On event-based PI control of first-order processes

Ubaldo Tiberi, José Araújo, Karl Henrik Johansson

ACCESS Linnaeus Center, KTH Royal Institute of Technology, Stockholm,

Sweden (e-mail:{ubaldot,araujo,kallej}@ee.kth.se)

Abstract: In this paper the design of an event-based proportional-integral (PI) control scheme for
stable first-order processes is considered. A novel triggering mechanism which decides the transmission
instants based on an estimate of the PI control signal is proposed. This mechanism addresses some side-
effects that have been discovered in previous event-triggered PI proposals, which trigger on the process
output. In the proposed scheme, the classic PI controller is further replaced with PIDPLUS, a promising
version of PI controller for networked control systems. Although PIDPLUS has been introduced to deal
with packet losses and time delays, and, to the best of our knowledge, a stability analysis of the closed-
loop system where such a controller is used has never been performed, here the performance of such
a controller in an event-based fashion are analyzed, and a stability analysis is further provided. The
proposed event-based scheme ensures set-point tracking and disturbance rejection as in classic time-
periodic implementations of PI controller, while greatly reducing the number of sensor transmissions.
The theoretical results are validated by simulations, where the benefits in using PIDPLUS in combination
with the proposed PI event-based triggering rule are shown.

Keywords: PI controller, Event-Based Control, Networked Control Systems (NCS), Adaptive Control.

