Sunteți pe pagina 1din 27

INTERNATIONAL JOURNAL OF ROBUST AND NONLINEAR CONTROL

Int. J. Robust Nonlinear Control 2016; 26:4008–4034


Published online 31 March 2016 in Wiley Online Library (wileyonlinelibrary.com). DOI: 10.1002/rnc.3546

Performance improvement of SISO linear control systems by


hybrid state resetting and sector confinement of trajectories

José Antonio González1,*,† , Antonio Barreiro2 , Sebastián Dormido3 and Pablo Falcón2
1 Centro Tecnológico de Automoción de Galicia, Pontevedra, Spain
2 Departamento de Ingeniería de Sistemas y Automática, E.T.S.I.Industriales, University of Vigo, Vigo, Spain
3 Departamento de Informática y Automática, UNED , Madrid, Spain

SUMMARY
Reset control techniques have been proposed to overcome fundamental limitations of linear controllers by
means of their transformation into hybrid models that combine continuous flow and discrete jump dynamics.
The hybrid nature in the control loop involves some difficulties when analyzing the performance of the
controller and some drawbacks on the controller design related to the stability conditions. The technique
that we propose is based on sector confined target dynamics of the continuous flow mode by means of the
application of the discrete reset jumps. This behavior, in the error plane .e.t/; e.t//,
P is correlated with certain
preferred sectors that lead to fast and over-damped responses. The paper studies how to design a hybrid
resetting version of a linear controller that achieves the required fast and over-damped responses to arbitrary
references. Copyright © 2016 John Wiley & Sons, Ltd.

Received 27 June 2013; Revised 1 March 2016; Accepted 2 March 2016

KEY WORDS: hybrid reset control; state reset; confinement dynamics; Lyapunov stability

1. INTRODUCTION

The idea of reset control was originated with the Clegg integrator [1] and formalized by I. Horowitz
and coworkers in [2] and [3] in the seventies. Since those early works, it was clear that the ben-
efit was to outperform linear solutions and to overcome fundamental limitations by means of an
extremely simple action: resetting to zero the states of the controller when the tracking error is zero.
Somewhat later, new formalized results were presented by Beker, Chait, Hollot and coworkers in
[4–7] or [8] that include the study for general reset systems with arbitrary dimension based on the
Horowitz’s reset condition (zero crossing of the error).
After that, new strategies for reset control using a reset condition based on the sign of the input
and output were proposed and studied in [9–13], stability analysis and limit cycles with reset band
conditions in [14–16], stability analysis in systems with time delays in [17, 18] and [19] or the
frequency properties of multiple anticipative reset systems in [20].
A summary of the principles and general approaches to reset control has already been recently
published in [21], and open problems have been recently presented in [22], which could be consulted
for obtaining a general glimpse into the field. Although the previous references show that there
already exists a significant body of knowledge on reset control, looking at it in perspective, it is still
a young field of research, with many open problems and much work to be carried out.
The development of hybrid reset design methodologies for specific control problem models
is one of those open fields. Although some reset control applications have been presented in
[23, 24] or [25], there is no too much literature about the development of general hybrid reset

*Correspondence to: José Antonio González Prieto, Centro Tecnológico de Automoción de Galicia, Pontevedra, Spain.
† E-mail: jose.gonzalez@ctag.com

Copyright © 2016 John Wiley & Sons, Ltd.


HYBRID STATE RESETTING AND SECTOR CONFINEMENT OF TRAJECTORIES 4009

control (HRC) design strategies. Usually, in these works, the design procedure is based on the search
of Lyapunov functions and the application of relaxation techniques as in [26] to obtain the solutions
as LMI expressions.
Instead, we apply a different approach by using the classical linear design, a Lyapunov function as
a geometrical constructor (also related with the error settling time performance) and the properties
achieved by the combination of some continuous flow and discrete jump restrictions to set up the
convergence evolution of the modified error dynamics. We propose and study a novel reset technique
based on confinement of the trajectories within a conic sector with the purpose of overcoming the
linear limitations, that is, achieving fast and over-damped responses.
Following this basic idea, we start with a linear control system, with an initial tuning, that gives
a fast but under-damped and overshooting behavior and after that we transform it into a hybrid
reset controller that eliminates the overshooting by means of the confinement of the modified error
dynamics trajectory inside some preferred conic sectors.
This approach has some points in common with other reset techniques, but there are significant
differences between them. Compared with other hybrid reset techniques that use conic flow sets, our
approach computes the after-reset value for the controller state, xc .t  / ! xc .t /, in order to travel
between boundaries of the confinement sector, that is, the jump is explicitly computed to travel
across the sector.
In addition, the flow region (instead of general sectors.xp .t /; xc .t //T M.xp .t /; xc .t // 6 0) is
.k/ .kC1/
explicitly particularized to sectors in the s .t /; s .t / planes, where s.t / is the modified error,
which has a certain resemblance to the sliding mode control (SMC) and is more transparent to the
designer, which still has some degrees of freedom to shape the desired responses.
The influence of arbitrary exogenous signals (references, disturbances) and how to deal with them
is fully treated within the design procedure, so general tracking of reference signals is also taken
into account in the procedure, which is validated on several simulation examples.
For this purpose, the paper is organized as follows. In Section 2, we introduce the problem state-
ment and set up the dynamics involved in the hybrid closed loop model. Next, in Section 3, we
introduce some assumptions needed to implement the proposed confinement technique. Sections 4
and 5 present the analysis of the continuous flow and the discrete jump modes in order to formulate
the restrictions that ensure the converge properties; Section 6 summarizes the design procedure, and
finally, in Section 7, we present an extended example analysis of this design procedure.

2. PROBLEM STATEMENT

Consider the linear plant (P)


²
xP p .t / D Ap xp .t / C Bp u.t /
P W (1)
yp .t / D Cp xp .t /

with xp .t / 2 Rp the plant states, u.t / 2 R as the control signal , yp .t / 2 R the plant output,
Ap 2 Rpp , Bp 2 Rp and Cp 2 R1p .
In order to solve the classical control problem where the plant output yp .t / must follow an input
reference signal r.t /, we create a SISO feedback closed loop where the controller (C) is given by
²
xP c .t / D Ac xc .t / C Bc e.t /
C W (2)
u.t / D Cc xc .t / C Dc e.t /

with xc .t / 2 Rc the controller states, e.t / D yp .t /  r.t / 2 R as the error signal, Ac 2 Rcc ,
Bc 2 Rc , Cc 2 R1c and u.t / 2 R as the control signal being applied over the plant previously
defined in (1).
Notice that (1) and (2) define a linear and output feedback control loop. Later, the loop will be
hybridized (reset actions will be introduced) thus becoming nonlinear. Furthermore, the reset trigger
will be based on full state information thus becoming a state feedback control system.

Copyright © 2016 John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control 2016; 26:4008–4034
DOI: 10.1002/rnc
4010 J. A. GONZÁLEZ ET AL.

Assumption 2.1
We assume that the input reference r.t / is a continuous smooth signal and it is possible to obtain
any of its higher order derivatives.
The LTI closed loop model is as follows:
PO / D AO1 x.t
x.t O / C BO 1 r.t /
(3)
O / D CO 1 x.t
y.t O /
where
 T
O / D xpT .t / xcT .t /
x.t 2 Rn (4)

is the closed loop state space, n D c C p and the matrices AO1 , BO 1 , and CO 1 can be obtained by means
of the feedback connection:

e.t / D r.t /  Cp xp .t / 2 R (5)

Remark 2.1
The definition of the augmented spaces and details (such as matrix definitions for AOk , BO k , . . . ) of
the dynamics equations related to the closed loop states, error, and the modified error dynamics can
be found in Appendix.
The transformation of the linear closed loop model into a hybrid closed loop model is based
on the application of reset actions to the controller states at the reset time instants, that is, the
continuous dynamics of the linear closed loop system is transformed into a hybrid dynamics that
contain continuous flow and discrete jump (active at the reset instants) dynamic modes.
This approach raises two basic questions that have to be solved in order to apply the hybrid reset
transformation:
 When apply reset? : This amounts to defining the reset trigger condition. This triggering can
be implemented at fixed reset times or at state-dependent times. In this case, it is also necessary
to choose which signal is used to determine the reset instant (the input, the product of the input
and the output, ...) and the signal condition to execute the jump (zero crossing, anticipative with
reset band, multiple reset band, dynamic anticipation, ...).
 How to apply reset? : This amounts to defining the after-reset value for the system state, that
is, the definition of the discrete control law that must be applied at the reset instants.
In this work, we use state-dependent times to determine the reset trigger instants (t ,  2 N), so
the hybrid reset design creates two different regions at the state space:
 The flow set F, where the system dynamics flows in the continuous mode.
 The jump set J, where the system dynamics jumps in the discrete mode.
where we assume that F [ J D Rn .
Before introducing the hybrid closed loop transformation, let us define the modified error
as follows:
q
X .j /
s.t / D ˛j e .t / (6)
j D0

with ˛j 2 R and 0 6 q < p.


