CEAI, Vol.20, No. 1, pp.

36-47, 2018 Printed in Romania

Event-Based PID Control:

Application to a Mini Quadrotor Helicopter

Sylvain Durand ∗ Bruno Boisseau ∗∗ Nicolas Marchand ∗∗

J. Fermi Guerrero-Castellanos ∗∗∗
∗
ICube, INSA Strasbourg, Univ. Strasbourg, CNRS (UMR 7357),
Strasbourg, France
(e-mail: [email protected]).
∗∗
GIPSA-lab, Univ. Grenoble Alpes, CNRS (UMR 5216),
Grenoble, France.
∗∗∗
Autonomous University of Puebla (BUAP), Faculty of Electronics,
Puebla, Mexico.

Abstract: Although periodicity simplifies design and analysis in control theory, it is no more
adapted for embedded systems because it results in a conservative usage of resources. Indeed,
the control signal is computed and updated at the same rate regardless whether is really required
or not. On the other hand, event-driven sampling calls for resources whenever they are indeed
necessary. Event-based PID controllers are proposed in this paper as an alternative to classical
PID approaches, with the same performance but reducing the control updates. The algorithms
are built here without safety limit condition contrary to the seminal event-based PID setup
that was originally proposed by Årzén (1999), in order to reduce the periodicity even more.
Both integral and derivative terms are considered. The different approaches are tested for
controlling the position of a real-time mini quadrotor helicopter. A reduction of the computing
and communication resources utilization is demonstrated for similar final performance.

Keywords: Event-based control, PID control, mini quadrotor helicopter.

1. INTRODUCTION (for instance when the dynamics of the controlled system

varies). Typical event-detection mechanisms are functions
Advances in fabrication and design of VLSI (very-large- on the variation of the state (or at least the output)
scale integration) circuits have resulted in the availability of the system, like in Årzén (1999); Durand and Marc-
of low-cost, low-power, small-sized computational elements hand (2009); Sandee et al. (2005); Sánchez et al. (2009a);
that are able to communicate via shared and possibly wire- Åström and Bernhardsson (2002); Heemels et al. (2009);
less communication network. Furthermore, development in Lunze and Lehmann (2010); Eqtami et al. (2010). Al-
MEMS (microelectromechanical systems), which provide though event-based control is well-motivated and theoret-
solid-state sensors and actuators, complements these ad- ical results have been stated in the literature in the last
vances. The net result is ubiquitous sensing, computation, decades, only few works report practical implementation.
actuation and communication that allow the development It has notably been shown in Åström and Bernhardsson
of the so called cyber-physical systems (CPS). These CPS (2002) that the control law can be updated less frequently
represent an integration of computing devices with phys- than with a periodic scheme while still ensuring the same
ical processes. In practice, embedded computers and net- performance. Stabilization of linear and nonlinear systems
works monitor and control physical processes (usually with is analyzed in Velasco et al. (2009); Tabuada (2007); Marc-
feedback loops) which, in return, affect computations and hand et al. (2013); Durand et al. (2014), where the events
communications. The use of small-sized computational are related to the variation of a Lyapunov function or the
elements emerges as an obvious trend to save space, weight time derivative of a Lyapunov function (and, consequently,
and energy. However, their implementation can result in to the state too).
additional challenges since the traditional feedback loop
that operates in continuous-time or at a fixed sampling Proportional-integral-derivative (PID) controllers are the
rate cannot be used anymore. Therefore, resource-aware most employed controllers in industry or even in hightech
implementations are required. devices because of their capability to provide a satisfac-
tory performance for many processes with a relative easy
In this context, recent works addressed alternative frame- design and by the availability of a large number of tuning
works where the control law is event-driven. Whereas the rules. However, the growing decentralization of current
control law is computed and updated at the same rate processes and the number of involved computational el-
regardless whether is really required or not in the classical ements demand the redesign of PID control schema. For
time-triggered approach, the event-based paradigm relaxes this reason, event-triggered PID controllers were addressed
the periodicity of computations and communications in by numerous researches in the last years. An original
calling for resources whenever they are indeed necessary
Control Engineering and Applied Informatics 37

and simple event-based PID control architecture was pro-

time instants of triggering
posed in Årzén (1999). The suggested scheme updates
the control signal only when the system output crosses
a certain threshold. Whereas an event was enforced with ē signal

a mix of level crossings and the use of a timer in order to

bound sampling period (for stability reason) in the initial hk+1
approach, this bounded period was then removed in Du- hk
rand and Marchand (2009) because, in fact, the Nyquist- tk−1 tk tk+1
time

Shannon sampling condition is no more consistent in the (a) Representation of the time instants and the sampling intervals.
asynchronous framework. Nevertheless, a safety limit is
uselessly applied in reported works like in Sánchez et al.
(2009b); Heemels et al. (2009); Mounier et al. (2011). Re-
cently, different event-based PID and PI algorithms were level-crossing detection

also developed for open-loop stable first-order systems ē

h̄
with delay, where simulations and experimental results
verify the effectiveness, see Beschi et al. (2012); Ruiz et al. hk
(2014). As evidenced by the above reviewed literature, very time

little attention has been dedicated to design simple and tk−1 tk = tj

tj−1
computational efficiently event-triggered PID controllers
“without safety limit condition”. Therefore, it is proposed (b) Sequences of tj and tk .
here to clearly highlight the efficiency of such an approach
by implementing a controller for an open-loop unstable Fig. 1. Event-driven control scheme.
system, that is an unmanned aerial vehicle, and testing instant for the event-based scheme really occurs at time
the proposed approach in real-time. tk = tj (and not t), as represented in Fig. 1(b). This will
The rest of the document is organized as follows. Classical be further discussed in section 3, however, next figures will
(time-triggered) PID controllers and the original event- not be as detailed anymore for the sake of not overload.
based version from Årzén (1999) are introduced in sec-
tions 2 and 3. Event-based PID controllers without safety 2. CLASSICAL (DISCRETE-TIME) PID CONTROL
limit condition are then detailed in section 4. Different
algorithms are presented but, whereas only the integral Equations of the continuous-time PID controller are
term was addressed in Durand and Marchand (2009) —
as well as in Årzén (1999)) — both integral and derivative u(t) = up (t) + ui (t) + ud (t) (1)

terms are concerned in the present paper. Furthermore, the up (t) = Kp e(t)
Z t
proposed algorithms are now based on different approxi-
u (t) = K e(t)dt
mation methods. A sketch of stability analysis is given with i i
t0
in section 5. The experimental platform is depicted in
section 6. The different approaches are tested for control- de(t)
ud (t) = Kd
ling the position of a real-time mini quadrotor helicopter dt
using a motion capture system with deported controller. where u is the control signal and e is the error between the
Experimental results highlight the capabilities of the pro- measurement and a given reference to track, up , ui and ud
posed approaches in reducing the control/communication are the proportional, integral and derivative parts of the
updates while maintaining similar performance. Conclu- PID controller, Kp , Ki and Kd are tunable parameters.
sions and future works finally end the paper in section 7.
In frequency (Laplace) domain, this gives