1. INTRODUCTION eventual implementations in real industrial systems, because

the utilization of such a control paradigm may lead to several
The proportional-integral-derivative (PID) controller has been problems as we discuss next.
applied to solve many control problems. Even though many The implementation of improper sampling techniques and con-
controller choices are currently available, PID controllers are troller structures may give rise to large oscillations of the pro-
still by far the most widely used form of feedback control.
In process industry it is know that more than 90% of the cess output, as observed by Årzen [1999], Cervin and Åström
[2007], Vasyutynskyy and Kabitzsch [2007], Durand and Marc-
control loops are regulated by PID controllers, Åström and hand [2009]. A common conclusion of previous studies has
Hägglund [2006]. Most of such controllers are Proportional- been that a large variation of integral component of the con-
Integral (PI), since the derivative part is usually not used in troller, due to long time-intervals between control updates,
practice, Åström and Hägglund [2006]. In traditional control appear to be the cause of large oscillations. Another poten-
schemes, the implementation of PI controllers has always been tial issue arising from the implementation of event-based PI
performed by assuming that sensing, computation and actuation controllers is that a sticking effect may occur. This effect is
are performed periodically. However, with the introduction of characterized by the absence of new events, even when the plant
networked control systems (NCSs), classic design techniques output is far from the desired set point, incurring in a non-zero
may no longer used. This originates from the fact that the steady-state error and no more controller updates.
network may introduce large communication delays and loss
of information, which greatly influences the controller perfor- Previous works have proposed methods to enhance event-based
mance (see Eriksson [2008] for an overview of design methods PI controllers performance. In Årzen [1999], Vasyutynskyy
of PID controllers for NCSs). Additionally, when the network is and Kabitzsch [2007], Rabi and Johansson [2008] and Durand
wireless, the control system designer should take into account and Marchand [2009] several sampling methods and controller
bandwidth usage and energy consumption in the control loop adaptations were proposed to improve the transient perfor-
design, Willig [2008]. Hence, new controller structures and PI mance of event-based PID controllers. Sánchez et al. [2011]
tunings methods are required. and Lehmann [2011] proposed event-based PI controllers that
rely on the knowledge of the plant model for sampling and con-
To cope with these problems, event-based techniques for trol. However, in an industrial perspective, the derivation of an
control were recently introduced, Åström and Bernhardsson accurate mathematical model of the process may be expensive,
[1999], Tabuada [2007]. Such techniques allow an efficient uti- thus such requirements may not be met in real implementations.
lization of the network resources, while ensuring a desired be- Moreover, in all the aforementioned papers, the steady-state
havior of the closed-loop system. This is achieved by exchang- analysis has not been addressed. However, when an event-based
ing information between sensor, controllers and actuators only control scheme is used, the design presents two degrees of
when relevant information is available. The use of these tech- freedom: one is represented by the choice of the sampling rule,
niques has attracted much attention from the area NCSs, Wang and the other is represented by the choice of the controller.
and Lemmon [2011], and applied for PI or PID control Vasyu-
tynskyy and Kabitzsch [2007], Rabi and Johansson [2008], In this paper, a novel event-based control scheme for stable
Durand and Marchand [2009], Sánchez et al. [2011], Lehmann first-order processes controlled by PI controllers is presented.
[2011]. However, no definite solution was yet achieved for The proposed scheme aim at canceling both the oscillations
around the set-point and the sticking effect by jointly con-
1 The work of the authors was supported by the by the VINNOVA project sidering an appropriate event-based rule and an adaptive PI
WiComPI, the Knut and Alice Wallenberg Foundation and the FeedNetBack controller. More precisely, we consider PIDPLUS, Song et al.
and Hycon2 EU projects.
Sensor r where δ > 0 and τ : R+ → R+ is a function that is reset to
zero at any sampling-time, i.e. τ (t+