The role of the auxiliary variable s.t / is that of a modified error e.t / defined to facilitate the
treatment of the convergence e.t / ! 0. It has a similar role to that of the sliding variable in SMC.
For example, if our target is an error following eP  e, it is equivalent to s  0 with the definition
s D e C e.P
The definition of the flow/jump surfaces that determine the reset trigger conditions and the dis-
crete control law applied at the jump mode are developed by means of a sector confinement of the

Copyright © 2016 John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control 2016; 26:4008–4034
DOI: 10.1002/rnc
HYBRID STATE RESETTING AND SECTOR CONFINEMENT OF TRAJECTORIES 4011

 
.k/ .kC1/
trajectory in the s .t /; s .t / planes for k D 0; 1; : : : ; c   1, where c  (with c  6 c) is
the number of controller states that can be used at the discrete jump mode to change the trajectory
direction as explained in Section 5.
We consider the augmented modified error sN .t / D Œs.t /; sP .t /; sR .t /; : : :T as the output of the
hybrid closed loop model, so from (3) and (44), the formulation for the hybrid closed loop model is
obtained as follows:
 Continuous dynamics : t ¤ t

PO / D AO1 x.t
x.t O / C BO 1 r.t / (7)

sN .t / D SN x.t
O / C N s .t / (8)
 Discrete dynamics : t D t
    
xp .t / xp t
D (9)
xc .t /  t
  
xp t
sN .t / D SN C N s .t / (10)
 t
where
 t with  2 N sets the reset time instants, so the pre-reset conditions are defined at t and the
 conditions at t .
after-reset
  t is the discrete control law that generates the reset jump in the controller states.
From the proposed model, we can observe that at the continuous flow dynamics the control law
is computed by means of a linear controller that uses the error information to generate the control
signal, but in the case of the discrete jump mode, the situation could be quite different. Although it
is possible to generate the reset trigger conditions and the discrete control law using an input/output
approach, in this work, we assume that the plant states are accessible, so the discrete jump condi-
tions are obtained with this information. In this sense, the hybrid design includes a mixing between
an input/output-based dynamic linear controller at the continuous flow mode and a state feedback
control related to the discrete jump mode (detection and execution of the jumps).
Note that if state feedback can be used, it is possible to design a non-overshoot controller for
linear plants ([27] and [28]), our proposal achieves the same objectives, for SISO systems, with
some differences:
 In the hybrid design, the state feedback control is applied only at the reset instants so, although
the detection of the trigger conditions is continuously evaluated, this could be an advantage
with respect to state feedback control.
 The hybrid design includes, at the continuous flow mode, a dynamic controller with integration
process. This implies that the hybrid design naturally includes dynamics that can be used to
reach zero-steady-state-error for step tracking.
 As shown in the next sections, the hybrid design can be obtained with the choice of very few
parameters that are directly related to the parameters of the closed loop linear system.
 It is possible to obtain equivalent tracking performance using two different linear controllers
when the same confinement is being applied, pointing out the possibility of testing the
confinement reset technique with a vast collection of predesignated linear controllers.
The proposed hybrid closed loop model, without external references, is described by (45) with
Q s .t / D 0, that is, by
sPN .t / D SQ sN .t /

Copyright © 2016 John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control 2016; 26:4008–4034
DOI: 10.1002/rnc
4012 J. A. GONZÁLEZ ET AL.

D 0, the state sN .t / is given by sN .t / D SN x.t


where, from (44) with N s .t /
O /. Assuming invertibility
of S (that is, observability of SO0 ; AO1 , see Appendix) both states sN .t / and x.t
N O / are equivalent. The
stability properties of this hybrid system can be ensured by the classical Lyapunov theorem [5]:
Theorem 2.1
The origin of the hybrid closed loop system given by (7), (8), (9), and (10) is a globally uniformly
asymptotically stable equilibrium point if there exists a function V W Rn ! R > 0 such that
V .Ns .t // > 0; 8Ns .t / ¤ 0 ^ V .0/ D 0
VP .Ns .t // < 0; 8Ns .t / 2 F
V .Ns .t // 6 0; sN .t / 2 J
kNs .t /k ! 1 H) V .Ns .t // ! 1
with
 
V .Ns .t // D V .Ns .t //  V sN t 6 0


In case that SN is not invertible, then SO0 ; AO1 is not observable (Appendix) and the modified
error vector sN .t / does not give full information on the internal state x.t
O /. In this case, stabiliz-
ing only the output sN .t / may induce hidden internal dynamics. Therefore, in order to develop an
applicable control design procedure, some additional assumptions have to be introduced in the
following sections.

3. SECTOR CONFINED TARGET DYNAMICS

Artstein’s theorem [29] states that a dynamical system has a differentiable control-Lyapunov func-
tion if and only if there exists a regular stabilizing feedback. A direct application of this theorem is
commonly used to design stabilization control laws as in [30, 31], and [32].
In this work, a different approach is considered
 by applying the theorem to define the flow/jump
.k/ .kC1/
regions of a hybrid dynamical system in the s .t /; s .t / planes for k D 0; 1; : : : ; c   1, so
the reset actions handle the confinement task preventing the system trajectory to get out of the flow
regions F.k/ at each plane.
s

Remark 3.1
Note
 that we use  F.k/
s
and J.k/ to denote the flow and jump regions that are defined in the plane
s
.k1/ .k/
s .t /; s .t / .

To simplify the presentation, we consider the design criterion to define the flow and jump regions:
FsP and JsP , related to the .s.t /; sP .t // plane. The technique proposed is based on the confine-
ment of the trajectories within the conic flow sector FsP where the tracking error achieves fast and
over-damped responses for a pre-designed linear controller with an initial tuning. Once the conic
flow sector is defined, the linear controller preserves the speed of response on the flow mode and
eliminates the overshoots by means of the reset actions.
To show how this property is achieved, let us suppose for a moment that in the modified error we
set s.t / D e.t /. As said before, the motivating idea is to pursue overcoming of linear limitations,
that is, achieving fast and over-damped responses. This behavior, in the error plane .e.t /; e.t P //, is
correlated with certain preferred sectors. The sectors as shown in Figure 1(a), bounded by e.t P /D
cF e.t / and e.t
P / D cZ e.t /, give rise to fast and under-damped responses, when cF , cZ are large.
Because this cannot be produced in the continuous mode, but requires resetting, if s.t / D e.t /,
the typical time evolution of e.t / would be the one in Figure 1(b), with discontinuities that are not
desirable for smooth tracking responses.
If we set s.t / D e.t / C ˛ e.t
P / the tracking error e.t / is obtained from s.t / by means of the filter
1
F .& / D .1C˛p/ . Thus, if we impose to s.t / the sector performance in Figure 1, we would obtain the
continuous tracking error in Figure 1(c).

Copyright © 2016 John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control 2016; 26:4008–4034
DOI: 10.1002/rnc
HYBRID STATE RESETTING AND SECTOR CONFINEMENT OF TRAJECTORIES 4013

Figure 1. Sector confined target dynamics. (a) In the phase plane .s.t/; sP .t// (compared with the linear
response in solid line), (b) Typical trajectory s(t), and (c) Filtered trajectory e(t).

This proposal enjoys some nice properties:


1. Sign invariance: if s.t / > 0 then e.t / > 0 . This is important regarding the suppression of
overshoot. The sign invariance can be easily ensured for more general modified errors. Let
.q/
s D ˛0 e C ˛1 eP C : : : ˛q e be the modified error, ˛.& / D ˛0 C ˛1 & C : : : ˛q & q the associated
polynomial, F .& / D 1=˛.& / the related filter, and h˛ .t / its impulse response. Then,
Z t
e.t / D h˛ .t   /s. /d :
0

Thus, if the filter F .& / D 1=˛.& / has a positive impulse response, h˛ .t / > 0, 8t >
0, then we have sign invariance: s.t / > 0 ) e.t / > 0. Thus, if s.t / does not overshoot,
neither does e.t /. The condition h˛ .t / > 0 can be guaranteed in different ways. For example,
if F .& / D 1=˛.& / D 1=..& C 1 / : : : .& C q // with k > 0, by interpreting F .& / as
a series connection of first order systems 1=.s C k /, it can be seen that it has a positive
impulse response. These structures for F .& / D 1=˛.& / leave still great flexibility in the choice
of ˛.& /.
2. Speed of response: If s.t / is made faster, then also e.t / becomes faster. This is achieved mak-
ing cF , cZ larger. Consider that s.t / is controlled to achieve an approximate response s.t / D
s0 exp. t /, with Laplace transform s0 =.& C /. Then, the error is E.& / D F .& /s0 =.s C /.
It contains contributions of the filter poles k and of the pole  . When  is dominant
(slowest) then, by retuning the control and making s.t / faster, we consequently obtain a faster
response for e.t /. In general, by tuning both F .& / D 1=˛.& / and the s.t / dynamics, it is
possible to obtain a desired response for e.t /. More details will be discussed in Section 4.
3. If the sector is made very narrow (taking cF  cZ ), the related behavior has a certain resem-
blance to the SMC, but here, the confinement is obtained by resetting the controller state xc
(not by bang-bang control).

