Computers and Chemical Engineering 31 (2006) 66–77

Robustness analysis and synthesis of a multi-PID controller

based on an uncertain multimodel representation
R. Toscano ∗ , P. Lyonnet
Laboratoire de Tribologie et de Dynamique des Systèmes, CNRS UMR5513, ECL/ENISE, 58 rue Jean Parot, 42023 Saint-Etienne Cedex 2, France
Received 1 February 2005; received in revised form 2 February 2006; accepted 8 May 2006
Available online 27 June 2006

Abstract
This paper presents an effective method for robustness analysis and synthesis of a multi-PID controller for non-linear systems where desirable
robustness and performance properties must be maintained across a large range of operating conditions. The robustness analysis problem is solved
using an uncertain multimodel of the original non-linear system. The model of uncertainties used is an interval matrix modeled by a stochastic
matrix which gives poor conservatism in the analysis of stability robustness. Moreover, the robust stability margin is interpreted as a smallest interval
matrix that causes instability. This stability margin is evaluated using a random search algorithm. Simulation studies are used to demonstrate the
effectiveness of the proposed method.
© 2006 Elsevier Ltd. All rights reserved.

Keywords: PID controller; Multimodel; Multi-PID; Parametric uncertainties; Robust stability; Random search algorithm

1. Introduction

The proportional-integral-derivative (PID) controller is the industrial standard for the control process. The popularity of the PID
controller can be attributed partly to their performance which is satisfactory in many applications and partly to their functional
simplicity, which allows engineers to operate them in a simple and straightforward manner. For these reasons, the majority of the
controllers used in industry are of PI or PID type.
Most of real plant operate in a wide range of operating conditions, the robustness is then an important feature of the closed-loop
system. When this is the case, the controller has to be able to stabilize the system for all operating conditions. In this way, numerous
progress has been made to improve the performances of the PI/PID control (Hang, Astrom, & Wang, 2002). In particular, tuning
methods based on optimization approach have recently received more attention in the literature, with the aim of ensuring good
stability robustness of the controlled system (Ge, Cliiu, & Wang, 2002; Hwang & Hsiao, 2002). However these new methods are not
very effective in the case of a strongly non-linear system evolving on a large range of operating condition. Indeed, it is well known
that a controller designed around a specific operating point may not be able to accommodate large variations in process dynamics.
This is due mainly by the presence of system non-linearities, which cause different dynamic behaviors from an operating point to
another. This kind of problems can be solved by linearizing the system equations around several operating points and designing
a linear controller for each region of operation. The resulting controllers are then combined by interpolation in order to obtain an
appropriate controller for the original non-linear system, this is the well known gain scheduling approach (Rugh & Shamma, 2000).
This procedure is time consuming and expensive, but is well accepted and gives satisfactory results in many applications. Another
similar approach facilitates the controllers interpolation through the use of validity or membership functions. By this method local
controllers are selected as a function of the current state of the process (Foss, Johasen, & Sorensen, 1995; Howitt & Luss, 1993;
Toscano, 2000). These approaches leads naturally to the approximation of a non-linear system by a family of linear systems correctly

∗ Corresponding author. Tel.: +33 477 43 84 84; fax: +33 477 43 84 99.
E-mail address: toscano@enise.fr (R. Toscano).

0098-1354/$ – see front matter © 2006 Elsevier Ltd. All rights reserved.
doi:10.1016/j.compchemeng.2006.05.002
 R. Toscano, P. Lyonnet / Computers and Chemical Engineering 31 (2006) 66–77 67

Nomenclature
A(i, j) element in row i and column j of the matrix A
Ai , Bi , C state matrices of the local linear model on the domain D
Āi , B̄i upper bounds of the matrices Ai and Bi , respectively
A- i, B- i lower bounds of the matrices Ai and Bi ,0respectively
A0i , Bi0 medium matrices, A0i = 21 (A - i + Āi ), Bi = 2 (B- i + B̄i )
1

Ai , Bi deviation matrices, Ai = 2 (Āi + A

1 1 1 1
- i ), Bi = 2 (B̄i + B- i )
1 1

Am gain margin and phase crossover frequency

D operating domain of the non-linear system
Di sub-operating domain number i of D
f(·) state function of the non-linear system
Gi (s) transfer function of the nominal local model number i
h(·) output function of the non-linear system
i, j, k, q indices
ki , kd , kp tuning parameters of the single PID controller
k1 , k2 , k3 tuning parameters of the local PID controller number i
l number of local models
M1 ⊗ M2 product element-by-element of the matrices M1 and M2
M1 ≤e M2 inequality element-by-element of the matrices M1 and M2
n order of the model
Pr{·} probability of the event {·}
r reference input
Re{·}, Im{·} real and imaginary part, respectively, of the complex number {·}
Ri (s) transfer function of the local PID controller number i
s Laplace variable
u input variable of the system
vi (·) interpolation function of the multi-PID controller
V(·), Vj (·) Lyapunov functions
wi (·) interpolation function of the multimodel
x, y state vector and output of the system, respectively
T
xa augmented state vector xa = [ xT ξ T ]
(·)T transpose of the vector or the matrix (·)