k ) = 0, where tk is the event
y(t) Event instant defined as
Actuator Plant Detector τ (t) tk = min{t : t ≥ tk−1 |τ (t) > δ},
u(tk ) y(tk ) and where h k := tk+1 − tk is the inter-event time. The
function τ (t) generally depends on the accessible variables of
the control system, like the output or the input signal, and the
PI Controller (β(t)) joint selection of τ (t) and δ encodes the desired behavior of the
closed-loop system.
r Given this setup, the problem we address in this paper is
formalized as follows:
Fig. 1. The proposed event-based PI control system architecture
Problem 2.1. Given the process (1) controlled by a PI con-
[2006] in combination with a PI-based triggering rule that troller of the form (2), determine:
schedule the measurement transmissions from the sensor to (1) an aperiodic event-based sampling rule τ (t) ≤ δ;
the controller. Although PIDPLUS was introduced to deal with (2) a dynamic integral update rule β(tk );
packet losses and time delays, and its stability properties and
performances are tested mainly by simulation or with experi- such that the closed-loop system exhibits zero steady-state error
mental test-beds, Ungan [2010], Kaltiokallio et al. [2010], here for constant reference signal and eventual constant external
we adopt such a controller in an event-based fashion, further disturbances, while reducing the number of transmissions from
providing stability conditions in a rigorous way. the sensor node to the controller as much as possible. ⊳
The remainder of the paper is organized as follows. In the
next section the control system architecture we consider is
presented. In Section 3 potential issues in using event-based 3. ISSUES WITH A NAIVE EVENT-BASED PI
techniques in PI controlled systems are discussed. In Section 4 CONTROLLER
we introduce the proposed event-triggered PI controller, and in
Section 5 we investigate its stability properties. In Section 6 When PI controllers are used in event-based fashion, we may
the theoretical results are validated by simulation, and finally, a have the sticking effect or we may experience unacceptable
discussion in Section 7 concludes the paper. oscillations of the output around the set-point. To illustrate
these drawbacks we give an example. Consider the process (1)
controlled by (2), with a = −0.7, b = 1, Kp = 0.23, Ti = 3,
2. EVENT-BASED PI CONTROL SYSTEM and consider the triggering implicitly defined by
τ (t) = |y(t) − y(tk )| ≤ δ , (3)
The system architecture for the event-based PI control we where δ = 0.03. We compare this event-based implementation
consider is depicted in Fig. 1. It comprises a plant, a sensor, a PI with a periodic implementation of period h = 0.3 s.
controller and an actuator. The sensor continuously measures
the plant output, and it has an event-detector implemented 3.1 Sticking effect
on-board. The event-detector decides when the plant output
should be transmitted to the controller. The controller generates Figure 2(a) represents the step response and the inter-event
the input signal based on the received measurement from the times. After t = 25 s, the controller is no longer updated, de-
sensor, which is then sent to the actuator and it is applied to the spite there is a steady-state error of ≃ 0.15, and the system gets
plant. We assume that a new control signal is computed and the stuck. The benefit of achieving zero steady state error for con-
actuator is updated at the same time and correspondently to the stant references offered by continuous or periodic discrete-time
reception of a new output measurement. PI controllers are clearly lost. This is due because the process
The process is a first-order system of the form: is stable and the constant control inputs applied between sensor
transmissions are not strong enough to fulfill condition (3).
ẋp (t) = axp (t) + bu(t) ,
(1) In the sequel we refer the problem of having steady-state error
y(t) = xp (t) ,
in addition to do not performing any control updates as sticking.
where xp (t) ∈ R is the process state, u(t) ∈ R is the As the reader may argue, a trivial method to avoid sticking is
control signal and y(t) ∈ R is the process output. We assume to add a time out to the sampling rule, so that the sensor is
asymptotic stability of the process, i.e., a < 0. The controller is enforced to send a new measurement to the controller when-
a PI controller, for which we consider its digital implementation ever the closed-loop system gets stuck. Nonetheless, several
given by problems arise also by adopting this simple trick, as we discuss
xc (tk+1 ) = xc (tk ) + β(tk )(r − y(tk )) , next.
(2)