Up (s) = Kp E(s)
Notation: Afterwards, let j ∈ Z denote the beginning

time of the current control sample in the periodic sampling Ui (s) = Ki E(s) (2)
scheme, that gives tj := j h̄, where h̄ is the (constant) s

sampling period. Also, let k ∈ Z denote the beginning Ud (s) = Kd sE(s)
time of the current control sample in the non-uniform where s is the Laplace operator. Moreover, a low-pass filter
event-based sampling scheme, that is tk , while hk := is added in the derivative term (to avoid problems with
tk − tk−1 denotes the (varying) sampling interval. In high frequency measurement noise), which yields
the present paper, an event is enforced using a level-
crossing detection mechanism, that is when the signal Kd s
crosses a given (relative or absolute) threshold ē, also Ud (s) = E(s) (3)
1 + Tf s
called level. The different notations for the non-uniform
sampling scheme are represented in Fig. 1(a). where Tf is the filter’s time constant.
Remark 1.1. Fig. 1(a) is intentionally not thorough since, The control architecture is depicted in Fig. 2(a), where H
actually, an event can only be detected periodically, at a is the plant to control, y its sensed output and ysp a given
discrete instant tj (and not at any instant t) by construc- reference to track. Cc is the continuous-time controller,
tion of the event-based mechanism. Typically, if an event derived from equations (2)-(3). It is based on the error
occurs (because a level crossing is detected) at a given time signal e := ysp − y to compute the control signal u that
t between two sampling instants tj−1 and tj , the sampling acts on the plant input.
38 Control Engineering and Applied Informatics

+
The forward approximation gives
ysp (t) e(t) u(t) y(t)
Cc H
− ui (tj ) = ui (tj−1 ) + Ki h̄e(tj−1 )


continuous-time u (t ) = Tf − h̄ u (t ) + Kd e(t ) − e(t ) (6)
PID controller d j Tf
d j−1
Tf
j j−1

(a) Continuous-time PID controller.

2.2. The backward difference approximation (or implicit
Euler’s method ) is based on the first-order Taylor
+
ysp (tj ) e(tj )
Cd
u(tj ) DA + u(t)
H y(t) 1 − z −1
ZOH series expansion given by s = .
− h̄
y(tj ) sampler The backward approximation gives
discrete-time + AD
PID controller
ui (tj ) = ui (tj−1 ) + Ki h̄e(tj )
(b) Discrete-time PID controller.

ud (tj ) = Tf ud (tj−1 ) + Kd e(tj ) − e(tj−1 ) (7)
Tf + h̄ Tf + h̄
Fig. 2. Classical PID controller setups.
2.3. The bilinear approximation (or Tustin’s method ),
Fundamental differences for a discrete-time version are based on the trapezoidal rule, is a first-order ap-
that i) the discrete controller operates on samples of the proximation of the natural logarithm function, that
sensed plant output rather than on continuous signals, i.e. is an exact mapping of the z-plane to the s-plane:
tj instead of t, and ii) the dynamics are implemented by 2 1 − z −1
algebraic recursive equations instead of differential equa- s= .
tions (Franklin et al. (1997)). The architecture is depicted h̄ 1 + z −1
in Fig. 2(b), where Cd is the discrete-time controller. The bilinear approximation gives
An analog-to-digital (AD) converter changes a physical

variable into a binary number (error of quantization is ui (tj ) = ui (tj−1 ) + Ki h̄ e(tj ) + e(tj−1 )
neglected in the present paper). The conversion occurs 2

repetitively every constant sampling period h̄, the sampled (8)
signal is y(tj ). The reference is directly considered as a ud (tj ) = 2Tf − h̄ ud (tj−1 ) + 2Kd e(tj ) − e(tj−1 )
2Tf + h̄ 2Tf + h̄
discrete signal ysp (tj ) here, but it could also be converted
and sampled. The error is e(tj ) := ysp (tj )−y(tj ). The con- These three first-order approximation methods are illus-
troller then feeds the control signal u(tj ), that is changed trated in Fig. 3 for the integral part computing. Note that
to a continuous variable u(t) by a digital-to-analog (DA) the different approximations work for smooth signals and
converter, with a constant value throughout the sampling are more accurate for smaller sampling period h̄. This
period h̄ thanks to a zero-order hold (ZOH). is why some improvements will be mandatory in a non-
Algebraic recursive equations for the digital controller can uniform sampling scheme where the (varying) sampling
be obtained making some approximations on the mapping interval can increase. This will be further explained in
of the s-plane in (2)-(3) to the z-plane. The resulting section 3.
algorithm in discrete-time domain is y(t) y(tj−1 ) y(t) y(t) y(tj−1 )
y(tj ) y(tj )

u(tj ) = up (tj ) + ui (tj ) + ud (tj ) (4)

where tj denotes the (periodic) sampling instants. tj−1 tj t tj−1 tj t tj−1 tj t
h̄ h̄ h̄
Remark 2.1. The control signal (4) is sent to the actuator (a) Forward approxi- (b) Backward ap- (c) Bilinear approxi-
of the plant as soon as it is available, assuming the mation. proximation. mation.
time delay is minimized by making the calculations of
the control law as short as possible (Åström and Murray Fig. 3. Comparison of three first-order numerical methods
(2008)). This explains why tj can be in both left and right of approximation used for discretization.
sides of formulas in the sequel.
Remark 2.2. Whereas the backward and bilinear methods
The proportional part is easily straightforward from (2) are numerically stable, the forward one can be numerically
unstable for some h̄: the numerical solution can grow very
up (tj ) = Kp e(tj ) (5) large for equations where the exact solution does not,
especially for stiff equations. Intuitively, this is because
However, several solutions exist for the integral and deriva- the forward approximation only considers the last value of
tive parts. In this section, first-order approximation meth- the signal to integrate (and not the current one). For this
ods are applied. Three cases are treated in particular. More reason, the forward method should not be a good solution
complex approximation methods could be applied (but when the sampling interval can become huge.
this is not the aim of the present paper).
Remark 2.3. The bilinear method reduces the error of
2.1. The forward difference approximation (or explicit approximation but is more complex in return. This is
Euler’s method ) is based on the first-order Taylor why this method will not be detailed in the sequel.
1 − z −1 Nevertheless, an extension to the techniques proposed
series expansion given by s = . afterwards is trivial.
h̄z −1
Control Engineering and Applied Informatics 39

