Professional Documents
Culture Documents
Automatica: Richard Seeber Markus Reichhartinger
Automatica: Richard Seeber Markus Reichhartinger
Automatica
journal homepage: www.elsevier.com/locate/automatica
Brief paper
article info a b s t r a c t
Article history: The super-twisting algorithm is a robust sliding-mode control law, which in the case of a saturated
Received 29 November 2018 control input suffers from a windup effect. This contribution proposes a modified control law that
Received in revised form 23 January 2020 is obtained by mitigating windup by means of the conditioning technique. Closed-loop stability is
Accepted 21 February 2020
analyzed for bounded and Lipschitz continuous perturbations acting on a first-order linear time-
Available online xxxx
invariant plant. Advantages compared to existing approaches are demonstrated both analytically, in
Keywords: terms of stability conditions, and numerically, in the course of simulations. An experiment considering
Sliding-mode control the control of a servomotor demonstrates the practical usability, straightforward implementation and
Robust control simple tuning of the proposed controller.
Control input saturation © 2020 Elsevier Ltd. All rights reserved.
Windup
Conditioning technique
Servomotors
https://doi.org/10.1016/j.automatica.2020.108921
0005-1098/© 2020 Elsevier Ltd. All rights reserved.
2 R. Seeber and M. Reichhartinger / Automatica 116 (2020) 108921
see, e.g., Hippe (2006) for an overview, but to the best knowledge given by U > 0 and the control input before saturation is denoted
of the authors their application or extension to the STA in the by u. One can see that u∗ satisfies |u∗ | ≤ U. The perturbation w
presence of disturbances has not been analyzed before. For the and its time-derivative ẇ are assumed to be bounded by
super-twisting algorithm, a few approaches in that regard have
recently been proposed in literature, see Castillo, Steinberger, |w| ≤ W , |ẇ| ≤ L, (6)
Fridman, Moreno, and Horn (2016a, 2016b) and Golkani, Koch, with the maximum perturbation amplitude W and the perturba-
Reichhartinger, and Horn (2018). For each approach it is proven tion’s Lipschitz constant L being non-negative constants.
that its application to a first-order integrating plant leads to Starting from any initial state x0 , the state x is to be steered
a globally asymptotically (even finite-time) stable closed loop. towards some given reference r, which is assumed to be bounded
However, the modifications also change the behavior of the con- by
troller when the actuator saturation is inactive. One of the mod-
ifications, which is based on switching from a relay controller | r | ≤ R. (7)
to the STA, furthermore may introduce a discontinuity into the
For this task to be feasible, the assumptions
otherwise continuous control signal.
The present paper contributes a modification of the STA, which a ≥ 0, (8a)
does not have these drawbacks, and a stability analysis of its
application to a first-order linear plant, which includes the first- U > W + aR (8b)
order integrator as a special case. In particular, the proposed are made. If (8a) or (8b) do not hold, there exist either initial
control law yields a continuous control signal and deviates from states, e.g., x0 = −a−1 (U + W ), or disturbance and reference
the unmodified STA only when the control input saturation is ac- values, e.g., w = −W , r = R, respectively, which make it
tive. It is constructed using the conditioning technique proposed impossible to steer x towards r.
in Hanus, Kinnaert, and Henrotte (1987) and is therefore called
the conditioned super-twisting algorithm (conditioned STA). 4. Conditioned super-twisting algorithm
The paper is structured as follows: After introducing some
definitions and notations in Section 2, the considered problem is
For the purpose of steering x towards r and keeping it there
stated in Section 3. The conditioned STA is derived in Section 4,
despite the perturbation w , the super-twisting control law
and its stability properties are analyzed in Section 5. The ap-
1
proach is then compared to existing approaches in Section 6 and u = k1 ⌊r − x⌉ 2 + v, (9a)
applied in the course of an experiment in Section 7. Section 8,
finally, draws conclusions and gives a brief outlook. An Appendix v̇ = k2 ⌊r − x⌉ 0
(9b)
contains proofs of lemmas and propositions. may be used. It can be interpreted as a non-linear PI-type con-
trol law and may hence suffer from a windup effect when the
2. Preliminaries and notation control input u is saturated. To prevent this windup, the condi-
tioning technique proposed in Hanus et al. (1987) is applied. This
A few notational conventions and definitions are first dis- technique consists of computing a modified, so-called realizable
cussed. Throughout the paper, common use of the abbreviation
reference r ∗ , such that applying r ∗ to the controller instead of r
⌊y⌉p = |y|p sign(y) (2) yields u = u∗ . As argued in Hanus et al. (1987), this mitigates
0
windup, because the controller with reference r ∗ then behaves
is made. In particular, ⌊y⌉ denotes the sign of y. It is worth as if no actuator saturation were present.