Greek symbols
α a positive real scalar
αmax robust stability margin
αi (ω), β(ω) real and imaginary part of Gi (jω), respectively
δ probability of non-detection of an unstable system
 probability of instability
ε difference between the reference input and the output, ε = r − y
η number of iterations necessary to detect an unstable system
λ[·] set of eigenvalues of the matrix [·]
λmax [·] largest eigenvalue of the symmetric matrix [·]
μi (·) validity function of the local model number i on the domain D
T
ξ state vector of the PID controller ξ = [ ξ1 ξ2 ]
Ξ A (t),Ξ B (t) stochastic matrices modeling the uncertainties on matrices Ai and Bi
(i) (i)
ΞA , ΞB the sample, number i, of the stochastic matrices Ξ A (t), and Ξ B (t)
τd time constant of the filtered derivative action of the PID
ω frequency (rad/s)
ωπ phase crossover frequency
68 R. Toscano, P. Lyonnet / Computers and Chemical Engineering 31 (2006) 66–77

combined between them. This concept is not new and was first developed in an elegant manner by Takagi and Sugeno within the
framework of the fuzzy set theory (Takagi & Sugeno, 1985). Takagi–Sugeno fuzzy models are non-linear systems described by a set
of if-then rules which gives local linear representation of an underlying system. Practically, such multi-linear modeling (multimodel)
technique can be used to extend the well known linear controller design tools to complex non-linear systems (Palm & Rehfuess,
1997; Tanaka & Sugeno, 1992).
The main objective of this paper is to provide a simple and practical method for the evaluation of the stability robustness of a given
PID controller in the context of parametric uncertainties. Contrary to existing solutions, the proposed method does not require the
resolution of LMIs or BMIs (Feron, Boyd, El Ghaoui, & Balakrishnan, 1994; VanAntwerp & Braatz, 2000). This is more interesting
from a practical point of view, because in the context of robust control design, the LMI solution generally requires to simultaneously
solve a number of convex inequalities which is exponential according to the number of parameters. Thus the LMI approach is
computationally critical for a large number of uncertain parameters. The same is true for BMIs but in addition no efficient algorithm
exists to solve BMIs.
In this work we propose to solve the robustness analysis problem by using a random search approach. To this end, a multimodel
representation is used which is able to represent a given non-linear process. The proposed multimodel takes into account parametric
uncertainties, which result of the process linearisation or identification around a family of operating point, by the use of a stochastic
matrix. On the basis of this uncertain multimodel, a random search algorithm is proposed to find an estimate of the largest parametric
uncertainties before instability. Using this algorithm, a practical design method of a multi-PID controller is proposed which allows
us to evaluate the robustness of the closed-loop system.
The paper is organized as follows. Section 2 presents the construction of the uncertain multimodel in order to obtain on the
operating range a correct representation of the behavior of the original non-linear system. Section 3 is devoted to the design of single
PID controller that is able to stabilize the uncertain multimodel. Sufficient conditions for the robust asymptotic stability are given,
which can be used to determine, for a given PID controller, the largest parametric uncertainties before instability (or equivalently the
smallest parametric uncertainty that causes instability). In Section 4 the single PID controller is generalized to a multi-PID controller
and a design methodology is presented. Simulation studies are conducted in Section 5, finally Section 6 concludes this paper.

2. The multimodel used for the synthesis

Consider the class of non-linear single-input single-output (SISO) plants expressed in the following form:

ż = f (z, u), y = h(z) (1)

where z ∈ Rnz denote the state vector, u ∈ R is the control input, y ∈ R is the measured output, and f and h are smooth functions on Rnz
and R, respectively. The design of an output feedback controller that stabilize the non-linear system (1) remains relatively difficult.
However, it is well known that this system can be correctly represented by an appropriate combination of linear local models.
Let D the desired operating domain of the underlying system. This domain can be divided into l local domains Di where the
system (1) can be represented by a local linear model.
Assumption 2.1. On each local domain Di , system (1) can be described by the following local linear state space representation:

ẋ(t) = Ai x(t) + Bi u(t), y(t) = Cx(t) (2)

where x ∈ Rn denote the state vector, u ∈ R is the control input, y ∈ R is the measured output, Ai , Bi and C are constant matrices
with appropriate dimensions. Note that it is always possible to put a given local model (strictly proper) in the form (2).
This local model can be obtained by an appropriate identification of the system around an operating point y0i ∈ Di possibly by a
local first-order plus dead time model or a second-order plus dead time model (see Remark 2.1). In the case where de non-linear
process model (1) is known, this local representation can be obtained by linearisation via first-order Taylor series expansion of the
non-linear functions f and h.
Remark 2.1. It is well known that a very large class of industrial process can be represented, around a given operating point, by a first-
order plus dead time model G(s) = k e−t0 s /(1 + τs), or a second-order plus dead time model G(s) = kω02 e−t0 s /(s2 + 2ζω0 s + ω02 ).
Using the approximation e−t0 s ≈ 1/(1 + (t0 /α)s)β , where the constant β is choozing in order to obtain a good accuracy of the time
delay, the transfer function of the process model is then given by G(s) = k/(sn + · · · + a1 s + a0 ), with n = β + 1 (first-order case), or
n = β + 2 (second-order case). It can be verified that (2) is a state space realisation of this transfer function.
However, the local model (2) is valid only around the operating point y0i ∈ Di , and is called the nominal local linear model of the
domain Di . For y = y0i and y ∈ Di , the corresponding dynamic behavior of the process can be very different of that of the nominal
local linear model. In fact, the system matrices (Ai , Bi ) varies on the domain Di . These variations can be seen as a parameters
uncertainties. In order to take into account these uncertainties, it is necessary to consider an uncertain linear local model.
 R. Toscano, P. Lyonnet / Computers and Chemical Engineering 31 (2006) 66–77 69