3. EVENT-BASED PID CONTROL ui (tk+1 ) = ui (tk ) + Ki hk e(tk ) (11)

Tf Kd
3.1 Årzén’s event-based PID controller ud (tk ) = ud (tk−1 ) + e(tk ) − e(tk−1 )
Tf + hk Tf + hk
An original event-based PID controller was proposed for (12)
the first time in Årzén (1999). The basic setup consists in where the varying interval hk replaces the constant period
two parts: a time-triggered event detector and an event- h̄ in the conventional equations (6) and (7).
triggered PID controller (Ceb ), as depicted in Fig. 4. Remark 3.2. Instead of calculating the integral part at the
e(tk ) current time ui (tk ) based on the previous value of the
ysp (tj ) + e(tj ) time-triggered u(tk ) DA + u(t)
Ceb signal e(tk−1 ) — as this was the case in (6) — in (11)
event detector request ZOH
− it is pre-calculated at the current sampling time tk for the
y(tj ) sampler y(t) next sample (in order to reduce the control delay) using the
event-based + AD current value of the signal e(tk ). Then, the next integral
PID controller part ui (tk+1 ) will be used in the PID control law (4) at
the next event tk+1 . This is possible because the change
Fig. 4. Event-based PID controller setup. of the error remains lower than ē between two successive
events by construction, thanks to condition (9).
Time-triggered event detector:
In the sequel, the initial Årzén’s proposal is firstly im-
An event detector is used for level-crossing detection. This proved correcting a forgetfulness in the discretization of
first part is periodically sampled with the period h̄ (that the integral part (see section 3.2). Then, the safety limit
is the same as for the corresponding conventional time- condition (10) is removed and the event-based PID con-
triggered PID). Typically, it runs like a zero-order holder. troller is modified in consequence (see section 4), changing
However, the holding instants (afterwards called events) either the event condition (that is the way to generate
are non uniform in time and decided with respect to the an event in order to calculate a new control signal) or
input dynamics. Therefore, the aim of the event logic is the control equations (adapting the conventional control
to determine the events. Then, the output holds the input strategy to an event-driven framework).
signal constant between two successive events. A request
is also generated at each event to trigger the second part 3.2 Discretization improvement for the integral part
(i.e. the event-triggered PID controller).
The method of discretization has to be carefully selected
In the original Årzén’s setup, an event occurs:
in a non-uniform sampling interval scheme. For instance,
i) either when the relative measurement crosses a given a forgetfulness was done in the original Årzén’s approach,
level ē, that is when as noticed in Durand and Marchand (2009). Indeed, it is
shown in Fig. 5(a) that the forward difference approxima-
e(tj ) − e(tk−1 ) ≥ ē (9) tion needs the next sampling interval hk+1 to calculate
ii) or if the maximal sampling period is achieved, that is the next integral part, whereas the current interval hk is
when applied in the Årzén’s mix up (11). Actually, the correct
(non-uniform sampling) forward equation should be
tj − tk ≥ hmax (10)
This second condition was added in order to ensure stabil- ui (tk+1 ) = ui (tk ) + Ki hk+1 e(tk ) (13)
ity by fulfilling the Nyquist-Shannon sampling condition. However, the next value hk+1 is not a priori known
Furthermore, as noticed in section 2, the different methods (since it varies). If one still wants to use the forward
of approximation are more accurate for a smaller sampling approximation, a solution may be to post-calculate the
period. In this sense, hmax allows to limit the increase of current integral part by shifting the time instants in
the sampling interval hk := tk − tk−1 . (13), that gives ui (tk ) = ui (tk−1 ) + Ki hk e(tk−1 ). But the
Remark 3.1. Conditions (9) and (10) are only verified at interest in reducing the control delay is no more reached
a periodic instant tj (since the event detector is time and a old value of the error is applied. Furthermore,
triggered with a constant sampling period). the forward approximation is not numerically stable (as
noticed in Remark 2.2).
Event-triggered PID controller:
Calculating the integral part with a more recent value
The second part calculates and updates the control signal of the error seems a better solution. For this reason, the
only when a request is received from the first part, that backward difference approximation is generally preferred.
is when an event was detected by the event detector. The Then, the integral part is calculated with the current
length of the (varying) sampling intervals hk is defined by sampling period hk and error e(tk ), as represented in
two successive events. The control signal is kept constant Fig. 5(b). From (7), the (non-uniform sampling) backward
during the interval hk . equation for the integral part is

For the event-based PID control law, the forward and ui (tk ) = ui (tk−1 ) + Ki hk e(tk ) (14)
backward difference approximation methods are applied
for the integral and derivative parts respectively in Årzén Remark 3.3. In order to simplify the understanding,
(1999), that gives Fig. 5(a) and Fig. 5(b) illustrate idealized behaviors of
40 Control Engineering and Applied Informatics

ysp
y
forward approximation
ē signal

hk+1
hk time
tk−1 tk
time
tk−1 tk tk+1
(a) Forward approximation. e = ysp − y
ē

ē signal time
backward approximation
hk − h̄
hk+1 h̄
hk
time
tk−1 tk tk+1 Fig. 6. Boundary of the integration increment between two
(b) Backward approximation. successive events after a large steady-state interval in
the non-uniform sampling scheme.
Fig. 5. Comparison between forward and backward ap- Actually, issues come from the calculation of the integral
proximations in the non-uniform sampling scheme. part (14) that involves the product hk e(tk ) — afterwards
called the integration increment Γk . Indeed, the integra-
the event-based scheme since they do not consider the tion increment is over-estimated in the current version of
synchronization of the events with the event detector sam- the control algorithm. However, in fact the time interval
pling instants tj . Indeed, the time of level crossing could be between two successive events after a large steady state
between two successive time-triggered sampling instants, can be divided into two parts, as drawn in Fig. 6, that are
that are tj − h̄ and tj , because the event detector is time i) the time interval where the signal is really in a steady
driven with the constant sampling period h̄ by construc- state and ii) the time interval required to detect a new
tion (see section 3.1). Nevertheless, the right behavior will level crossing.
be discussed and taken into account in the sequel.
i) The first part starts the last time a control signal was
computed, that is at the sampling instant tk−1 , and
4. ANALYSIS AND MODIFICATION OF THE finishes just before the reference changes. Because of
EVENT-BASED SCHEME the time-triggered event detector, this time is tk − h̄.
During this interval, i.e. hk − h̄, the error remains very
In order to simplify the original Årzén’s event-based setup, small, it is lower than the detection level ē else this is
the safety limit condition (10) — initially introduced not the steady state.
for stability reason — is removed here. However, several ii) Then, because the error becomes higher than the
modifications are mandatory with such a simplification. detection level ē, the second part starts. A request
is sent and a new control signal is finally calculated
4.1 Boundary of the integration increment at time tk . This instant does not necessarily occurs
exactly when the reference changes because of the
periodic behavior of the time-triggered event detec-
Fig. 6 depicts an example of the error e between the tor (see Fig. 1(b) and Remark 3.1), nonetheless the
measured signal y (i.e. output of the controlled system) interval to detect the level crossing is bounded by the
and a given reference to track ysp , in order to visualize sampling period h̄.
how it behaves during transient and steady-state intervals
(note that the error signal is zoomed but proportions are To summarize, the time interval between two successive
respected). Typically, many events should occur when the events can be divided into i) a first part where the
measured signal is converging to the reference (i.e. during sampling interval increases but the error remains small
transients) since the error is varying. Conversely, no event and ii) a second part where the error could become large
should be enforced once the measured signal vanishes but only during a short instant. Taking into account such
close to the reference (i.e. during steaty-state intervals). considerations makes the integration increment Γk :=
By removing the safety condition (10), the duration of a hk e(tk ) will not explode anymore, since both h(·) and e(·)
steady-state interval can increase as far as the error does compensate themselves each other. Finally, the integral
not cross the threshold level ē — because of (9) — whereas part (14) is hence bounded in practice
it was limited to hmax before. This allows to reduce even
more the control updates but new problems have to be ui (tk ) = ui (tk−1 ) + Ki Γk (15)
considered, in particular when the reference drastically
changes after a large steady-state interval. This can lead to with Γk ≤ (hk − h̄)ē + h̄e(tk )
overshoots whereas the system output is over-corrected in and this boundary has to be integrated in the event-based
case of huge sampling interval hk and/or huge error e(tk ). control algorithms without safety limit condition that are
Such a case is deeply analyzed here. then developed in section 4.5 based on that result.
Control Engineering and Applied Informatics 41