Copyright © 2016 John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control 2016; 26:4008–4034
DOI: 10.1002/rnc
4014 J. A. GONZÁLEZ ET AL.

In order to obtain the target performance in Figure 1(a), some assumptions have to be made. Let
G.& / D N.& /=D.& / be the transfer function of the plant and suppose that N.& /, D.& / have no
common roots (minimal realization) with degrees deg.N / D m, deg.D/ D n, and relative degree
rd D n  m > 0.
Assumption 3.1
The plant G.& / D N.& /=D.& / has relative degree rd 6 q C 1 and is minimum phase. The reasons
for imposing these assumptions are as follows. Suppose that q D 1, so s.t / D e.t / C ˛ e.t
P /. If we
transform the plant in the normal form [33], we obtain the plant model:

P D P
C Q
.r/
y .t / D R
C S  C Ku

h iT  T
.r1/
where
D y.t / y.t
P / : : : y .t / and the complete state vector is
;  of dimension r C
m D n. From this, it is clear that instantaneous jumps in the control u have no influence in the
.r1/
output and its first derivatives y.t /; y.t
P /; : : : ; y .t /.
Thus, if we intend to produce the jumps in Figure 1(a) on the signals s.t / D e.t / C ˛ e.t P / or
sP .t / D e.t
P / C ˛ e.t
R / (on at least one of them, on s or on sP ), then we need to be able to produce
jumps in e.t /; e.t
P / or e.t
R / (necessarily on at least one of the three). Because the reference r.t / does
not jump, then we necessarily need to produce a jump on y.t /; y.t P / or y.t
R / (on at least one of the
three). This could be possible only with relative degree rd D 2, that is q C 1 at most. Notice that, in
contrast, if rd D 3, the first derivative of the output that could jump is y«, thus s.t / and sP .t / would
be continuous, and no jumps could be produced in the plane of Figure 1(a).
On the other hand, regarding stability, if the control design only deals with the objective of
Figure 1(a), that is, making zero the s.t / and its derivatives, as a consequence, in the absence of
external references and disturbances, it is equivalent to make zero the signal y.t / and its derivatives.
But from the normal form decomposition, this amounts to stabilizing only
, without making clear
the stability of . If we are not to worry about this, we can impose the internal or zero dynamics
P D Q to be stable, that is, impose G.& / to be minimum phase.

Remark 3.2
Relaxation of the assumption. The previous assumption is taken for simplicity, to facilitate the
exploration of the strategy of the sector bounding reset control. The assumption could be removed
at the expense of a more involved or constrained treatment. If the plant is non-minimum phase, we
must impose, in addition to sector bounding, conditions for the internal stability of the whole state
of the system, under this hybrid control. This could restrict the class of sectors that are suitable
for confinement.

4. CONTINUOUS FLOW MODE

Let us consider quadrant IV in Figure 2 and assume that we confine the trajectories within the sector
given by the lines sP .t / D s.t / with  2 ŒF ; Z .
We will now give design guidelines to relate the design parameters (slope  and error coeffi-
cient ˛) to the approximate behavior of the error e.t /, without external references in the case of
s.t / D e.t / C ˛ e.t P /. To obtain the approximate relation and for simplicity, assume that the trajec-
tory in .s.t /; sP .t // moves precisely along the line sP .t / D s.t / (the central slope  D M is a
good choice for ). Both sP .t / D s.t / and s.t / D e.t / C ˛ e.t
P / imply the property:

e.t
R / C h1 e.t
P / C h0 e.t / D 0

with

Copyright © 2016 John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control 2016; 26:4008–4034
DOI: 10.1002/rnc
HYBRID STATE RESETTING AND SECTOR CONFINEMENT OF TRAJECTORIES 4015

Figure 2. Sector target confinement dynamics. Continuous and discrete modes.

1 
h1 D C h0 D
˛ ˛
At the after-reset instant t , the error equations satisfy

e.t
R  / C h1 e.t
P  / C h0 e.t / D 0

R  /Ch1 e.t
Taking the limit trajectory e.t P  /Ch0 e.t / D 0 as a guide of design, it can be expressed
in the Laplace domain as follows:
Œh1 e.t / C e.t
P  / C &e.t /
E.& / D 2
& C h1 & C h0
Remark 4.1
The notation uses the symbol & as the Laplace operator in order to avoid confusion with the modified
error s.t /.
The choice of h1 and h2 that gives rise to a double real pole is the solution leading to fast
trajectories without overshoots is obtained when h21 D 4h0 , with a double real root  D  h21 , so

Œh1 e.t / C e.t


P  / C &e.t / A1 A2
E.& / D 2
D C
& C h1 & C h0 & C  .& C  /2
where
1
˛D  D 2

h0 D 2 h1 D 2
A1 D e.t /A2 D  e.t / C e.t
P /

In the next section, we will show that the number of planes where we can apply the sector
confinement technique depends on the number of control states that can be used at the discrete

 conic sectionin c 6 c
jump mode. If we assume that it is possible to create a confinement
.k/ .kC1/
planes, we can develop a selection criterion for Fk and Zk for each s .t /; s .t / plane with
k D 0; 1; : : : ; c  1 using the same technique as we have shown for the .s.t /; sP .t // plane. Therefore,
the following parameters

Copyright © 2016 John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control 2016; 26:4008–4034
DOI: 10.1002/rnc
4016 J. A. GONZÁLEZ ET AL.

2
 Mk D
˛
 Fk D cF  Mk
 Zk D cZ  Mk
with cF < 1, cZ > 1 and k D 0; 1; : : : ; c   1 will be employed when the modified error is defined
as s.t / D e.t / 
C ˛ e.t
P /. Note that
 the previous criterion can be used to select flow conic regions
.k/ .kC1/
in higher order s .t /; s .t / planes because the relation between the variables at each plane is
always the same (derivation).
If the modified error includes high order terms of the error (e.t
R /; «
e .t /; : : : ) we need to intro-
duce a criterion to compute the parameters ˛i that define the modified error (6). The expression
s.t / D e.t / C ˛1 e.t
P / remains a first order anticipative Taylor approximation of e.t C ˛/, and the
modified error can be understood as a time advanced error signal, so if we choose to apply a Taylor
series expansion
1
X ˛ i .i /
˛2 ˛3
s.t /  e.t C ˛/ D e.t / C ˛ e.t
P /C e.t
R /C «
e .t / C : : : D e .t / (11)
2Š 3Š iŠ
i D0

it is possible to obtain the parameters that define the modified error as follows:
˛i
˛i D

˛2
As an example, if s.t / D e.t / C ˛ e.t
P /C 2
e.t
R /, the condition sP .t / D s.t / implies that

e .t / C h2 e.t
« R  / C h1 e.t
P  / C h0 e.t / D 0
with

2 C ˛
h2 D
˛
2 C 2˛
h1 D
˛2
2
h0 D 2
˛
so we need to select the parameters  and ˛ to obtain the values of h2 ; h1 ; h2 that determine a stable
and, if possible, non-overshooting error trajectory. Note that if we select a more general criterion
than the Taylor series expansion, for example, by taking s.t / D e.t / C ˛1 e.t P / C ˛2 e.t
R /, we have
now three parameters to compute in order to fit the stability requirements in the previous error
dynamic equation.
Now that we have chosen a criterion for the reset design parameters, we need to set the expression
that determines the stability properties of the hybrid system. Let us define a Lyapunov function
as follows:
V .t / D sN T .t /IV sN .t / (12)
where IV 2 Rnn is a positive definite matrix.
Remark 4.2
We will use V .t / instead of V .Ns .t // without loss of generality in order to reduce the notation.
From (12) and (45), it is obtained:
VP .t / D sN T .t /IV sN .t / C 2Ns T .t /IV Q s .t / (13)

