Sunteți pe pagina 1din 7

A Comparison of Adaptive PID Methodologies Controlling a DC Motor With a

Varying Load

Luís Osório, Jérôme Mendes, Rui Araújo, and Tiago Matias


Institute for Systems and Robotics (ISR-UC), and
Department of Electrical and Computer Engineering (DEEC-UC),
University of Coimbra, Pólo II, PT-3030-290 Coimbra
lbica@isr.uc.pt, jermendes@isr.uc.pt, rui@isr.uc.pt, tmatias@isr.uc.pt

Abstract for streamlining in the areas of theory and application, and


for increasing reliability and robustness [3]. The work
This work addresses the problem of controlling un- of Kolavennu et al. [6] shows that in many real-world
known and time varing plants for industrial aplications. processes where a nonadaptive controller is sufficient, an
To deal with such problem several Self-Tuning Controllers adaptive controller can achieve an even better quality of
with a Proportional Integral and Derivative (PID) struc- control. Other example is given in [12] where the use of
ture have been chosen. The selected controllers are based an adaptive controller decreased fuel consumption signif-
on different methodologies, and some use implicit identifi- icantly.
cation techniques (Single Neuron and Support Vector Ma- Adaptive controllers follow three basic approaches:
chine) while the others use explicit identification (Dahlin, the Model Reference Adaptive Systems (MRAS), the
Pole placement, Deadbeat and Ziegler-Nichols) based in Heuristic Approach (HA), and the Self-Tuning Con-
the Least Squares Method. The controllers were tested on trollers (STC). The MRAS controllers use one or multiple
a real DC motor with a varying load. The results have system models to determine the difference between the
shown that all the tested methods were able to properly output of the adjustable system and the output of a refer-
control an unknown plant with varying dynamics. ence model, and adjust the parameters of the adjustable
system or generate a suitable input signal [4]. The meth-
ods based on HA do not require determining the optimum
1 Introduction solution of a problem, ignoring whether the solution can
be proven to be correct, provided that it produces a good
Because of its simplicity and good performance, the result. Such methods are based on expert human experi-
Proportional Integral and Derivative (PID) controller is by ence [1]. STC are based on the recursive estimation of
far the most popular feedback controller in the automatic the characteristics of the system. Once the system is de-
control field. In industrial processes the classical PID con- termined, appropriate methods can be employed to design
troller was employed in about 90% or more of control an adequate controller [11].
loops [2]. Generally, engineers tune the optimal param- The main objective of this work is to test PID algo-
eters of a PID controller to match the operating condition rithms that can get close to the concept of “plug and play”
and such parameters remain fixed during the whole op- (algorithms that do not require information about the plant
eration [14]. The problem when using fixed parameter to be controlled and must be able to auto-adapt their con-
controllers is that most of the processes met in industrial trol parameters taking in account the variations of the
practice have dynamics that are not modeled or that can plant). Controllers based on MRAS require the knowl-
change over time. In such cases, the classical controller edge of an approximate model of the plant to control, and
with fixed parameters may became unstable and would be HA controllers are experience-based techniques for learn-
required to be adequately re-tuned to retain robust con- ing the control laws, meaning that both these approaches
trol performance. To overcome this difficulty, adaptive require previous information about the plant. Thus, only
algorithms were developed, which extends the area of real controllers based in STC will be considered.
situations in which high quality control can be achieved. Dahlin’s PID Controller [8] was selected for its low or-
According to Bobal et al. [5] the development of adap- der, the Pole Placement Controller [13] for having very
tive control started in the 1950s with simple analogue low computation, the Deadbeat controller of second and
techniques since the computing equipment had not the re- third orders [7] for having no parameters to be adjusted,
quired performance to execute the most sophisticated al- the Ziegler-Nichols controller [14] to verify how an older
gorithms that were already proven in theory. Later in the controller could be compared to newer ones, the Single
1980s, as the microprocessors became faster and cheaper, Neuron Controller [11] for beeing a method based based
it evolved to discrete-time control and the theory devel- on biological Systems and the Support Vector Machine
oped in the early years was finally be applied. At the controllers [10][9] for beeing based on machine learning.
present there is yet much unused potential in mass appli- To compare the performance of the control algorithms