pointing out that the relations From (9a) one can see that the realizable reference r ∗ satisfies
d⌊y⌉p d|y|p ⌉ 21
= p |y|p−1 , = p ⌊y⌉p−1 (3) u∗ = k1 r ∗ − x + v.
⌊
(10)
dy dy
hold for all y ̸ = 0. The so-called conditioned control law is then obtained by replac-
The saturation function, whose output is limited to ±M with ing r by r ∗ in the dynamic part of the controller; (9b) is thus
M > 0, is denoted by satM and is defined as replaced by
⌉0
v̇ = k2 r ∗ − x .
{ ⌊
y |y| ≤ M (11)
satM (y) := (4)
M ⌊y⌉0 |y| > M . By subtracting v and taking the sign on both sides of (10), one
The time derivative of a function y(t) is denoted by ẏ(t). The so- can see that ⌊r ∗ − x⌉0 = ⌊u∗ − v⌉0 holds. Hence, the conditioned
lutions of all systems of differential equations with discontinuous STA is obtained as1
right-hand side, such as, e.g., (1b), are understood in the sense of 1
u = k1 ⌊r − x⌉ 2 + v, (12a)
Filippov, see Filippov (1988). ⌉0
v̇ = k2 u − v ∗
,
⌊
A solution of a system is called finite-time stable, if it is stable in (12b)
the sense of Lyapunov and all trajectories starting in some vicinity
u = satU (u).
∗
(12c)
of that solution converge to it in finite time. The solution is called
globally finite-time stable, if this is true for all trajectories. Fig. 1 depicts block diagrams of the classical (unconditioned)
and the conditioned STA. One can see that the conditioned STA is
3. Problem statement straightforward to implement, because, compared to the classical
STA, only the term in the sign function of the controller’s dynamic
Consider a first-order plant of the form part needs to be changed.
ẋ = −ax + u∗ + w (5a) In practical implementations, the super-twisting control law
typically is implemented in a discrete-time environment. The
u∗ = satU (u) (5b)
with a constant parameter a, the state variable x, the saturated 1 It is noteworthy that the conditioned control law (12) is obtained without
control input u∗ , and a perturbation w . The control input bound is explicitly calculating r ∗ .
R. Seeber and M. Reichhartinger / Automatica 116 (2020) 108921 3
Fig. 1. Block diagrams of classical STA (9) and conditioned STA (12).
simplest approach for this purpose is forward Euler discretization. Corollary 2. If the positive parameters k1 , k2 and the non-negative
For more systematic ways of discretizing the STA, which maintain Lipschitz constant L satisfy (17), then system (16) with |δ| ≤ L is
several desirable properties of the continuous time controller, the globally finite-time stable.
reader is referred to Brogliato, Polyakov, and Efimov (2020), Koch
and Reichhartinger (2019) and references therein. Proof. Without input saturation, u = u∗ always holds. Thus,
Proposition 1 guarantees finite-time convergence of all trajecto-
5. Stability analysis ries to the Lyapunov-stable origin. □
In order to investigate the stability properties of the closed 5.2. Saturated control input
loop, the control error e := r − x and abbreviations ∆ := w − ar
and δ := ∆ ˙ = ẇ are introduced. The control error is then
The closed loop with control input saturation is now consid-
governed by the differential equation
ered. In this case, the closed loop formed by the interconnection
ė = −ae − u∗ − ∆. (13) of plant (5) and controller (12) in terms of the state variables (15)
is given by
One can see that ∆ acts as a disturbance on the error system;
according to (6), (7) it and its time-derivative δ are bounded by
1
ẋ1 = −ax1 + satU (x2 − ∆ − k1 ⌊x1 ⌉ 2 ) + ∆ (18a)
|∆| ≤ W + aR, |δ| ≤ L.
