See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/325652997

Model-free sliding mode control, theory and application

Article  in  Proceedings of the Institution of Mechanical Engineers Part I Journal of Systems and Control Engineering · June 2018
DOI: 10.1177/0959651818780597

CITATIONS READS

10 761

3 authors, including:

Amin Ramezani
Tarbiat Modares University
110 PUBLICATIONS   456 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Application of Model Predictive controller (MPC) in Linear Time Varying systems (LPVs) with Input-Output framework View project

Deep Visual-based Model Predictive Controller using Recurrent Neural Networks for Wheeled Mobile Robots View project

All content following this page was uploaded by Amin Ramezani on 05 October 2020.

The user has requested enhancement of the downloaded file.

Original Article

Proc IMechE Part I:

J Systems and Control Engineering
1–10
Model-free sliding mode control, Ó IMechE 2018
Reprints and permissions:
theory and application sagepub.co.uk/journalsPermissions.nav
DOI: 10.1177/0959651818780597
journals.sagepub.com/home/pii

Nahid Ebrahimi, Sadjaad Ozgoli and Amin Ramezani

Abstract
In this article, a novel data-driven sliding mode controller for a single-input single-output nonlinear system is designed
from a new perspective. The proposed controller is model-free, that is, it is based on just input and output data.
Therefore, it is suitable for systems with unknown models. The approach to design the controller is based on an optimi-
zation procedure. First, a linear regression estimation is assumed to exist for the system behavior. Then an optimal con-
troller is designed for this estimated model. The cost function is proposed in a way that minimization of it, could
guarantee that the sliding function and its first derivative converge to zero. Based on rigorous theoretical analysis,
boundedness of the tracking error is then proved. Uncertainty is then considered and the control law is modified to
cope with it. To demonstrate the validity and the performance of the proposed method in different situations, different
computer simulations and experimental tests have been provided. Results show the effectiveness of the proposed
method for different systems in different situations.

Keywords
Model-free adaptive control, sliding mode control, data-driven control, tracking error convergence, optimization

Date received: 10 December 2017; accepted: 22 April 2018

Introduction without explicit usage of its mathematical model. In

general, DDCs are categorized into two classes based
During the past decades, sliding mode control (SMC) on their controller structure: the first class is data-
has received considerable attention among researchers driven methods with a pre-specified controller struc-
owning to its several advantages such as robustness to ture. In this class, the controller parameters are calcu-
uncertainty and external disturbances, fast convergence lated based on some optimization procedures.7 Iterative
rate and good transient response.1–3 feedback tuning,8,9 virtual reference feedback tun-
In general, the SMC design is composed of two ing10,11 and unfalsified control12–14 are some examples
steps: (1) define a sliding surface and (2) design the of this category. The main obstacle for applying this
proper control law in order to force the system trajec- class is to determine the proper controller structure. In
tories to converge to the predefined surface. In the clas- addition, this category is not supported by stability
sic SMC design procedure, the input–output or state- analysis.6
space model of the process is needed. However, in In the second type, the controller structure is not
many cases, the rigorous process model is not available. needed to be known. Iterative learning control,15–17
Even if available, it is often a complicated model, may subspace predictive control18,19 and model-free adap-
be along with uncertainties and time-varying behavior. tive control6 are some examples of this type. Compared
Complicated models lead to complex high-order con- with the first class, these methods are more robust when
trollers which are difficult to design, to implement, to they are applied to practical systems.7
maintain and to diagnose.4,5
On the other hand, development of sensor technol-
ogy has enabled us to produce and to store a large Faculty of Electrical & Computer Engineering, Tarbiat Modares University,
amount of valuable data, including useful information Tehran, Iran
of equipment and operations. So researchers motivate
Corresponding author:
to develop data-driven controllers (DDCs).6 Data-dri- Sadjaad Ozgoli, Faculty of Electrical & Computer Engineering, Tarbiat
ven control is the synthesis of a controller using merely Modares University, Tehran, Iran.
online input and output data of the control system Email: [email protected]
2 Proc IMechE Part I: J Systems and Control Engineering 00(0)
   