Copyright © 2016 John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control 2016; 26:4008–4034
DOI: 10.1002/rnc
HYBRID STATE RESETTING AND SECTOR CONFINEMENT OF TRAJECTORIES 4017

where

IV D SQ T IV C IV SQ (14)

At each kplane with k D 0; 1; :::; c   1, the Lyapunov stability condition is constrained within
  
.k/ .k1/ .k/ .k1/
s .t / C Zk s .t / s .t / C Fk s .t / 6 0

that is, at each kplane, the condition (13) is subject to


h i h iT
.k1/ .k/ .k1/ .k/
I
s .t / s .t / k s .t / s .t / 60 (15)

with
 
 Fk  Zk  Zk
I k D 2 R22 (16)
 Fk 1
The continuous flow Lyapunov stability restriction can be formulated as follows:
sN T .t /IV sN .t / C 2Ns T .t /IV Q s .t / < 0
subject to (17)
N
sN .t /Ik sN .t / 6 0; k D 0; 1; :::; c   1
T

where
2 3
0 0 (k+1-col) # 0
6 0 0 ::: 07
INk D4 2 Rnn
(k+1-row) 7! ::: I k ::: 5
0 0 ::: 0
Let us assume that Q s .t / D 0 (system without exogenous inputs and some cases with constant
input reference), so applying the S-procedure the stability condition (17) is achieved if there exists
0 ; 1 ; :::; c  1 > 0 2 R such that
 1
cX
IV  D IV  k INk 6 0 (18)
kD0

As a particular simplification, we can choose


2 3
1 0 0 ::: 0
6 0 V1 0 ::: 0 7
6 7
IV D 6 0 0 V 2 ::: 0 7 2 Rnn (19)
4 ::: ::: ::: ::: ::: 5
0 0 0 ::: Vn1
where Vi > 08i D 1; 2; : : : ; n  1.
This implies the following:
.n1/
V .t / D s.t /2 C V1 sP .t /2 C V2 sR .t /2 C ::: C Vn1 s .t /2
.n1/ .n/
VP .t / D 2s.t /Ps .t / C 2V1 sP .t /Rs .t / C 2V2 sR .t /«
s .t / C ::: C 2Vn1 s .t / s .t /

In this case, a sufficient, but not necessary, condition to achieve the stability requirement VP .t / < 0
is as follows:
.k/ .kC1/
s .t / s .t / < 0; k D 0; 1; : : : ; n  1 (20)

Copyright © 2016 John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control 2016; 26:4008–4034
DOI: 10.1002/rnc
4018 J. A. GONZÁLEZ ET AL.

 
.k/ .kC1/
so the trajectories in the s .t /; s .t / planes for k D 0; 1; : : : ; n  1 are restricted to quad-
 
.k/ .kC1/
rants II and IV with s .t / s .t / 6 0 . Note that this condition is sufficient but not necessary,
so
 it is possible to find stable solutions that flow at some time intervals in quadrants I and II
.k/ .kC1/
with s .t / s .t / > 0 by means of an adequate selection of the terms Vi .
This result gives us a nice objective to reach from the point of view of the qualitative trajectory
analysis that is in line with the idea of flow conic confinement. It also indicates that if the sector
confined target dynamics can be extended to higher order planes, we can then develop more sophis-
ticated strategies that allow us to confine the dynamics of the higher order terms of sN .t / into stable
flow regions.

5. DISCRETE JUMP MODE

This section addresses the implementation of the conic confinement of the modified error vector
sN .t / by means of jumps or resets applied on the controller states xc .t /. This conic confinement is
combined with the standard stability restriction that the Lyapunov function V .t / D sN T .t /IV sN .t /
should not increase after these discrete jumps.
The simplest case is illustrated in Figure 2, where the desired jump actions can be described
in the following way: When the trajectory in the continuous mode impacts the sector boundaries
in Figure 2 (that is, the lines sP .t / D F s.t / or sP .t / D Z s.t /), find the instantaneous jump
xc .t / that has to be applied to the controller states xc .t / such that, after the jump:
 (i) the Lyapunov increment is nonnegative, that is, V .t / 6 0
 (ii) the modified error goes back to the sector center line, sP .t / D M s.t /.

5.1. Condition (i) on the Lyapunov function


We have the following:

V .t / D V .t /  V t 6 0

where
  
V t D sN T t IV sN t
V .t / D sN T .t /IV sN .t /
and therefore,
 T 
V .t / D sN .t /T IV sN .t /  sN t IV sN t (21)

Lets define

SNp D SN .1 W n; 1 W p/ 2 Rnp (22)

SNc D SN .1 W n; p C 1 W n/ 2 Rnc (23)

From (44), the reset and after-reset conditions can be written as follows:
   
sN t D SNc xc t C SNp xp t C N s t
sN .t / D SNc  .t / C SNp xp .t / C N s .t /
Because in the discrete jump mode, the plant states and exogenous inputs remain unchanged, and
it is possible to set
 
SNp xp t C N s t D SNp xp .t / C N s .t /

Copyright © 2016 John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control 2016; 26:4008–4034
DOI: 10.1002/rnc
HYBRID STATE RESETTING AND SECTOR CONFINEMENT OF TRAJECTORIES 4019

so the expression for the discrete jump in the modified error augmented state space is as follows:
   
sN .t / D sN t C SNc  t  xc t (24)
Applying (21) and (24), the discrete jump stability condition is as follows:
 T   
xj t QJ xj t C 2Ns T t QS xj t 6 0 (25)
with
  
xj t D  t  xc t
QJ D SNcT IV SNc 2 Rcc
QS D IV SNc 2 Rnc
We have that in the following:

 xjT .t /Qxj xj .t / > 0 so 2Ns T t Qs xj .t / must be as negative as possible in order to obtain
the appropriate result. The design of the discrete control law has to be accomplished by means

of a balance between the position  sN t of the augmented modified error and the size of the
jump has to be executed xj t  .
 that
 sN T t 7! 0 implies that xj t 7! 0 if (25) has to be achieved. This means that the reset
jump distance decreases and it becomes null in the steady state, that is, the continuous and
discrete control laws are equal at the stationary equilibrium origin.

5.2. Condition (ii) on the modified error jump.


Let us assume that we have chosen to apply the reset confinement at a number of planes  equal

to the number of controller states (that is c  D c). At the jump instants t , the state xO t D
 T   T   T 
xp t  xc t may freely change its last c components, in any direction xc D  t 
xc t .
From Equation (44), sN .t / D SN x.tO / C N s .t /, and because N s depends only on the reference deriva-
tives, that are assumed to be continuous, the instantaneous jump Ns in sN is produced only by the
 >  >
jump xO D 0> ; xc> , that is, Ns D SN 0> ; xc> . 
Taking SNc from (23) and using theprevious reasoning,
 we can move the state from a point
sN t 2
Rn to an after-reset point N .t  / D sN t C Ns t 2 Rn performing any jump Ns t that can be
 s
expressed, for any xc t , in the form:
 
Ns t D SNc xc t
On the other hand, if we use 0; 2; : : : k  1, planes for conic confinement, the after-reset values
should belong to the central lines, that is,
sP .t / C M0 s.t / D 0; sR .t / C M1 sP .t / D 0; : : :
Equivalently,
0 1
0 1 s.t /
 M0 1 0 : : : : : : B sP .t / C
@0 B C
 M1 1 0 : : : A B sR .t / C D 0;
0 0  M2 1 ::: @ A
::
:
or, in compact form:
ƒNs .t / D 0;
where ƒ 2 Rkn has rank k. Combining the previous ideas, the computation of the adequate
controller jump xc is as follows. The jump equation is

Copyright © 2016 John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control 2016; 26:4008–4034
DOI: 10.1002/rnc
4020 J. A. GONZÁLEZ ET AL.

      
sN .t / D SN xO t C xO t C N s t D SN xO t C SNc xc t C N s t :
The condition that the after-reset value belongs to the central line of all planes is as follows:
  
0 D ƒNs .t / D ƒSN xO t C ƒSNc xc t C ƒN s t ;
or
  
ƒSNc xc t D ƒSN xO t  ƒN s t (26)

formc Mxc D v, with M D ƒSc 2 R . In


This equation has the N kc
order that there exists a
jump solution xc t 2 R for all possible values of the vector v 2 Rk , the condition is



rank ƒSNc D k (27)
Because the sizes of the factors ƒ; SNc are .k  n/ and .n  c/, then it can be concluded that it
is an easy to satisfy condition whenever c > k, that is, when the number k of chosen planes is not
larger than the number c of controller states.