Furthermore, note that the inequality (15), which was ini- For the integral term, an exponential function is chosen
tially built for the steady-state intervals, is finally correct such that
for the whole running. Indeed, (15) becomes (14) during a
transient phase (where an event is enforced at each time- hiexp (hk ) = hk e−αi (hk −h̄) (17)
triggered sampling period) since hk = h̄ in this case.
where αi > 0 is a degree of freedom to increase/decrease
the exponential sampling interval (17). This forgetting
4.2 Sampling interval in the derivative term factor allows to have hiexp (hk ) close to the sampling
interval value hk when it is small, whereas it exponentially
The derivative part of the event-based PID controller can vanishes when hk increases.
be canceled when removing the safety limit condition (10),
because a huge sampling interval hk makes vanishing (12). The dynamics of (16) and (17) with respect to hk is
depicted in Fig. 7.
In practice, the derivative term calculates the slope of the
error over time in order to predict the system behavior.
0.4
As a consequence, if this rate of change vanishes, the
derivative action will not be applied anymore. For this 0.3

reason, a modification is also required. The two parts of 0.2

the time interval between two successive events after a 0.1

large steady state (see Fig. 6) are also analyzed here: 0
0 1 2 3 4 5 6 7 8
i) During the first part, the time derivative of the error
remains small because the error is lower than ē, the
reference does not change and the control is kept con- Fig. 7. Evolution of the exponential forgetting factor with
stant. The only reason to have an important derivative respect to the sampling interval, for h̄ = 0.05 s and
is in case of perturbation, but the perturbation is αi = αd = 1 s−1 .
assumed to be small enough (else it will enforce a new
event). For these reasons, the derivative term can be 4.4 Level-crossing mechanism
neglected in the first part of the steady state.
ii) During the second part, the time derivative of the Årzén suggested to calculate and update the control signal
error becomes really important because of the refer- using a relative measurement (9), that is when the error
ence change. Nevertheless, whereas hk is considered
changes enough from its last value, i.e. e(tk )−e(tk−1 ) ≥
in the existing approach (12), only the interval h̄ can ē (see section 3.1). However, such a setup creates troubles
be considered because, in practice, the rate of change because the system can reach a final (undesired) value
can only occur during the level detection interval. which is different of the reference. In this case, no event
To summarize, only the last sampling period h̄ is taken occurs anymore whereas the system is not stabilized. For
into account in the derivative part, which hence remains this reason, an absolute measurement is preferred here as
as defined in (7) for the backward case. in Durand and Marchand (2009). A new control signal is
thus calculated as soon as the current error crosses the
4.3 Forgetting factor of the sampling interval detection level, that is when

e(tj ) ≥ ē (18)
Another solution when removing the safety limit condi-
tion (10) consists in adding a forgetting factor of the Remark 4.1. With the proposed method, the number of
sampling interval so that, after a long steady-state interval, samples should inevitably increase during the transients
the value of hk is reduced enough to not impact the control but, at least, the error between the system output and
signal too much. This is true for both the integral as the reference will be lower than ē during the steady-
well as the derivative parts. The idea is to replace the state intervals. This was not the case before. In fact, this
linearly varying sampling interval hk in (14) and (12) by an modification was not required in the original Årzén’s setup
exponentially decreasing one. The approach proposed here because the system always reached the reference thanks to
is somehow similar to the anti-windup mechanism used in the safety limit condition (10).
control theory, where the error induced by the saturation Remark 4.2. Whereas the level-crossing detection mecha-
has to be compensated. nisms in (9) and (18) are different (relative vs. absolute),
For the derivative term, an exponential function is chosen the meaning of the constant ē is the same: it is a threshold
such that the impact of the sampling interval decreases as used to enforce the events. Therefore, the setup detailed
the elapsed steady-state time increases, that is before, in particular in Fig. 6, is still true.

4.5 Event-based PID control without safety limit condition

hdexp (hk ) = h̄ + (hk − h̄)e−αd (hk −h̄) (16)
where αd > 0 is a degree of freedom to increase/decrease Based on the suggested modifications above, different
the exponential sampling interval (16). Such a forgetting strategies are proposed: the first one where nothing else
factor allows to have hdexp (hk ) close to the sampling is done than removing the safety limit condition (10), and
interval value hk when it is small, whereas only the the other ones where the integral and/or derivative parts
last sampling period h̄ is considered when it is large (as of the control law are modified in order to reduce their
suggested in section 4.2). impact after a large steady-state interval.
42 Control Engineering and Applied Informatics

Algorithm 1 – Only without safety limit condition:

15

This case corresponds to the Årzén’s proposal where 0.4

the safety limit condition (10) is removed without doing 10 0.3

anything else. Important overshoots are expected because 5

0.2

0.1
2 2

of the principle described in section 4.1. 0 0.5

1
1.5
0 0.5
1
1.5

8 7 6 5 4 8 7 6 5 4
3 2 3 2

Remark 4.3. This first algorithm is presented in order to

1 1

show that the modifications which follow are required. (a) Algorithm 1. (b) Algorithm 2.
Algorithm 2 – Saturation of the integration increment:
0.7

This strategy consists in bounding the integration incre- 0.6

0.1

0.08

ment Γk after a large steady-state interval in order to

0.5

0.4 0.06

reduce the overshoots (as discussed in section 4.1). The 0.3

0.2
0.04

2 0.02 2

product between h(·) and e(·) is thus saturated according 0.1

0 0.5
1
1.5
0 0.5
1
1.5

to (15), which yields

8 7 6 5 4 8 7 6 5 4
3 2 1 3 2 1

(c) Algorithm 3. (d) Algorithm 4.

ui (tk ) = ui (tk−1 ) + Ki Γsat (tk ) (19)
with Γsat (tk ) = (hk − h̄)ē + h̄e(tk ) Fig. 8. Evolution of the integration increment for the
Only the last sample is taken into account for the deriva- different algorithms, with h̄ = 0.05 s, ē = 0.01 and
tive term, as defined for the classical approach in (7) (see αi = 1 s−1 .
section 4.2 for further details). On the other hand, the exponential forgetting factor of
Algorithm 3 – Exponential forgetting factor of the sampling the sampling interval is applied in the derivative term, as
interval: previously defined in (21).