Model-free adaptive control (MFAC) is a new yðk + 1Þ = f yðkÞ, . . . , y k  ny , uðkÞ, . . . , uðk  nu Þ
approach in this class. The main idea of MFAC is ð1Þ
based on applying a virtual equivalent dynamic linear
model at each operating point instead of identifying a where ny and nu are unknown orders of the system out-
global nonlinear model.7 This method uses a concept put, y and the control input, u. The mapping
called pseudo partial derivative (PPD), which estimates f(:) : Rny + nu + 2 ! R is an unknown nonlinear function.
the mentioned model using only input–output data. Throughout the article, we consider that the nonlinear
In comparison to other data-driven methods with system equation (1) satisfies the following assumptions:
unknown controller structure, MFAC has several
advantages that make it more suitable for practical Assumption 1. System equation (1) satisfies the general-
applications: it depends only on the real-time input– ized Lipschitz condition, that is, jDy(k + 1)j ł bjDu(k)j
output data of the systems, that is, it does not need any for any fixed k and jDu(k)j 6¼ 0, where b is a positive
external testing signal and training process; and it is constant.
simple and can be easily applied.6 Since now, it is
implemented to several nonlinear and complex systems
such as solid oxide fuel cell,20 wide area power system21 Assumption 2. The partial derivatives of f(:) are continu-
and artificial heart control.22 ous with respect to control input, u(k).
In this article, a novel model-free sliding mode con-
troller is proposed based on an approach similar to the Remark 1. It should be noted that these two assumptions,
MFAC approach. The sliding controller is designed are not restricting and are reasonable from a practical
from a new viewpoint, the optimization. For this, a point of view. Both of them are satisfied in many practi-
suitable cost function is considered that forces the slid- cal applications and are typical constraints accepted in
ing variable and its first derivative to vanish. In order the field of data-driven control design. Assumption 1 is
to approximate the model, MFAC is employed. The actually a typical condition met in several design methods
proposed method only depends on input and output for nonlinear systems. Assumption 2 poses an upper lim-
measurement data to build a regression model. itation on the change rate of the system output, driven by
Therefore, the system model is not a requirement for the change of the control input. From an energy perspec-
the controller design procedure. The uncertainty effect tive, the energy rate increasing inside a system cannot go
is also considered and the controller is designed in a to infinity if the energy change of the input is finite.6
manner that copes with it. The main advantage of the
proposed method over others is that the proposed
method in this article provides a novel approach for Lemma 1.7 Consider that the nonlinear system equation
designing model-free SMC. This approach is based on (1) satisfies assumptions 1 and 2, if jDu(k)j 6¼ 0, then
optimizing a cost function that its parameters and its there exists a time-varying parameter c (k) 2 R called
structure are ‘‘design parameters,’’ that is, the design the PPD such that the system equation (1) can be trans-
engineer can get other forms of SMC by changing the formed in the compact form dynamic linearization
cost function. For example, in order to further improve, (CFDL), that is
the designer can select the coefficients of cost function
to be time varying or adaptive with respect to SMC. Dyðk + 1Þ = c ðkÞðuðkÞ  uðk  1ÞÞ ð2Þ
Based on theoretical analysis, the boundedness of where c (k) is a bounded parameter and
the tracking error is guaranteed. Different numerical Dy(k + 1) = y(k + 1)  y(k).
simulations and experimental tests are done to validate
the effectiveness of the proposed method.
The rest of the article is organized as follows: in sec- Proof. See Hou and Wang.7
tion ‘‘Problem statement and preliminaries,’’ problem Since the system model is unknown, the time-varying
formulation and some preliminaries are presented. PPD is unknown and must be estimated. The estima-
Section ‘‘Design of proposed controller’’ explains the tion algorithm is obtained by minimizing the following
design process of novel optimal model-free SMC and cost function with respect to c (k).6
the corresponding stability analysis. The simulation
and experimental results are reported in ‘‘Simulation J1 ðc ðkÞÞ = jyðkÞ  yðk  1Þ  c ðkÞDuðk  1Þj2
2
studies’’ and ‘‘Experimental speed control of a DC + mjc ðkÞ  ^c ðk  1Þj
motor’’ sections. Finally, concluding remarks are made ð3Þ
in section ‘‘Conclusion.’’
where m . 0 is a weighting factor, and ^c (k) is the esti-
mation of c (k).
Problem statement and preliminaries
By minimizing the cost function equation (3) with
A general single-input single-output nonlinear system is respect to c (k) and setting it to be zero, the estimation
considered as is derived as follows6
Ebrahimi et al. 3

