Estimation of Parameters and Tuning of a Speed

PI of Permanent Magnet DC Motor Using

Differential Evolution
Maycon Chimini Bosco1, Jacqueline Jordan Guedes2, Marcelo Favoretto Castoldi3, Alessandro Goedtel4,
Emerson Ravazzi Pires da Silva5, Luiz Francisco Sanches Buzachero6
1,2,3,4,5,6
Department of Electrical Engineering, Federal Technological University of Paraná - UTFPR, Cornélio Procópio, Brazil
[email protected], [email protected], [email protected], [email protected],
[email protected], [email protected]

Abstract— In this work, a system of parameter estimation and the systems design solution [6], linear systems solution [7] and
tuning of a speed PI controller of a permanent magnet DC motor robot design Manipulator [8].
using differential evolution is presented. Initially the algorithm
estimates permanent magnet DC motor parameters, they are,
II. LITERATURE REVIEW
armature resistance, armature inductance, constant of torque,
coefficient of friction and moment of inertia. Later, with the
parameters of the plant estimated, the algorithm tuned the best A. Permanent Magnet Direct Current Motor
set of parameters of the PI controller, using the performance DC motors are one of the most commonly used
index of controllers ITAE. Experimental results are presented to components in industries today to perform motion control in
validate the proposed method. small power applications. Permanent magnet motors stand out
over other DC motors because they do not require an external
Keywords— Estimation of parameters; DC motor; Differential field circuit, thus eliminating losses in the field circuit copper.
evolution; ITAE; Speed control; PI Controller. This motor has as a disadvantage the dependence of permanent
magnets which do not produce high flux density compared to
I. INTRODUCTION some other DC motor models. Thus, the motor has lesser
Industries are increasingly investing in the automation of induced torque per armature current amperage than a shunt DC
the production process and the tasks performed by employees motor. Still, permanent magnet DC motors are at risk of
are carried out by robots. These machines are programmed to demagnetizing [9].
perform quick and precise movements. In most cases, small From the equivalent circuit of the DC motor, manipulations
DC motors do not have board data displaying their parameters. are made with the motor equations to obtain its dynamic
In this way, at the moment of designing a speed controller or model, which is represented in state space in (1), where Ra is
position, one has the problem of not knowing the parameters of the armature resistance, La is the armature inductance, K b is
the plant. Thus, the controller is often designed using empirical
techniques, which can lead to a high tuning time, in addition to the induced voltage constant, J is the moment of inertia, B is
an inadequate choice of parameters [1]. the friction coefficient, V is the applied voltage, ωm is the
angular speed, ia is the armature current and T is the load
Citing some of the estimation methods already developed,
there is the method proposed by [2] use an online method using torque. This model and all manipulation are found in [10]. The
two segments of recursive least squares algorithms, one fast state space will be used in the algorithm to simulate the
and one slow, are combined uniquely in real time. In [3] operation of the DC motor.
particle swarm optimization (PSO) is used in their system. The ª ia º ª − Ra / La − Kb / La º ª ia º ª1 / La 0 º ªV º
reference [4] was successful in its estimation using Extended « » = « + (1)
Kalman Filtering. ¬ ωm ¼ ¬ K b / J − B / J »¼ «¬ωm »¼ «¬ 0 −1 / J »¼ «¬T »¼
This work aims to contribute to the development of
methods for estimation of parameters of DC motors and the B. Differential Evolution
tuning of controllers. The system is designed to estimate the Differential Evolution (DE) is a stochastic optimization
parameters of the permanent magnet DC motor and provide the algorithm, based on a population of solutions, which operates
user with the best values of the PI controller parameters to be through computational steps similar to those employed by
used in this motor. For this reason the differential evolution most Genetic Algorithms (GA). It was proposed in 1995 by
algorithm will be used. According to [5], this strategy consists Rainer Storn and Kenneth Price for solving optimization
of a simple optimization method and has been applied problems with continuous variables [11]. The success of DE
successfully in many branches of science, citing as examples over other methods of optimization is due to the fact that its