A forgetting factor of the sampling interval is added in 4.6 Simulation results

such a way it does not impact the control signal too much
after a long steady state (as discussed in section 4.3). Consider a continuous-time second-order system in its
The exponentially decreasing sampling intervals hdexp (hk ) general form
in (16) and hiexp (hk ) in (17) are applied in the derivative κ
and integral parts respectively. This gives H(s) = (23)
2ζ 1
s + 2 s2 1+
ωn ωn
ui (tk ) = ui (tk−1 ) + Ki Γexp (tk ) (20)
where κ is the static gain, ζ is the damping ratio and ωn
with Γexp (tk ) = hiexp (hk )e(tk ) is the natural frequency. Consider also the PID controller
Tf in (2) and the desired closed-loop system as a first-order
ud (tk ) = ud (tk−1 ) system of the form
Tf + hdexp (hk )
Kd 1
+ e(tk ) − e(tk−1 ) (21) Hcl (s) = (24)
Tf + hdexp (hk ) 1 + τs
Algorithm 4 – Hybrid strategy: for a given time constant τ . Then, a naive solution (using
pole compensation) for the control parameters is
This is a mix of the previous algorithms. In fact, in the
2ζ 1 1
first case the integration increment increases with respect Kp = , Ki = , Kd = (25)
to h(·) and e(·), as drawn in Fig. 8(a). The second strategy κτ ωn κτ κτ ωn2
minimizes the impact of the integration increment but whereas the filter’s time constant Tf in (3) is chosen with
it still increases with respect to both h(·) and e(·), as respect to the value of the sampling period h̄.
shown in Fig. 8(b). Finally, the third algorithm adds an
Remark 4.4. The values of the control parameters (25)
exponential forgetting factor of the sampling interval in
are obtained by pole placement of the closed-loop system
such a way that the integration increment decreases but, as
considering the time-triggered PID controller. Then, the
a result, the product between h(·) and e(·) is higher when
event-based PID controllers are designed with these same
the sampling interval is small, as represented in Fig. 8(c).
values. Note that the aim here is to find the best control
The idea here is to have a small impact of the sampling
synthesis (otherwise pole compensation would clearly not
interval all the time. For this reason, it is proposed to apply
be chosen), but to compare the proposed event-based
the exponential forgetting factor into the algorithm with
approaches with the classical time-triggered one. In other
saturation, which leads
words, the event-based scheme tries to be as close as
possible of the time-triggered closed-loop shaping.
ui (tk ) = ui (tk−1 ) + Ki Γhybrid (tk ) (22)
Remark 4.5. The control parameters are computed in the
with Γhybrid (tk ) = hiexp (hk )
− h̄ ē + h̄e(tk ) continuous-time domain in (25). Another solution could
The evolution of the integration increment of this “hybrid” be to directly calculate them in discrete-time domain, this
algorithm is represented in Fig. 8(d). is detailed in Durand and Guerrero-Castellanos (2015).
Control Engineering and Applied Informatics 43

Consider the application case above, with: implement the bilinear approximation). However, whereas
only absolute measurement (18) enforces events in the
• The continuous-time second-order system (23), where
proposal case, events are enforced either with the relative
κ = 0.2, ζ = 0.8 and ωn = 0.5 s−1 ;
measurement (9) or when the safety limit condition (10)
• The desired closed-loop system as a first-order system
of the form (24), where τ = 1.98 s; is satisfied with Årzén. On the one hand, it can be noticed
• The sampling period h̄ = 0.1 s; periodic events because of such a safety limit condition. On
• The PID control parameters are as defined in (25) the other hand, the proposal only generates events during
(using pole compensation), that gives Kp = 8.08, transients and then no event during steady states.
Ki = 2.53 s−1 , Kd = 10.11 s and Tf ≈ 5h̄; Comparison with Arzen's PID (bilinear approximation)

• The exponential decay αi = αd = 10 s−1 and detec-

1.2

tion limit ē = 0.01 in the event-based controllers.

system output
0.8

0.6
The different event-based PID algorithms without safety 0.4
limit condition are compared in Fig. 9 for both backward 0.2
setpoint
Arzen

and bilinear approximation methods. Note the reference 0

algo 4

0 5 10 15 20 25
changes (from 0 to 1) at time t = 5 s. On the one hand,

updates
1
0.5
the frequency of updates is reduced in all event-based cases 0
0 5 10 15 20 25
compared to the classical time-triggered strategy. This is time [s]

depicted in the (extra) bottom plot, where ‘1’ means the

control law is calculated and updated during the sampling Fig. 10. Simulation results: comparison with Årzén’s event-
period h̄, ‘0’ means the control is kept constant. Further- based PID controller (bilinear approximation).
more, when the safety limit condition (10) is removed
(algorithm 1), an important overshoot occurs after a large 5. STABILITY ANALYSIS
steady-state interval as expected. Such an overshoot is
reduced when saturating the integration increment (al- Only a sketch of analysis is given here, because the
gorithm 2), applying an exponential forgetting factor of principle is trivial: if the system becomes unstable due
the sampling interval (algorithm 3) or mixing both latter to the fact that between two successive events it runs
strategies (algorithm 4). On the other hand, the bilinear in open loop, then the dynamic of the error activates
approximation method in Fig. 9(b) gives better results the event condition and, hence, the system is driven in
than the backward one in Fig. 9(a) in terms of frequency a stable regime. In other words, the event-based scheme
updates and tracking performance (but with a little bit consists in generating events when the system becomes
more complex control algorithm in return). unstable (that is when the error grows), while keeping the
Event-based PID (backward approximation)
control signal constant when the system is stable (when
2 the measure is closed to the reference value). The closed-
1.5
loop stability is based on the stability property of the
PID controller (assuming the control parameters lead to a
system output

1
setpoint
stable closed-loop system). Note that the controller runs in
algo 1
0.5 algo 2
algo 3
open loop between two successive events (e.g. tk−1 and tk ),
0
algo 4
but in fact it is a closed-loop system thanks to the event
0 5 10 15 20 25
detector which periodically monitors the system output
updates

1
0.5
0
(at each periodic sampling time tj ) and decides when to
0 5 10
time [s]
15 20 25
update the control signal. In the worst case (in term of
(a) Backward approximation. frequency of events), the event-based feedback works as
when applying the classical discrete-time controller: the
Event-based PID (bilinear approximation) events occur periodically, with h̄ as the sampling period
2
(the minimum sampling time by construction). Conversely,
1.5 when the interval between two successive events increases,
system output

1
stability is still ensured because the error remains lower
setpoint
algo 1
than ē by construction, thanks to the triggering condition
0.5 algo 2
algo 3
algo 4
(18), then allowing a practical stability.
0
0 5 10 15 20 25