hDuðk  1Þ By considering equation (7), the cost function can be

^
c ðkÞ = ^
c ð k  1Þ +
m + Duðk  1Þ2 ð4Þ rewritten as
ðDyðkÞ  ^
c ðk  1ÞDuðk  1ÞÞ
J2 = ðð1 + lÞeðk + 1Þ  ð2 + lÞeðkÞ + eðk  1ÞÞ2
where h 2 (0, 2) is added to make the estimation algo- + bðð1 + lÞeðk + 1Þ  eðkÞÞ2 ð9Þ
rithm more general and to have more choice. In addi-
tion, it will facilitate the convergence proof of the Consider the output-tracking error e(k + 1) =
estimation algorithm.6 In order to make the parameter y(k + 1)  yd (k + 1) and equation (2), by finding the
estimation algorithm equation (4) have strong ability to optimal controller for this, that is, ∂J=∂u(k) = 0, the
track time-varying parameter, a reset mechanism is optimal control law can be obtained as stated in equation
considered as (6). Substituting this control law into the error equation
^ and considering equation (2), it can be concluded that
c ðkÞ = ^c ð0Þ if j^c ðkÞj\e or jDuðkÞj\e or signð^c ðkÞÞ
6¼ signð^c ð 0Þ Þ ð2 + b + lÞ 1
eðk + 1Þ = eðkÞ  eðk  1Þ
ð5Þ ð1 + bÞð1 + lÞ ð1 + bÞð1 + lÞ
ð10Þ
Where e . 0 is a small positive constant.
Note that the reset algorithm is intended to keep the Then by considering the sliding surface equation (7)
scheme alive at all time. In addition, it is added to the and according to equation (10), it can be deduced that
estimation algorithm in order to make the condition
SðkÞ
Du(k) 6¼ 0 be satisfied. Sðk + 1Þ = ð11Þ
1+b

Design of the proposed controller Since b . 0, we have

In this section, the main work of the article is expressed jSðk + 1Þj \ jSðkÞj ð12Þ
via the following two theorems. In the first theorem, a which leads to the convergence of S(k) to zero. Now, if
model-free controller is proposed based on optimiza- we consider equation (7), it can be deduced that
tion of a proper cost function. This cost function is
designed to force a pre-specified sliding variable and its e ð k  1Þ
first derivative to vanish. In the second theorem, the eðkÞ = ð13Þ
1+l
robust version of the controller is designed, and it is
shown that the tracking error is ultimately bounded. Therefore, as k tends to infinity, the tracking error
converges to zero.

Design of optimal model-free controller

Design of robust controller
Theorem 1. The following control law
The proposed model-free controller equation (6) as
yðkÞ  yd ðk + 1Þ designed in the previous section leads to the convergence
uopt ðkÞ = uopt ðk  1Þ  of tracking error to the origin. However, it is designed
^
c ð k Þ
ð6Þ without considering the uncertainty effects. In order to
ð2 + b + lÞeðkÞ  eðk  1Þ
+ cope with the uncertainties effect, it is assumed that the
ð1 + bÞð1 + lÞ^c ð k Þ
effect of uncertainties and disturbances can be modeled as
guarantees that the states of nonlinear system equation the following form by adding d(k) to the equation (2)
(1) satisfying assumptions 1 and 2 converge to the fol-
Dyðk + 1Þ = c ðkÞðuðkÞ  uðk  1ÞÞ + dðkÞ ð14Þ
lowing surface
where d(:) is considered as a bounded signal satisfying
SðkÞ = ðeðkÞ  eðk  1ÞÞ + leðkÞ = 0 ð7Þ jd(k)jhD, 8ki0, and D is a positive constant.
where l is a positive constant, e(k) = y(k)  yd (k) is The main result is stated in theorem 2 and the fol-
output-tracking error, and yd (k) is the desired output. lowing Lemma is used to prove it.
In addition, the convergence of the output-tracking
error to zero is guaranteed. Lemma 2.23 Consider the scalar dynamic system