cations and there are still opportunities for improvements, a real experimental setup composed of two coupled DC
motors with varying load, was build and used. is updated at each iteration, k, using equation (4)
The paper is organized as follows. Section 2 presents C(k − 1)Φ(k)
the algorithms used to perform the identification and the Θ(k) = Θ(k − 1) + (y(k) − Θ(k − 1)T Φ(k)),
1+ξ
control of the plants. Section 3 is dedicated to the analy- (4)
sis and discussion of the results. Finally, section 4 makes where ξ = Φ(k)T C(k − 1)Φ(k), and C(k) is the co-
concluding remarks. variance matrix of the regression vector Φ(k) which is
updated at each iteration, k, using equation (5)
2 STC Methodologies C(k − 1)Φ(k)Φ(k)T C(k − 1)
C(k) = C(k − 1) − , (5)
ε−1 + ξ
STC algorithms can be divided in two categories. If the
identification is explicit then controllers that use the trans- where ε = ϕ(k − 1) − 1−ϕ(k−1) ξ and ϕ(k − 1) is the
fer function to determine the gains of the controller can forgetting factor at iteration (k − 1).
be applied. This means that the identification algorithm The adaption of ϕ is performed as follows:
and the controller algorithm can be chosen independently.
On the other hand, implicit controllers do not translate the 1
ϕ(k) = o,
plant’s dynamics into a transfer function, and that means
n h i
(ν(k)+1)η ξ
1 + (1 + ρ) ln(1 + ξ) + 1+ξ+η
−1 1+ξ
that the controller must be created specifically to the out- (6)
put of that identification algorithm. The advantage of im- where ν(k) = ϕ(k − 1)(ν(k − 1) + 1),
plicit algorithms is that they require less processor time. (y(k)−ΘT (k−1)Φ(k))2
In this paper r(k) represents the input reference and the η = λ(k) , λ(k) = ϕ(k −
h i
tracking error is given by e(k) = r(k) − y(k). 1) λ(k − 1) + (y(k)−Θ(k−1)T Φ(k))2
, and ρ is posi-
1+ξ
tive constant.
2.1 Explicit Identification for STCs
In LSMadf, the forgetting factor ϕ(k) and the variables
When using explicit STCs, it is necessary to estimate
λ(k) and ν(k) are automatically adjusted, so the initial
the plant’s transfer function in real time. If this is per-
values of this variables do not have much impact in the
formed recursively it allows the model of the plant to
identification process. In any case, they should be set be-
adapt whenever the real plant’s dynamics change. In [5]
tween zero and one.
the LSM identification algorithm with adaptive directional
forgetting (LSMadf) is presented, which uses a forget-
2.2 Control Algorithms for Explicit Identification
ting factor that is automatically adjusted depending on the
changes of the input and output signals. A brief overview of the five tested STC controllers is
The methods based on LSM perform discrete on-line presented in the following items:
explicit identification of a plant producing a transfer func- • Dahlin PID Controller [8]:
tion of the form
This algorithm is based on a transfer function with the
B(z −1 ) b1 z −1 + b2 z −2 + ... + bm z −m form of (1) with n = 2 and m = 1. Thus, the estima-
G(z) = = z −d , tion vector is Θ(k − 1) = [â1 , â2 , b̂1 ]T and the regression
A(z −1 ) 1 + a1 z −1 + a2 z −2 + ... + an z −n
(1) vector is Φ(k) = [−y(k − 1), −y(k − 2), u(k − 1)]T . The
where m, n ∈ N are the input and output orders of the control law of the Dahlin’s algorithm is given by
system, respectively, and d ∈ N is the time-delay. Thus, 
T0
u(k) = Kp + e(k) − e(k − 1) + e(k)+
A(z −1 )y(k) = B(z −1 )u(k), (2) TI
 (7)
TD
where u(·) : N → R and y(·) : N → R are the process + [e(k) − 2e(k − 1) + e(k − 2)] + u(k − 1),
T0
input and output, respectively.
The estimated output of the identified plant is given by where T0 is the sampling interval, and Kp , TI , TD are
the proportional gain, the integral time constant, and the
ŷ(k) =ΘT (k − 1)Φ(k) = −â1 y(k − 1) − ... − ân y(k − n)+ differential time constant, respectively, which depend of
the model parameters as follows:
+ b̂1 u(k − d − 1) + ... + b̂m u(k − d − m), (3)
(â1 + 2â2 ) Q
Kp = , (8)
where vector Θ(k−1) = [â1 , ..., ân , b̂1 , ..., b̂m ]T contains b1
the estimate of the process’s parameters from the last iter- T0
TI = − 1 TD
, (9)
ation, and Φ(k) = [−y(k − 1), ..., −y(k − n), u(k − d − â1 +2â2
+1+ T0
1), ..., u(k − d − m)]T is the regression vector which con- T0 â2 Q
tains the input and output information. TD = , (10)
KP b̂1
T0
• Least Squares Method With Adaptive Directional For- where Q = 1 − e− B and B is a positive constant. In
getting [5]: this algorithm, B is an adjustment factor that specifies the
The LSMadf is an evolved form of LSM where a forget- dominant time constant of the transfer function according
ting factor is used to give less weight to older data, and to changes made to the process output of a closed control
this forgetting factor is automatically updated at each iter- loop. The smaller the B gets, the quicker the step response
ation. In this method the vector of parameter estimations of the closed control loop becomes.
• Pole Placement [13]: where the controller’s coefficients p1 , p2 , q0 , q1 and q2 are
given by
This Pole Placement algorithm requires that the user ad-
justs the natural frequency (ωn ) and damping factor (ξ)   1 0 b̂1 0 0
−1 
p1 −â1