1 3.2 Oscillatory behavior
u(tk ) = Kp (r − y(tk )) + xc (tk ) ,
Ti
where xc ∈ R is the integrator state, Kp ∈ R is the proportional Consider again the previous example, and, in addition to the
event rule defined by (3), consider a time-out that enforce the
gain, Ti ∈ R is the integration time, β(tk ) represents the sensor to send a packet if no events are detected for hmax
integrator update rate, r ∈ R is the reference signal assumed units of time. Such time-out is added to the sampling rule (3)
to be constant and tk is the time instant in which a new to avoid sticking. In the example we set hmax = 35 s. The
control input is computed. In traditional implementations, the output response is depicted in Figure 2(b). By considering the
parameter β(tk ) is constant and equal to the sampling interval. event-based rule (3) plus the time-out, we obtain a large output
In event-based schemes, it is common to define an event as the oscillations around the set-point.
violation of a triggering rule having the following form Such oscillations are due to the large value of hmax . Because the
τ (t) ≤ δ , integrator update rate β(tk ) is equal to the inter-event times hk ,
2 4.1 PI-based triggering rule
Event−based
1.5 Periodic PI (h=0.3 s)
Samples Inspired by the deadband triggering-rule based on the output
y(t)

0.5
signal Otanez et al. [2002], our intuition is to consider the
deadband triggering rule on an appropriate filtered version of
the output signal. For instance, by considering the control input
0
0 20 40 60 80 100 120 140 160 180 200
Time [s]
1.5
as a filtered version of the output, our idea is to consider a
Event−based deadband sampling on a input-like signal. By denoting
Periodic PI (h=0.3 s)
1
1 t
Z
u(t)

0.5
ũ(t) = K̃p (r − y(t)) + (r − y(s)) ds + x̃c (tk ) ,
T̃i tk
0
0 20 40 60 80 100 120 140 160 180 200
(4)
Time [s] where K̃p , T̃i ∈ R are two sampling parameters, and x̃c ∈ R is
100 the state of the integrator implemented on the sensor, the event-
based rule we propose is implicitly defined by the condition
tk+1 −tk [s]

50


τ (t) = ũ(tk ) − K̃p (r − y(t))
0
0 20 40 60 80 100 120 140 160 180 200 Z t (5)
Time [s] 1
− (r − y(s)) ds − x̃c (tk ) ≤ δ ,
(a) Sticking effect T̃i tk
2 We denote (5) as PI-based triggering rule. Whenever the sys-
1.5
Event−based
Periodic PI (T=0.3 s)
tem gets stuck, the integral term in (5) grows unbounded, en-
Samples forcing the sensor to send a new measurement to the controller,
y(t)

1
and the sticking is avoided. Moreover, the controller is no
0.5
longer updated if and only if yk∗ = r, that yields ũ(tk+1 ) =
0
0 20 40 60 80 100 120 140 160 180 200 ũ(tk ) for all k greater than k ∗ . That way, when the system gets
Time [s]
stuck, the sampling rule imposes a time-out that depends on the
1.5
Event−based
distance of yk from the desired set-point r.
Periodic PI (T=0.3 s)
1
Notice that by using such a triggering rule, the sensor can
u(t)