978-1-5090-4281-4/17/$31.00 ©2017 IEEE

Authorized licensed use limited to: KTH Royal Institute of Technology. Downloaded on October 28,2024 at 13:52:17 UTC from IEEE Xplore. Restrictions apply.
implementation is relatively simple, to have a high speed of the stopping criterion established by the user, which can be
convergence, to have few control parameters and, mainly, to defined by means of a fixed number of iterations or by a
have a powerful search method [6]. desired fitness value. Fig. 1 shows the block diagram of the
proposed algorithm.
The algorithm is started by creating a population with
random values covering the entire search space. This generated
population is an array of NP rows, where NP is the number
of chromosomes (vectors) solution and the number of columns
depends on the amount of constants that the problem in which
search optimization possesses, called genes. After the
population is created, each chromosome is used as a possible
solution. In this way, the value that each chromosome has in
solving the problem, called fitness stage, is verified. After this
procedure arise the operators of the differential evolution that
are divided into 3 stages: mutation, crossover and selection [5].
Mutation is the stage responsible for maintaining the
diversity of the population. At this stage that the creation of
new individuals occurs, these individuals are called donors and
are created by adding the weighted difference between two
solutions to a third solution. These solutions are chosen
randomly from the candidate solutions. These solutions are
chosen randomly from the candidate solutions. Equation (2)
shows how this process is performed:

V(
q +1)
(
= X αq + F X βq − X γq ) (2)

where V ( ) represents the donor vector, X α , X β and X γ

q +1

are vectors of the population chosen randomly and different

from each other, F is a value ranging from 0 to 1, called
scaling factor, which is a parameter of the DE algorithm that
controls the amplitude of the weighted difference. The index q
represents the vectors that were created in the previous
generation, while the index q + 1 refers to a vector that is being Fig. 1. Differential Evolution block diagram.
developed in the current mutation process. What will differ between the two steps of the algorithm is
To begin the process of crossing must first select which the objective function that is used in them. For the estimation
will be the target, that is, the population vector that will be of parameters will be used an objective function based on the
combined with the donor created previously. It is necessary method of the least squares and for the stage of tuning of the PI
that each individual of the population is the target at least once will be used the performance index of ITAE controllers.
in the interaction. This step aims to increase the diversity of
individuals in the population. This procedure is represented by C. Proportional Integrative Controller
(3): The PI is a controller that operates with feedback and
performs a control of the plant using the weighted sum of the
­° V ( q +1) , if ri ≤ CR error with the integral of this value. The same is widely used in
i
U tq +1 = ® (3) industries because of its relatively simple structure that can be
q
°̄ d ,i , if ri > CR
X understood and applied to practice [12]. The PI can be
represented by (4):
where U t represents the experimental vector, ri is a random
t
number [0,1], X d ,i is the target vector to be combined with the
donor and CR is called a reproduction constant and represents
u ( t ) = K p e ( t ) + Ki ³ e ( t ) dt
0
(4)

the probability of occurrence of the crossing, that is the

where u ( t ) is the reference used, K p is the proportional gain,
probability of the experimental vector to inherit the values of
the variables of the donor vector. K i is the integral gain and e ( t ) is the error obtained by the
In the selection step it is verified if the experimental vector difference between the desired value and the measured value.
has a cost solution better than the target vector. Thus, if it has a The use of this controller is due to the fact that it is possible to
better solution cost, the experimental vector is housed in the improve the transient response with the contribution of
population of the new generation, otherwise the target vector proportional action, while the integral action corrects the
will occupy the position. After this process the first generation steady state error [13].
is created. The number of generations or iterations depends on

