2020 American Control Conference

Denver, CO, USA, July 1-3, 2020

LQG-Optimal versus Simple Event-Based PID Controllers

Anton Cervin and Marcus Thelander Andrén

Abstract— In this paper, we study event-based PID control v

from an optimal stochastic control perspective. The purpose is
to better understand what implementation features are critical u y
for achieving good event-based PID performance. For this end, Generalized Generalized
hold Process sampler
we formulate an LQG control design problem for a double
integrator process with an integral disturbance, where the
solution is an ideal PID controller. We then consider the trade-
off between LQG cost and average sampling rate and give an
interpretation of the optimal sampled-data controller and event- Event-
based sampling policy in terms of PID control. Based on insights based
from the optimal solution, we finally discuss how suboptimal but controller
simple event-based PID controllers can be implemented. The
proposed implementation is evaluated in a simulation study Fig. 1. An event-based controller structure for a single-input–single-output
and compared to earlier work in event-based PID control. process. The solid lines represent continuous signal transmission, while the
The results highlight the importance of considering both the dashed lines indicate event-based communication.
triggering rule and the transmitted information in order to
obtain an event-based PID controller with good performance.
and (ii) control signal generation in between events. The
I. I NTRODUCTION design of (i) is the topic of several works, ranging from
Event-based feedback control has a history reaching back simple but well-known rules such as send-on-delta [9]–[11]
to at least the middle of the last century, but the field has and integral triggering [12], [13], to more complex model-
received renewed interest since the publication of [1] and dependent rules [7], [14]. While model-based rules often
[2], now two decades ago. The former paper studied optimal achieve better performance in theory, they are also more
event-based impulse control of a first-order stochastic sys- difficult to implement in practice. This is also true for (ii)
tem, and analyzed the trade-off between average sampling where options range from a simple zero-order hold to more
rate and output variance. The latter proposed a simple involved signal generators (a.k.a. generalized holds) [15],
event-based PID controller and evaluated its performance [16]. One of the main motivators of this paper is the study
in simulations. Since then, several theoretical and practical of reasonable choices of (i) and (ii) for event-based PID,
research studies of event-based control have been performed, striking a balance between performance and simplicity of
see the survey papers [3] and [4]. implementation.
One possible structure for an event-based controller is The main contributions of this paper are
shown in Fig. 1. Typically, we are interested in the trade-
off between regulatory performance and the average event • the formulation of an LQG design problem, where the
rate (corresponding to, e.g., network usage) in the loop. optimal solution is an ideal PID controller;
A proposal to close the gap between theoretical, optimal • an interpretation of the optimal sampled-data controller
event-based control and practical, heuristic event-based PID and event-based sampling policy in terms of PID con-
control via a simple benchmark problem was put forward in trol;
[5]. In this paper, we continue that development and study • a numerical evaluation of some common heuristic trig-
a modified benchmark problem from both LQG and PID gering rules and control generators in comparison to the
perspectives. The work is prompted by recent theoretical optimal solution.
breakthroughs by Mirkin and co-authors [6]–[8], who have
provided an LQG-optimal controller structure under inter- The rest of this paper is outlined as follows. In Sec. II, we
mittent sampling. The resulting optimal controller may in formulate the LQG design problem that results in an ideal
general not be practical to design nor to implement, but it PID controller. In Sec. III we review Mirkin’s LQG-optimal
provides a lower bound on the achievable LQG cost with sampled-data controller structure and give an interpretation
any event-based controller. of it for the considered design problem in terms of PID con-
There are two main aspects to consider in the design trol. In Sec. IV we discuss various heuristic event-based PID
of event-based control: (i) the rule for triggering events implementations and how they relate to the optimal solution.
A performance comparison is presented in Sec. V, where
This work has been supported by the Swedish Research Council, grant the benefit of different controller structures and sampling
no. 2017-04491. The authors are with the Department of Automatic Control,
Lund University, Sweden, and are members of the ELLIIT Excellence policies are evaluated. Finally, Sec. VI concludes the paper
Center. E-mail: {anton,marcus}@control.lth.se and suggests some future work.