zðk + 1Þ = zðkÞ + gðkÞ  esignðzðkÞÞ ð15Þ

Proof. Assume the following cost function
If jg(k)j \ d, d . 0 and d \ e, then there is a finite
J2 = ðSðk + 1Þ  SðkÞÞ2 + bSðk + 1Þ2 ð8Þ time number, K . 0 such that
4 Proc IMechE Part I: J Systems and Control Engineering 00(0)

jzðkÞj ł e + d \ 2e 8k ø K ð16Þ So, if the right-hand side of the inequality is

considered

Theorem 2. For the nonlinear system equation (1), satis- ð1 + lÞeðkÞ  eðk  1Þ ł M ð24Þ
fying assumptions 1 and 2 and by considering system Then this recursive equation leads to
uncertainties modeled as equation (14), if we define the !
sliding surface as equation (7) and apply the following X k
M e ð 0Þ
control law eðkÞ ł j + k
8k ø K ð25Þ
j=1 ð1 + l Þ ð1 + l Þ
yð k Þ  yd ð k + 1Þ
uð k Þ = uð k  1Þ  Now, since l . 0, when k tends to infinity, we have
^
c ðkÞ
ð2 + b + lÞeðkÞ  eðk  1Þ M
+ eð‘Þ ł ð26Þ
ð1 + bÞð1 + lÞ^ c ðkÞ l
b G For the case that (1 + l)e(k)  e(k  1) ø  M, the
+ SðkÞ  signðSðkÞÞ
ð1 + bÞð1 + lÞ^ c ð k Þ ^
c ð k Þ proof is similar and yields to e(‘) ø (  M=l), so it can
ð17Þ be concluded that

where G = (g=(1 + l)) + D and g is a positive con- M

jeð‘Þj ł ð27Þ
stant, then the tracking error is ultimately bounded. l
So the tracking error is ultimately bounded.
Proof. Consider the output-tracking error e(k + 1), by
applying control law equation (17), we obtain Remark 2. According to Hou and Jin,6 the parameters
h and m affect the convergence of estimation algorithm.
ð2 + l + bÞ 1
eðk + 1Þ = eðkÞ  In Hou and Jin,6 it is proved that the estimation error,
ð1 + lÞð1 + bÞ ð1 + lÞð1 + bÞ
~c (k) = ^c (k)  c (k), satisfies the following equation
b
eðk  1Þ+ SðkÞ  GsignðSðkÞÞ+dðkÞ
ð1 + lÞð1 + bÞ j~c ðkÞj ł d1 j^c ðk  1Þj + 2b ł . . . ł d k1 ~
1 j c ð 1Þ j
ð18Þ  
2b 1  d k11
+
As a consequence 1  d1
ð28Þ
ð 2 + l + bÞ 1
Sðk + 1Þ = eð k Þ  e ð k  1Þ
ð 1 + bÞ ð 1 + bÞ where
b
+ SðkÞ  Gð1 + lÞsignðSðkÞÞ hDu2 ðk  1Þ he21
ð 1 + bÞ 0 ł d1 ł 1  2
ł1  ð29Þ
m + jDuðk  1Þj m + e21
+ ð1 + lÞdðkÞ  eðkÞ
ð19Þ and b is the Lipschitz constant. So, it seems that by
selecting smaller m and bigger h, estimation error will
So decrease and faster convergence is obtained. However,
it may result in more oscillations.
1 b According to sliding surface equation (7), it can be
Sðk + 1Þ = SðkÞ + SðkÞ
ð 1 + bÞ ð1 + b Þ ð20Þ seen that choosing a larger l can decrease the reaching
 Gð1 + lÞsignðSðkÞÞ + ð1 + lÞdðkÞ time, but it can yield in more oscillations also. In addi-
tion, the parameter G can be selected according to
Then upper bound of uncertainties. Actually, larger values
for G can yield in faster tracking, but since it is the
Sðk+1Þ  SðkÞ =  ðg + Dð1 + lÞÞsignðSðkÞÞ
coefficient of the sign(.) term, it yields in more oscilla-
+ ð1 + lÞdðkÞ tions or the chattering phenomena.
ð21Þ