6. DESIGN PROCEDURE

In the previous sections, we have analyzed different aspects of the proposal. Now, these parts can
be combined in the form of an algorithmic procedure to design all the related parameters.
Step 1 Start with a model of the plant as in (2), a minimal realization .Ap ; Bp ; Cp / of dimension
p. Obtain the transfer function G.& / D N.& /=D.& / and its relative degree rd , with 1 6
rd 6 p.
Step 2 Choose the base linear controller as in (2), .Ac ; Bc ; Cc ; Dc /. It is suggested to detune the
controller such that the closed loop is stable and fast but oscillatory and overshooting.
 > > >

vector in (4) xO D xp ; xc , of dimension n D p C c. Obtain the
Step 3 Form the complete state
closed loop matrices AO1 ; BO 1 ; CO 1 as in (32).
Step 4 Define the class of references r.t / to be tracked. This can be implemented with a reference
generator block that provides in a noise-free way the required derivatives .r; P r;
R «r ; : : :).
Pq .j /
Step 5 Choose coefficients ˛0 ; ˛1 ; : : : ˛q defining the modified error s.t / D j D0 ˛j e .t /; with
˛q ¤ 0 and where e.t / D r.t /  yp .t /. Take at least q > .rd  1/ (Assumption 3.1). The
Taylor approximation proposed  in Section 4 is a good start point to develop this choice.
>
.n1/
Step 6 Introduce the new state sN D s; sP ; : : : ; s , obtain the input-dependent state transfor-
mation (8)(44) and the related differential equation (45):
sN .t / D SN x.t
O / C N s .t /; sPN .t / D SQ sN .t / C Q s .t /:
  
.c 1/ .c  /
Step 7 Choose the number c  of planes .s; sP /; .Ps ; sR /; : : : s ; s where conic confinement is
going to be implemented. With c controller states, use at most c  6 c planes.
Step 8 Fix the central slopes of the sectors, sP D M0 s, sR D M1 sP , : : :, for example, by taking
Mk D ˛2k , and choose the lower and upper slopes bounding the sectors: Fk < Mk
<  Zk .
Step 9 Check the LMI (18) SQ > I C I SQ  †k k INk 6 0, (Lyapunov condition in the flow
mode). If the LMI is found feasible, continue. Otherwise, go back to Steps 8, 7, or 5 and
redefine choices.
Step 10 Check conditions (25)(26)(27) (Lyapunov conditions in the jump mode). In particular, if
the rank condition (26) holds, then the jump is feasible and the resulting Lyapunov test is
(25) xj> QJ xj C 2Ns > QS xj 6 0, to be applied to points sN .t / on any of the sector bounding
lines, and for jump values given by (26). If the resulting condition (25) is found unfeasible,
go back to Steps 8, 7, or 5 and redefine choices. If it is feasible, finish the procedure.

Copyright © 2016 John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control 2016; 26:4008–4034
DOI: 10.1002/rnc
HYBRID STATE RESETTING AND SECTOR CONFINEMENT OF TRAJECTORIES 4021

7. EXAMPLE APPLICATIONS

7.1. Second order plant reset design procedure


Let us consider the control of a second order plant given by the following:
xP p1 .t / D xp2 .t /
xP p2 .t / D u.t /
yp .t / D xp1 .t /
In this example, we focus our attention on the design procedure of reset controllers, and the results
will be compared with the performance obtained with a sliding mode controller with a super-twisting
approach given by
s.t / D e.t / C ˛ e.t
P /
1
u.t / D cjs.t /j 2 sign.s.t // C !.t /
!.t
P / D bsign.s.t //
with parameters c D 10:0, b D 5:0 when the reference profile is given by r.t / D 2cos.t /.
The design of the first order reset controller is based on a classic PI with transfer function:
Ki
Cf:o: .& / D Kp C
&
and parameters:
Ki D 0:1 Kp D 5:0
The sector confined target dynamics parameters are selected by means of the Taylor approach
presented in Section 4, so we have set the following:
2
˛ D 0:6 M0 D D 3:33 F0 D 0:5M0 D 1:67 Z0 D 2:0M0 D 6:67
˛
From (26), it is obtained the discrete control law as follows:
     
Ki ˛xc t D xp1 t ˛.2  Kp /  rP t  ˛ rR t C xp2 t 2˛ 2 C 1
    
 2˛ r t C ˛ rP t C Kp ˛r t C Ki ˛xc t
In the case of the second order reset system, the base linear controller is given as follows:
²
xP c .t / D Ac xc .t / C Bc e.t /
Cs:o: W
u.t / D Cc xc .t / C Dc e.t /
with
   
01 0  
Ac D Bc D Cc D Ki 0:1Ki Dc D Kp
00 1
At the second order reset system, we take the same parameters of the first order reset controller
for the first plane of confinement and set
 M1 D  M0  F1 D  F0  Z1 D  Z0
in order to define the conic section in the second confinement plane .Ps .t /; sR .t //.
The matrix M D ƒSNc in (26) that determines the reset jumps in the first two planes is as follows:
 
Ki ˛ 0:1Ki ˛
M D
Ki .2˛ 2 C 1/ 0:1Ki .2˛ 2 C 10˛ C 1/
with rank .M / D 2.

Copyright © 2016 John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control 2016; 26:4008–4034
DOI: 10.1002/rnc
4022 J. A. GONZÁLEZ ET AL.

Figure 3. Plant output and control effort with reference profile r.t/ D 2cos.t/.

Figure 3 shows the results obtained with the proposed controllers. This figure includes a detailed
view of the control signal when the systems achieve the steady state, where it is clear the reduction
of the chattering in the case of the reset controllers with respect to the sliding mode design.
Figure 4 shows the trajectories in .s.t /; sP .t // and .Ps .t /; sR .t // where we can observe the difference
in the confinement properties between the first order (with confinement only in the first plane) and
the second order reset controller (with confinement in both planes).
Figure 5 shows the responses obtained when the reference frequency is multiplied by two. We
can see that the SMC needs to be recalculated in order to obtain a zero error response at the steady
state but the reset controllers keep this property without any change. This invariant property with
respect to the reference frequency is related to the confinement properties of the hybrid design, but
this result will be formalized in future works.

7.2. PI closed loop stabilization and loop shaping reset control


In this section, we explore the application of the hybrid design with the second order plant proposed
in [34] where a generalized CNF control technique is used to track several references. The transfer
function of the second order plant is as follows:
100
P .& / D
&2  5& C 10
Let us emphasize some important remarks about this example:
 The plant has two complex unstable roots at & D 2:5 ˙ 1:93j .
 As specified in [34], we use a saturation value of usat D 2 that limits the maximum control
effort that can be applied to the plant, so we can test the robustness of the bybrid design with
respect to this nonlinear limitation in the closed loop system.
 It is not possible to design a PI controller that stabilizes the linear closed loop. However,
we use the hybrid transformation with a PI controller and a linear unstable-based closed
loop response to show the properties of the reset design in order to obtain asymptotic stable
tracking responses.

Copyright © 2016 John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control 2016; 26:4008–4034
DOI: 10.1002/rnc
HYBRID STATE RESETTING AND SECTOR CONFINEMENT OF TRAJECTORIES 4023

Figure 4. Trajectories in .s.t/; sP .t// and .Ps .t/; sR .t// with reference profile r.t/ D 2cos.t/.

Figure 5. Plant output and control effort with reference profile r.t/ D 2cos.2t/.

 We also design, employing the ©Matlab sisotool, a second order loop shaping controller that
achieves a stable response to a step reference, but with a significant error at the steady state, so
we can test if the hybrid design corrects this problem.
Let us start with the design of the controllers that we want to compare:
 PI control
We consider a PI controller with
Kc D 1:03 Ki D 0:18

Copyright © 2016 John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control 2016; 26:4008–4034
DOI: 10.1002/rnc
4024 J. A. GONZÁLEZ ET AL.

giving rise to a second order closed loop linear with transfer function
102:9& C 18:02
T .& / D
&3  5& 2 C 112:9& C 18:02
which has two unstable poles, so the base linear system with the PI controller is unstable.

Figure 6. Second order plant. Step reference.

 

Figure 7. Second order plant. Sinusoidal reference r.t/ D sin 2 t C 6
.