978-1-5386-8266-1/$31.00 ©2020 AACC 3678

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY ROURKELA. Downloaded on January 16,2025 at 15:41:08 UTC from IEEE Xplore. Restrictions apply.
 vz where the optimal Kalman gain kz is obtained by solving
the associated Riccati equation, yielding
1 √
kz = rz . (5)
s
vx The complete LQG controller is then given by
z
u P w P x y ẑ˙ = kz (ÿ/k 2 − u − ẑ),
k k (6)
s s u = −lx ẏ/k − ly y − ẑ.
At first glance, it may seem that the controller needs access
Fig. 2. Double integrator process with integral input disturbance. to ÿ. However, in input–output form the controller can be
written as
II. T HE LQG D ESIGN P ROBLEM
(kz +klx )p2 + (k 2 ly +kkz lx )p + k 2 kz ly
As a benchmark problem, consider the system in Fig. 2, u=− y, (7)
k2 p
consisting of a double integrator and an input disturbance
d
integrator. This is a simple control model, but relevant for, where p := dt is the differential operator. Comparing this to
e.g., mechanical systems such as a satellite or a cart on an ideal PID controller in parallel form,
a rail. We will first show how an LQG design problem  1 
for the system can be formulated such that the solution u = −K 1 + + pTd y,
pTi
is an ideal PID controller. While a similar problem for a
marginally stable process was proposed in [5], a drawback we obtain the algebraic relationships
of that setup was the very complicated expressions for the ly k + k z lx
resulting PID controller with a derivative filter. Further, not K= , (8)
k
all PID controllers could be interpreted as LQG controllers, ly k + k z lx
i.e., the inverse problem was not well-defined. Here we Ti = , (9)
k z ly k
choose a different process and cost function, which produces k z + lx k
a simpler solution and also permits inverse calculations (i.e., Td = . (10)
ly k 2 + k z lx k
from given PID parameters to an LQG problem) in all cases.
The derivation of the LQG controller below is trivial and we B. Interpretation of the LQG Controller
refer to, e.g., [17] for further details. The LQG controller (6) is not suitable for implementation
as it stands since it has ÿ as an input. A change of variables,
A. Translation between LQG and PID
k z lx kz
For the system in Fig. 2, assume that k > 0 is a scalar gain ẑi = ẑ − y − 2 ẏ,
parameter and that vz and vx are independent continuous- k k
time white noise processes with intensities rz > 0 and 1, separates out the integrator state, ẑi , and moves the direct
respectively. The control objective is to minimize the cost terms into the state feedback law. The resulting observer can
function n o be written as
J = E qy y 2 + 2qyw yw + w2 , (1)  k z lx 
ẑ˙i = −kz upi + y + ẑi , (11)
2
k
where qy > 0 and qyw ≤ qy are scalar weights. Note that
where the control signal u = upi + ud has been split into a
we penalize w = z + u rather than the control signal u in
PI part and a D part with
order to allow the controller to have true integral action.
Assuming that the states z, x, and y are available for upi = −Ky − ẑi ,
(12)
continuous feedback, the linear-quadratic control law (e.g., ud = −KTd ẏ.
[17]) is given by
In this formulation ẏ does not enter the integrator, which is
u = −lx x − ly y − z, reasonable.
The observer (11) can be further modified by introducing
where the feedback gains lx and ly are given by the solution
a parameter 0 ≤ α ≤ 1 to split the PI control signal as
of the associated algebraic Riccati equation, yielding
q
√
upi = αupi + (1 − α)upi = −αKy − αẑi + (1 − α)upi ,
lx = 2 qy − qyw , (2)
√ yielding the family of possible observers
ly = q y . (3) 