Since jd(k)j \ D and based on Lemma 2 and consid- Simulation studies

ering equation (21), it can be deduced that there is a In this section, in order to show the applicability of the
finite time number K . 0 such that proposed method for different systems several simula-
jSðkÞj ł M 8k ø K ð22Þ tions are provided. First, to show robustness of the pro-
posed method for uncertain systems, an uncertain
where M = g + 2(1 + l)D. Therefore nonlinear system with parametric uncertainty is simu-
lated. Then the performance of the proposed method for
 M ł ð1 + lÞeðkÞ  eðk  1Þ ł M 8k ø K ð23Þ controlling complex biological systems is demonstrated
Ebrahimi et al. 5

Figure 1. Tracking performance for Example 1. Figure 2. Control signal for Example 1.

by a real example: the tumor growth control for different Figure 3 illustrates the sliding function for this exam-
patients. Another study is also provided in the next sec- ple. As it can be seen, the sliding variable converges to
tion. It is an implementation of the proposed method on the origin and sliding mode is achieved.
a real setup for speed control of a DC motor. This is pro- In order to validate the proposed method, the simu-
vided in order to illustrate the effective performance of lation results are compared with proportional–integral–
the proposed method for practical applications. derivative (PID) controller. Figure 4 shows the tracking
performance of the proposed method compared to PID
controller.
Example 1 As it can be seen, the proposed method has a better
tracking performance with smaller tracking error and
Consider an uncertain nonlinear system as follows

yðkÞyðk  1Þyðk  2Þuðk  1ÞðyðkÞ  1Þ + ð1 + aðkÞÞuðkÞ

yð k + 1Þ = ð30Þ
1 + y2 ðkÞ + y2 ðk  1Þ + y2 ðk  2Þ

The time-varying parameter is assumed to be

a(k) = 4round(k=500) + sin(k=100).
smaller overshoots. To have a better comparison of the
The desired signal is considered to be
mentioned methods, the values of jjejj2 are also listed in
8  kp  Table 1.
>
< 0:3 sin 100 k ł 200
round ð k
Þ Table 1 shows quantitatively that tracking perfor-
yd ðk + 1Þ = 0:3ð1Þ 100 200 \ k ł 500
: 0:3 sin kp 
> mance is improved by the proposed method in compar-
100 k . 500 ison to PID controller.
ð31Þ

The controller parameters are set to be Example 2: A more realistic case

m = 0:9, h = 0:001, b = 0:9, l = 3, G = 0:001 and Cancer is one of the leading causes of death in the
the initial value of the PPD estimation is considered to world, and the necessity of tumor growth control is
be ^
(0) = 4. obvious. To date, different control methods are used
The simulation results are shown in Figures 1 and 2. for this purpose such as super-twisting algorithm,24
Figure 1 shows the tracking performance of the pro- robust model-predictive algorithm25 and optimal bang-
posed method, and Figure 2 shows the control signal. bang control.26 Among these methods, SMC is one of
Obviously, in spite of the time-varying and uncertain the most powerful methods, which has good tracking
structure, the proposed controller has a good tracking performance, high accuracy and disturbance rejection
performance, and the system output can track the ability. On the other hand, most of the biological sys-
desired output signal perfectly. In addition, the control tems have complicated behavior, and it is difficult to
signal has smooth behavior with no chattering. establish their accurate model. So in this section, the
6 Proc IMechE Part I: J Systems and Control Engineering 00(0)

Table 1. Norm-2 of tracking error jjejj2 for Example 1.

