LNCS SEMMCO Non Frsgile PI Controller

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 12

Non-fragile Robust PI Controller Design Using

Co-variance Matrix Adaptation Evolutionary Strategy

K. Mohaideen Abdul Kadhar and S. Baskar

Electrical & Electronics Engineering department


Thiagarajar College of Engineering
{makeee,sbeee}@tce.edu

Abstract. This paper discusses the application of Co-variance Matrix


Evolutionary Strategy (CMA-ES) in the design of non-fragile robust PI
controller. The desired maximum sensitivity of the closed loop system is
considered as an objective and success rate of stability under probabilistic
controller uncertainty is taken as a constraint for non-fragile robust PI controller
design problem. Success rate of stability is calculated using Monte Carlo
simulation (MCS) under probabilistic controller perturbation. CMA-ES finds
the optimal controller parameter based on robustness objective and non-
fragileness constraint. The Single Input Single Output (SISO) first order sugar
cane raw juice neutralization process and second order Irrigation canal systems
are considered as a test systems. The performance of the CMA-ES designed
non-fragile robust PI controller is compared with the flat phase concept based
PI controller and Astrom suggested PI controller for both test systems.
Simulation results demonstrated that CMA-ES based non-fragile robust PI
controller has better performance in robustness as well as non-fragileness.

Keywords: CMA-ES, non-fragile robust PI controller, maximum sensitivity,


probabilistic parametric perturbation.

1 Introduction

Robustness is the ability of the controller which makes the closed loop system to be
stable even under the effect of uncertainties. Many of the robust controller design
methods considered only the plant side uncertainty and doesn’t consider controller
side uncertainty. Because, it is assumed that there are no uncertain parameters
occurred in the controller side [1]. But in reality, there are some unavoidable amount
of uncertainty exists in the controller due to the dynamics of analogue controller and
the effect of quantization error in digital controllers [2, 3].
L. H. Keel and Bhattacharya proposed the importance of considering controller
uncertainty in robust controller design based on the results of various robust
controllers’ instability due to controller uncertainty. If the uncertainties in the
controller parameters make the closed loop system unstable then the controller is said
to be a fragile controller [1, 3]. Whidbhone et al [2] proposed a method for the
reduction of controller’s fragility by minimizing closed loop pole sensitivity without

B.K. Panigrahi et al. (Eds.): SEMCCO 2013, Part II, LNCS 8298, pp. 32–43, 2013.
© Springer International Publishing Switzerland 2013
Non-fragile Robust PI Controller Design Using CMA-ES 33

considering the performance of the closed loop system. Zhisheng et al designed the
non-fragile H∞ robust controller considering both controller and plant perturbations
[4]. L. H. Keel, Bhattacharya, Whidbhone and Zhisheng are analyzed the fragility
characteristic in higher order robust controllers.
Most of the industries have implemented PI and PID type controllers because of its
simplicity, easy implementation and good performance [5]. Thus, it is imperative to
analyze the fragility of PID controllers. Datta et al., Ho , designed the less fragile PID
controllers based on the maximum l2 norm parametric stability margin without
considering the performance characteristics of the closed loop system [6, 7]. Irinel-
Constantin et al suggested a geometric approach to find the stability region of the PID
controller, based on stability crossing boundaries for SISO systems and they found the
maximum deviation of the controller parameters within the closed loop stability
boundary [15].Alfaro suggested a Fragility Index (FI) for classifying the various PID
controllers based on the Maximum sensitivity (Ms) of the closed loop system. Also,
Alfaro classified the PID controllers in to three categories namely i) fragile controller
ii) Non-fragile controller and iii) Resilient controller, based on Fragility Index (FI)
[3].
Based on the literature survey, it is found that the design of Non-fragile robust PID
controller that can tolerate plant uncertainty as well as controller uncertainty is of
immediate necessity. Many researchers analyzed the fragility of the designed PID
controller [6, 7, 15], but they are not including the controller’s fragility criterion in the
robust PID controller design. By considering the efficiency of evolutionary algorithms
in controller design, non-fragile robust PI controller design is formulated as an
optimization problem. Recently, CMA-ES algorithm is successfully applied as an
efficient optimization tool for optimizing nonlinear, multi-modal real world
optimization problems.
In this paper, Co-variance Matrix Adaptation Evolutionary Strategy is used for the
design of non-fragile robust PI controller. The desired maximum sensitivity (Ms=1.4)
is considered as an objective and the success rate of stability under probabilistic
controller perturbation is taken as a constraint for non-fragile robust PI controller
design problem. The Single Input Single Output first order sugar cane raw juice
neutralization process and second order Irrigation canal system[8] are chosen as a test
systems for the design of CMA-ES based non-fragile robust PI controller.
Consideration of robustness as well as non-fragileness in a controller design which
gives the better robust controller.
The rest of the paper is organized as follows, Section 2 discusses the problem
formulation; Section 3 describes the implementation of CMA-ES in non-fragile robust
PI controller design; Simulation results are given in Section 4; Conclusion is on
Section 5.
34 K.M. Abdul Kadhar and S. Baskar