Next we consider optimal estimation of the state vector. ẑ˙i = −kz (1 − α)(upi + ẑi ) + kzklx − αK y .
Since there is no measurement noise on y, we immediately With α = 0 we retain (11), while with α = 1 we recover
have x = ẏ/k. The lack of process noise on x allows us to the standard integrator formula that only uses y:
formulate a reduced-order Kalman–Bucy filter for z as
K
ẑ˙ = kz (ÿ/k 2 − u − ẑ), (4) ẑ˙i = y. (13)
Ti

3679
Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY ROURKELA. Downloaded on January 16,2025 at 15:41:08 UTC from IEEE Xplore. Restrictions apply.
 vz vx v
ud := u̇ P w P x y
1 k k
s s s ud θ Kalman
LQR, ηa Process, η
filter, η̂
Fig. 3. Modified version of the system in Fig. 2, where the input disturbance
and control signal have been merged into a single, controllable state.

k z lx Fig. 4. Representation of the LQG-optimal sampled-data controller

With α = kK we obtain the observer
structure. The vectors η, ηa and η̂ are the state vectors in each subsystem
1 respectively. Solid and dashed lines represent continuous and sampled
ẑ˙i = −


upi + ẑi , (14) signals respectively.
Ti
which can be recognized as the classical “automatic reset” With these modifications the design problem is well-
realization of integral action (see, e.g., [18]). posed and can be summarized by the following generalized
In a continuous feedback setting, all of the above observers process with state vector η = [x, y, w]⊺ , noise vector v =
behave identically, but in an event-based implementation they [vx , vy , vz ]⊺ , cost vector ξ and system output θ:
may yield different results depending on when and how the 
variables are communicated between the sensor, controller, η̇ = Aη + Bv v + Bud ud ,

and actuator. ξ˙ = Cξ η + Dξud ud , (15)

To further guide the design of event-based implementa- 
θ = Cθ η + Dθv v.
tions, we continue in the next section with a review of
Mirkin’s LQG-optimal sampled-data controller. The system parameters in (15) are given by
     
0 0 k k 0 0 0
III. M IRKIN ’ S LQG-O PTIMAL C ONTROLLER
A = k 0 0 , Bv = 0 0 0 , Bud = 0 ,
The LQG-optimal controller under any given sampling 0 0 0 0 0 1 1
sequence was originally derived by Mirkin in [6], and 
0 q qyw 1
  
subsequently adapted to event-based sampling in [7], [8]. 0
Cξ = 0
 2
qy − qyw 0 , Dξud =  0  ,

The optimal solution retains elements of the continuous-time √
implementation in the form of an LQ control law and a 0 0 0 qud
Kalman-Bucy filter. We will here review the optimal design
   