to control a second order plant with n = 2 and m = 2 p2   â1 1 b̂2 b̂1 0 −â 
which means that this algorithm’s estimation vector is    2
(23)
  
 q0  =  â 2 â1 b̂3 b̂2 b̂1  −â3  ,
Θ(k − 1) = [â1 , â2 , b̂1 , b̂2 ]T and the regression vector q      0 
1 â3 â2 0 b̂3 b̂2 
is Φ(k) = [−y(k − 1), −y(k − 2), u(k − 1), u(k − 2)]T . q2 0 â3 0 0 b̂ 0
3
The control law is given by
and r0 = 1/(b̂1 + b̂2 + b̂3 ).
u(k) =q0 e(k) + q1 e(k − 1) + q2 e(k − 2)+
(11) • Ziegler-Nichols with Forward Rectangular Discretiza-
+ (1 − γ)u(k − 1) + γu(k − 2),
tion (ZN) [14]:
where the coefficients q0 , q1 and q2 can be calculated by The experimental tuning of parameters for a continuous-
time PID controller designed by Ziegler and Nichols 70
1 years ago is still a good option. The algorithm is based on
q0 = (d1 + 1 − â1 − γ), (12)
b̂1 a third order system with n = 3 and m = 3. Thus, the
! estimation vector is Θ(k − 1) = [â1 , â2 , â3 , b̂1 , b̂2 , b̂3 ]T
q1 =
â2
− q2 −
â1 b̂1
+1 , (13)
and the regression vector is Φ(k) = [−y(k − 1), −y(k −
b̂2 b̂2 â2 2), −y(k − 3), u(k − 1), u(k − 2), u(k − 3)]T . The control
s1 law is given by
q2 = , (14)
r1 u(k) = q0 e(k) + q1 e(k − 1) + q2 e(k − 2) + u(k − 1), (24)
where where the controller’s coefficients q0 , q1 and q2 are given
(
−2e −ξωn T0
p
cos(ωn T0 p1 − ξ 2 ), if ξ ≤ 1,
by
d1 = (15)
−ξωn T0
if ξ > 1,
 