2 Problem Formulation

In this paper, an ideal PI controller structure given in equation (1) is chosen for the
design of non-fragile robust PI controller. The maximum sensitivity (MS) of the closed
loop system given in equation (1) is considered as a design parameter for the design
of non-fragile robust PI controller [4, 7]. MS has a good geometrical interpretation
with Nyquist diagram and the inverse of MS (i.e ) is the shortest distance from the
critical point -1+j0 to the Nyquist curve of the open loop transfer function(L=GK) as
shown in figure 1.

L(s) plane

Fig. 1. Nyquist curve Loop transfer function

1 ; max (1)

where, G(s) is the plant to be controlled, x is controller parameter vector (i.e


, ). Maximum sensitivity (MS) of the closed loop system gives the
information about the closed loop system’s sensitivity under the effect of parameter
variations in closed loop system. The maximum sensitivity as (MS =1.4) which gives
the good phase and gain margin for maximum robustness [3, 5]. Hence, desired MS
value is chosen as 1.4 for robustness objective in non-fragile robust PI controller
design.

2.1 Controller Uncertainty Modeling

The controller parameters (kp and Ti) are perturbed as shown in equation (2).
1 ; 1 (2)

where, and are the perturbed controller parameters and δ is the parametric
perturbation. If the perturbed controller parameter vector arranged as ,
then, the perturbed controller and the maximum sensitivity ) are given in
equation (3).
Non-fragile Robust PI Controller Design Using CMA-ES 35

1 ; max (3)

Usually, perturbation as in the equation (3) is considered as a step change in the


plant or controller parameters i.e δ is a single scalar [3]. But this type of uncertain
behavior doesn’t mean the actual real time perturbations. Hence, probabilistic
perturbation is used in this paper to mimic the real time uncertain behaviors of
controller parameters [11].

2.2 Probabilistic Approach

The maximum sensitivity bound in the range of 1.2 to 2 which gives the guaranteed
robustness to the closed loop system [3, 5]. If the MS value is in between the above
specified range (i.e. MS =1.2 to 2) then the closed loop system is stable. Lower MS≈1.2
gives better robustness and Larger MS≈2 gives aggressive and less acceptable
robustness [7]. By using this maximum sensitivity bound, success rate of stability
under probabilistic controller perturbation is evaluated using Monte Carlo Simulation
(MCS).

2.3 Evaluation of Non-fragileness Constraint

MCS is a method for iteratively evaluating a deterministic model using sets of random
numbers as inputs [9, 10]. In this paper, q×n number of random samples are generated
by using normal distribution with zero mean and 0.1 standard deviation as N (0, 0.1)
for uncertain parameters. where, 0.1 represent the 10% of uncertain parameters. The
generated random samples are added to the controller parameters as per the equation
(2). Under random variation of controller parameters, the stability of the closed loop
system is represented using a discrete limit state function based on the maximum
sensitivity bound and given in the equation 4 [9,10].