Cθ = 0 1 0 , Dθv = 0 1 0 .
applied to the double integrator system in Fig. 2, where
the controller elements can be interpreted in terms of PID The cost in the modified design problem is given by
control.
n o n o
Jmod = E ξ ⊺ ξ = E qy y 2 +2qyw yw+w2 +qud u2d . (16)
A. Preliminaries
B. The Optimal Controller
The design of the optimal controller assumes a well-posed
While the optimal solution has a realization divided into a
H2 design problem according to the conditions in [19, Sec.
discrete-time controller and a generalized sampler and hold
14.5]. However, the system in Fig. 2 does not satisfy these
[7, Remark 2], we will here instead opt for the more intuitive
conditions due to the uncontrollable state z and lack of
realization shown in Fig. 4. The controller is then divided
measurement noise. We will therefore instead consider the
into a Kalman-Bucy filter on the sensor side,
design for a slightly modified problem, which in the limit
will be equivalent to the original one. η̂˙ = Aη̂ + Bud ud + Ks (θ − Cθ η̂), (17)
The first modification is to transform the system in Fig. 2
into the system shown in Fig. 3. This is done by regarding the and an LQR signal generator on the actuator side. The
controller integrator as part of the process, and then merging control signal is based on a continuous-time simulation of
the input disturbance state z and the control signal u into the the process, whose state vector ηa is reset to the current
controllable state w = z + u. The “new” input to the system state-estimate η̂ at each sampling time ti :
is then the derivative of the original input, ud := u̇. Note
(
η̇a = (A − Bud La )ηa , ηa (ti ) = η̂(ti ),
that, so far the original and modified design problems are (18)
ud = −La ηa .
equivalent, since regardless of whether the system in Fig. 2
or 3 is considered, the optimal control design will still result The vectors Ks and La are the Kalman–Bucy filter and LQR
in the same closed-loop system and cost. gains respectively from the corresponding continuous-time
Secondly, we proceed similarly to [7] and add an (arti- LQG controller.
ficial) white noise signal vy with intensity ry > 0 to the As ry , qud → 0+ we retain the original design problem.
system output y and a small artificial penalty qud > 0 to the The simulated LQR together with the input integrator on the
input signal ud . As long as ry and qud are close to zero, the actuator side are then reduced to an ideal PID controller
modified design problem will only differ slightly, and in the
limit ry , qud → 0+ the original problem is recovered. u = −Kya − zi,a − KTd ẏa ,

3680
Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY ROURKELA. Downloaded on January 16,2025 at 15:41:08 UTC from IEEE Xplore. Restrictions apply.
 where the “measurement” ya and integral action zi,a are
generated by the intermittently reset simulation
ÿa + klx ẏa + k 2 ly ya = 0, ya (ti ) = y(ti ), ẏa (ti ) = ẏ(ti ),
K
żi,a = ya , zi,a (ti ) = ẑi (ti ).
Ti
(19)
On the sensor side, the signals y and ẏ are directly available,
while ẑi is given by (13). At sampling times {ti }, the data
(y, ẏ, ẑi ) is transmitted to the actuator side and resets the
simulation according to (19).
Naturally, this continuous-time scheme is difficult to re-
alize in practice, but from a theoretical point of view it Fig. 5. Numerical approximation of the optimal sampling threshold (blue
provides a useful performance baseline for comparisons to surface) for the LQG-optimal controller described in Sec. III. A line parallel
more practical implementations. to the LQ feedback gain vector [lx , ly , 1] (red) is plotted for reference. The
threshold was obtained using the method described in [20].
C. Event-Based Sampling
Define the error η̃ := ηa − η̂. Whenever η̃ = 0 holds, signal generators infeasible. Arguably, this is the case for the
the optimal sampled-data controller will behave identically optimal controller of the previous section, which motivates
to its continuous-time counterpart. This is the case just after the need for suboptimal but simpler implementations. In
sampling, but in between sampling actions the error will drift this section we will highlight some features of previously
due to disturbances in the system. The dynamics of η̃ will proposed event-based PID controllers from the literature, and
fundamentally determine the closed-loop performance, and discuss which features are useful yet practical to achieve
it can be shown that (16) can be re-expressed as [8, Thm. 1] good LQG performance.