Assumption 2.2. On the domain Di , the system matrices (Ai , Bi ) verifies the following inequalities:

- i ≤e Ai ≤e Āi ,
A B- i ≤e Bi ≤e B̄i (3)

- i , Āi , B
where the matrices A - i and B̄i are known bounds of the nominal matrices Ai and Bi , respectively.
Using this assumption, the matrices Ai and Bi can be rewritten as follows:

Ai = A0i + ΞA (t) ⊗ A1i , Bi = Bi0 + ΞB (t) ⊗ Bi1 with |ΞA (k, q)(t)| ≤ 1, |ΞB (k, q)(t)| ≤ 1, A0i = 21 (A
- i + Āi ), Ai
1

= 21 (Āi − A
- i ), Bi = 2 (B- i + B̄i ), Bi = 2 (B̄i − B- i )
0 1 1 1
(4)

Remark 2.2. With this new formulation of uncertainties, the matrices Ai and Bi are not quelconque, for instance, each element
of Ai is such that Ai (k, q) ∈ [A
- i (k, q), Āi (k, q)]. This representation is then more realistic compared to bounded uncertainties which
lead to very conservative results. Indeed, bounded uncertainties include matrices which never appear in the evolution of the real
system on the domain Di .
The uncertain linear local model on the domain Di is then written as follows:

ẋ(t) = (A0i + ΞA (t) ⊗ A1i )x(t) + (Bi0 + ΞB (t) ⊗ Bi1 )u(t), y(t) = Cx(t), y ∈ Di , i = 1, . . . , l (5)

We have thus a family of l local models allowing to represent validly the behavior of the system on each local areas. If y ∈ Di
then the local model number i describe correctly the system. This idea can be formalized in the following manner, let μi (y) > 0 a
function allowing to indicate the validity of the local model number i on the domain Di :

μi (y) : D = ∪i Di → [0, 1] (6)

such that μi (y) ≈ 1 for y ∈ Di and decreasing rapidly to zero beyond Di . The non-linear process model (1) can then be approximated
by interpolation of the linear uncertain local models (5):
l

ẋ(t) = wi (y)[(A0i + ΞA (t) ⊗ A1i )x(t) + (Bi0 + ΞB (t) ⊗ Bi1 )u(t)],
i=1
μi (y)
y(t) = Cx(t), y ∈ D = ∪i Di , wi (y) = l (7)
i=1 μi (y)

where wi (y) is the interpolation function which connect smoothly the local models together in order to form, on the domain D a
global model of the non-linear system (1).
Example 2.1. Consider the model of a stirred tank reactor:
q q Hk0 ρc Cpc
ĊA = (CAf − CA ) − k0 CA e−E/RT , Ṫ = (Tf − T ) − CA e−E/RT + qc (1 − e−hA /ρc Cpc qc )(Tcf − T ) (8)
V V ρCp ρCp V
whose variables, parameters and nominal values are the same as defined in Ge et al. (2002) and reproduced below.

Parameter Notation Value

Process flow rate q 100 l/min

Feed concentration CAf 1 mol/l
Feed temperature Tf 350 K
Coolant inlet temperature Tcf 350 K
Reactor volume V 100 l
Heat transfer coefficient hA 7 × 105 cal/min/K
Reaction rate constant k0 7.2 × 1010 min−1
Activation energy term E/R 1 × 104 K
Heat of reaction H −2 × 108 cal/mol
Liquid densities ρ, ρc 1 × 103 g/l
Specific heat Cp , Cpc 1 cal/g/K

In this example, CA is the measured output, qc is the control variable and CAf is the disturbance. Consider the operating range
defined as D = {(CA , T, qc ) : CA ∈ [0.06, 0.13]}, for the operating points CA
1 = 0.06, C 2 = 0.1 and C 3 = 0.13, the corresponding
A A
70 R. Toscano, P. Lyonnet / Computers and Chemical Engineering 31 (2006) 66–77

Fig. 1. Open-loop responses to successive step changes in the flow rate.

nominal local linear models are

     
−16.67 −0.047 −10 −0.047 −7.69 −0.046
A1 = , A2 = , A3 = , B1T = [ 0 −0.99 ],
3133.33 7.42 1800 7.33 1338.46 7.19

B2T = [ 0 −0.88 ], B3T = [ 0 −0.82 ], T 1 = 449.47, qc1 = 89.03, T 2 = 438.54, qc2 = 103.41,
T 3 = 432.92, qc3 = 110.03