⌉0
(14)
⌊ 1
ẋ2 = −k2 x2 − ∆ − satU (x2 − ∆ − k1 ⌊x1 ⌉ ) 2 + δ. (18b)
5.1. Unsaturated control input The following theorem gives conditions that guarantee global
finite-time stability of the closed loop with the conditioned con-
Finite-time stability of the closed loop without control input
troller and with saturation:
saturation, i.e., for u = u∗ , is first shown. The obtained results are
then used to prove global finite-time stability of the closed loop
Theorem 3 (Stability of the conditioned STA). Let the perturbation
with control input saturation.
w(t) and the reference r satisfy (6), (7) with given non-negative
Introducing the state variables
bounds W , L, and R. Suppose that the plant parameter a and the
x1 := −e = x − r , x2 := v + ∆, (15) control input bound U satisfy (8) and that the controller parameters
k1 , k2 fulfill the inequalities
the closed loop in the case u = u∗ is given by √
1 U + W + aR
ẋ1 = −ax1 − k1 ⌊x1 ⌉ + x2 ,
2 (16a) k1 > 2k2 , k2 > L. (19)
U − W − aR
ẋ2 = −k2 ⌊x1 ⌉ + δ.
0
(16b)
Then, the origin of the closed-loop system (18) or, equivalently, the
Stability of this system is investigated in Orlov et al. (2011), and solution x(t) = r, v (t) = ar − w (t) of the interconnection of plant
for a = 0 in Moreno and Osorio (2012) and Seeber and Horn (5) and controller (12) is globally finite-time stable.
(2017), for example. Here, the Lyapunov based approach pro-
posed in Seeber and Horn (2017) is used with a little modification Remark 4. Condition (19) is equivalent to the inequalities
to obtain the following stability condition.
k21 − 2k2
W < U − aR, L < k2 , (20)
Proposition 1. Suppose that the perturbation w (t) satisfies (6) and k21 + 2k2
that the parameters k1 and k2 satisfy which give an upper bound on the perturbation amplitude W in
√
k1 > k 2 + L, k2 > L. (17) terms of control input bound U and maximum reference value R.
Suppose, furthermore, that for all trajectories of the closed-loop Remark 5. The conditioned STA (12) is independent of the pa-
system formed by the interconnection of plant (5) and controller (12) rameter a. Therefore, it is also robust against uncertain knowledge
there exists a time instant T , after which the control saturation is of this parameter, as long as all possible values of a satisfy (19)
inactive, i.e., such that u(t) = u∗ (t) holds for all t ≥ T . Then, x1 (t)
or, equivalently, (20).
and x2 (t) as defined in (15) converge to the origin in finite time, and
the origin is Lyapunov-stable. √
Remark 6. Note that k1 > 2k2 is a necessary condition for
the theorem’s conditions to be fulfilled. Since large values of k1
Proof. Given in the Appendix. □
are known to lead to increased chattering, this condition seems
The assumptions of Proposition 1 are stated in a rather in- restrictive. In Pérez-Ventura and Fridman (2019), √ it is shown,
volved fashion in order to enable its use in the following section, however, that the parameter setting k1 ≈ 2 k2 minimizes
where, essentially, the existence of the time instant T will be the amplitude of chattering with second-order parasitic actua-
proven. The following corollary, which is an immediate conse- tor dynamics. Furthermore, Zhang and Reger (2018) show that
√
quence, restates the stability result more concisely in terms of having k1 ≥ 2 k2 is desirable if the impact of brief violations
system (16) without control input saturation. of the assumed bound on the disturbance’s time-derivative is to
4 R. Seeber and M. Reichhartinger / Automatica 116 (2020) 108921
Remark 8. One can see from this proof that under the conditions
of the theorem, the set N is a domain of attraction also for the
classical (i.e., unconditioned) STA (9) applied to the plant (5) with
control input saturation. This is obvious from the fact that, for any
trajectories starting in N , the relations |u| ≤ U and, thus, u = u∗
hold for all time.
(a) If k1 and k2 are positive, then the set M is forward invariant 6.2. Saturated super-twisting algorithm
and any trajectory enters M in finite time.