Proposed controller PID controller

Example 1 0.7956 2.0658

Table 2. Parameters value of tumor growth model.

Parameters Value

§ 0.5301 per day

b 5.8502 per day
d 0.0011 per mm2 per day
p0 500 mm3
q0 625 mm3
N 0.1027
C 0.184
a 0.0016 per day

Figure 3. Sliding variable for Example 1. where p is the tumor’s volume, q is the carrying capacity
of the vasculature. b and d are the stimulation and inha-
bitation parameters for the endothelial support, respec-
tively. § denotes a tumor growth parameter. The three
terms with negative signs in the second equation show
several kinds of inhibition. aq indicates the loss of vas-
cular support via natural causes. In general, a is negligi-
ble compared to other parameters. The endogenous
inhibitor is modeled by dqp2=3 that impacts its vascular
support. The last negative term Cqu models vascular
support loss because of giving outside anti-angiogenic
agents, and the variable u is the control signal which
indicates angiogenic dose rate. The variable C is the
angiogenic killing parameter.
Improvement process of Ledzewicz model includes
the derivation of the model coefficients by maximum gra-
dient algorithm and adding a linear term Np in model
structure to adapt with gathered experimental data.
The nominal system’s parameters are listed in Table 2.
To validate and check the robustness of the pro-
posed controller, different parameters (for § and b) for
different patients have been used. The parameter values
for different patients used in simulations are given in
Figure 4. Tracking performance of the proposed method and Table 3.
PID controller.
The simulation results are shown in Figures 5–7.
Figure 5 illustrates control signal which is applied for
proposed method is used to control the volume of a
one of the patients. Figure 6 shows the changes in the
cancer tumor. The goal is to impose tumor to decay. tumor’s volume that have been achieved by applying
Since the proposed controller does not require a
the proposed controller. Figure 7 shows endothelial
mathematical model of the controlled system, the fol-
volume cells.
lowing anti-angiogenic model is merely used to produce As seen, while the transient responses of different
input and output data in simulation. It is an improved patients are not the same, but the controller perfor-
version of the Ledzewicz model introduced in mance of the proposed method for all of them is satis-
Ledzewicz et al.27 factory and has a low tracking error without any
It is as the following oscillations; in other words, the tumor’s volume con-
(  verges to the healthy equilibrium point and the cancer
p_ =  §pln pq  Np tumor is destroyed in all cases.
2
ð32Þ
q_ = bp  dqp3  aq  Cqu In order to validate the tracking performance of the
proposed method, the simulation results of the
Ebrahimi et al. 7

Table 3. Parameter values for different patients.

§ b

Patient 1 0.1572 5.5760

Patient 2 0.2055 5.8497
Patient 3 0.2055 7.1984
Patient 4 0.2401 4.2397
Patient 5 0.197 5.8497

Figure 7. Volume of the endothelial cells.

Figure 5. Control signal.

Figure 8. Tumor’s volume achieved by the proposed method

and PID controller.

As seen, the proposed method has faster response

with smaller tracking error in comparison to PID
controller. To have a better comparison of the men-
tioned methods, the values of jjejj2 are also calculated
in Table 4.
The norm-2 of tracking error listed in Table 4 also
confirms better tracking performance of the proposed
method.
Figure 6. Changes in the tumor’s volume achieved by applying
the proposed controller.
Experimental speed control of a DC
motor
proposed method and PID controller are illustrated in In this section, to show that the proposed controller
Figures 8 and 9. is also applicable for real industrial systems, it is
8 Proc IMechE Part I: J Systems and Control Engineering 00(0)

Figure 11. Block diagram of the control system.

Figure 9. Volume of the endothelial cells by the proposed

method and PID controller.

Table 4. Norm-2 of tracking error jjejj2 for Example 2.