−2e cosh(ωn T0 ξ 2 − 1), T0 TD
q0 = KP 1+ + , (25)
TI T0
d2 = e−2ξωn T0 , (16)  
TD
r1 = (b̂1 + b̂2 )(â1 b̂1 b̂2 − â2 b̂21 − b̂22 ), (17) q1 = −KP 1 + 2 , (26)
T0
s1 = â2 [(b̂1 + b̂2 )(â1 b̂2 − â2 b̂1 )+ TD
q2 = KP , (27)
+ b̂2 (b̂1 d2 − b̂2 d1 − b̂2 )], (18) T0

b̂2 where the proportional gain is KP = 0.6KPu , the integral


γ = q2 , (19) time constant is TI = 0.5Tu and the differential time con-
â2
stant is TD = 0.125Tu . This is a Ziegler-Nichols based
and T0 is the sampling interval. algorithm, thus it is required to determine the ultimate pro-
portional gain KPu and the ultimate period of oscillations
• Deadbeat Controller of Second Order (DB2) [7]: Tu . Figure 1 explains how these parameters can be calcu-
This controller is based on a second order plant with n = lated.
2 and m = 2 which means that this algorithm’s estimation
vector is Θ(k − 1) = [â1 , â2 , b̂1 , b̂2 ]T and the regression 2.3 Implicit STC
vector is Φ(k) = [−y(k − 1), −y(k − 2), u(k − 1), u(k − A brief overview of the three implicit STC controllers
2)]T . The control law is given by tested is presented in the following items:
u(k) = r0 r(k) − q0 y(k) − q1 y(k − 1) − p1 u(k − 1). (20) • Single Neuron (SN) [11]:
The Single Neuron algorithm here described is a self adap-
where the controller’s coefficients q0 , q1 and p1 are given tive PID controller that has a simple structure and requires
by few computation effort. The control law is given by
   −1  
p1 1 b̂1 0 −â1
u(k) = u(k − 1) + KP x1 (k) + KI x2 (k) + KD x3 (k), (28)
q0  = â1 b2 b̂1  −â2  , (21)
q1 â2 0 b̂2 0 where
and r0 = 1/(b̂1 + b̂2 ). x1 (k) = e(k), x2 (k) = ∆e(k), x3 (k) = ∆2 e(k). (29)

• Deadbeat Controller of Third Order (DB3) [7]: The proportional gain KP , the integral gain KI , and the
differential gain KD are given by
For Deadbeat control on a third order system with
n = 3 and m = 3, the estimation vector is Θ(k − KP = Kw1 (k), KI = Kw2 (k), KD = Kw3 (k), (30)
1) = [â1 , â2 , â3 , b̂1 , b̂2 , b̂3 ]T , and the regression vector is where K is a positive scale parameter that can be in-
Φ(k) = [−y(k−1), −y(k−2), −y(k−3), u(k−1), u(k− creased/decreased to adjust the responsiveness of the con-
2), u(k − 3)]T . The control law is given by troller. The coefficients wi (k) are given by
u(k) =r0 r(k) − q0 y(k) − q1 y(k − 1)− wi (k)
(22) wi (k) = P3 , (31)
− q2 y(k − 2) − p1 u(k − 1) − p2 u(k − 2), i=1 |wi (k)|
where 0 < η < 1 is the learning rate,

k−1
X αi−k+L+1 (k)(u(k) − u(i))K(x(k), x(i))
∂ ŷ
(k) = ,
∂u σ2
i=k−L
(42)
where L is the size of the sliding window, sample u(i)
is the command signal stored at component n + 1 of the
sliding window vector at instant i, x(i) (44),

− kx(i) − x(j)k2
 
K(x(i), x(j)) = exp , (43)
σ2

is the RBF used in the kernel function of the LSSVM, σ


is the bandwidth of the RBF,

x(k) = [y(k), ..., y(k −n+1), u(k), ..., u(k −m+1)]T , (44)