For gaussian validity functions, the nominal multimodel is given by:

     
ĊA (t) 3
CA (t) − CA i
i
= wi (CA ) Ai + Bi (qc (t) − qc ) ,
Ṫ (t) i=1
T (t) − T i
  
μi (CA ) 1 CA − CA i 2
wi (CA ) = 3 , μi = exp −
j=1 μj (CA )
2 σi

where the parameters σ i are chosen in order to cover the totality of the operating range (in this example σ i = 0.01, i = 1, 3). Comparison
results are shown in Fig. 1 for successive step changes in the flow rate that varies between qc = 89.03l min−1 and qc = 110.031 min−1 .
One can see that in the whole operating range CA ∈ [ 0.06 0.13 ], the multimodel is a good approximation of the non-linear system
but, of course, not represent exactly the real system. If in this nominal model is incorporated the proposed model of uncertainties,
one defines in fact a set of model which include the evolution of real system. Thus the stabilization of the uncertain multimodel
implies the stabilization of the real system. This aspect is studied in the next section.

3. Robust stabilization of the uncertain multimodel

Consider a nth-order multimodel with parametric uncertainties, which is described by the following state space equation:
l
 μi (y)
ẋ(t) = wi (y)[(A0i + ΞA (t) ⊗ A1i )x(t) + (Bi0 + ΞB (t) ⊗ Bi1 )u(t), y(t) = Cx(t), wi (y) = l (9)
i=1 i=1 μi (y)

The objective is to design a PID controller for robust stabilization of (9), the control law is in the following standard form:

t
1 1 kd
ξ̇1 (t) = − ξ1 (t) + ε(t), u(t) = ki ε(τ) dτ + (ε(t) − ξ1 (t)) + kp ε(t) (10)
τd τd 0 τd
where ε = r − y is the error, r the reference input and y the measured output. Note that this representation in the time domain,
corresponds to the following transfer function u(s)/ε(s) = kp + ki /s + kd s/(1 + τd s), which is proper, this PID is then physically
realisable. Let ξ̇2 (t) = r(t) − y(t), the control input is then written as follows:
1 1
ξ̇1 (t) = − ξ1 (t) + ε(t), ξ̇2 (t) = r(t) − Cz(t), u(t) = Kxa (t) + Kr r(t) (11)
τd τd
 R. Toscano, P. Lyonnet / Computers and Chemical Engineering 31 (2006) 66–77 71
 
with K = −(kd /τd + kp )C −kd /τd ki , Kr = (kd /τd + kp ), and xaT = [ xT ξ1 ξ2 ]. The closed-loop system of (9) and (10)
is then
l

ẋa (t) = wi (y){[Aic + Ãic + (Bci + B̃ci )K]xa (t) + Bri r(t)}, y(t) = [ C 01×2 ]xa (t) (12)
i=1

with

(13)

The main result for the asymptotic stability of the closed-loop uncertain multimodel (12), is summarized in the following theorem.
 
Theorem 3.1. If there exist a symmetric and positive definite matrix P, a matrix K = −(kd /τd + kp )C −kd /τd ki and a real
number α such that:

(14)

for i = 1, . . ., l, then the origin of the closed-loop system (12) is an asymptotically stable equilibrium point, for all parametric
uncertainties satisfying:

A0i − αA1i ≤e Ai ≤e A0i + αA1i , Bi0 − αBi1 ≤e Bi ≤e Bi0 + αBi1 , i = 1, . . . , l (15)

Proof. For the simplicity of notation, define Q̃i = Aic + Bci K + Δi . Consider the Lyapunov function candidate V (xa ) = xaT Pxa ,
where P is a time invariant, symmetric and positive definite matrix. By substituting (12) with r = 0, into the time derivative of V(xa ),
V̇ (xa ) = ẋaT Pxa + xaT P ẋa , we obtain
l T l  l 
  
V̇ (xa ) = xaT wi (y)Q̃i Pxa + xaT P wi (y)Q̃i xa , V̇ (xa ) = xaT wi (y)(Q̃Ti P + P Q̃i ) xa ,
i=1 i=1 i=1
l 

V̇ (xa ) = xaT wi (y)(Qi + ΔTi P + PΔi ) xa
i=1

Let λmax [·] the largest eigenvalue of the symmetric matrix [·], if

λmax [Qi + ΔTi P + PΔi ] < 0

for i = 1,. . ., l and ΞA (k, q)(t), ΞB (k, q)(t) ∈ [−α, α], then V̇ (xa ) < 0. The uncertain closed-loop multimodel is then asymptotically
stable on the domain D = ∪j Dj for all uncertainties satisfying (15).
Theorem 3.1 cannot be used to compute directly the matrix gain K, however the following design procedure can be adopted for
this task.