(b) If k1 and k2 satisfy (19), then the set N is forward invariant In Castillo et al. (2016b), the so-called Saturated Super-
and any trajectory enters N in finite time. Twisting Algorithm (saturated STA) is proposed for dealing with a
constrained control input. It consists in applying a relay controller
Proof. Given in the Appendix. □ up to the first time instant T when the switching condition
Once trajectories stay in the set N , the control input saturation
|r − x(T )| ≤ γ 2 with a non-negative switching level parameter
is inactive, i.e., |u| ≤ U holds, due to (21c). The behavior of γ is satisfied. Starting at t = T , the STA (9) with v (T ) = 0 is then
the trajectories in this set is, hence, determined by the stability applied to the plant. The control law may be written as
properties of the closed loop without saturation, i.e., of system
{
U ⌊r − x⌉0 t<T
(16). Using these considerations, Theorem 3 can now be proven. u= 1 (28a)
k1 ⌊r − x⌉ 2 + v t ≥ T,
Proof of Theorem 3. Since k1 and k2 satisfy (19), all trajecto-
t<T
{
0
ries enter the forward invariant set N in finite time by virtue v̇ = (28b)
of Lemma 7. In this set, |u| ≤ U and, thus, u = u∗ holds; k2 ⌊r − x⌉0 t≥T
furthermore, one has T = inf{t : |r − x(t)| ≤ γ 2 } (28c)
√ √
k1 > 2k2 > k 2 + L, k2 > L. (23) with initial condition v (0) = 0.
Thus, the conditions of Proposition 1 are fulfilled and any trajec- Closed-loop stability with this controller is proven in Castillo
tory x1 (t), x2 (t) of system (18) or, equivalently, of the interconnec- et al. (2016b) using a Lyapunov function from a family of
tion of (5) and (12), converges to the origin in finite time. System quadratic functions that was originally proposed in Moreno and
(18) is therefore globally finite-time stable. □ Osorio (2012). In Seeber and Horn (2019) it is shown how to
R. Seeber and M. Reichhartinger / Automatica 116 (2020) 108921 5
Table 1
Parameters of the mathematical model (35) for the experimental setup depicted
in Fig. 5.
Parameter Symbol Unit Value
Motor inertia Im kg m2 3.87 × 10−7
Load inertia Ia kg m2 1.60 × 10−3
Gear ratio ηg – 70
Motor torque constant km Vs 7.67 × 10−3
Winding resistance Rw 2.44
Table 2
Parameters of the control law (39) as well as maximum control input and
reference speed used in the experiments.
Parameter Symbol Unit Value
Maximum motor voltage U V 8
Maximum desired speed ωd,max rad
√ s
−1
4
STA gain k1 V s −1 9
STA integrator gain k2 V s−1 20
where km , ηg and Rw denote the dc-motor’s torque constant, the According to (20), this tuning permits for a maximum perturba-
gear ratio, and the resistance of the dc-motor’s coil, respectively. tion amplitude W and Lipschitz constant L of
The total moment of inertia I is given by k21 − 2k2
W < U − aR ≈ 0.563 V, (42a)
I = ηg2 Im + Ia , (36) k21 + 2k2
L < k2 = 20 V s−1 . (42b)
with Im and Ia being the motor and additional load inertia, re-
spectively. Friction effects, mainly introduced by the bearings and In order to demonstrate how windup is mitigated by applying
the slip rings, which are used for signal transmission between the proposed concept, a temporary blockade of the rotating load
load and motor, are neglected. Note that this does not change the is considered. To realistically simulate this event without actually
model structure and has only a minor impact on the final model’s mechanically blocking the rotor, which is difficult to do in a
reproducible way, a voltage of zero is temporarily applied to the
parameters. The parameters of the setup are given in Table 1.
motor instead of the control signal u∗ .