Copyright © 2016 John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control 2016; 26:4008–4034
DOI: 10.1002/rnc
HYBRID STATE RESETTING AND SECTOR CONFINEMENT OF TRAJECTORIES 4025

 Loop shaping control


Using ©Matlab sisotool, we design the following second order controller:

0:2& 2 C 2768& C 7895


C.& / D
& 2 C 170& C 11706
that stabilizes the response but with an error at the steady state for a step reference.
 State feedback control
In this case, the control law is given by

u.t / D Nr r.t /  Kxp .t /

with Nr D 2:240 and K D Œ2:14; 0:35 obtained with poles placement at psf1 D 14 and
psf2 D 16.
 Reset control
In this case, the plant has a relative degree rd D 2 so q D r  1 D 1, and we have that

s.t / D e.t / C ˛ e.t


P /

Using the same procedure as in the previous example for the selection of the reset
parameters, we have set

˛ D 0:1 M D 20 F D 15 Z D 25

Note that in the case of the second order loop shaping controller we apply the confinement
in the first two planes.
Next, we summarize the results obtained in the present example:
 Step response. Figure 6 shows the results for a step reference. We can see that all the controllers
achieve an asymptotic error convergence, but the loop shaping linear controller has an error at
the steady state that is corrected in its hybrid transformation.

 

Figure 8. Second order plant. Multiple sinusoidal reference r.t/ D 1 C 0:3 sin 2 t C 4
C 0:1 sin.6 /.

Copyright © 2016 John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control 2016; 26:4008–4034
DOI: 10.1002/rnc
4026 J. A. GONZÁLEZ ET AL.

Figure 9. Second order plant. Transcendental reference r.t/ D e sin.2 t / .

 Sinusoidal response. Figure 7 shows the tracking response for sinusoidal reference


r.t / D sin 2 t C
6
where the hybrid controllers achieve a rise time similar to the CNF example. Figure 8 shows
the results for the case of a multiple sinusoidal reference.
 Transcendental response. Finally, if we consider the transcendental reference problem proposed
in [34], we have that

r.t / D e sin.2 t /

The results for this reference can be consulted in Figure 9.

8. CONCLUSIONS AND FUTURE WORKS

This paper presents a new proposal for HRC. The proposal belongs to the class of reset controllers
having the reset triggered by input-output sector conditions (such as e.t /  u.t / 6 0). The main
differences are that our reset conditions are not based on the input-output, but on the error alone e.t /
or, more precisely, on a certain modified error (typically s.t / D e.t / C ˛ e.t P /) and its derivatives
sN D .s; sP ; : : :/. Furthermore, the controller state xc is not reset to zero, but to a value that moves
back sN to the center of the sectors specified for .s; sP ; sR ; : : :/. The freedom in both the definition of the
modified error s.t / and on the choice of the preferred sectors gives enough design flexibility and, at
the same time, preserves conceptual simplicity.
The objective of achieving a tracking response with non-overshooting (and non-undershooting)
systems (NOUS) has already been addressed in the literature, see [27] and references therein. There
exist powerful NOUS techniques that allow to obtain these special class of transient responses,
useful in many applications, such as electromechanics safe positioning systems.
The NOUS approach is related to the so-called Byrnes–Isidori regulator for asymptotic track-
ing, see for example [35]. The idea is to represent the reference r(t) by an exosystem dw/dt= Sw,
r=Dw and force perfect tracking. By the internal model principle, the controller has to reproduce

Copyright © 2016 John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control 2016; 26:4008–4034
DOI: 10.1002/rnc
HYBRID STATE RESETTING AND SECTOR CONFINEMENT OF TRAJECTORIES 4027

the reference in its internal dynamics, this is obtained via a certain regulator equation in its lin-
ear and nonlinear versions [35]. The NOUS specialization is based on adding extra conditions
that ensure this constrained error response. There exist elegant solutions and design software for
NOUS controllers.
On the other hand, HRC seeks for non-overshooting errors as it is a relevant objective for plants
subject to fundamental limitations. The original Horowitz idea [3] was to reset over-tuned oscillatory
controllers. Thus, the principles in HRC and NOUS are different. NOUS is more systematic in
achieving the desired responses, but it requires to specify the exosytem dw/dt=Sw, and the value of
S affects the controller. HRC is more independent of the reference, and the exosystem does not have
to be modeled at all.
Both HRC and NOUS are state feedback (in principle, but could include observers), but the
smoothness of the NOUS solution u=Fx+Gw [27] is at the expense of direct dependence on state
feedback, Fx, whereas HRC is less dependent on the state: the flow mode is output feedback, and
the state is only used to detect the sector boundaries. Thus, some degree of inaccuracy in the state
could be allowed.
In summary, HRC and NOUS are coincident in pursuing constrained tracking errors, but because
they are techniques coming from different fields, they consequently have many different aspects.
Their combination may be an interesting topic for future research. Perhaps the powerful NOUS tools
could be adapted to manage the sector constraints in the linear flow mode, and the jump resets could
be used to address plant uncertainty or unexpected references.
Our reset control proposal, as commented in Section 3 and shown in Figure 3, has certain resem-
blance to the SMC, although there are many differences between them. In SMC, the design starts
with the plant, in the form dx=dt D Ax C Bu C f . (in the case of state feedback, y D C x D x).
The uncertainty f is assumed to satisfy a matching condition f D B
. In our HRC technique, we
assume a well-known linear plant, without uncertainty. The so-called reaching and sliding phases in
SMC have a different form in HRC. The reaching phase takes finite time in SMC, whereas in HRC,
it can be considered instantaneous: if conditions in Section 5 are met, the central lines of the sectors
can be reached by a single jump. After the reaching phase, the resulting dynamics are quite different
as well: in SMC, the state is forced to lie in some sliding surface F x D 0, persistently commut-
ing the level of the control action (in HOSMC this effect is injected inside the control dynamics).
In HRC, the state is constrained to conic sectors and is allowed to flow continuously, jumping back
only when beating on the boundary. Only in the special case of very narrow sectors in HRC a per-
sistent or frequent jump mode might appear. Even in this extreme case, the commutation rationale
is different.
The reason for the many differences is obvious: our approach does not come from the field of
SMC by proposing changes in its design principles. In contrast, our HRC proposal comes from the
different field of reset control and, by forcing some conic confinement into the flow mode, it exhibits
certain similarity to SMC. In any case, after comparison of the two approaches, the detected simi-
larities suggest, for further research, to study the potentials of merging these techniques to improve
benefits towards uncertain plants subject to fundamental limitations. As preliminary explorations,
the example section includes numerical comparisons by simulations using super-twisting SMC.
The purpose of the paper has been to clearly present and explain the new ideas, to find conditions
for Lyapunov stability, and to validate and compare the proposal on simulation examples. For ease
of reading, the derivations of many state-space equations (the controller is input–ouput in the flow
mode, but it requires state information for the discrete mode) have been moved to the Appendix.
In Sections 4 and 5, we have presented algebraic conditions for Lyapunov stability and solvability
of the sector confinement problem. The examples show successful performance of the proposal,
compared with a variety of alternative controllers.
As a main line of future work, we consider addressing the issue of robustness in a more formal
way, now explored only by means of simulation examples. For this purpose, the effect of exoge-
nous inputs should be properly bounded. It seems that if the stable confinement problem without
external signals is robustly solvable, then the same problem will be solvable with properly bounded
exogenous signals.

Copyright © 2016 John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control 2016; 26:4008–4034
DOI: 10.1002/rnc
4028 J. A. GONZÁLEZ ET AL.

A second line of future work is the issue of well-posedness. This could be solved in this way: if
the exogenous signal (reference r.t /) is small enough and the error sN D .s; sP ; : : :/ is large enough,
notice that in the flow mode we have a linear response, so the time required to travel from the sector
center to its boundary can be lower bounded, thus excluding Zeno behavior. When the error sN is
small and close to zero, we can adopt a simple-time regularization.
A third line of future work is to formalize the overcoming of linear limitation by means of this
new reset control technique and compare it with other nonlinear solutions, like SMC. This is, among
the three future works, probably the most difficult one. Fundamental linear limitations are expressed
in the form of time-domain or frequency-domain trade-offs between speed of response, robustness,
and so on. The approximate nature of their formulations makes it difficult to formally prove their
overcoming. Some related work has been carried out in [20], for a different class of reset controllers.

APPENDIX

In this appendix, we define, using a recursive formulation, the state dynamics, the error, and the
modified error dynamics for the SISO linear closed loop model (3).