2 A. Årzén’s Simple Event-Based PID Controller
Jmod = γ0 + qud E{ La η̃ }, (20)
Most proposals of event-based PID controllers in the
where γ0 is continuous-time LQG optimal cost. The value literature can be traced back to the seminal paper of Årzén
of the second term in (20) is determined by the choice of [2]. We therefore start by giving a brief review of the
sampling policy, which ideally should be as small as possible algorithm here.
for a given average sampling rate. On the sensor side, the system output y is monitored
As detailed in [14], the event-based sampling policy that periodically with a fixed, short period hnom , and the decision
minimizes (20) is in the form of a threshold on η̃. Finding to transmit data to the controller on the actuator side is based
the optimal threshold for a given setup generally requires on a simple send-on-delta condition combined with a time-
computationally demanding numerical methods. However, as out hmax . The sensor operation is described by the following
seen in Fig. 5, the optimal threshold can in this case be pseudo-code, in which h act denotes the actual sampling
well approximated by two parallel hyperplanes, orthogonal period:
to the vector [lx , ly , 1]⊺ . This approximation corresponds to y := AnalogIn();
a policy which triggers sampling whenever h_act := h_act + h_nom;
IF abs(y - y_old) >= delta OR h_act >= h_max THEN
|ũ| > ∆. (21) Send(y);
y_old := y;
where ũ := K(y − ya ) + KTd (ẏ − ẏa ) + (ẑi − zi,a ) is the ENDIF
difference in control signal between two ideal PID controllers An on-the-fly discretized version of the PID algorithm is
with feedback from the true process and the simulation (19) then implemented on the actuator side. It runs at each sensor
respectively. event and is represented by the following pseudo code:
We will use (21) as the (near optimal) event-based thresh-
y := Receive();
old policy for the optimal controller structure, where the h_act := Time() - time_old;
design parameter ∆ > 0 is chosen as a trade-off between u_p := -K * y;
LQG-cost and average sampling rate. a_d := T_d / (N * h_act + T_d);
u_d := a_d * u_d - K * N * a_d * (y - y_old);
IV. S IMPLE E VENT-BASED PID I MPLEMENTATIONS u := u_p + u_i + u_d;
AnalogOut(u);
Event-based implementations of PID controllers are usu- u_i := u_i - K / T_i * h_act * y;
ally motivated by improved resource efficiency, especially y_old := y;
time_old := Time();
in networked control systems, where savings in energy
and bandwidth can be achieved by transmitting data less Here, N represents the maximum derivative gain in the
often. However, the computational capacity in the sensor controller. If there is no measurement noise, we can let
and actuator nodes are usually limited in embedded imple- N → ∞ and the derivative part becomes a pure backward
mentations, which makes complex triggering conditions and difference.