and
α(k) = U(k)(Y(k) − 1v b(k)), (45)
th
where αi−k+L+1 (k) is the (i − k + L + 1) element of
vector α(k),
Figure 1: Ziegler-Nichols method: algorithm to determine the
ultimate proportional gain KPu and the ultimate period of oscil- 1Tv U(k)Y(k)
b(k) = , (46)
lations Tu . 1Tv U(k)1v
where 1v = [1, . . . , 1]L×1 , Y(k) = [y(k), . . . , y(k − L +
and are obtained through normalization of the weight co- 1)]T ,
efficients
   −1
∂y(k) A(k) H
wi (k) = wi (k − 1) + ηi Ke(k)xi (k − 1)sgn , (32) U(k) = , (47)
∂i∗ (k) HT h
where ηi is the learning rate of the weight coefficient H = [K(x(k − L), x(k − 1)), · · · ,
wi (k), and sgn(·) is a signal function. The current ref- K(x(k − L), x(k − L + 1))]T , (48)
erence of the single neuron i∗ (k) is given by
3
X where h = K(x(k − L), x(k − L)) + C −1 , and A(k) is
i∗ (k) = i∗ (k − 1) + K w̄i (k)xi (k). (33) given by (54). C is a positive regularization factor, and if
i=1 its value is low, then the outlier points are deemphasized.
and ∂y(k)/∂i (k) = (y(k)−y(k−1))/(i∗ (k)−i∗ (k−1)).

• Least Squares Support Vector Machine with Kernel


• Least Squares Support Vector Machine [10]: Tuning [9]:
In the Least Squares Support Vector Machine (LSSVM)
adaptive PID Controller, the PID parameters are adjusted The Least Squares Support Vector Machine with Kernel
using the gradient information of LSSVM to perform on- Tuning (LSSVMKT) adaptive PID controller is an evolu-
line implicit identification. The control law of this method tion of the LSSVM controller. The main difference is the
is given by ability to adjust the LSSVMKT kernel bandwidth (σ) as
follows:
u(k) = u(k−1)+KP xc1 (k)+KI xc2 (k)+KD xc3 (k), (34)
σ(k + 1) = σ(k) + ∆σ(k), (49)
where,
where
xc1 (k) = ∆e(k), xc2 (k) = e(k), xc3 (k) = ∆2 e(k). (35) ∂ ŷ(k)
∆σ(k) = η(k)êm (k) , (50)
∂σ(k)
The proportional gain KP (k+1), the integral gain KI (k+
1), and the derivative gain KD (k + 1) are given by k−1
X  αi−k+L+1 (k)K(x(k), x(i))
∂ ŷ(k)
KP (k + 1) = KP (k) + ∆KP (k), (36) = (x(k)−
∂σ(k) σ(k)3
KI (k + 1) = KI (k) + ∆KI (k), (37)
i=k−L
 (51)
KD (k + 1) = KD (k) + ∆KD (k), (38) T
− x(i)) (x(k) − x(i)) ,
where
∂ ŷ êm (k) = y(k) − ŷ(k), (52)
∆KP (k) = ηe(k) (k)xc1 (k), (39)
∂u k−1
∂ ŷ X
∆KI (k) = ηe(k) (k)xc2 (k), (40) ŷ(k + 1) = αi−k+L+1 (k)K(x(k), x(i)) + b(k). (53)
∂u i=k−L
∂ ŷ
∆KD (k) = ηe(k) (k)xc3 (k), (41)
∂u
 
K(x(k − 1), x(k − 1)) + C −1 ··· K(x(k − L + 1), x(k − 1))
A(k) =  .. .. .. (54)
.
 
. . 
K(x(k − 1), x(k − L + 1)) ··· K(x(k − L + 1), x(k − L + 1)) + C −1

Table 1: Statistical comparison between all controllers studied


in this work.
IAE ITAE ISE RMS Points
Dahlin 872 (2) 117667 (3) 27860 (2) 166.9 (2) 9 (2)
Pole Placement 973 (5) 124279 (5) 28703 (3) 169.4 (3) 16 (3)
Communication DB2 867 (1) 117412 (2) 27717 (1) 166.5 (1) 5 (1)
DB3 994 (7) 153091 (8) 28906 (5) 170.0 (5) 25 (7)
ZN 1113 (8) 121935 (4) 30349 (7) 174.2 (7) 26 (8)
SN 974 (6) 112145 (1) 33386 (8) 182.7 (8) 23 (5)
PLC LSSVM 961 (4) 142252 (7) 29949 (6) 173.1 (6) 23 (5)
LSSVMKT 917 (3) 138017 (6) 28891 (4) 170.0 (4) 17 (4)