Augmented space definitions

Let us define the augmented spaces as follows:


h iT
.n1/
r.t
N / D r.t / r.t P / r.t R /« r .t / ::: r .t / 2 Rn1
h iT
.n1/
e.t
N / D e.t / e.t P / e.t R /« e .t / ::: e .t / 2 Rn1
h iT
.n1/
sN .t / D s.t / sP .t / sR .t / «
s .t / ::: s .t / 2 Rn1
where n D p C c.
States dynamics

The first step in the dynamics formulation corresponds to the states dynamics equations. Our
goal here is to obtain an iterative formulation for the k-order states dynamics that begins with
the following:
xP p .t / D Mp xp .t / C Mc xc .t / C Bp Dc r.t / (28)

xP c .t / D Np xp .t / C Nc xc .t / C Bc r.t / (29)
where
Mp D Ap  Bp Dc Cp 2 Rpp Mc D Bp Cc 2 Rpc
Np D Bc Cp 2 Rcp Nc D Ac 2 Rcc
The closed loop states dynamics can be formulated as follows:
PO / D AO1 x.t
x.t O / C BO 1 r.t / (30)

O / D CO 1 x.t
y.t O / (31)
where
 
Mp Mc
AO1 D 2 Rnn (32)
Np Nc
 
Bp Dc
BO 1 D 2 Rn1 (33)
Bc

Copyright © 2016 John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control 2016; 26:4008–4034
DOI: 10.1002/rnc
HYBRID STATE RESETTING AND SECTOR CONFINEMENT OF TRAJECTORIES 4029

 
CO 1 D Cp 0 2 R1n (34)

Deriving the state terms,


RO / D AO2 x.t
x.t O / C BO 2 r.t / C BO 1 r.t
P / (35)

with
AO2 D AO21 2 Rnn
BO 2 D AO1 BO 1 2 Rn1
Deriving the state terms up to the k-order, we obtain
.k/
xO .t / D AOk x.t
O / C BN k r.t
N / (36)

where
 
BN k D BO k BO k1 : : : BO 1 0 : : : 0 2 Rnn

and
AOk D AOk1 2 Rnn
BO k D AOk1 BO 1 2 Rn1
Error dynamics

The error signal e.t / D r.t /  yp .t / can be obtained as

e.t / D LO 0 x.t
O / C r.t /

where
 
O 0 D Cp 0 2 R1n
L

The derivation of the error term leads to

P / D LO 1 x.t
e.t O / C FO1 r.t / C r.t
P /

where
LO 1 D LO 0 AO1 2 R1n
FO1 D LO 0 BO 1 2 R11
Applying the iterative procedure, as in the previous states dynamics section, it is obtained:
.k/
e .t / D LO k x.t
O / C FNk r.t
N / (37)

where
 
FNk D FOk FOk1 : : : FO1 1 0 : : : 0 2 R1n

and
LO k D LO k1 AO1 2 R1n
FOk D LO k1 BO 1 2 R11
We have that

N / D LN x.t
e.t O / C FN r.t
N / (38)

Copyright © 2016 John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control 2016; 26:4008–4034
DOI: 10.1002/rnc
4030 J. A. GONZÁLEZ ET AL.

where
2 3
2 3 1 0 0 0 ::: 0
LO 0 6 FO1
6 O 7 6 1 0 0 ::: 07 7
N D 6 L1 7 2 Rnn 6 07 nn
72R
L N
F D 6 F2 FO1
O 1 0 :::
4 ::: 5
4 ::: ::: ::: ::: ::: ::: 5
LO n1
FOn1 FOn2 ::: FO2 FO1 1
Applying (38), it is obtained:
O / D LN 1 .e.t
x.t N /  FN r.t
N // (39)
Differentiating (38), we obtain the dynamics equation related to the error augmented space as
follows:
PN / D LQ e.t
e.t N /  LQ FN r.t
N / C LN BN 1 r.t / C FN r.t
PN /

where
LQ D LN AO1 LN 1 2 Rnn


Because LN is the observability matrix of AO1 ; LO 0 , assuming observability, LN 1 is well defined,
and the matrix LQ has the structure:
2 3
0 1 0 0 ::: 0
60 0 1 0 ::: 0 7
LQ D 6
4 ::: ::: ::: ::: ::: ::: 5
7

lQ0 lQ1 lQ2 ::: ::: lQn1

where lQ0 ; lQ1 ; : : : are the coefficients of the characteristic polynomial of AO1 . After some calculation,
we have that
PN / D LQ e.t
e.t N / C Q e .t / (40)
with
2 3
0
6 0 7
6 7
Q e .t / D 6
6
::: 7
7
4 0 5
.n/
r .t / C N rN
and
 
N D 0 1 ::: n1 2 R1n
defined for k D 0; 1; ; : : : ; n  1 as follows:
n1
X
k D FOnk  lQi FOi k 2 R
i Dk

where FO0 D 1. The error dynamics can be formulated in polynomial form as follows:
.n/ .n1/ .n/
e .t /  lQn1 e .t /  :::  lQ0 e.t / D r .t / C N rN (41)
Modified error dynamics

From definition (6), the modified error is as follows:


s.t / D ˛N e.t
N / (42)

Copyright © 2016 John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control 2016; 26:4008–4034
DOI: 10.1002/rnc
HYBRID STATE RESETTING AND SECTOR CONFINEMENT OF TRAJECTORIES 4031

where
 
˛N D ˛0 ˛1 : : : ˛q 0 : : : 0 2 R1n

with ˛j 2 R. We have that

s.t / D SO0 x.t


O / C s .t /

where
SO0 D ˛N LN 2 R1n
N /2R
s .t / D ˛N FN r.t
The derivation of the modified error term leads to

sP .t / D SO1 x.t
O / C HO 1 r.t / C P s .t /

where
SO1 D SO0 AO1 2 R1n
HO 1 D SO0 BO 1 2 R
Applying the iterative procedure, it is obtained:
.k/ .k/
s .t / D SOk x.t
O / C HN k r.t
N / C  s .t / (43)

where
SOk D SOk1 AO1 2 R1n
HO k D SOk1 BO 1 2 R
and
 
HN k D HO k HO k1 : : : HO 1 0 : : : 0 2 R1n

The dynamics equation with respect to the closed loop states is as follows:

sN .t / D SN x.t
O / C N s .t / (44)

where
2 3 2 3
SO0 s .t /
6 SO1 7 6 P s .t / 7
SN D 6 7
4 ::: 5 2 R
nn
N / C O s .t / 2 Rn
N s .t / D HN r.t O s .t / D 6
4 :::
7 2 Rn
5
.n1/
SOn1  s .t /
and
2 3
0 0 0 0 ::: 0
6 HO 1 0 0 0 ::: 07
6 7
HN D 6
6 H O 2 HO 1 0 0 ::: 07 72R
nn

4 ::: ::: ::: ::: ::: ::: 5


HO n1 HO n2 ::: HO 2 HO 1 0
Applying the procedure used in the error, we have that the dynamics for the modified error are
as follows:

sPN .t / D SQ sN .t / C Q s .t / (45)

where

Copyright © 2016 John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control 2016; 26:4008–4034
DOI: 10.1002/rnc
4032 J. A. GONZÁLEZ ET AL.

2 3
0 1 0 0 ::: 0
60 0 1 0 ::: 0 7
SQ D SN AO1 SN 1 D4 2 Rnn
::: ::: ::: ::: ::: ::: 5
sQ0 sQ1 sQ2 ::: ::: sQn1

2 3
0
6 0 7
6 7
Q s .t / D 6
6
::: 7
7
4 0 5
.n/
N r.t
N / C  s .t /  O O s .t /
with
 
N D 0 1 ::: n1 2 R1n
defined for k D 0; 1; ; : : : ; n  1 as follows:
n1
X
k D HO nk  sQi HO i k 2 R
i Dk

and
 
O D sQ0 sQ1 sQ2 : : : : : : sQn1 2 R1n
Q
is the last row of the matrix S.
Assuming that SN is invertible, the previous expression is well defined. Notice that both LQ and SQ
are similar to AO1 and thus lQk D sQk . The modified error dynamics can be formulated in polynomial
form as
.n/ .n1/ .n/
s .t /  sQn1 s .t /  : : :  sQ0 s.t / D N r.t
N / C  s .t /  O O s .t / (46)
Well-posedness of the state transformations
Previously, we have used three different state vectors: the natural coordinates x, O the error coordi-
P : : :/> , and the modified error sN D .s; sP ; : : :/> . They are related by Equations (44)
nates eN D .e; e;
and (38); without external inputs, they take the form:
N x;
eN D L O sN D SN x:
O
N SN are invertible. Let L.& / D NL .&/
In this section, we show that under mild conditions L; DL .&/
be the
(strictly proper) open loop (base linear) transfer function. Then,
NL .&/
 LN 1 exists if and only if DL .&/
is minimal.
To see this, note that from (38) and previous relations
0 1 0 1
LO 0 LO 0
B LO 1 C B LO 0 AO1 C
B C B C
bar L D B : C D B :: C
@ :: A @ : A
O
Ln1 O
L0 A O n1
1

where AO1 is the state matrix of the base linear closed loop system and L O 0 D .Cp ; 0/ is the


output matrix for the error e.t / D r.t /  yp .t /. Then, L is invertible if the pair AO1 ; LO 0
N
is observable, if and only if the transfer function representations of the closed and open loops,
NL .& /=.DL .& / C NL .& // and NL .& /=DL .& / do not contain zero-pole cancellations, that is, if we
use minimal realizations.