Authorized licensed use limited to: KTH Royal Institute of Technology. Downloaded on October 28,2024 at 13:52:17 UTC from IEEE Xplore. Restrictions apply.
 In order to perform the controller tuning, the ITAE performed for tests in no load and in short time, where the
(Integral Time Absolut Error) performance controller index, samples used do not exceed two seconds of motor operation, in
which was developed by [14] and is an area not yet explored, this way, the temperature variation will not interfere in the
will be used, since the results obtained through classical results of the experiment.
methods still satisfy almost in all cases the user's need.
Therefore, the ITAE was chosen to be used in the objective To estimate the motor parameters, a comparison between
function of the differential evolution algorithm. The ITAE the measured and estimated speed and current signals is
index is defined by (5): performed. For this, the algorithm should minimize the
difference between the measured and estimated signals of the
t armature current and the motor speed. Equation (6) shows the
ITAE = ³ t e ( t ) dt
0
(5) objective function to be minimized:
N 2 N 2
The ITAE was chosen to be used in tuning due to its 1 § iai − iˆai · K § ωmi − ω
ˆ mi ·
characteristic of having a small overshoot in the response and PE =
N ¦ ¨
i =1 ©
iai ¹
¸ +
N ¦ ¨
i =1 ©
ωmi
¸
¹
(6)
having the oscillations damped. This behavior is possible due
to its structure, because as the absolute error is weighted by where PE is the objective function to be minimized, N is the
time, for high initial errors there is a small weighting, since for number of points of armature current and motor speed signals,
errors that take a longer duration there is a large weighting
[15]. ia is the measured armature current, iˆa is the estimated
armature current, ωm is the measured speed, ω ˆ m is the
III. RESEARCH METHOD estimated speed and K is a weighting factor.
As mentioned before, this paper proposes the parameters To carry out the estimation the constants of the DE
estimation of a permanent magnet DC motor in conjunction algorithm (scale factor F , and crossover constant CR ) were
with its optimized speed controller, using an unique DE set F = 0.9 and CR = 0.5 . The size of the population was
algorithm. In first stage, the algorithm estimates the parameters fixed in 15 chromosomes and the number of genes used was 5,
of the motor and so, these results are input of the second stage, each one belonging respectively to armature resistance,
which calculates the best PI for the plant. Fig. 2 presents the armature inductance, torque constant, moment of inertia and
proposed method. friction coefficient. For this step, T = 0 was also used, because
in this way, by analyzing the dynamic model of the motor, it is
possible to eliminate a term of this system, facilitating the
estimation of the moment of inertia J and friction coefficient
B.
To verify the efficiency of the implemented algorithm, the
model of a permanent magnet DC motor was simulated
through Simulink software in order to estimate the parameters
of this motor. Twenty estimates were used, maintaining as a
criterion of stopping a fitness value considered satisfactory to
obtain a good estimation. Table I shows the parameters of the
simulated motor and the average of the estimates made. The
simulated DC motor model parameters are in the standard
Simulink software model.

TABLE I. RESULTS OF SIMULATED TEST

Ra La Kb J B Generations
Simulated 3.120 0.012 0.300 0.007 0.003 -
Estimated 3.120 0.012 0.300 0.007 0.003 123
Fig. 2. Proposed methodology.

A. Parameters Estimation By means of the results of the 20 estimates it was observed

The motor parameters estimated are: armature resistance, that the performance of the algorithm for this method is
armature inductance, torque constant, induced voltage satisfactory, since the values of the parameters converge to the
coefficient, moment of inertia and friction coefficient. The same value with precision and with a low amount of
estimation is performed using the speed and armature current generations.
curves as input data, and the estimation will not be performed
in real time. B. Controller Tuning
Real-time parameter estimates may be affected by Considering the results of the estimation of the simulated
temperature variation or due to magnetic variation, but in this DC motor parameters, it is important to validate the method
case real-time estimation will not be used. Thus, they will be by means of experimental tests. For this, the permanent

Authorized licensed use limited to: KTH Royal Institute of Technology. Downloaded on October 28,2024 at 13:52:17 UTC from IEEE Xplore. Restrictions apply.
magnet DC motor present in the 2208 servomechanism module
developed by Datapool Electronic, and presented by Fig. 3, is
used with a sample rate of 150 samples/second. After the
acquisition of the armature current and speed curves, the
filtering of the signal was carried out, in order to eliminate the
noises emitted by the servomechanism kit.