0.5 potentially compute the new control input ũ(tk+1 ) and send
this information straight to the actuator. However, whenever the
0
0 20 40 60 80 100 120 140 160 180 200 utilization of (5) cancels the sticking problem, the controller
Time [s]
would be updated with ũ(tk+1 ) = ũ(tk ) ± δ. This control
100
update rule leads to limit cycles that may generate unacceptable
oscillations of the output. To avoid such oscillations, we let the
tk+1 −tk [s]

50 sensor to verify when (5) is violated, and we let it to transmit

to the controller the value of yp (tk ) instead of the value of
0
0 20 40 60 80 100 120 140 160 180 200
ũ(tk+1 ). The controller updates the input signal according to
Time [s] the received measurement yp (tk ) and to the elapsed inter-event
(b) Oscillations around the set-point times.
Remark 4.1. In general, it is possible to use completely differ-
Fig. 2. Issues with a naive event-based PI controller. ent tuning of the parameters K̃p and T̃i at the sensor and Kp and
Ti at the controller. This fact should not provide any concerns,
because constraining the PI sampler and the PI controller pa-
if hmax is too large so it is β(tk ). Then, the control input input rameters to have the same tuning, then the sensor manufactures
applied at time tk+1 = tk + hmax may be too aggressive and would be constrained to produce ad-hoc sensors depending on
it potentially triggers an oscillatory behavior of the output. On which controller is used by a certain customer. ⊳
the other hand, by choosing small values for hmax to reduce the
strong variation of the integral state, we may lose the benefits 4.2 Integrator update rate adaptation
of using an event-based scheme since more transmissions are
required.
The integration update rate β(tk ) that we use is the same as
The design of an event-based PI controller presents then two PIDPLUS Song et al. [2006]. After some calculation, we get
degrees of freedom: one degree of freedom is represented by that the PIDPLUS can be rewritten in the form (2), where
the selection of an appropriate design of the integrator update hk−1

rate β(tk ), and the other concerns the choice of a suitable event- β(tk ) = −Ti (1 − e Ti ) . (6)
based rule τ (t) ≤ δ. In the sequel we consider the formulation of the PIDPLUS
as (2), with β(tk ) defined as in (6), and we will show that
to guarantee asymptotic stability of the closed-loop system a
condition on the controller’s proportional gain Kp must be
4. PROPOSED EVENT-BASED PI CONTROL SCHEME fulfilled.

To solve the sticking problem, we need to enforce the sensor 5. STABILITY ANALYSIS
to send a new measurement to the controller whenever the
system gets stuck, and we have to adapt the controller to avoid In this section we study the stability property of the controlled
oscillations around the set-point. These are the arguments of the process when β(tk ) is chosen according to (6), and τ (t) ≤ δ
following two sections. is chosen according to (5). Before addressing the general case
of aperiodic controller updates, we first show how the adapta-
tion (6) ensures asymptotic stability of the controlled-process !x(t)!
for any fixed controller update rate of period h > 0.