1 1.2 2
(4)
0

By using MCS, the probability of success rate ( ) for stability is calculated using
the q number of binary data from the limit state function as in the equation (5).

∑ , , (5)

where, q is the total number of samples , limit state function at the ith
sample, n is the number of controller parameters . This probability of success rate for
stability is used as a constraint in the design of non-fragile robust PI controller.
36 K.M. Abdul Kadhar and S. Baskar

2.4 Objective Function


An error function J(x) is used as an objective to achieve the desired maximum
sensitivity in non-fragile robust PI controller design and it is given in equation (6).
| | (6)

where, Ms(x) is the actual maximum sensitivity and Msd is the desired maximum
sensitivity value (Msd =1.4). CMA-ES finds the optimal controller parameter vector
x= [kc, Ti] by minimizing the error function J(x) while satisfying the non-fragileness
constraint as follows.

Minimize
Subject to

∑ , , 0.8 (7)

In equation (7), non-fragileness constraint is set as 0.8 for achieving at least 80% of
success rate in closed loop stability under probabilistic controller perturbation.

2.5 Test Systems

The dynamic model of the sugar cane raw juice neutralization process [8] is given in
equation (8).
.
(8)

The second order irrigation canal system [8] model is shown in equation (9).

.
(9)

3 CMA-ES in Non-fragile Robust PI Controller Design

CMA-ES was proposed by Hansen and Ostermeier in 2001. The basic idea in this
stragety is moving the population in the form of multivariate normal distribution. The
distribution takes the form of hyper ellipsoid and search history of the population is
used for moving the population to optimum one [11, 12]. CMA-ES algorithm code is
downloaded from [11] and the CMA-ES code is customized for handling constraints
for the design of non-fragile robust PI controller.

3.1 Constraint Relaxation and Fitness Function


By using εc constraint relaxation method, the constraint violation was relaxed up to
the generation control limit. The εc level is set to zero up to the generation control
limit gc, to obtain the solutions with no constraint violation.
Non-fragile Robust PI Controller Design Using CMA-ES 37

0 (10)

After the generation control limit gc, constraint is violated as follows.


0 1 ,0 g g
g (11)
0 , g g

Where is the top ith individual and i= (0.05* λ).The recommended parameter
ranges are [13] :g 0.1g , 0.8g 2,10 . Due to this constraint
relaxation, the individual with best objective are considered for next generation even
it has high constraint violation. After constraint relaxation, penalty parameter less
constraint handling scheme is employed to find the fitness function. The fitness
function for any solution x is given as follows:

(12)
g
where,
F(x) - Fitness function,
J(x) - Objective function
Jmax - objective function value of the worst feasible solution in the population

3.2 Initialization
The initial CMA-ES parameters are set as population size(λ)=10; max_Feval=1000,
m=0, σ=0, 0.25 x , , . where, x , are upper
and lower bounds of controller parameters respectively.

3.3 Controller Parameter Setting


The solution vector for SISO system is arranged as x= [kc, Ti].
The lower and upper bounds of the solution vector are set as follows.
Sugar cane raw juice neutralization process x = [0, 0]; = [5, 10].
Irrigation canal system x = [0, 0]; = [1, 500];
The dimension for SISO non-fragile robust PI controller: n=2.

3.4 Co-variance Matrix Adaptation Evolutionary Strategy

Step 1: Generate λ×n random controller parameter matrix of x as follows.

x σ N 0, CM where 1 (13)

Step 2: Stop if the termination criteria is met as in the section 3.5.


38 K.M. Abdul Kadhar and S. Baskar

Step 3: Calculate the fitness value (F) based on objective and constraint violation as
in equation (12).
Step 4: Sort the solutions based on fitness value and Select the best µ number of
individuals ( ) from the λ population.