Proposed controller PID controller

Example 2 2:6 3 103 2:8 3 103

Figure 12. Tracking performance for Example 3.

(PCI1711u) is installed to send the control signal u(k)

and to receive the output signal y(k). Figure 11 illus-
trates the block diagram of the DC motor and its
controller.
The desired output is considered as yd (k+1) = 500
r= min , and the controller parameter are set to be
m = 0:5, h = 0:6, b = 0:1, l = 0:2 and G = 0:04.
The experimental results are shown in Figures 12
and 13. Figure 12 illustrates the tracking performance,
and Figure 13 shows the control signal.
As seen, in few seconds, good tracking performance
is obtained, and the motor’s speed receives to its desired
Figure 10. Experimental setup of the proposed method. value. In addition, it is obvious that the control signal is
a bounded and smooth signal.
In the next step, to evaluate the performance of
applied experimentally for speed control of a DC the proposed method in the presence of external distur-
motor (Figure 10). bances, a mechanical load is applied on the motor by a
The DC motor generates a torque about 0:059 Nm magnetic brake. Tracking performance and control sig-
at 2 A input current. The moment inertia is about nal in this case are shown in Figures 14 and 15.
3 3 105 kgm2 . The controller algorithm is applied by It is apparent that the proposed method acts
a Pentium Dual-Core, 2.7 GHz computer with robustly against external disturbance and can adapt
Windows 7. A tachometer measures the motor’s speed with variations. The control signal is smooth and
as the output signal. A data acquisition card bounded in this case too.
Ebrahimi et al. 9

Figure 13. Control signal for Example 3.

Figure 15. Control signal for Example 3 in the presence of
external disturbance.

the method in practice. The simulation and experimen-

tal results demonstrate that the proposed controller has
good performance for tracking and can cope with exter-
nal disturbances and practical uncertainties.

Acknowledgements
The authors would like to show their gratitude to Dr
Mojtaba Barkhordari from Shahid Bahonar University
who provided his Lab setup and his time for the imple-
mentation of the proposed method.

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest
with respect to the research, authorship and/or publica-
tion of this article.
Figure 14. Tracking performance for Example 3 in the
presence of external disturbance.
Funding
The author(s) received no financial support for the
Conclusion research, authorship and/or publication of this article.

In this article, an optimal model-free sliding mode con-

troller is proposed. The proposed controller is designed ORCID iD
based on optimization on a linear approximation of the Sadjaad Ozgoli https://orcid.org/0000-0001-8567-
system. The linear approximation is extracted as done 1679
in the well-known MFAC method. Unlike most of the
sliding mode controllers that use an accurate model of References
the system, this method merely uses the input and out-
1. Pai MC. Discrete-time output feedback quasi-sliding
put data of the system. The tracking performance of the
mode control for robust tracking and model following of
proposed method is analyzed, and the convergence of
uncertain systems. J Frankl Inst 2014; 351(5): 2623–2639.
the tracking error is proved. 2. Lu YS and Chiu CW. Global sliding-mode control with
To make sense with the performance of the proposed generalized sliding dynamics. Asian J Control 2009;
method, two simulation studies are provided. In addi- 11(4): 449–456.
tion, one experimental test (speed control of a DC 3. Pai MC. Robust discrete-time sliding mode control for
motor) has been performed to show the effectiveness of multi-input uncertain time-delay systems. Proc IMechE,
10 Proc IMechE Part I: J Systems and Control Engineering 00(0)