Fig. 5. Speed with estimated parameters.

Fig. 3. Servo mechanism module Datapool Electronic 2208.

To perform the controller tuning, a system with two blocks

in series (controller and plant system) and a feedback is used.
The plant block contains the estimated parameters of the
permanent magnet DC motor and the PI block is composed of Fig. 6. Armature current with estimated parameters.
a controller discretized in time, therefore, the sampling
frequency of the data acquisition board must be taken into In relation to the results, it can be observed that in the speed
account. Two controllers were tuned, one for the motor running signals the difference between the experimental and the
on no load and one for the motor operating with load estimated signals, both in the transient and in the steady state,
variations. The PI controller has the function of maintaining is small (1% maximum error). In relation to the current, the
the motor speed with a value of 3 rad/s. The system is approximation between the curves shows a steady state error of
represented by the block diagram in Fig. 4. 1%. However, in the transient, the result does not have the
same performance, which can aggravate the presence of errors
in the tuning step of the speed controller.

B. Controllers Tuning
The first tuned controller, where the motor operated in no
load condition, reached values of 0.7089 for K p and 23.7907
Fig. 4. Block diagram of the control system used. for K i . For the second controller, where the motor operates
with load, the tuned values were 0.7212 for K p and 25.7570
IV. RESULTS for K i . The results of the tests carried out with the controllers
A. Experimental Estimation of Motor Parameters are presented through comparisons between the speed and
current curves obtained through Simulink software and the
The algorithm is executed 20 times and the parameters curves obtained from the experimental tests. Fig. 7 and 8 show
chosen to be used in the system plant were obtained through the response of PI Controller 1 operating with load variation.
the arithmetic mean of the estimates. In relation to the
estimates, it can be observed that the parameters presented a By making the analysis of the speed response, it is noticed
low variation. Thus, we can conclude that the algorithm is that the periods of permanent regime are satisfactory and in
converging to the same value in all estimates. Fig. 5 and 6 relation to the transient ones the answer also presents good
show the speed and armature current generated by means of the results. The maximum error occurs when the load is removed,
estimated parameters. where the error has a maximum value of 8%.
The speed presented has a steady state amplitude of 3.3
rad/s. The armature current presented has peak values of
0.59 A for the real signal and 0.50 A for the estimated signal.
On a permanent basis the current has an amplitude of 0.34A.

Authorized licensed use limited to: KTH Royal Institute of Technology. Downloaded on October 28,2024 at 13:52:17 UTC from IEEE Xplore. Restrictions apply.
Fig. 7. PI controller 1 tuning results: speed. Fig. 10. PI controller 2 tuning results: armature current.