3681
Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY ROURKELA. Downloaded on January 16,2025 at 15:41:08 UTC from IEEE Xplore. Restrictions apply.
 B. Choice of Measurement Filter and Triggering Rule vz
Årzén’s event-based controller and many subsequent ones
uses the send-on-delta rule [9], possibly after filtering out 1
measurement noise [10]. It has been pointed out in several s
works, e.g., [12], [13] that integral sampling is less sensitive z vx
to noise and also eliminates the deadband effect. u P w P 1 y
The optimal solution contains a Kalman-Bucy filter at the (s + 1)3
sensor that filters out measurement noise and estimates the
full state vector. It triggers on the difference in control signal Fig. 6. Process considered for Setup B in the performance comparisons.
between the sensor and actuator side, hence utilizing the
entire state vector. utilize a simplified model in the actuator [15]. A further
For PID control, it would seem like a reasonable com- possible solution is to use ZOH but adapt the feedback gain
promise to trigger on the PD part of the control signal, i.e., according to the recently experienced hold intervals. We will
whenever refer to this heuristic method as adaptive PD gain.
K(y − yold ) + KTd (ẏ − ẏold ) > ∆. V. P ERFORMANCE E VALUATION
This would also be relatively simple to implement in analog In this section we evaluate the performance of several
electronics, in conjunction with a second-order anti-aliasing combinations of the heuristic event-based methods discussed
filter (see [18]). We will refer to this option as PD triggering. in the previous section and compare the results to Mirkin’s
To avoid stationary errors in the case of zero process noise optimal controller with event-based sampling. Higher re-
and no events, the trigger needs to be combined with a time- source efficiency being one of the main motivators for event-
out hmax , similar to Årzén’s solution. based control, we focus on the trade-off between average
sampling rate (equivalent to mean network or CPU usage)
C. Choice of Data to Communicate and the LQG cost as measured by (1). Two different setups
Most heuristic methods only communicate the sensor are considered:
value y to the actuator side at events, although some works A. The double integrator process in Fig. 2 with the gain,
have proposed separate triggers and transmissions for the cost, and noise parameters k = 1, qy = 4, qyw = 0,
different parts of the PID controller, e.g., [11]. rz = 1. The LQG-optimal controller is a PID controller
The optimal solution, however, transmits an estimate of with the parameters K = 4, Ti = 2, Td = 0.75.
the full state vector. Sending a few extra bytes in a network
B. The stable third-order process in Fig. 6 with the cost
packet costs very little, and, if PD triggering is used, both
and noise parameters qy = 5.5, qyw = 0, rz = 0.1. For
y and ẏ are already available in the sensor node and should
this higher-order process, the LQG-optimal controller
be communicated to the actuator.
is not a PID controller. We can however find the best
D. Choice of Integrator Implementation possible PID parameters using nonlinear optimization,
Årzén’s integrator implementation does not work well for yielding K = 2.15, Ti = 2.67, Td = 1.23.
long inter-event times. Durand & Marchand therefore pro- Setup A matches the studied problem exactly, while Setup B
posed to include a forgetting factor to alleviate the problem is representative of a lag-dominated stable process from
[21]. Another solution is to implement the integrator in the process industry. In both cases, the LQG costs found in the
form of an automatic reset, Eq. (14). This has for instance evaluation have been normalized so that the continuous PID
been adopted in the PIDplus commercial controller [22] and controller has a relative cost of 1.0.
has been proven to work well for event-based PI control [13]. The performance of the different controllers were evalu-
The optimal solution estimates the integral state as part of ated by Monte Carlo simulations in TrueTime [23]. Through-
the full state vector. It is however of practical advantage to out, hnom = 0.01 s was used as the simulation timestep
have the integrator separately, since it becomes easier to deal (and hence smallest possible event detection interval). For
with practical issues such as anti-reset windup and controller each scenario, a 1000 s simulation was run using the same
mode switches [18]. noise input sequences. For the event-triggered algorithms,
the trigger parameter ∆ was swept over a range of values,
E. Choice of Control Signal Generator generating different average sampling rates and LQG costs.
At the actuator side, zero-order hold (ZOH) between
events is often assumed in heuristic implementations. Setting A. Triggers and Sent Information
the correct feedback gain for each output however requires We first study how the choice of event trigger (send-on-
knowledge of the next hold interval. Better choices may delta or PD trigger) and the information sent between sensor
therefore be impulse generators or general control signal and actuator (only y or y and ẏ) impact the performance of
generators as discussed in [16]. the event-based PID controller. The rest of the PID controller
The optimal solution includes a full-state plant model at is implemented like Årzén’s (see Sec. IV-A). Results for
the actuator as a control generator. One compromise is to Case A are reported in Fig. 7. It is seen that PD triggering

3682
Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY ROURKELA. Downloaded on January 16,2025 at 15:41:08 UTC from IEEE Xplore. Restrictions apply.
 Fig. 7. Trade-off between LQG cost and average sampling rate for Setup A Fig. 9. Trade-off between LQG cost and average sampling rate for Setup A
using different triggers and sent information. using different integrator algorithms and fixed or adaptive feedback gain.

