.5 I
t=O sec
7
,
/ ',
0.18-
0 18- /---

I
, i' ,
/
\ 'i t=2000 sec

&
-
I
0 16- /

2 ' / t=400 sec ' -

2 0.14
0 14- I 1.
I i
i '
I ---1--___-____
00.12
12;' '
t=O sec

0 20 40 60 80 49 60 80 100
alpha [Yo] alpha [%I
Figure 5 0 Closed loop response for a nonlinear first order process

Figure 5.0 (graph Set point / Process variable) shows, between t=O
and t-400 secs, that the initial setting of the controller produces an I Parameters I Expressions I
unstable closed loop behavior In this period, the adaptive algorithm
changes the Proportional and lntegrarke Gains to modi@ the closed &/200+ 1.0
loop response. However, at the 60% operating point the controller is 3.0sec
tuned around [=I300 sec. At the time of t =2000 sec the PI L

parameters have converged to the right setting. The Proportional 8 5.0 sec
Gain and Integrative Gain graphics show the nonlinear fbnctions Sample Time 1.0 sec
found by the adaptive dgorithm, and the fast convergence of the The initial parameters of the associative memory were set to the same
parameters values, such that the setting of the PID controller gave an unstable
closed loop behavior The centers of the base fbnctions were selected
The second system has a gain and phase characteristic which is (0.0, IO,20, 30,40, 50, 60,70, 80, 90,100 ] %.
af3kted by the input amplitude.

Ha = kfa) PlD Paramrters Value

(rl(a)s+I/(ns+I) Proportional Band I80 %
Integnltive Time 8.0 sec
Where the parameters of the transfer hnction are characterized by
Derivative Time 0.0 sec
the following expressions .
Noise Rand (nb) 0.5 %
WmaX 35.0 sec
I Da 0.25

406
 ___ -- Setpoint and Process Variable
100 - 7

,
I
A
80 i-
I

i
201 -

0.6-
0
i

t
200

I
-___ ----A--

400

Pro oftional Gain

.p I -
'
1
600 800 IO00
time [sec]
-.,-____

0.25 -
1200

-I
I400

Integrative
,
-
1600

Gain
_-_

--
I
__
1800 2000

I
I

02 -

I' / ; '.
/ /- - t=2000 sec
; I

I
i
p . 1
-Y
03r hi
\

-----
02- 1 4 0 0 sec
t=2000 sec
,
O I L _____ _ A _ __-___
0 20 40 60 80 100
alpha [%I
Figure 6.0 Closed loop response for a nonlinear second order process

Figure 6 0 (graph Set point / Process variable) shows, between t=O between the parameters and the sheduling variable can be used to
and 1-400,M unstable closed loop behavior for operating points over program flexible single loop commercial PID controllas, or it can be
Wh In this period, the adaptive algorithm changes the Proportional implemented over a distributed control systems with some
and Integrative Gains to modi@ the close loop response At t ~ 2 0 0 0 development environment .
sec the PI parameters have converged to the ideal setting The
Propottional Gain and Integrative Gain graphics show the nonlinear
functions found at different tunes by the adaptive algorithm, and the REFERENCES
fast convergence to the desired parameters
111 W J Rugh, "Analytical framework for gain scheduling", IEEE
6 CONCLUSIONS 91
i ' w t r d M u g ~ r ~ rJan
e.
[2] J P Segovia, L Pavesi, L Zambrano, R Palma, "Development of
This paper has presented a new adaptive nonlinear PID controller, a self tuning PID controller based on pattern recognition" (in
which can be used to control highly nonlinear processes. or as an Spanish), Proceedings of the XI Elecfncal Eng. Congress, Valdivia,
ctfcctive tool to tune commercial PID scheduling controllers The 1992
results obtained are promising. and hrther testing on real industrial (31 J Wang and W J R u b "Linearized Model matching for single
processes are under way The algorithm has few initial parameters to input nonlinear systems", I F X TAC. vol 33. pp 793-796,August
be set, the definition of the centers of the base hnctions can done I988
using some prior information about the nonlinearities of the process [4] J Wang and W J Rugh, "Feedback linearization families for
The use of a scheduling variable enhances the robustness and nonlinear systems", IEEE TAC, ~0132.pp 935-940,October 1987
performance of the adaptive controller, when this controller is applied [SI J G Ziegler and N B Nichols,"Optimum setting for automatic
to control a time invariant nonlinear plant. The tinal relationship controllers", Trans. ASME, VOW,pp 645-651,1942

407