1. Compute the matrix gain K for the worst-case local model. For this, one can use, for example, the method described in Toscano
(2005).
2. Verifies that the following inequality maxi Re{λ[(Aic + Bci )K]} < 0 is satisfied.
3. Search the largest number α = αmax such that

max Re{λ[Aic + Ãic + (Bci + B̃ci )K]} < 0 (16)

i,α

For α ≤ αmax , the PID controller stabilize the multimodel for all uncertainties satisfying (15).
72 R. Toscano, P. Lyonnet / Computers and Chemical Engineering 31 (2006) 66–77

3.1. Robustness analysis

It is now necessary to give some developments concerning the step three of the procedure given above. Note that the value
αmax can be seen as a robust stability margin, that is the smallest symmetric uncertainty that causes instability. Indeed, consider a
sufficiently small α > 0 such that the closed-loop system is stable. Next increase α until αmax so that the closed-loop system becomes
unstable. Thus αmax is the robust stability margin. The determination of this stability margin is not an easy task because this problem
is related to the stability of an interval matrix which is an NP-hard problem (Vidyasagar & Blondel, 2001). An interesting approach
to solve this kind of problem is to use a random search algorithm. For this, consider, for given matrices P, K and a number α the
following function:

1 if max Re{λ[Aic + Ãic + (Bci + B̃ci )K]} < 0
Ψ (ΞA (t), ΞB (t)) = i (17)
0 otherwise

the main result to check the robust stability of the closed-loop uncertain multimodel (12), is summarized in the following theorem.

Theorem 3.2. If the algorithm

1. Choose a number of iterations η, and a size of uncertainties α.

2. Generate the matrices ΞA and ΞB , with random uniformly distributed elements on the interval [−α, α].
3. If Ψ (ΞA , ΞB ) = 0 then stop.
4. If the number of iterations is incomplete go to step 2, otherwise stop.

is stopped without to have Ψ (ΞA , ΞB ) = 0 after a number of iterations η such that η ≥ ln(δ)/ln(1 − ε), then, with a confidence
1 − δ, the closed-loop system can be declared stable for all system parameters satisfying (15), except possibly for those belonging
to a set of measure no larger than ε.

Proof. For a given probability of instability  = Pr{ψ = 0}, which is considered as an improbable and indesirable event, we want to
determine the number of iterations in order to detect the apparition of this indesirable event with a given probability 1 − δ. Consider
a total number of η experiments (or iterations),
η the probability so that we have one closed-loop system instable after successive
closed-loop systems stable is given by ε + k=2 ε(1 − ε)k−1 . We want to determine the number of iterations in order to detect
the apparition
η of an instable closed-loop system with a probability at least 1 − δ. The problem is then to determine η such that
ε + k=2 ε(1 − ε)k−1 ≥ 1 − δ, which gives:

ln(δ)
η≥ (18)
ln(1 − ε)

Finally, if the probability of instability is “known” to be , then the algorithm will detect at least an unstable instance within
η ≥ ln(δ)/ln(1 − ), with a probability larger than 1 − δ. Now assume that the algorithm runs up to all η iterations without detecting
instability. This means only that true probability is less than or equal to , but not necessarily that the system is stable in all cases.
Indeed, let  be the true but unknown probability of instability. For η iterations, we have not detected unstable system, this imply
that the true probability of detection of an unstable system after successive stable systems, is such that:
η
 η

   K−1
ε + ε (1 − ε ) ≤ε+ ε(1 − ε)K−1
K=2 K=2

which imply that  ≤ . Note that, for a probability of instability , we have:

 η


k−1
1− ε+ ε(1 − ε) ≤δ
k=2

which means that the largest probability of non-detection of an unstable system is δ, consequently the smallest probability of detection
of stable systems is 1 − δ. Finally, if after η ≥ ln(δ)/ln(1 − ), iterations, only stable plants are generated, it can be asserted, with a
confidence 1 − δ, that the system is stable for all system parameters satisfying (15), except possibly for those belonging to a set of
measure no larger than .
Based on this result, an estimate α̂max of αmax , can be determined using the following random search algorithm.
 R. Toscano, P. Lyonnet / Computers and Chemical Engineering 31 (2006) 66–77 73

Algorithm 3.1.

1. For a given  ∈ (0,1) and δ ∈ (0,1), choose a number of iterations η such that η ≥ ln(δ)/ln(1 − ), a lower bound αI = αinf , an upper
bound αS = αsup such that αinf ≤ αmax ≤ αsup , and a precision ρ.
2. Compute α = αI +α 2
S

(1) (1) (1) (1)

3. Generate η i.i.d. samples (ΞA , ΞB ), . . . , (ΞA , ΞB ), with random uniformly distributed elements on the interval [−α, α].
η (1) (1)
4. If i=1 Ψ (ΞA , ΞB ) < η then αS = α go to step 2, otherwise αI = α.
5. If αS − αI > 2ραI go to step 2, otherwise stop.