V. FINAL CONSIDERATIONS
It was possible to observe that the parameter estimation
obtained satisfactory results. Both curves obtained the
approximation of the real curves with greater error in the
transient of the current curve. The controllers tuned from the
estimated parameters obtained good results in the speed
control. The speed curves showed that the settling time was
considerably low and the transient error remained below 10%
in practically all tests, and in the steady state the error was 2%.
On the other hand, current curves in all cases presented high
errors in the transient, which did not happen in the steady state
Fig. 8. PI controller 1 tuning results: armature current. periods. The errors caused by the estimation step were loaded
until the step of controlling and adjusting the PI parameters.
In relation to the current curve we can observe that the
worst response occurs in the transients with the error between
the curves exceeding in some points the margin of 10%. In the REFERENCES
regime, the biggest difference occurs when the motor operates [1] A. E. Fitzgerald, C. Kingsley Jr, and S. D. Umans. Electric machinery,
under load, where the maximum error between the curves is 6th ed. Mcgraw-Hill, 2006.
5%. The next test is performed using the PI Controller 2, which [2] S. J. Underwood, and I. Husain, “Online parameter estimation and
adaptive control of permanent-magnet synchronous machines,” IEEE
was tuned from the motor operating with load variation. The Transactions on Industrial Electronics, v. 57, n. 7, p. 2435-2443, 2010.
curves present in Fig. 9 and 10 show the curves of the tests
[3] V. P. Sakthivel, R. Bhuvaneswari, and S. Subramanian, “Multi-objective
performed. parameter estimation of induction motor using particle swarm
optimization,” Engineering Applications of Artificial Intelligence, v. 23,
By means of the presented curves it is noticed that both n. 3, p. 302-312, 2010.
have performance in permanent regime similar to the simulated [4] S. Aksoy, A. Mühürcü, and H. Kizma, “ State and parameter estimation
curves. In relation to the transients, the curves have errors of in induction motor using the Extended Kalman Filtering algorithm,” in
8% to 14% between the speed curves and the transient current Proceedings of the IEEE International Symposium of Modern Electric
of 5% to 30%. Power Systems, 2010, pp. 1-5.
[5] S. Das, and P. N. Suganthan, “Differential evolution: a survey of the
However, in relation to the moment that the load variation state-of-the-art,” IEEE transactions on evolutionary computation, v. 15,
occurs, it can be observed that the current has a behavior n. 1, pp. 4-31, 2011.
similar to the simulated curve. [6] R. Storn, “System design by constraint adaptation and differential
evolution,” IEEE Transl. J. Evolutionary Computation, vol. 3.1, pp. 22–
34, 1999.
[7] S. Cheng, and C. Hwang, "Optimal approximation of linear systems by a
differential evolution algorithm," IEEE Trans1. Systems, Man and
Cybernetics, Part A: Systems and Humans, vol.31.6 pp. 698-707, 2001.
[8] C. Gonzales, D. Blanco, and L. Moreno, “Optimum robot manipulator
path generation using Differential Evolution,” in proceedings of IEEE
Congress on Evolutionary Comptation, 2009, pp. 3322-3329.
[9] S. J. Champan. Electric machinery Fundamentals, 5 th ed. Mcgraw-Hill
Education, 2006.
[10] R. Krishnan. Electric motor drives: modeling, analysis, and control.
Prentice Hall, 2001.
[11] R. Storn, and K. Price, "Differential evolution—a simple and efficient
adaptive scheme for global optimization over continuous spaces,”
Fig. 9. PI controller 2 tuning results: speed. International Computer Science Institute, 1995.

Authorized licensed use limited to: KTH Royal Institute of Technology. Downloaded on October 28,2024 at 13:52:17 UTC from IEEE Xplore. Restrictions apply.
[12] N. Özturk, “Speed control for DC motor drive based on fuzzy and
genetic PI controller–a comparative study,” Elektronika ir
Elektrotechnika, v. 123, n. 7, pp. 43-48, 2012.
[13] K. Ogata; Y. Yang. Modern control engineering, 5 th ed. Pearson, 2009.
[14] D. Graham, and R. C. Lathrop, "The synthesis of optimum transient
response: Criteria and standard forms," American Institute of Electrical
Engineers, Part II: Applications and Industry, Trans. of the, vol. 72.5,
pp. 273-288, 1953.
[15] D. Maiti, A. Acharya, M. Chakraborty, A. Konar, R. Janarthanan,
“Tuning PID and FOPID controllers using the integral time absolute
error criterion,” In 4th International Conference on Information and
Automation for Sustainability, 2008, pp. 457-462.

Authorized licensed use limited to: KTH Royal Institute of Technology. Downloaded on October 28,2024 at 13:52:17 UTC from IEEE Xplore. Restrictions apply.