Copyright © 2016 John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control 2016; 26:4008–4034
DOI: 10.1002/rnc
HYBRID STATE RESETTING AND SECTOR CONFINEMENT OF TRAJECTORIES 4033

On the other hand, let .& / D & n C n1 & n1 C : : : C 1 & C 0 be the closed loop characteristic
equation and ˛.& / D ˛0 C ˛1 & C : : : ˛q & q the polynomial that defines the modified error (6).
˛.&/
 SN 1 exists if and only if .&/
is minimal.
To see this, note that from (44) and previous relations
0 1 0 N 1
SO0 ˛N L
B SO1 C B ˛N LN AO1 C
B C B C
SN D B : C D B :: C
@ :: A @ : A
SOn1 ˛N LN AO n1
1

Thus,
0 1
˛N
B ˛N AOe C
B C
SN LN 1 D B :: C ; with N AO1 LN 1
Ae D L
@ : A
˛N AOen1

It is seen from the structure of the error vector eN D P : : :/> that the error equations ePN D Ae eN
.e; e;
have a state matrix:
0 1
0 1 ::: 0 0
B 0 0 ::: 0 0 C
B : :: :: ::C
Ae D B B :: : ::: : :
C
C
@ 0 0 ::: 0 1 A
0 1 : : : n2 n1
This matrix allows us to define the state equations:
0 1
0
B :: C
B C
e.t N / C B : C n .t /; s.t / D .˛0 ; ˛1 ; : : : ˛q ; 0; : : : ; 0/e.t
NP / D Ae e.t N /
@0A
1

with n .t / being an auxiliary input depending on the reference and on its derivatives. Note that this
system has a transfer function ˛.& /=.& /. Collecting the previous relations, we conclude that SN is
invertible if and only if the pair .Ae ; ˛/
N is observable, if and only if the polynomial ˛.& / does not
contain a common root with the closed loop characteristic polynomial .& /.

ACKNOWLEDGEMENTS

This work was supported in part by the Spanish Ministry of Economy and Competitiveness under
the Projects DPI2012-31303 and DPI2013-47100-C2-2-P.

REFERENCES
1. Clegg J. A nonlinear integrator for servomechanism. Transactions A.I.E.E.m, Part II 1958; 77:41–42.
2. Horowitz I, Rosenbaum P. Nonlinear design for cost of feedback reduction in systems with large parameter
uncertainty. International Journal of Control 1975; 24(6):977–1001.
3. Krishnan K, Horowitz I. Synthesis of a nonlinear feedback system with significant plant-ignorance for prescribed
system tolerances. International Journal of Control 1974; 19(4):689–706.
4. Beker O. Analysis of Reset Control Systems PHD Thesis. University of Massachussetts, Department of Electrical and
Computer Engineering, 2001.
5. Beker O, Hollot C, Chait Y, Han H. Fundamental properties of reset control systems. Automatica 2004; 40:905–915.

Copyright © 2016 John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control 2016; 26:4008–4034
DOI: 10.1002/rnc
4034 J. A. GONZÁLEZ ET AL.

6. Beker O, Hollot C, Chait Y. Plant with integrator: an example of reset control overcoming limitations of linear
systems. IEEE Transactions on Automatic Control 2001; 46(11):1797–1799.
7. Guo Y, Wang Y, Xie L, Zheng J. Stability analysis and design of reset systems: theory and an application. Automatica
2009; 45:492–497.
8. Chait Y, Hollot C. On Horowitz’s contributions to reset control. International Journal of Robust and Nonlinear
Control 2002; 12:335–355.
9. Nešić D, Zaccarian L, Teel A. Stability properties of reset systems. Automatica 2008; 44(8):2019–2026.
10. Loquen T, Tarbouriech S, Prieur C. Stability of reset control systems with nonzero reference. Proceedings of the 47th
IEEE Conference on Decision and Control, Cancun, Mexico, December 2008; 9–11.
11. Loquen T, Tarbouriech S, Prieur C. Stability of reset control systems with nonzero reference. Proceedings of the 47th
IEEE Conference on Decision and Control, Cancun, Mexico, December 2008; 9–11.
12. Tarbouriech S, Da Silva J, Garcia G. Delay-dependent anti-windup strategy for linear systems with saturating inputs
and delayed outputs. International Journal of Robust and Nonlinear Control 2004; 14:665–682.
13. Aangenent W, Witvoet G, Heemels W, van de Molengraft M, Steinbuch M. Performance analysis of reset control
systems. International Journal of Robust and Nonlinear Control 2009; 20(11):1213-1233.
14. Baños A, Dormido S, Barreiro A. Limit cycles analysis in reset control systems with reset band. Nonlinear Analysis:
Hybrid Systems 2011; 5(2):163–173.
15. Baños A, Dormido S, Barreiro A. Stability analysis of reset control systems with reset band. 3rd IFAC Conference
on Analysis and Design of Hybrid Systems, Zaragoza, Spain, 2009; 180–185.
16. Baños A, Dormido S, Barreiro A. Reset control systems with reset band: well-posedness and limit cycle analysis.
19th Mediterranean Conference on Control and Automation, Corfu, Greece, June 2011; 20–23.
17. Baños A, Barreiro A. Delay-independent stability of reset systems. IEEE Transactions on Automatic Control
2009; 54(2):341–346.
18. Baños A, Carrasco J, Barreiro A. Reset times-dependent stability of reset control with unstable base systems. ISIE
Conference, Vigo, Spain, June 2007; 4–7.
19. Baños A, Barreiro A. Delay-dependent stability of reset control systems. Proceedings of the 2007 American Control
Conference, New York, USA, July 2007; 11–13.
20. Prieto J, Dormido S, Barreiro A. Frequency domain properties of reset systems with multiple reset anticipations.
Control Theory & Applications, IET 2013; 7(6):796-809.
21. Baños A, Barreiro A. Reset Control Systems. Springer-Verlag London Limited, 2012.
22. Zhao G, Neṡic D, Tan Y, Wang J. Open problems in reset control. 52nd IEEE Conference on Decision and Control,
Florence, Italy, December 2013; 3326–3331.
23. Jostein B, Johansen T, Smogeli O, Sorensen A. Lyapunov-based integrator resetting with application to marine
thruster control. IEEE Transactions on Control Systems Technology 2008; 16(5):908–917.
24. Guo Y, Wang Y, Xie L, Zheng J. Frequency-domain properties of reset systems with application in hard-disk-drive
systems. IEEE Transactions on Control Systems Technology 2009; 17(6):1446–1453.
25. Zheng Y, Chait Y, Hollot C, Steinbuch M, Norg M. Experimental demostration of reset control systems. Control
Engineering Practice 2000; 8:113–120.
26. Nešić D, Zaccarian L, Teel A. First order reset elements and the Clegg integrator revisited. Proceedings American
Control Conference 2005; 1:563–568.
27. Schmid R, Ntogramatzidis L. A unified method for the design of nonovershooting linear multivariable state-feedback
tracking controllers. Automatica 2010; 46(nı 2):312–321.
28. Krstic M, Bement M. Nonovershooting control of strict-feedback nonlinear systemss. IEEE Transactions on
Automatic Control 2006; 51(nı 12):1938–1943.
29. Artstein Z. Stabilization with relaxed controls. Nonlinear Analysis: Theory, Methods & Applications 1983;
7(11):1163-1173.
30. Lin Y, Sontag E. A universal formula for stabilization with bounded controls. Systems & Control Letters 1991;
6(16):393–397.
31. Sontag E. A universal construction of Artstein’s theorem on nonlinear stabilization. Systems & Control Letters
1989; 13(2):117-123.
32. Cai X, Han Z. Universal construction of control Lyapunov functions for linear systems. Latin American Applied
Research 2006; 36(1):15–22.
33. Slotine J, Li W. Applied Nonlinear Control. Prentice-Hall, 1991.
34. Cheng G, Peng K, Chen B, Lee T. Improving transient performance in tracking general references using composite
nonlinear feedback control and its application to high-speed-table positioning mechanism. IEEE Transactions on
Industrial Electronics 2007; 54(nı 2):1039–1051.
35. Sastry S. Nonlinear Systems: Analysis, Stability, and Control, Vol. 10. Springer Science & Business Media, 2013.

Copyright © 2016 John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control 2016; 26:4008–4034
DOI: 10.1002/rnc

S-ar putea să vă placă și