Lemma 5.1. Consider the system (1) controller by (2), where α(!xk−1 !)
β(tk ) is given by (6). Then, the controlled system is asymptot-
ically stable for any constant sampling interval h > 0 if, and α(!xk !)
only if 0 < −bKp /a < 1. ⊳ α(!xk+1 !)
Proof: We start the proof by assuming r = 0. To capture α(!xk+2 !)
all the model details of the closed loop system, let us denote α(!xk+3 !)
x = [xp xc ]T , and consider the hybrid model, Goebel et al. t
[2009] tk tk+1 tk+2 tk+3 tk+4
ẋ(t) = Ac x(t) + w , if x ∈ C ,
(7)
x(tk+1 ) = Ad (β(tk ))x(tk ) , if x ∈ D , Fig. 3. Stability analysis of the closed-loop system. The value
where of kx(t)k is bounded by a decreasing bound α(kxk k) in
any time interval [tk , tk+1 ).
a b −bKp xp (tk )
Ac = 0 0 , w = 0 (8)
Condition 2 is always verified, while conditions 1–3 are verified
1 0 for all h > 0 if, and only if 0 < −bKp /a < 1. Then, by
!
Ad (β(tk )) = Kp , (9) assumption, we have that Φ is Schur for any constant time
− β(tk ) 1 h > 0. The Schur property of Φ implies the convergence to
Ti
zero of the sequence x(tk ), and since kx(t)k ≤ α(kx(tk )k) for
and where the flow and the jump sets are defined respectively t ∈ (tk , tk+1 ), then asymptotic stability of the origin follows.
as C = {x ∈ R2 : τ (t) ≤ δ} and D = {x ∈ R2 : τ (t) > δ},
where τ (t) and δ are defined in (5). Given the system (7), we In the case r 6= 0, we have that the equilibrium point of the
want to study the stability property of the origin. Under the controlled system is in xeq = [r −aTi /(Kp b)·r]T . Hence, the
assumption a < 0, and because the input w is constant and set-point tracking can be studied by considering a coordinate
it acts only on the asymptotically stable part of (7), i.e. it acts change that translates xeq into the origin, and then by studying
only on the dynamics of xp , then kx(t)k ≤ α(kx(tk )k) for the stability of the origin in the new coordinates as done in the
all t ∈ (tk , tk+1 ), and for a certain K-class function α(·) 2 , case r = 0. When there are external disturbances, it is possible
see Fig. 3. To achieve asymptotic stability of the origin, it is to proceed in the same way by considering xeq = [r −
then enough to show the convergence to zero of the sequence aTi /(Kp b) · r − d]T as equilibrium point. ⋄
x(tk ). Because it holds xp (t+ k+1 ) = xp (tk ) and ẋc (t) = 0, It is well known that while a common method to test instability
we can study the behavior of x(tk ) only correspondently to is to verify the position of the eigenvalues of the discretized
the transmissions times t = tk . By observing the particular controlled process outside the unit circle, to prove stability
structure of (7), it is enough to study the stability properties of the continuous-time process, this method should be used
of the following system with caution because no information is given on what occurs
x(tk+1 ) = Φx(tk ) , (10) between consecutive discretization instants. It can happen for
where the matrix example that at every discretization instant it holds xp (tk ) = 0

b ah b ah
 but the output is oscillating between the discretization points.
ah However, by resorting to hybrid models, we proved that be-
 e − Kp a (e − 1) a (e − 1)  tween two consecutive sampling instants, the continuous-time
Φ= Kp , (11)
− β(tk ) 1 dynamics are also upper bounded with a class K function α(·),
Ti and such a bound converges to zero, see Fig. 3.
is obtained by considering the exact discretization of the con- Remark 5.1. The reader may argue that since PIDPLUS en-
tinuous time process (1) controlled by (2) and by considering sures asymptotic stability for any constant sampling period, a
constant time intervals of the form [tk , tk+1 ). Note that under natural method to reduce the amount of communication be-
constant sampling, the matrix Φ is time invariant. The polyno- tween the sensor and the controller would be to use a large
mial characteristic of Φ is given by constant sampling period. However, if a disturbance suddenly
p(λ) = λ2 − tr(Φ)λ + det(Φ) , enters the system, the performance may drastically deteriorate
since it will be detected only at the next sampling instant that
where tr(Φ) and det(Φ) denote the trace and the determinant of may be far, while an event-based control scheme would react
the matrix Φ respectively. By applying the Jury criterion we get immediately. ⊳
the following necessary and sufficient conditions for asymptotic
stability In the case of event-based control, the stability analysis is more
1. | det(Φ)| < 1 , involved, since the inter-event times are varying, namely the
2. 1 − tr(Φ) + det(Φ) > 0 , sampling intervals hk are not constant. This implies that the
3. 1 + tr(Φ) + det(Φ) > 0 . matrices Φ are time-varying, and the controlled system can be
rewritten as
By using PIDPLUS, the above conditions become
x(tk+1 ) = Φ(hk , hk−1 )x(tk ) , (12)
h
1. |eah − ab Kp e Ti (eah − 1)| < 1 , where
 
2. b ah
h
b ahk b ahk
a Kp (e − 1)(1 − e Ti ) > 0 , ahk
e − K p (e − 1) (e − 1)
h
Φ(hk , hk−1 ) =  a hk−1 a .
 