Indeed, we always have αI ≤ αmax ≤ αS . Moreover, after N iterations, αI − αS =2−N (αsup − αinf ). Thus, when the algorithm is stopped,
(αI + αS )/2 is guaranteed to approximate αmax within a relative accuracy of ρ, that is |(αI + αS )/2 − αmax |≤ραmax .
By this approach, the dynamic performance of the closed-loop system varies on each local domain Di because we have only
one controller. If we want to obtain constant performance on the global domain D = ∪i Di , it is necessary to adopt a multi-PID
controller, this is the purpose of the following section.

4. The multi-PID controller approach

Constant performance on the global domain D = ∪i Di can be obtained using the following multi-PID controller:
l 
t 
1 1  k2i 1 if y(t) ∈ Di
i i
ξ̇(t) = ξ1 (t) + ε(t), u(t) = υi (y) k1 ε(τ) dτ + (ε(t) − ξ1 (t)) + k3 ε(t) , υi (y) =
τd τd 0 τ d 0 if y(t) ∈
/ Di
i=1
(19)

where the domain Di is defined as follows:

Di = {y : μi (y) ≥ μj (y), j = 1, . . . , l, j = i}, i = 1, . . . , l (20)

with μi (y) defined as in (6). In the relation (19), ε = r − y is the error, r the reference input and y the measured output. Let ξ̇2 (t) =
r(t) − y(t), the control input is then written as follows:
l

1 1  1 if y(t) ∈ Di
ξ̇1 (t) = ξ1 (t) + ε(t), ξ̇2 (t) = r(t) − Cz (t), u(t) = υi (y)(Ki xa (t) + Kri r(t)), υi (y) =
τd τd 0 if y(t) ∈
/ Di
i=1
(21)

with Ki = [ −(k2i /τd + k3i )C −(k2i /τd ) k1i ], Kri = (k2i /τd + k3i ), and xaT = [ xT ξ1 ξ2 ]. The closed-loop system of (9) and (19)
is then
l 
 l
ẋa (t) = ωi (y)υj (y){[Aic + Ãic + (Bci + B̃ci )Kj ]xa (t) + Brij r(t)}, y(t) = [ C 01×2 ]xa (t) (22)
i=1 j=1

ij j
where Aic , Ãic , Bci and B̃ci are defined in (13), Br is the same as Bri but Kr becomes Kr (see relation (13)). The main result for the
asymptotic stability of the closed-loop uncertain multimodel (22), is summarized in the following theorem.
Theorem 4.1. If there exist a set of symmetric and positive definite matrices Pj , a set of matrices Kj =
[ −(k2j /τd + k3j )C −k2j /τd k1j ] and a set of numbers αj (with j = 1, . . ., l) such that:

(23)

for j = 1, . . ., l, then the origin of the closed-loop system (22) is an asymptotically stable equilibrium point, for all parametric
uncertainties satisfying:

A0j − αj A1j ≤e Aj ≤e A0j + αj A1j , Bj0 − αj Bj1 ≤e Bj ≤e Bj0 + αj Bj1 , j = 1, . . . , l (24)

74 R. Toscano, P. Lyonnet / Computers and Chemical Engineering 31 (2006) 66–77

Proof. Suppose that the system evolve on the domain Di , on this domain only two local models are active, the models of indice
j − 1 and j or j and j + 1, thus on the local domain Dj the closed-loop multimodel can be written as follows:

j+1

ẋa (t) = ωi (y){[Aic + Ãic + (Bci + B̃ci )Kj ]xa (t) + Brij r(t)} (25)
i=j−1

note that the interval matrix generated with this expression is larger than the real interval matrix, and thus gives more conservative
results for the stability studies. Let us recall that, in reality, when the system evolve on the domain Di the system matrices (Aj ,
Bj ) is such that A j
- j ≤e Aj ≤e Āj and B- j ≤e Bj ≤e B̄j thus the state matrix of the closed-loop system is such that A
- j + B- j K ≤e Aj +
j j
Bj K ≤e Āj + B̄j K . For a weak conservatism, it is then convenient to study the stability on the domain Dj for the system defined
as follows:

ẋa (t) = [Ajc + Ãic + (Bci + B̃ci )Kj ]xa (t) + Brij r(t) (26)

Consider the Lyapunov function candidate Vj = (xa ) = xaT Pj xa , where Pj is a time invariant, symmetric and positive definite matrix.
By substituting (26) with r = 0, into the time derivative of V̇j (xa ), V̇j (xa ) = ẋaT Pj xa + xaT Pj ẋa , we obtain

V̇j (xa ) = ẋaT [Qj + ΔTj Pj + Pj Δj ]xa

Let λmax [·] the largest eigenvalue of the symmetric matrix [·], if the following condition is satisfied:

∀ ΞA (k, q)(t), ΞB (k, q)(t) ∈ [−αj , αj ], λmax [Qj + ΔTj Pj + Pj Δj ] < 0

for j = 1, . . ., l then V̇j (x) < 0 (j = 1, . . . , l). The uncertain closed-loop multimodel is then asymptotically stable on the domain
D = ∪j Dj for all uncertainties satisfying (24).
Theorem 4.1 cannot be used to compute directly the set of feedback gains (K1 , . . ., Kl ), however the following design procedure
can be adopted.