6. EXPERIMENTAL RESULTS: APPLICATION TO A

updates

1
0.5
0
0 5 10 15 20 25
MINI QUADROTOR HELICOPTER
time [s]

(b) Bilinear approximation. Unmanned aerial vehicles (UAVs), and particularly the
mini quadrotor helicopters, give rise to great enthusiasm
Fig. 9. Simulation results: closed-loop system responses in research because of its high manoeuvrability, its pay-
with event-based PID controllers using the different load capacity and its ability to hover, see Castillo et al.
algorithms without safety limit condition. (2004). However, the quadricopter is nonlinear, unstable
and under-actuated with only four input forces (voltage
Finally, the hybrid algorithm 4 with bilinear approxima- of each rotor) for six output coordinates to control (roll
tion is compared with the Årzén’s algorithm in Fig. 10 (the φ, pitch θ and yaw ψ for the attitude; x, y and z for the
original Årzén’s algorithm was modified in order to also position). Fortunately, its model can be broken down into
44 Control Engineering and Applied Informatics

two subsystems: one defining the translation movement the quadcopter) and, consequently, x and y positions too.
(position) and the other one the rotation movement (atti- The aim of the control strategy is to stabilize the x and y
tude). Both are coupled in cascade since the translational positions of the quadrotor helicopter to a given reference.
subsystem depends on the rotational one, but the rota- The yaw angle ψ and altitude z are not addressed here
tional subsystem is independent of the translational one. and will be constant. In other words, the control objective
consists in calculating the angles φ and θ as well as the
The algorithms proposed in the present document will be
thrust T (all needed by the internal loop) that will drive
tested in practice for such a control architecture. Event-
the quadcopter to a given position x and y.
based control is quite new for UAVs systems, since the
system has to be actively actuated to remain stable. 6.2 System model for position control
Nevertheless, attitude control was addressed in Téllez-
Guzmán et al. (2012); Guerrero-Castellanos et al. (2013, Let consider the dynamical system model
2014) and position control is now addressed here.
mξ¨ = mg + RT b3 (26)
6.1 Experimental platform T T
where ξ = (x y z) , m is the mass, g = [0 0 g] with g the
The system is a 18 grams Blade Nano QX quadricopter gravitational constant, T is the thrust of the quadricopter.
(see Fig. 11). In order to release the platform from de- b3 is a unit vector of the body fixed frame and R is the
cision making related to guidance and navigation, posi- rotation matrix to express the orientation of the body fixed
tion and orientation are calculated by a Vicon motion frame in the inertial frame
capture system, where the movement of the vehicle is " #" #" #
cos ψ − sin ψ 0 cos θ 0 sin θ 1 0 0
processed through T40s high-resolution cameras (based R = sin ψ cos ψ 0 0 1 0 0 cos φ − sin φ
on the principle of inverse projection and triangulation). 0 0 1 − sin θ 0 cos θ 0 sin φ cos φ
Measurements are sent to the control unit through a UDP | {z }| {z }| {z }
Rz Ry Rx
frame every 2ms. Control algorithms are programmed in
Matlab/Simulink and implemented in the control unit side Developing these equations gives the model
in real time at 200Hz to a target computer using xPC
target toolbox. Finally, control variables are sent to the mẍ = T (cos ψ sin θ cos φ + sin ψ sin φ) (27)
quadricopter through a built-in bridge that converts UDP mÿ = T (sin ψ sin θ cos φ − cos ψ sin φ) (28)
frames to DSMX 2.4Ghz protocol. An overview of this
architecture is presented in Fig. 12. mz̈ = T cos θ cos φ − mg (29)
z 6.3 From nonlinear to linear control
ψ
x The tricky idea consists in dividing the controller into
y two cascaded parts, as suggested in Hably et al. (2006):
a first one that contains a linear control law (i.e. a PID
θ
controller here) and a second part used to compensate
φ
for the nonlinearities of the system. The second part is
calculated in such a way that the model obtained from
its input to the output of the system behaves like a simple
Fig. 11. Representation of the different control variables in and linear system (i.e. a double integrator here, as detailed
the mini quadrotor helicopter. below). Therefore, the PID controller acts on the (virtual)
double integrator system, while the whole controller (i.e.
Vicon unit
both PID and nonlinear parts) drives the real system.
Motion capture In practice, reformulating (29) gives

T cos θ cos φ = m(z̈ + g) (30)

Blade Nano QX
Vicon Tracker It is assumed that the quadrotor helicopter will not flip in
DSMX UDP frames order to avoid a situation where φ = ±π/2 or θ = ±π/2.
Control unit Multiplying (27) and (28) by sin ψ and cos ψ respectively,
and dividing the difference by (30) gives the relation
UDP frames for φ. Similarly, multiplying (27) and (28) by cos ψ and
sin ψ respectively, and dividing the sum by (30) gives
iRC: Ethernet to DSMX bridge
Matlab/Simulink xPC target
the relation for θ. The relation for T is easily obtained
from (30). This yields
Fig. 12. Experiment architecture.
ẍ sin ψ − ÿ cos ψ
The leveling control of the mini quadricopter (the control φ = tan−1 cos θ
z̈ + g
of pitch θ and roll φ) is done in a (non accessible) internal
ẍ cos ψ + ÿ sin ψ
loop. Only the control of x, y, z positions and yaw ψ angle θ = tan−1
is possible. It is assumed that all the control variables can z̈ + g
be independently controlled. It is also assumed that both z̈ + g
pitch and roll similarly behave (thanks to the symmetry of T =m
cos θ cos φ
Control Engineering and Applied Informatics 45

Then choosing The performance indexes obtained for the different exper-
iments (detailed below) are summarized in Table 1 for x
−1 rx sin ψ − ry cos ψ and y positions. Results are discussed in section 6.5.
φ = tan cos θ (31)
rz + g Table 1. Performance indexes for different ex-