Relay and was set to achieve a reference speed of 100 [pp/(0.25


seg)] (pulses per 250 milliseconds). After 20 seconds the
Lamps
Power
reference speed changed to 120 [pp/(0.25 seg)], and at 60
Source seconds it changed again to 90 [pp/(0.25 seg)]. The relay
Controlled Load
Motor
was turned on at 40 seconds (increasing the load of the
Motor
generator), and was turned off at 80 seconds.
Figure 2: Photo of the setup used to perform the experiments.
3.2 Control Algorithms Comparison
Figure 3 shows the output speed of the real DC mo-
3 Results and Discussion tor under the control of the studied control algorithms. It
shows that all the controllers were able to properly fol-
This section discusses the results obtained when the low reference changes and that they were able to com-
adaptive algorithms were set to control a real plant. The pensate variations on the load of the motor. Since all the
performances of the controllers are compared using four controllers performed similarly the IAE, ITAE, ISE, RMS
different statistical indices, the Integral Absolute Error numerical indices, eqs. (54), were used to compare the
(IAE), the Integral Time-weighted Absolute Error (ITAE), controllers performances.
the Integral Square Error (ISE), and the Root Mean Square Table 1 presents the results of the application of these
(RMS), which are defined as follows: indices for all control algorithms. Each controller received
N
X N
X a score for each numerical index based on its performance
IAE = |e(k)|, IT AE = k|e(k)|, (the best received 1 and the worst received 8) and the
k=1 k=1 best controller was the one which summed least points.
With just 5 points, the Deadbeat controller of second order
v 
u P
N 2
k=1 e(k)
N
achieved the best score. Figures 4(a) and 4(b) shown the
u
X
e(k)2 , (54)
t
ISE = RM S = ,
N results of the Deadbeat controller of second order. Figure
k=1
4(a) shows how the output of the plant and the control sig-
where N is the number of samples (time instants). nal change when the reference changes, and when a vari-
ation on the motor load is introduced. Figure 4(b) shows
3.1 Plant the time evolution of the plant’s estimated parameters.
A system composed of two motors, a shaft coupler, a Besides controller performance, simplicity of tunning
motor driver, a relay, two lamps, a programmable logic is another important feature that was pursued. The explicit
controller (PLC), a computer (running Scilab) and a power identification algorithms LSMadf have two variables that
source was used to test the control algorithms. The com- need to be tuned, the initial gain of the covariance matrix,
puter and the PLC were connected using the OPC (OLE and the forgetting factor ρ. Neither of them is much sen-
(Object Linking and Embedding) for Process Control) sitive and a satisfactory tuning of these variables is easy
communication protocol. Figure 2 outlines the connec- to obtain. The Deadbeat algorithms (of second and third
tions between all the components of the setup. One of the orders) and Ziegler-Nichols do not have any variable to be
motors receives command signals, and the other works as adjusted (obviously the variables from the explicit iden-
a generator. The control signal can be varied in the in- tifications still need to be adjusted), which means they
terval from 0 to 100 (percentage), which corresponds to a are easier to install. The Dahlin and Single Neuron algo-
variation from 0 to 12 Volts. The lamps are connected to rithms, both have a scale parameter to increase/decrease
the terminals of the generator and since they consume en- the responsiveness of the controller which is also easy to
ergy, they increase its load. The relay is used to turn on/off adjust. The Pole Placement algorithm has two variables
the lamps/load. The tests consisted of running all the con- that need to be adjusted, the natural frequency ωn , and
trol algorithms during 100 seconds with a sampling inter- the damping factor ξ, which makes it a bit more chal-
val of 250 milliseconds. The motor always started in rest lenging for the installer. The algorithms LSSVM and
Figure 3: Result of the test with all the algorithms controlling a real DC Motor with a varying load.

Data” (reference: SCIAD/2011/21531) co-financed by


QREN, in the framework of the “Mais Centro - Regional
Operational Program of the Centro”, and by the European
Union through the European Regional Development Fund
(ERDF).