3. 2(1 + e ) − ab Kp (eah − 1)(1 − e Ti ) > 0 ,
ah
−Kp 1 − e Ti 1
2 A continuous function α : [0, a) → R
≥0 , a > 0 , is said to be of class K if
it is strictly increasing and α(0) = 0, Khalil [2002]. (13)
However, by using PIDPLUS, it is still possible to verify the The results are depicted in Figure 4. By comparing Figure 4
stability condition under any event-based rule, as stated in the with Figure 2, we can appreciate how there are no oscillations
next result. of the output, and how the system does not stick. Within this
simulation we obtained N = 28 number of transmissions
Theorem 5.1. Consider the system (12), and assume 0 < and IAE=13.45 for the event-based implementation, while the
−bKp /a < 1. Let 0 < hmin < hmax two arbitrary positive periodically sampled PI controller generated N = 1500 trans-
constants. If there exists a matrix P = P T > 0 that sastisfies missions and IAE=8.97. In order to generate the same number
of transmissions as the event-based, the periodically sampled PI
Φ(hmin , hmin )P ΦT (hmin , hmin ) − P < 0, (14) controller requires a transmission period h = 16 s which would
T
Φ(hmin , hmax )P Φ (hmin , hmax ) − P < 0, (15) render the closed-loop system unstable. However, if using the
T PIDPLUS with period h = 16 s, the system is stable according
Φ(hmax , hmin )P Φ (hmax , hmin ) − P < 0, (16) to Lemma 5.1, but its transient response is slower than the
T
Φ(hmax , hmax )P Φ (hmax , hmax ) − P < 0, (17) proposed scheme, with an IAE=22.31 s.
then, the origin is asymptotically stable for any hmin ≤ hk ≤ We also test the disturbance rejection of the proposed scheme.
hmax . ⊳ The simulation result is illustrated in Figure 5, where a dis-
Proof: The system (12) can be viewed as a discrete-time turbance of amplitude 0.2 is added to the input of the plant at
system with time-varying uncertainty. Let the set A := t = 200 s. As it can be seen, the proposed scheme efficiently
{Φ(hk , hk−1 )|hmin ≤ hk ≤ hmax }k∈N . It is easy to verify rejects the disturbance with a small number of samples. With
that every matrix that belongs to A can be expressed as convex our method we experienced N = 47 transmissions with an
combination of the four matrices Φ(hk , hk−1 ) obtained when IAE=23.08, while the periodically sampled PI with h = 0.3
hk , hk−1 are equal either to hmin or hmax , and then the un- achieves an IAE=15.4 with N = 1500 number of transmis-
certainty is polytopic. Hence, asymptotic stability of (12) is sions. The PIDPLUS with period h = 16 s is able to reject
achieved if there exists a matrix P that satisfies (14)–(17), see the disturbance, but its response is slow, where the disturbance
for example Amato [2006]. ⋄ being only detected at approximately t = 205 s, 5 s after it
occurred. Moreover, it generated N = 28 transmissions with
The previous theorem states that the PIDPLUS controller en- an IAE=36.40.
sures asymptotic stability of the closed-loop system, no matter
when the controller is updated, provided that hmin ≤ hk ≤
hmax . Then, the set-point is asymptotically tracked and eventual 6.2 Example 2: First-order process with delay
undesirable oscillations around it are canceled if the controller
receives an infinity number of measurements. Hence, accord- We now evaluate our event-based scheme for the control of
ing to Theorem 5.1, we have to slightly modify the sampling a first-order process with delay. We remark that the stability
rule (5) by using hk = hmin or hk = hmax if the inter-sampling analysis provided in Section 5 are no longer valid for a plant
times given by (5) are too short or too long respectively. with delay. This example serves the purpose of demonstrating
the robustness of the proposed control scheme to delays.
The fact that an infinite number of measurements are required
to achieve asymptotic tracking of the set-point should not We consider the same example as above, with the addition of
mislead about the efficiency of the proposed control scheme. an actuation delay of 5 s. The results are depicted in Figure
The efficiency of the proposed method relies on the fact that 6. As it can be seen, the response becomes oscillatory with
the transmissions can be performed at any time, provided that the introduction of the actuation delay, whereas the proposed
hmin ≤ hk ≤ hmax , where hmax can be very large. control scheme successfully tracks the set-point and rejects the
disturbance. Moreover, the performance is very close to the
6. SIMULATION RESULTS periodically sampled PI controller with h = 0.3, where our
scheme provided IAE=26.41, while the periodic PI provided
We illustrate the performance of the proposed scheme when IAE=26.83. As an added benefit, the number of required trans-
controlling a first-order plant of the form (1). Then, just for the missions is significantly reduced to 62.
sake of investigation, we simulate the case in which there is a
time delay between the controller and the actuator. For steady- 7. CONCLUSIONS
state condition analysis we look at the number of transmitted
packets N from the sensor to the controller. Moreover, we use When event-based techniques are used in PI control scheme,
the Integral of the Absolute Error (IAE) parameter as a general drawbacks as sticking or output oscillation may arise. To cope
indicator for both transient and steady-state performance. The with these problems we proposed a novel event-based scheme
IAE value is calculated as Z that provides a PI-based triggering used in combination with
∞
IAE = |r − y(s)|ds. (18) PIDPLUS. Despite PIDPLUS was introduced to deal with
0
network imperfections like packet losses and time delays, here
The simulations were performed using Simulink in combina- we used such a controller in an event-based scheme, further
tion with Truetime, Cervin et al. [2003]. analyzing the stability property of the closed-loop system.
Simulations results show how the utilization of PIDPLUS in
6.1 Example 1: First-order process combination with the PI-based triggering rule is capable to
achieve asymptotic set-point tracking and disturbance rejection
as classic PI controller, while drastically reducing the number
We consider the same setup as in Section 3 and we compare the of transmissions from the sensor to the controller.
performance of the proposed scheme with a periodic implemen-
tation of period h = 0.3 s. The PI-based sampling scheme is set Future work include the extension to processes with delay
with Kp and Ti as for the controller. Note that the assumptions and the extension to multi dimensional systems. Moreover, the
of Lemma 5.1 are satisfied. effect of the derivative part of the PIDPLUS when used in
an event-based scheme is worth of investigation. Finally, the
0.0956 −0.090 optimal choice of δ in the PI-based triggering rule to achieve
For this system we find the matrix P = −0.090 0.1944
a trade-off between performance of the closed-loop system and
to fulfill Theorem 5.1 for hmin = 0.3 s and hmax = 1010 s. number of transmission is another future research topic.
2 2
Event−based Event−based
1.5 Periodic PI (h=0.3 s) 1.5 Periodic PI (h=0.3 s)
Periodic PIDPLUS (h=8.5 s) Samples
y(t)