Step 6: Update the mean m(g+1) using ∑ : (14)

where, w ∑
µ 1 ; (15)

This weighted average gives the new mean to move the population towards optimal
solution. The variance of effective selection mass [12] is calculated from the equation

Step7: Adaptation of global step size using evolution path

g+1 g g g+1 g
1 . 2 g (16)

PG
σ σ exp 1 (17)
E ,I

Step 8: Update Co-variance matrix using the evolution path

μ
p 1 C P C 2 C . μ μ (18)
σ

Then the covariance matrix CM(g) is extended from the evolution path

T 1 T
CM(g+1) = 1- .C g
+ Pcg+1 Pcg+1 +(1- ) 2 ∑μi=1 (xg+1 g g+1 g
i -<x>μ ) (xi -<x>μ ) )
σ(g)
(19)

Step 9: The updated CMA-ES parameters , σ , CM(g+1) are used to


generate λ×n random controller parameter matrix of x using the equation (13) for the
next generation (g+1), set g=g+1, return to step 3. This procedure is repeated until the
stopping criteria met.

3.5 Stopping Criteria


The recommendations of stopping criteria in the design of non-fragile robust PI
controller are given as below.
Non-fragile Robust PI Controller Design Using CMA-ES 39

• max_Feval: stop if maximum function evaluation is reached.


• TolFun: Tolerance value for objectives is assumed to 10-5. Stop if all the fitness
function (F) values of the last 20 generation are below TolFun.
• TolX: Tolerance value for co-ordinates is assumed to 10-5[12].

4 Simulation Results

All the simulations are done in MATLAB R2009b software on a 32 bit core2duo
processor PC operating at 2.93 GHz with 3GB RAM. Initialization and controller
parameter settings are done as in the section 3.

4.1 Controllers
The CMA-ES designed non-fragile robust PI controller for both sugar cane raw juice
neutralization process and irrigation canal system are given in Table 1. Table 1 shows
the controller parameters and their corresponding maximum sensitivity, probability of
success rate under 20% probabilistic perturbations.

Table 1. Controller parameters and maximum sensitivity for both systems

Sugar cane raw juice Irrigation canal system


neutralization process
Controller Ms Ps(x) Controller Ms Ps(x)
parameters parameters
CMAES-PI kp=4.3365, 1.4 1 kp=0.1920, 1.4 1
ki=46.3593 ki=410.09.
C.A.Monje-PI kp=1.2749, 1.36 0.9 kp=0.5511, 1.83 0.6
[8] ki=24.51. ki=0.0008,
kd=80.13.
Astrom-PI [14] kp=1.76, 1.32 0.7 kp=0.00262, 1.58 0.9
ki=29.32. ki=5.86.

From Table 1, it is clear that CMA-ES based PI controller has more success rate of
stability under probabilistic perturbation for both systems.

4.2 Time Domain Analysis


Figure 2 and Figure 3 shows the closed loop step response of three controllers’ for
sugar cane raw juice neutralization process and Irrigation canal system respectively.
From the Figures 2 and 3, CMA-ES based non-fragile robust PI controller has less
over shoot and quick response than other PI controllers for both systems.
40 K.M. Abdul Kadhar and S. Baskar

Fig. 2. Closed loop step response of three controllers for sugar cane neutralization process

Fig. 3. Closed loop step response of three controllers for Irrigation canal system

4.3 Robustness Analysis

Figure 4 and 5 shows the response of the load disturbance rejection of the three
controllers for sugar cane raw juice neutralization process and Irrigation canal system
respectively. From figure 4 and 5, it is clear that CMA-ES designed Non-fragile
robust PI controller reject the input load disturbance quickly than other controllers for
both systems.
Non-fragile Robust PI Controller Design Using CMA-ES 41

Fig. 4. Load disturbance rejections of the controllers for sugar cane neutralization process

Fig. 5. Load disturbance rejections of the controllers for Irrigation canal system

4.4 Fragility Analysis


To check the fragile characteristic of the controller’s, the mean of random number is
increased from 0.1 to 0.5 and the corresponding probability of success rate for
stability is plotted against the mean of the perturbation level. Figure 5 and 6 shows the
comparison results of three controller’s success rate against the increasing
perturbation in controllers for sugar cane raw juice neutralization process and
Irrigation canal system respectively. From figure 5 and 6 CMA-ES designed non-
fragile robust PI controller has near 80% success rate of closed loop stability up to 0.5
mean of random perturbations. Astrom based PI controller and flat phase concept
based PI controller are not able to withstand its closed loop stability whenever the
perturbations increased.
42 K.M. Abdul Kadhar and S. Baskar