References
(a) Speed and control signal. (b) Identified coefficients. [1] A. Ajiboye and R. Weir. A heuristic fuzzy logic approach
to emg pattern recognition for multifunctional prosthesis
Figure 4: Result of the real test using the Deadbeat controller of control. IEEE Transactions on Neural Systems and Reha-
second order using LSM with adaptive directional forgetting. bilitation Engineering, 13(3):280–291, September 2005.
[2] K. J. Åström and T. Hägglund. PID Controllers: Theory,
Design, and Tuning. Instrument Society of America, Re-
search Triangle Park, NC, USA, 1995.
LSSVMKT revealed to be the most difficult to adjust. Not [3] K. J. Astrom and B. Wittenmark. Adaptive Control.
only both algorithms have six variables that need to be Addison-Wesley, Boston, MA, USA, 2nd edition, 1994.
adjusted (which means that the installer needs to have a [4] P. Bashivan and A. Fatehi. Improved switching for multi-
deeper understanding of the controller) but the calibration ple model adaptive controller in noisy environment. Jour-
of these variables also revealed to be more sensitive and nal of Process Control, 22(2):390–396, 2012.
difficult. [5] V. Bobál, J. Böhm, J. Fessl, and J. Macháček. Digital Self-
tuning Controllers: Algorithms, Implementation and Ap-
plications. Springer, London, UK, 2005.
4 Conclusions [6] P. K. Kolavennu, S. Palanki, D. A. Cartes, and J. C. Telotte.
Adaptive controller for tracking power profile in a fuel
In this work, several adaptive PID controllers, STCs cell powered automobile. Journal of Process Control,
with a PID structure, that can be used to control un- 18(6):558–567, 2008.
[7] V. Kučera. A dead-beat servo problem. International Jour-
known plants in industry were tested and compared. The
nal of Control, 32(1):107–113, 1980.
controllers were tested on a real DC motor with a vary- [8] V. Kučera. Analysis and Design of Discrete Linear Con-
ing load, and their performance was mathematically an- trol Systems. Prentice-Hall, Inc., Upper Saddle River, NJ,
alyzed. The tested algorithms were STCs with either USA, 1991.
implicit or explicit identification (the later requiring in- [9] K. Ucak and G. Oke. Adaptive pid controller based on on-
dependent identification algorithms). The employed ex- line lssvr with kernel tuning. In Proc. International Sym-
plicit identification method was the LSMadf, and had a posium on Innovations in Intelligent Systems and Applica-
good performance. Among the control algorithms, the one tions (INISTA 2011), pages 241–247, June 2011.
[10] S. Wanfeng, Z. Shengdun, and S. Yajing. Adaptive pid
which performed better was the Deadbeat of second order,
controller based on online lssvm identification. In Proc.
followed by the Dahlin’s controller, and the third best was IEEE/ASME International Conference on Advanced In-
the LSSVRKT. Besides having the best performance, the telligent Mechatronics (AIM 2008), pages 694–698, July
Deadbeat of second order and Dahlin, were also very easy 2008.
to tune to a satisfactory performance. The LSSVMKT was [11] M. Wang, G. Cheng, and X. Kong. A single neuron self-
much more difficult to tune. adaptive pid controller of brushless dc motor. In Proc.
Third International Conference on Measuring Technol-
ogy and Mechatronics Automation (ICMTMA 2011), vol-
Acknowledgment ume 1, pages 262–266, January 2011.
[12] P. E. E. Wellstead and M. B. Zarrop. Self-Tuning Systems:
This work was supported by Project SCIAD “Self- Control and Signal Processing. John Wiley & Sons, Inc.,
Learning Industrial Control Systems Through Process New York, NY, USA, 1st edition, 1991.
[13] B. Wittenmark. Self-tuning PID-controllers Based on Pole
Placement. Department of Automatic Control, Lund Insti-
tute of Technology, 1979.
[14] J. G. Ziegler and N. B. Nichols. Optimum settings for au-
tomatic controllers. Transactions of ASME, 64:759–768,
1942.