rx cos ψ + ry sin ψ periments with several PID control strategies.
θ = tan−1 (32)
rz + g (a) Control of the position along x.
rz + g Nb (%) IAE IAU
T =m (33)
cos θ cos φ Classical PID
backward method 100 1.78 12.42
bilinear method 100 1.59 12.59
with rx , ry and rz as virtual control inputs from the algo 1 94.92 1.88 12.74
position errors on x, y and z axes respectively. Therefore, backward
algo 2 88.46 1.86 11.94
the (virtual) system to control can be written as three algo 3 92.50 1.71 11.29
Event-based PID algo 4 88.61 1.79 12.22
independent double integrators (ē = 10 mm) algo 1 82.27 1.71 10.90
algo 2 97.88 2.75 17.84
bilinear
ξ¨ = r (34) algo 3
algo 4
90.61
91.19
1.64
1.69
10.81
10.81
T algo 1 63.70 2.37 15.31
where r = (rx ry rz ) . Eventually, considering the atti- algo 2 67.32 2.81 17.40
tude control dynamics on the x and y accelerations as a backward
algo 3 66.20 2.59 17.54
first-order system leads to the transfer functions Event-based PID algo 4 67.97 2.74 17.90
(ē = 50 mm) algo 1 63.70 2.26 13.99
ωx bilinear
algo 2 63.89 2.79 18.01
x= rx (35) algo 3 59.09 2.36 15.07
s2 (s + ωx ) algo 4 60.01 2.44 14.61
ωy
y= 2 ry (36) (b) Control of the position along y.
s (s + ωy ) Nb (%) IAE IAU
1 backward method 100 1.55 10.29
z = 2 rz (37) Classical PID
bilinear method 100 1.43 8.79
s
algo 1 99.26 2.25 15.09
To summarize, the position control of the quadcopter algo 2 89.81 1.74 10.90
backward
consists in controlling three independent (filtered) double algo 3 69.31 1.39 8.91
integrators (35)-(37). This will be fulfilled by a PID Event-based PID algo 4 77.50 1.61 11.00
(ē = 10 mm) algo 1 78.73 1.69 10.62
controller. Then, nonlinear equations (31)-(33) — i.e. the algo 2 99.26 2.73 18.67
second part of the controller — allow to calculate the bilinear
algo 3 78.58 1.54 9.68
variables needed for the orientation control. algo 4 81.43 1.50 8.85
algo 1 61.13 2.23 13.53
Remark 6.1. The altitude z is actually a double integrator algo 2 64.66 2.73 16.84
backward
system which gain decreases with respect to the battery algo 3 63.70 2.72 17.85
load. This is no more detailed in the sequel since only the x Event-based PID algo 4 65.39 2.66 17.52
(ē = 50 mm) algo 1 58.43 2.22 13.02
and y positions are addressed, nevertheless the battery will
algo 2 64.01 2.68 16.81
be charged before each experiment to reduce its impact. bilinear
algo 3 55.90 2.32 14.17
Remark 6.2. Since it is assumed that the drone similarly algo 4 55.01 2.04 12.36
behaves in x and y axes, then the control parameters will
be tuned only once and applied for both axes in the sequel.
6.5 Experimental results
6.4 Performance indexes The mini helicopter has to track three points in the space,
which coordinates are (x, y, z) = (0, 0, 1), (0, 1, 1), and
Performance indexes, introduced in Sánchez et al. (2009b), (1, 1, 1) respectively. The altitude is kept the same (i.e. 1 m
are recalled here. They allow to compare the different high) for the three points because the study only focuses
event-based proposals with respect to classical approaches: on x and y position control (z is not depicted in the plots).
• The number (Nb) of samples required to perform the Typically, the quadricopter has to draw a triangle in the
test bench (normalized and expressed in percentage); xy-plane. The system goes to one point to the following
• The IAE index, which gives information on the refer- point after a waiting time of 10 s in order to analyze both
ence tracking (and so on the performance): the reference tracking and the stabilization. This gives
different steps at different instants, which are filtered in
N
X order to avoid abrupt changes in the position references.
IAE = e(tj )dt
j=0 The system trajectory in the xy-plane are compared for
several strategies, for both backward and bilinear approx-
• By analogy, the IAU index gives information on the imation methods:
control effort:
N • Classical (time-triggered) PID controllers;
X • Event-based PID controllers without safety limit con-
IAU = u(tj )dt
dition, where the control signal is calculated and
j=0
updated only when the system output crosses a given
where N is the final simulation time. level ē:
46 Control Engineering and Applied Informatics

◦ With saturation of the integration increment (al-

xref
gorithm 2); 1 x

position [m]
◦ With exponential forgetting factor of the sam- 0.5
yref
y
pling interval (algorithm 3);
◦ With hybrid strategy (algorithm 4). 0

0 5 10 15 20 25

angle [deg]
20
roll
The control parameters (obtained by trial and error) are 0 pitch

Kp = 3, Ki = 0.1 s−1 and Kd = 2 s. The sampling −20

4
0 5 10 15 20 25

period is h̄ = 0.01 s, the derivative filter’s time constant 2

control
control x
0 control y

is Tf ≈ 2h̄. The event-based control parameters are αi = −2

−4
0 5 10 15 20 25

αd = 1 s−1 and ē = 10 or 50 mm. Note that the battery time [s]

quadrotor is charged before each new experiment. (a) Backward algorithm.

Experimental results are represented in Fig. 13 for the xref

time-triggered PID control when using either the back- 1 x

position [m]
y
ward or the bilinear approximation methods, for a 26 s 0.5 y
ref

interval time interval. The top plot shows the reference 0

and the measured positions. The bottom plots show the 0 5 10 15 20 25

angle [deg]
angles (roll φ and pitch θ) and control signals. One can 20
0
roll
pitch
see in Fig. 13(b) that the bilinear method gives better −20
0 5 10 15 20 25
performance than the backward one in Fig. 13(a), this is 4
2

control
control x
0
also verified when comparing the IAE and IAU indexes for −2
−4
control y

both methods in Table 1. 0 5 10

time [s]
15 20 25

Experimental results for the event-based control are repre- (b) Bilinear algorithm.
sented in Fig. 14 (only algorithm 4 with bilinear method is
depicted but all of them give more or less similar results). Fig. 13. Experimental results of the time-triggered PID
Extra bottom plots show the sampling instants in these controller (backward and bilinear approximation
event-based control schemes (‘1’ means the control law is methods).
calculated and updated during the sampling period h̄, ‘0’
means the control is kept constant). In the different cases xref
(even those not represented), the measured positions in 1 x
position [m]

yref
Fig. 14(a) and performance indexes in Table 1 are close 0.5 y

to the desired ones when the detection level is small, i.e. 0

ē = 10 mm, with a reduced frequency of updates (30% less 0 5 10 15 20 25
angle [deg]

20
of updates in the better case). When the detection level 0
roll
pitch
increases, i.e. ē = 50 mm in Fig. 14(b), the frequency of −20
0 5 10 15 20 25
updates decreases even more (until 45% less of updates) 4
2
control

control x
0 control y
but with weaker performance in return (because of the −2
−4
unstable behavior in open loop that occurs between two 0 5 10 15 20 25
updates updates

1
0.5 updates x
successive events, see section 5), so the IAE and IAU 0
0 5 10 15 20 25
1
indexes increase. 0.5
0
updates y

0 5 10 15 20 25
To summarize, the tradeoff between performance and fre- time [s]

quency of control updates is clearly highlighted. One can (a) ē = 10 mm.

hence imagine how the computing resources utilization can
be reduced in embedded systems (as well as communica- xref
1 x
tions in the case of a deported controller as in the present
position [m]

yref

application). 0.5 y

Remark 6.3. Performance can be increased with other 0

0 5 10 15 20 25
values of the control parameters, αi and αd for instance,
angle [deg]