Some small differences are visible, but overall the choice

of integrator and PD gain implementation has quite a small
impact on the performance. The automatic reset integrator
performs slightly better than the forgetting integrator, while
the adaptive gain seems to sometimes improve things and
sometimes not. The results for Setup B are very similar and
are omitted here.
C. Overall Comparison
In a final comparison, we study the range of possible
performance under the various event-based PID controllers
and the LQG-optimal periodic and event-based controllers.
We also include the optimal periodic sampled-data controller
[24] as a point of reference. The best heuristic event-based
PID controller is achieved by a combination of PD event
triggering, sending both y and ẏ to the actuator, an automatic
reset integrator, and a fixed PD gain. The results are shown in
Fig. 8. Trade-off between LQG cost and average sampling rate for Setup B
using different triggers and sent information.
Figs. 10 and 11. For the double integrator, the performance of
the best event-based PID controller is very close to Mirkin’s
gives a dramatic performance improvement over send-on- optimal controller with event-based sampling. For the stable
delta. The combination of PD triggering and sending both third-order system, a large performance improvement is also
y and ẏ gives close to optimal event-based performance possible, but the distance to optimality is not exactly known
for average event rates down to 1.2 Hz. The results for since the optimal event-based controller in this case only
Setup B are quite similar as seen in Fig. 8. Note however provides a lower bound on the achievable cost. Overall it
the performance gap between the continuous-time PID and is seen that simple event-based controllers can be made to
the LQG controllers in this case, as a PID controller cannot perform much better than periodic controllers if the proper
optimally control a third-order process under the given LQG implementation choices are made.
criterion. VI. C ONCLUSION AND F UTURE W ORK
B. Integrator Algorithm and Feedback Gain In this paper, we have built upon the work in [5] and
Keeping the controller with PD triggering and transmis- studied a modified benchmark problem for which the LQG-
sion of y from the previous subsection, we now try two optimal controller is a PID controller. Using the optimal
alternative integral implementations: Durand & Marchand’s sampled-data controller structure by Mirkin as a baseline, we
exponential forgetting integral and the automatic reset imple- have studied the LQG-optimal solution of the PID control
mentation. We also experiment with fixed PD feedback gain problem to gain new insights on how a good but simple
versus an adaptive feedback gain that is adjusted based on the event-based PID controller could be implemented.
previous sampling interval. Results for the double integrator The results suggest that a practical yet well-performing
are reported in Fig. 9. event-based PID should trigger events not only on the