y(t)
1 Samples 1

0.5 0.5

0 0
0 50 100 150 200 250 300 350 400 450 0 50 100 150 200 250 300 350 400 450
Time [s] Time [s]
1.5 1.5
Event−based Event−based
Periodic PI (h=0.3 s) Periodic PI (h=0.3 s)
1 1
Periodic PIDPLUS (h=8.5 s)
u(t)

u(t)
0.5 0.5

0 0
0 50 100 150 200 250 300 350 400 450 0 50 100 150 200 250 300 350 400 450
Time [s] Time [s]
100
100
−t [s]

tk+1 −tk [s]

k

50
k+1

50
t

0 0
0 50 100 150 200 250 300 350 400 450 0 50 100 150 200 250 300 350 400 450
Time [s] Time [s]

Fig. 4. Simulation result of a PIDPLUS controller with PI Fig. 6. Simulation result of a PIDPLUS controller with PI sam-
sampling in Example 1. The performance of the event- pling for disturbance rejection under an actuation delay
based controller is compared to a periodic implementation L = 8 s.
of a classic PI controller with period h = 0.3s and a
PIDPLUS controller with period h = 16s. S. Durand and N. Marchand. Further Results on Event-Based
PID Controller. In Proceedings of ECC 2009, pages 1979–
1984, Budapest, Hongrie, August 2009.
2
Event−based
L. Eriksson. PID Controller Design and Tuning in Networked
1.5 Periodic PI (h=0.3 s) Control Systems. PhD thesis, Helsinki University of Tech-
Periodic PIDPLUS (h=16 s)
nology, 2008.
y(t)

1 Samples
0.5 R. Goebel, R. Sanfelice, and A. Teel. Hybrid dynamical
0
systems. IEEE Control Systems Magazine, 29(2):28–93,
0 50 100 150 200
Time [s]
250 300 350 400 450
2009.
1.5 O. Kaltiokallio, L. M. Eriksson, and M. Bocca. On the perfor-
Event−based mance of the pidplus controller in wireless control systems.
In 18th MED, pages 707 –714, Jun. 2010.
Periodic PI (h=0.3 s)
1
Periodic PIDPLUS (h=16 s)
u(t)

0.5 H. Khalil. Nonlinear Systems. Pretience Hall, 3rd edition, 2002.

D. Lehmann. Event-based state-feedback control. PhD thesis,
0
0 50 100 150 200 250 300 350 400 450 Ruhr-Universitat Bochum, 2011.
Time [s]
P. G. Otanez, J. R. Moyne, and D. M. Tilbury. Using deadbands
100
to reduce communication in networked control systems. In
Americal Control Conference (ACC), volume 4, pages 3015–
tk+1 −tk [s]

50 3020, 2002.
M. Rabi and K. H. Johansson. Event-triggered strategies for
0 industrial control over wireless networks. In International
0 50 100 150 200
Time [s]
250 300 350 400 450
Conference on Wireless Internet, pages 1–7, 2008.
J. Sánchez, A. Visioli, and S. Dormido. A two-degree-of-
Fig. 5. Simulation result of a PIDPLUS controller with PI sam- freedom pi controller based on events. Journal of Process
pling for disturbance rejection. Performance comparisons Control, 21(4):639 – 651, 2011. ISSN 0959-1524.
as in Figure 4. J. Song, A. K. Mok, D. Chen, M. Nixon, T. Blevins, and W. Wo-
jsznis. Improving pid control with unreliable communica-
REFERENCES tions. In ISA EXPO Technical Conference, 2006.
P. Tabuada. Event-triggered real-time scheduling of stabilizing
F. Amato. Robust Control of Linear Systems subject to Uncer- control tasks. IEEE Transactions on Automatic Control, 52,
tain Time-varying Parameters. Springer, 2006. 2007.
K. E. Årzen. A simple event-based PID controller. In IFAC V. Ungan. Networked PID controllers for wireless sys-
World Congress, pages 423–428, Beijing, China, 1999. tems. Master’s thesis, Royal Institute of Technology (KTH),
K. J. Åström and B. Bernhardsson. Comparison of periodic November 2010.
and event-based sampling for first-order stochastic systems. V. Vasyutynskyy and K. Kabitzsch. Simple pid control algo-
In IFAC World Congress, 1999. rithm adapted to deadband sampling. In IEEE ETFA, pages
K.J. Åström and T. Hägglund. Advanced PID Control. ISA, 932 –940, Sep. 2007.
NC, 2006. X. Wang and M.D. Lemmon. Event-triggering in distributed
networked control systems. Automatic Control, IEEE Trans-
A. Cervin and K.J. Åström. On limit cycles in event-based actions on, 56(3):586 –601, march 2011. ISSN 0018-9286.
control systems. In 46th IEEE CDC, pages 3190 –3195, Dec. A. Willig. Recent and emerging topics in wireless industrial
2007. communication. IEEE Transactions on Industrial Informat-
A. Cervin, D. Henriksson, B. Lincoln, J. Eker, and K.-E. Arzen. ics, 4(2):102–124, 2008.
How does control timing affect performance? analysis and
simulation of timing using jitterbug and truetime. Control
Systems, IEEE, 23(3):16 – 30, june 2003. ISSN 1066-033X.