20
roll
but this is not the aim of the present work. 0
−20
pitch

0 5 10 15 20 25
4
2
control

control x
7. CONCLUSIONS AND FUTURE WORKS 0
−2
control y

−4
0 5 10 15 20 25
updates updates

1
This paper recalled the classical time-triggered PID con- 0.5
0
updates x

trol scheme using different methods of approximation. The 1

0.5
0 5 10 15 20 25
updates y
event-based paradigm was then introduced and event- 0
0 5 10 15 20 25

based PID control algorithms were developed. The pro- time [s]

posed approaches are built without safety limit condition, (b) ē = 50 mm.
contrary to the original setup in Årzén (1999), and both
integral and derivative terms were considered. The dif- Fig. 14. Experimental results of the event-based PID
ferent approaches were tested in simulation and then on controller (bilinear method, algorithm 4 for different
a real-time system: a mini quadrotor helicopter using a detection levels ē).
motion capture system to provide its position, where the
Control Engineering and Applied Informatics 47

controller is deported and communications are hence of Eqtami, A., Dimarogonas, D.V., and Kyriakopoulos, K.J.
high importance. Experimental results showed the effec- (2010). Event-triggered control for discrete-time sys-
tiveness of the proposals with a reduction of the computing tems. In Proceedings of the IEEE American Control
and communication resources utilization. The advantage Conference.
of an event-driven scheme was hence highlighted and the Franklin, G., Powell, J., and Workman, M. (1997). Digital
encouraging results strongly motivate to continue develop- Control of Dynamic Systems (3rd Edition). Addison-
ing event-based control strategies. Wesley.
Guerrero-Castellanos, J.F., Téllez-Guzmán, J.J., Durand,
Next step is to consider delays, as in Durand (2013).
Nonlinear strategies will also be a trajectory for future S., Marchand, N., and Álvarez Muñoz, J. (2013). Event-
works, with event-based control laws in the spirit of Marc- triggered nonlinear control for attitude stabilization of
hand et al. (2013); Durand et al. (2014). Eventually, a a quadrotor. In Proceedings of the IEEE International
cooperative approach, where several systems are controlled Conference on Unmanned Aircraft Systems.
together through a (wireless) network, will be the next Guerrero-Castellanos, J.F., Téllez-Guzmán, J.J., Durand,
application to highlight the interest of event-based tech- S., Marchand, N., Álvarez Muñoz, J., and González-
niques in reducing computation/communication resources. Dı́az, V. (2014). Attitude stabilization of a quadrotor
by means of event-triggered nonlinear control. Journal
ACKNOWLEDGMENT of Intelligent and Robotic Systems, 73, 123–135.
Hably, A., Kendoul, F., Marchand, N., and Castillo, P.
The authors would like to thank J. Dumon and J. Minet for (2006). Further results on global stabilization of the
their technical support as well as T. Da Silva, M. Fayet, D. PVTOL aircraft. In Proceedings of the Second Multidis-
Manica, P. Perraud for their contribution in the context of ciplinary International Symposium on Positive Systems
student projects (Ense3, Grenoble university). They also : Theory and Applications, Lecture Notes in Control and
would like to thank the anonymous reviewers for their Information Sciences LNCIS 341, p. 303–310. Springer.
valuable comments and suggestions to improve the quality Heemels, W., Sandee, J., and van den Bosch, P. (2009).
of the paper. Analysis of event-driven controllers for linear systems.
International journal of control, 81, 571–590.
This work has been partially supported by the LabEx Lunze, J. and Lehmann, D. (2010). A state-feedback
PERSYVAL-Lab (ANR–11-LABX-0025). The experimen- approach to event-based control. Automatica, 46, 211–
tal platform was designed by the GISPA-lab technical 215.
staff and partially funded by EquipEx Robotex (ANR– Marchand, N., Durand, S., and Guerrero-Castellanos, J.F.
10-EQPX-4401). (2013). A general formula for event-based stabilization
of nonlinear systems. IEEE Transactions on Automatic
REFERENCES Control, 58(5), 1332–1337.
Mounier, H., Cela, A., Niculescu, S.I., and Wang, J.
Årzén, K.E. (1999). A simple event-based PID controller.
(2011). Event driven intelligent PID controllers with
In Preprints of the 14th World Congress of IFAC.
applications to motion control. In Proceedings of the
Åström, K.J. and Bernhardsson, B. (2002). Comparison of
18th world congress of IFAC.
Riemann and Lebesque sampling for first order stochas-
Ruiz, A., Jimnez, J., Sanchez, J., and S., D. (2014).
tic systems. In Proceedings of the 41st IEEE Conference
A practical tuning methodology for event-based PI
on Decision and Control.
control. Journal of Process Control, 24, 278–295.
Åström, K.J. and Murray, R.M. (2008). Feedback Systems:
Sánchez, J., Guarnes, M., and Dormido, S. (2009a). On the
An Introduction for Scientists and Engineers. Princeton
application of different event-based sampling strategies
University Press.
to the control of a simple industrial process. Sensors, 9,
Beschi, M., Dormido, S., Sanchez, J., and Visioli, A.
6795–6818.
(2012). Characterization of symmetric send-on-delta PI
Sánchez, J., Guarnes, M., Dormido, S., and Visioli, A.
controllers. Journal of Process Control, 22, 1930–1945.
(2009b). Comparative study of event-based control
Castillo, P., Dzul, A., and Lozano, R. (2004). Real-time
strategies: An experimental approach on a simple tank.
stabilization and tracking of a four-rotor mini rotorcraft.
In Proceedings of the European Control Conference.
IEEE Transactions on Control Systems Technology, 12,
Sandee, J.H., Heemels, W.P.M.H., and van den Bosch,
p. 510–516.
P.P.J. (2005). Event-driven control as an opportunity
Durand, S. (2013). Event-based stabilization of linear sys-
in the multidisciplinary development of embedded con-
tem with communication delays in the measurements.
trollers. In Proceedings of the American Control Con-
In Proceedings of the American Control Conference.
ference, 1776–1781.
Durand, S. and Guerrero-Castellanos, J.F. (2015). Event-
Tabuada, P. (2007). Event-triggered real-time scheduling
based digital PID control. In Proceedings of the 1st
of stabilizing control tasks. IEEE Transactions on
IEEE International Conference on Event-Based Con-
Automatic Control, 52, 1680–1685.
trol, Communication, and Signal Processing.
Téllez-Guzmán, J.J., Guerrero-Castellanos, J.F., Durand,
Durand, S. and Marchand, N. (2009). Further results
S., and Marchand, N. (2012). Event-based LQR control
on event-based PID controller. In Proceedings of the
for attitude stabilization of a quadrotor. In Proceedings
European Control Conference.
of the 15th IFAC Latinamerican Control Conference.
Durand, S., Marchand, N., and Guerrero Castellanos,
Velasco, M., Martı́, P., and Bini, E. (2009). On Lyapunov
J.F. (2014). Event-based stabilization of nonlinear
sampling for event-driven controllers. In Proceedings of
time-delay systems. In Proceedings of the 19th World
the 48th IEEE Conference on Decision and Control.
Congress of IFAC.