Part I: J Systems and Control Engineering 2012; 226(7): manufacturing applications. Control Eng Pract 2017;
927–935. 64(1): 74–87.
4. Hou Z, Chi R and Gao H. An overview of dynamic-line- 17. Huang SN, Tan KK and Lee TH. Iterative learning algo-
arization-based data-driven control and applications. rithm with a quadratic criterion for linear time-varying
IEEE Trans Ind Electron 2017; 64(5): 4076–4090. systems. Proc IMechE, Part I: J Systems and Control
5. Zhu Y and Hou Z. Controller dynamic linearisation- Engineering 2002; 216(3): 309–316.
based model-free adaptive control framework for a class 18. Luo X. Subspace predictive control for continuous-time
of non-linear system. IET Control Theory A 2015; 9(7): systems. Int J Contr Autom 2015; 8(7): 175–184.
1162–1172. 19. Zhang L, Xu SZ and Zhao HT. Adaptive subspace pre-
6. Hou Z and Jin S. Model free adaptive control: theory and dictive control with time-varying forgetting factor. Int J
applications. Boca Raton, FL: CRC Press, 2013. Autom Comput 2014; 11(2): 205–209.
7. Hou Z and Wang Z. From model based control to data 20. Xu D, Jiang B and Liu F. Improved data driven model
driven control, survey, classification and perspective. free adaptive constrained control for a solid oxide fuel
Inform Sci 2013; 235: 3–35. cell. IET Control Theory A 2016; 10(12): 1412–1419.
8. Meng W, Xie SQ, Liu Q, et al. Robust iterative feedback 21. Lu C, Zhao Y, Men K, et al. Wide-area power system sta-
tuning control of a compliant rehabilitation robot for biliser based on model-free adaptive control. IET Control
repetitive ankle training. IEEE/ASME T Mechatron Theory A 2015; 9(13): 1996–2007.
2017; 22(1): 173–184. 22. Chang Y, Gao B and Gu K. A model-free adaptive con-
9. Huusom JK, Hjalmarsson H, Poulsen NK, et al. A trol to a blood pump based on heart rate. ASAIO J 2011;
design algorithm using external perturbation to improve 57(4): 262–267.
iterative feedback tuning convergence. Automatica 2011; 23. Li S, Du H and Yu X. Discrete-time terminal sliding
47(12): 2665–2670. mode control systems based on Euler’s discretization.
10. Radac MB, Precup RE and Roman RC. Model-free con- IEEE T Automat Contr 2014; 59(2): 546–552.
trol performance improvement using virtual reference 24. Zioui N, Tadjine M and Boucherit MS. Super-twisting
feedback tuning and reinforcement Q-learning. Int J Syst algorithm approach to control tumors growth. In: Pro-
Sci 2017; 48(5): 1071–1083. ceedings of the 3rd international conference on systems and
11. Campi M and Savaresi S. Direct nonlinear control design: control (ICSC), Algiers, Algeria, 29–31 October 2013,
the virtual reference feedback tuning approach. IEEE T pp.457–461. New York: IEEE.
Autom Control 2006; 51(1): 14–27. 25. Bumroongsri P and Kheawhom S. Optimal dosing of
12. Sánchez-Peña RS, Colmegna P and Bianchi F. Unfalsi- breast cancer chemotherapy using robust MPC based on
fied control based on the controller parameterisation. Int linear matrix inequalities. Eng J 2015; 19(1): 97–106.
J Syst Sci 2015; 46(15): 2820–2831. 26. Ledzewicz U and Schattler H. Optimal bang-bang con-
13. Safonov M and Tsao T. The unfalsified control concept: trols for a two-compartment model in cancer chemother-
a direct path from experiment to controller. Feedback apy. J Optimiz Theory App 2002; 114(3): 609–637.
Contr Nonlinear Syst Complex 2005; 202: 196–214. 27. Ledzewicz U, Schattler H and Berman A. On the struc-
14. Safonov MG and Tsao TC. Unfalsified control concept ture of optimal controls for a mathematical model of
and learning. IEEE T Automat Contr 1997; 42(6): 843–847. tumor anti-angiogenic therapy with linear pharmacoki-
15. Chi R, Liu X, Zhang R, et al. Constrained data-driven netics. In: Proceedings of the IEEE control applications,
optimal iterative learning control. J Process Control 2017; (CCA) & intelligent control, (ISIC), St. Petersburg, 8–10
55(1): 10–29. July 2009.
16. Lim I, Hoelzle DJ and Barton KL. A multi-objective
iterative learning control approach for additive

View publication stats