Fig. 6. Fragility analysis of different controllers for neutralization process

Fig. 7. Fragility analysis of different controllers for Irrigation canal system

5 Conclusion

In this paper, non-fragile robust PI controller is designed using Co-variance Matrix


Adaptation Evolutionary Strategy (CMA-ES) for sugar cane raw juice neutralization
process and Irrigation canal system. Parametric probabilistic controller perturbation is
used to check the robustness and fragileness of the designed controller. CMA-ES
based non-fragile robust PI controller has better command tracking and load
disturbance rejection characteristic than Astrom based PI controller and flat phase
concept based PI controller for both systems. In controller fragility, CMA-ES based
PI controller has maximum success rate of stability under the probabilistic controller
perturbation for both systems. CMA-ES based non-fragile robust PI controller is
Non-fragile Robust PI Controller Design Using CMA-ES 43

better in robustness as well as non-fragileness. Hence, CMA-ES based non-fragile


robust PI controller is suitable for real time applications where there is a fragility
problem.

Acknowledgment. This research was supported by Indian government via University


Grants Commissions’ Maulana Azad National Fellowship (MANF) scheme (MANF
File no: MANF-MUS-TAM-3045). The authors would like to thank the principal and
management of Thiagarajar College of Engineering for their continuous
encouragement and providing the facilities for the research.

References
1. Keel, L.H., Bhattacharyya, S.P.: Robust, Fragile or Optimal. IEEE Transactions on
Automatic Control 42(8) (August 1997)
2. Whidborne, J.F., Istepanian, R.S.H., Wu, J.: Reduction of Controller Fragility by Pole
Sensitivity Minimization. IEEE Transactions on Automatic Control 46(2) (February 2001)
3. Alfaro: PID controller fragility. ISA Transactions 46(4), 555–559 (2007)
4. Duan, Z., Huang, L., Wang, L.: Robustness analysis and synthesis of SISO systems under
both plant and controller perturbations. Systems & Control Letters 42, 201–216 (2001)
5. Astrom, K.J.: Control system Design (2002)
6. Datta, A., Ho, M.-T., Bhattacharyya, S.P.: Structure and Synthesis of PID Controllers.
Springer, London (2000)
7. Ho, M.T.: Non-fragile PID controller design. In: 39th IEEE Conference on Decision and
Control, Sydney, Australia (December 2000)
8. Monje, C.A., Calderon, A.J., Vinagre, B.M., Chen, Y., Feliu, V.: On Fractional PIλ
Controllers: Some Tuning Rules for Robustness to Plant Uncertainties. Nonlinear
Dynamics 38, 369–381 (2004)
9. Wittwer, J.W.: “Monte Carlo Simulation Basics From” Vertex42.com (June 1, 2004),
http://www.vertex42.com/ExcelArticles/mc/
MonteCarloSimulation.html
10. Ray, L., Stengel, R.F.: A Monte Carlo approach to the analysis of control system
robustness. Automatica 29(1), 229–236 (1993)
11. Hansen, https://www.lri.fr/~hansen/cmaesintro.html
12. Hansen, N.: The CMA Evolution Strategy: A Tutorial (June 28, 2011)
13. Takahama, T., Sakai, S., Iwane, N.: Solving Nonlinear Constrained Optimization Problems
by the ε Constrained Differential Evolution. In: 2006 IEEE Conf. on Systems, Man, and
Cybernetics, Taipei, October 8-11 (2006)
14. O’Dwyer, A.: Handbook of PI and PID Controller Tuning Rules. Imperial College Press
(2009)
15. Mor˘arescu, I.-C.: Stability Crossing Boundaries and Fragility Characterization of PID
Controllers for SISO Systems with I/O Delays. In: ACC 2011 (2011)

You might also like