1. For each local domain Dj , compute the matrix gain Kj for the nominal local model (see Remark 4.1).
j
2. Using Algorithm 3.1, find for each local domain Dj , the largest number αj = αmax such that

max Re{λ[Ajc + Ãjc + (Bcj + B̃cj )Kj ]} < 0 (27)

αj

j
Then, for all αj ≤ αmax , the multi-PID controller stabilize the uncertain multimodel for all uncertainties satisfying (24).

Remark 4.1. For low order local nominal models (n ≤ 2), the design of the PID controller can be easily done by a classical poles
placement. For high order nominal models (n > 3), the step 1 of the design procedure can be realised by using, for instance, the
method presented in Appendix A.

5. Simulation results

Consider the model of a stirred tank reactor given in Example 2.1. For the operating range D = {(CA , T, qc ) : CA ∈ [0.06, 0.13]},
we have, for example, the following local domains D1 = {(CA , T, qc ) : CA ∈ [0.06, 0.080]}, D2 = {(CA , T, qc ) : CA ∈ [0.08, 0.115]},
D3 = {(CA , T, qc ) : CA ∈ [0.115, 0.13]}. On the domain Di , the matrices Ai and Bi (i = 1, . . ., 3) can be written in the form Ai =
A0j + ΞA (t) ⊗ A1i and Bi = Bj0 + ΞB (t) ⊗ Bi1 , with:
       
−14.6 −0.047 0 2.1
0 0
A01= , B10= , A11= , B10= ,
2716.7 7.4 −0.96 416.7
0.015 0.032
       
−10.6 −0.047 0 1.9 0 0
A2 =
0
, B2 =
0
, A2 =
1
, B2 =
1
,
1919.6 7.3 −0.89 380.4 0.065 0.041
       
−8.2 −0.047 0 0.5 0 0
A3 =
0
, B3 =
0
, A3 =
1
, B3 =
1
1438.8 7.2 −0.83 100.3 0.035 0.005
∀ t, k, q, ΞA (k, q)(t) ∈ [−1, 1] and ΞB (k, q)(t) ∈ [−1, 1]
 R. Toscano, P. Lyonnet / Computers and Chemical Engineering 31 (2006) 66–77 75

Fig. 2. Robustness analysis and closed-loop response to successive step changes in the effluent concentration.

The controller parameters are determined by a classical poles placement in order to obtain a second-order dominant mode with
a first overshoot of 15% and a settling time of 1.5 s. The parameters are as follows:
∀ y ∈ D1 , k11 = 1642, k21 = 62.06, k31 = 303.49, τd = 0.01, ∀y ∈ D2 , k12 = 1771, k22 = 160.17,
k32 = 489.92, τd = 0.01, ∀y ∈ D3 , k13 = 1642, k23 = 62.06, k33 = 303.49, τd = 0.01
On each domain, the robust stability margin, evaluated for  = 0.005 and δ = 0.005, is α1max = 0.79, α2max = 0.85 and α3max = 3.31,
respectively. Fig. 2 shows the robustness analysis for each domain, and the closed-loop response for successive step changes in the
effluent concentration CA that varies between 0.06 and 0.14. It can be observed that in the whole operating regime, the dynamical
performances obtained are similar in each sub-domain.

6. Conclusion

In this paper an effective method for robustness analysis and synthesis of a multi-PID controller for non-linear systems was
developed via uncertain multimodel approach. Simulation studies was used to demonstrate the effectiveness of the proposed method.
The main results obtained in this paper can be easily generalized for multivariable PID controllers.

Appendix A

A.1. Robust design of a multi-PID controller for models of high degree

Let Li (s) = Ri (s)Gi (s) be the open-loop transfer function for the nominal local model number i. The transfer function of the local
PID controller Ri (s) and the transfer function of the local nominal model Gi (s) are defined as follows:
i s m + · · · + bi s + bi
bm k1i
Gi (s) = C(sI − Ai )−1 Bi = 1 0
, Ri (s) = + k1i s + k3i (28)
sn + an−1
1 + · · · + a1i s + a0i s

The consequences of the uncertainties on the parameters of the system are an uncertainty on the static gain (i.e. b0i /a0i ) and an
uncertainty on the dynamical behavior of the system (i.e. the location of its poles and zeros). The objective is then to find the local
76 R. Toscano, P. Lyonnet / Computers and Chemical Engineering 31 (2006) 66–77