Rewriting the dynamics captured by the model (35) in terms
The conditioned STA and an STA without anti-windup are
of the state variable applied to this setup using the parameters in Table 2. For this
IRw purpose, both control laws are implemented in a discrete-time
x := ω (37)
k m ηg environment with a sampling time of Ts = 5 ms and the forward
R. Seeber and M. Reichhartinger / Automatica 116 (2020) 108921 7
|v| > U, then v and v − u∗ have the same sign due to the fact k1 1
≤− |e|− 2 (U − W − aR) + k2 . (A.13)
that |u∗ | ≤ U. One thus obtains from (12b) 2
⌉0 Here, the fact is used that u, e, u∗ and u∗ − v have the same sign
v̇ = k2 u∗ − v = −k2 ⌊v⌉0 .
⌊
(A.5)
if (21b) is satisfied. Using (21a) one furthermore obtains
Clearly, if (21b) holds at any time instant t = T1 , it also holds 1
Orlov, Y., Aoustin, Y., & Chevallereau, C. (2011). Finite time stabilization of a Vazquez, C., Collado, J., & Fridman, L. (2014). Super twisting control of a
perturbed double integrator–Part I: continuous sliding mode-based output parametrically excited overhead crane. Journal of the Franklin Institute, 351(4),
feedback synthesis. IEEE Transactions on Automatic Control, 56(3), 614–618. 2283–2298.
Pérez-Ventura, U., & Fridman, L. (2019). Design of super-twisting control gains: Zhang, D., & Reger, J. (2018). H∞ optimal parameters for the super-twisting algo-
A describing function based methodology. Automatica, 99, 175–180. rithm with intermediate disturbance bound mismatch. In 15th international
Plestan, F., Evangelista, C., Puleston, P., & Guenoune, I. (2018). Control of a workshop on variable structure systems (pp. 303–308).
twin wind turbines system without wind velocity information. In 15th
international workshop on variable structure systems (pp. 150–155).
Polyakov, A., & Poznyak, A. (2009). Reaching time estimation for ‘‘super-twisting’’ Richard Seeber received his M.Sc. degree in Electrical
second order sliding mode controller via Lyapunov function designing. IEEE Engineering from Graz University of Technology in
Transactions on Automatic Control, 54(8), 1951–1955. 2012, and he completed his Ph.D. at the Institute of
Reichhartinger, M., Golkani, M. A., & Horn, M. (2015). Experimental evaluation Automation and Control, Graz University of Technol-
of observer-based throttle valve control using super-twisting algorithm. In ogy in 2017. He currently holds a Postdoc position
2015 European control conference (pp. 2144–2149). at the Christian Doppler Laboratory for Model Based
Ruiz-Zea, C. A., Jimenez-Rodriguez, E., Castaneda, J. M. C., Loukianov, A. G., Control of Complex Test Bed Systems at Graz University
& Sanchez-Torres, J. D. (2018). Second order sliding mode control of a of Technology, Austria. His research interests include
STATCOM with saturated inputs. In 15th international conference on electrical theory of sliding mode control systems, control of
engineering, computing science and automatic control. automotive test beds, and control of systems with
Seeber, R., & Horn, M. (2017). Stability proof for a well-established super-twisting actuator constraints.
parameter setting. Automatica, 84(10), 241–243.
Seeber, R., & Horn, M. (2019). Optimal tuning of the saturated super-twisting
algorithm. In 18th European control conference (pp. 3606–3612). Markus Reichhartinger currently is an Associate Pro-
Seeber, R., & Horn, M. (2019a). Guaranteeing disturbance rejection and control fessor at the Institute of Automation and Control at
signal continuity for the saturated super-twisting algorithm. IEEE Control Graz University of Technology, Graz, Austria. He re-
Systems Letters, 3(3), 715–720. ceived his Master of Science degree in Information and
Seeber, R., & Horn, M. (2019b). Optimal Lyapunov-based reaching time bounds Computer Engineering (Telematik) at Graz University
for the super-twisting algorithm. IEEE Control Systems Letters, 3(4), 924–929. of Technology and the Doctor degree in Information
Shtessel, Y. B., Edwards, C., Fridman, L., & Levant, A. (2014). Sliding mode control Technology at University of Klagenfurt, Austria in 2006
and observation. New York, NY: Springer. and 2011, respectively. His research interests include
Utkin, V. (2013). On convergence time and disturbance rejection of the design of robust feedback loops, sliding-mode
super-twisting control. IEEE Transactions on Automatic Control, 58(8), based concepts and applications of automotive and
2013–2017. mechatronic systems.