3683
Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY ROURKELA. Downloaded on January 16,2025 at 15:41:08 UTC from IEEE Xplore. Restrictions apply.
 R EFERENCES
[1] K. J. Åström and B. Bernhardsson, “Comparison of periodic and event
based sampling for first-order stochastic systems,” in Preprints 14th
World Congress of IFAC, Beijing, P.R. China, Jul. 1999.
[2] K.-E. Årzén, “A simple event-based PID controller,” in Preprints 14th
World Congress of IFAC, Beijing, P.R. China, Jul. 1999.
[3] W. P. M. H. Heemels, K. H. Johansson, and P. Tabuada, “An introduc-
tion to event-triggered and self-triggered control,” in IEEE Conference
on Decision and Control, Dec 2012.
[4] Q. Liu, Z. Wang, X. He, and D. Zhou, “A survey of event-based
strategies on control and estimation,” Systems Science & Control
Engineering, vol. 2, no. 1, pp. 90–97, 2014.
[5] A. Cervin, “LQG-optimal PI and PID control as benchmarks for event-
based control,” in Proc. 2nd International Conference on Event-Based
Control, Communication and Signal Processing, Krakow, Poland,
2016.
[6] L. Mirkin, “Intermittent redesign of analog controllers via the Youla
parameter,” IEEE Trans. Autom. Control, vol. 62, no. 4, pp. 1838–
1851, 2017.
[7] M. Braksmayer and L. Mirkin, “H2 optimization under intermit-
tent sampling and its application to event-triggered control,” IFAC-
PapersOnLine, vol. 50, no. 1, pp. 7869–7874, 2017, 20th IFAC World
Fig. 10. Trade-off between LQG cost and average sampling rate for Setup A Congress.
using different optimal and suboptimal PID and LQG controllers. [8] A. Goldenshluger and L. Mirkin, “On minimum-variance event-
triggered control,” IEEE Control Syst. Lett., vol. 1, no. 1, pp. 32–37,
2017.
[9] M. Miskowicz, “Send-on-delta concept: An event-based data reporting
strategy,” Sensors, vol. 6, no. 1, p. 49, 2006.
[10] V. Vasyutynskyy and K. Kabitzsch, “Implementation of PID controller
with send-on-delta sampling,” in International Conference Control,
2006, pp. 4423–4428.
[11] P. G. Otanez, J. R. Moyne, and D. M. Tilbury, “Using deadbands to
reduce communication in networked control systems,” in Proceedings
of the 2002 American Control Conference (IEEE Cat. No.CH37301),
vol. 4, 2002, pp. 3015–3020 vol.4.
[12] M. Miskowicz, “Asymptotic effectiveness of the event-based sampling
according to the integral criterion,” Sensors, vol. 7, no. 1, pp. 16–37,
2007.
[13] U. Tiberi, J. Araujo, and K. H. Johansson, “On event-based PI control
of first-order processes,” in 2nd IFAC Conference on Advances in PID
Control, 2012.
[14] M. T. Andrén, B. Bernhardsson, A. Cervin, and K. Soltesz, “On event-
based sampling for LQG-optimal control,” in 2017 IEEE 56th Annual
Conference on Decision and Control (CDC), Dec 2017, pp. 5438–
5444.
[15] V. Vasyutynskyy and K. Kabitzsch, “First order observers in event-
based PID controls,” in 2009 IEEE Conference on Emerging Tech-
nologies Factory Automation, Sep. 2009, pp. 1–8.
Fig. 11. Trade-off between LQG cost and average sampling rate for Setup B [16] K. Åström, “Event based control,” in Analysis and Design of Nonlinear
using different optimal and suboptimal PID and LQG controllers. Control Systems, A. Astolfi and L. Marconi, Eds. Springer, 2008,
pp. 127–147.
measurement y but also on its derivative ẏ, here referred [17] K. J. Åström, Introduction to stochastic control theory, ser. Mathe-
matics in science and engineering. Academic Press, 1970, vol. 70.
to as ”PD triggering”. Furthermore, the sensor should also [18] K. J. Åström and T. Hägglund, Advanced PID Control. ISA – The
transmit both y and ẏ to the controller at events. Small further Instrumentation, Systems and Automation Society, 2006.
improvements can be achieved if the integral action uses [19] K. Zhou, J. C. Doyle, and K. Glover, Robust and Optimal Control.
Englewood Cliffs, NJ: Prentice-Hall, 1996.
the automatic reset realization. Adapting the controller gain [20] M. T. Andrén, “Using radial basis functions to approximate the LQG-
based on the length of the recent sampling period is probably optimal event-based sampling policy,” in 2019 18th European Control
not worth the extra implementation complexity, however. Conference (ECC), June 2019, pp. 2832–2838.
[21] S. Durand and N. Marchand, “Further results on event-based PID
One possible direction for further performance improve- controller,” in Proc. European Control Conference, 2009.
ments is to experiment with other control signal generators [22] J. Song, A. K. Mok, D. Chen, M. Nixon, T. Blevins, and W. Wojsznis,
than ZOH. The downside of more sophisticated signal gener- “Improving PID control with unreliable communications,” in ISA
EXPO Technical Conference, 2006, pp. 17–19.
ators is that they require a process model to be implemented, [23] D. Henriksson, A. Cervin, and K.-E. Årzén, “TrueTime: Simulation of
which is often not available in applications where PID control loops under shared computer resources,” in Proceedings of the
control is considered. Another research direction would be to 15th IFAC World Congress on Automatic Control, Barcelona, Spain,
Jul. 2002.
examine how well PD triggering actually approximates the [24] K. J. Åström and B. Wittenmark, Computer-Controlled Systems (3rd
optimal event trigger for the full-model system. Ed.). Upper Saddle River, NJ: Prentice-Hall, 1997.

3684
Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY ROURKELA. Downloaded on January 16,2025 at 15:41:08 UTC from IEEE Xplore. Restrictions apply.