PID controller parameters (k1i , k2i , k3i ), so that the closed-loop system is not too sensitive to the model uncertainty and to have
acceptable dynamical performance. The sensitivity of the closed-loop system to the uncertainty on the static gain can be reduced by
an appropriate gain margin Am . By definition on the gain margin we must have
1
Li (jωπ ) = Ri (jωπ )Gi (jωπ ) = − (29)
Am
√
where ωπ is the phase crossover frequency of the loop, and j = −1. Relation (29) can be rewritten as follows:
 
ki 1 ki
αi (ωπ )k3i − βi (ωπ ) k2i ωπ − 1 = − , βi (ωπ )k3i − αi (ωπ ) k2i ωπ − 1 = 0 (30)
ωπ Am ωπ
where αi (ω) are the real part and imaginary part of Gi (jω), respectively. The solution of (30) is given by:
αi (ωπ ) βi (ωπ )ωπ
k3i = − , k1i = k2i ωπ2 − (31)
(αi (ωπ ) + (βi (ωπ )2 )Am
2
(αi (ωπ )2 + βi (ωπ )2 )Am
Thus for a given gain margin Am and a given ω␲ , one can compute the proportional gain k3i and the relation between the integral
gain k1i and the derivative gain k2i . Now we want to find k2i so that the closed-loop system is less sensitive as possible to the uncertainty
on the dynamical behavior and to obtain a good transient response and a good rejection on the load disturbance. As shown in (Toscano,
2005) these objectives can be reached by maximising the shortest distance from the Nyquist curve of the open-loop transfer function
(i.e. Li (jω)) to the critical point −1. Thus the derivative gain k2i is determined such that:

max min|1 + Li (jω)| (32)

k2i ω

Finally, given the gain margin Am and the frequency ωπ , the determination of the controller parameters (k1i , k2i , k3i ) is formulated as
the following optimization problem:
 
i i i i i i k1i
max min|1 + L (jω)|, L (jω) = R (jω)G (jω) = k3 + j k2 − (αi (ω) + jβi (ω)),
k2i >0 ω ω
αi (ωπ ) βi (ωπ )ωπ
k3i = − , k1i = k2i ωπ2 − (33)
(αi (ωπ ) + βi (ωπ )2 )Am
2
(αi (ωπ )2 + βi (ωπ )2 )Am
which is numerically easy to solve. There is no difficulty to consider the PID controller with filtered derivative action (i.e. using
k2i s/(1 + τd s) instead of k2i s). Choosing for example τd  ωπ−1 , has a minimal effect on the shape of Li (jω) for ω  ωπ. Alternatively,
the PID can be assumed of the form Ri (s) = k1i /s + k2i s/(1 + τd s) + k3i , with τ d fixed before solving the optimization problem (33).
In this case, we have:

ki Gi (s)
Li (s) = Ri (s)Gi (s) = (k3i + k1i τd ) + 1 + (k2i + k3i τd )s (34)
s 1 + τd s
The design procedure can then be applied by redefining the transfer function of the system as Gia (s) = [1/(1 + τd s)]Gi (s), and
considering the new PID parameters:
(k1i ) = k1i , (k2i ) = k2i + k3i τd , (k3i ) = k3i + k1i τd (35)
Note that, contrary to existing PID design methods, the proposed approach does not use any simplification on the model used for
the representation of the local behavior of the considered system.

References

Feron, E., Boyd, S., El Ghaoui, L., & Balakrishnan, V. (1994). Linear matrix inequalities in system and control theory. Studies in applied mathematics. Philadelphia,
PA: SIAM.
Foss, B. A., Johasen, T. A., & Sorensen, A. V. (1995). Nonlinear predictive control using local models applied to a batch fermentation process. Control Engineering
Practice, 3, 389–396.
Ge, M., Chiu, M.-S., & Wang, Q.-G. (2002). Robust pid controller design via lmi approach. Journal of Process Control, 12, 3–13.
Hang, C. C., Astrom, K. J., & Wang, Q. G. (2002). Relay feedback auto-tuning of process controllers—a tutorial review. Journal of Process Control, 12, 143–162.
Howitt, G. D., & Luss, R. (1993). Control of a collection of linear systems by linear state feedback control. International Journal of Control, 58, 143–162.
Hwang, C., & Hsiao, C.-Y. (2002). Solution of a non-convex optimization arising in pi/pid control design. Automatica, 38(11), 143–162.
Khalil, H. K. (2002). Nonlinear system. Prentice Hall.
Palm, R., & Rehfuess, U. (1997). Fuzzy controllers as gain scheduling approximators. Fuzzy Sets and Systems, 85, 233–246.
Rugh, W. J., & Shamma, J. S. (2000). Survey paper, research on gain scheduling. Automatica, 36, 1401–1425.
 R. Toscano, P. Lyonnet / Computers and Chemical Engineering 31 (2006) 66–77 77

Takagi, T., & Sugeno, M. (1985). Fuzzy identification of system and its applications to modeling and control. IEEE Transaction on Systems, Man and Cybernetics,
1, 116–132.
Tanaka, K., & Sugeno, M. (1992). Stability analysis and design of fuzzy control systems. Fuzzu Sets and Systems, 45, 135–156.
Toscano, R. (2000). Control of a nonlinear system, by fuzzy supervision, a local approach. In Proceedings of the second conference on management and control of
production and logistic.
Toscano, R. (2005). A simple robust PI/PID controller design via numerical optimization approach. Journal of Process Control, 15(1), 81–88.
VanAntwerp, J. G., & Braatz, R. D. (2000). A tutorial on linear and bilinear matrix inequalities. Journal of Process Control, 10, 363–385.
Vidyasagar, M., & Blondel, V. D. (2001). Probabilistic solutions to some np-hard matrix problems. Automatica, 37, 1397–1405.