Articulo Variable Sampling Rate Networked Control Systems
Articulo Variable Sampling Rate Networked Control Systems
Articulo Variable Sampling Rate Networked Control Systems
ABSTRACT
In this paper we propose a variable sampling rate approach
for networked control system. During control procedure we
regulate sampling rate in terms of system behavior intensity.
Input increment is used as indicator to reflect the system intensity. A regulator is placed at the controller side, with fullstate feedback it can predict the system behavior in the future
and determine the next sampling instant according our regulation rule. We implement our approach both over ideal network and average modeled dropout network. By simulation
we show that by using our approach we realize more effective usage of network bandwidth than using constant sampling
rate. We also present simulation method, wherein we can find
regulation rule that optimize the networked control systems.
1. INTRODUCTION
During the past two decades, the use of data network in control applications is rapidly increasing. Connecting control
systems over communication networks eliminates the restrictions of traditional point-to-point control architectures and offers many advantages in terms of low cost installation and
maintenance, and reconfigurability. In spite of the benefits,
the communication network exhibits characteristics which degrade control system performance. These characteristics include discretization, quantization effect, time-delay, and data
loss. Thus, the challenges arise, to design networked control
systems, which should take into account more factors than
traditional control systems.
A proper message transmission protocol is necessary to
guarantee the network quality of service (QoS). There are a
wide variety of different commercially available control network, such as ControlNet, DeviceNet, Profibus, WorldFIP,
emerging Ethernet and Wireless. In the work of Lian, et al.[1]
three classes of control networks are compared for their performance. In the further work [2], the impact of network architecture on control performance NCS, and design considerations related to control quality of performance as well as
This work was supported by CT, Simens AG
Plant
x(tx )
full-state x(tk )
x = Ax(x) + Bu(t)
Packetization
Sensor
Regulator
2
Plant
Controller
u(t)
Controller
Actuator
Packetization
x
(tk )
u(t+ ) = K x
(tk )
: packet.
hk = tk+1 tk .
2.1. Assumptions
We model our system as a sampled-data system, and the plant
behaves as a continuous time system. The network resides between sensor and controller as well as controller and actuator,
Fig.1. The data generated by sensor or controller at a time
instant tk will be encapsulated into single packet. No processing delay and waiting delay for encapsulation and are also
assumed. The following assumptions are also used throughout our work: (1)All sensor nodes are time-driven. (2)All
controller nodes are event-driven. (3) All actuator nodes are
event-driven.
2.2. NCS Model
For simplicity we only consider linear SISO control system
as described in the following:
x(t)
= Ax(t) + Bu(t)
y(t) = Cx(t)
(1)
with the full-state feedback control u(t) = Kx(t). Therefore, the closed-loop continuous system is given by
x(t)
= (A BK)x(t).
(2)
Now the feedback control loop is closed through a communication network. The full-state information is sent in one
packet, as shown in Fig.1. The system equation can be written
as:
x(t)
= Ax(t) BK x
(t), t [tk , tk+1 )
x
(t+
k ) = x(tk ), k = 0, 1, ...,
(3)
ttk
eAs B ds)x(tk ),
(4)
t [tk , tk+1 ).
3. VARIABLE SAMPLING RATE APPROACH IN NCS
3.1. Principle
Our approach is inspired from the Variable Bit Rate (VBR)
strategy [7], which is widely used in audio and video compression and transmission. We place an external regulator at
the controller side, see Fig.2, whose task is to select and tell
sensor node the next sampling instant tk + hi , hi H, when
the full-state feedback x(tk ) arrives. This information will be
attached together with control signal in a packet. It wont increase the packet size too much, e.g. if s = 2 (it is the case
we consider in the following), we only need one additional bit
in the packet.
Since sampling is the quantization in time, we could use
high resolution quantization when the plant behaves fast and
intensive, which means using small hi in set H, in order that
the trajectory will be smoothed. We will use input increment
as our index to reflect the intensity of system behavior, defined as:
uk = |u(tk+1 ) u(tk )|,
(5)
where tk and tk+1 are two successive sampling instant. With
full-state feedback the regulator can predict the next input increment and distinguish the system intensity in future. Then
k = 0, 1, . . .
Rav =
(9)
Because the controller is event-driven, the average transfer rate in both down-link (controller and actuator) and uplink (sensor and controller) is equivalent, in the following
the average transfer rate means explicitly the feedback rate
in up-link. Obviously, if data is sampled at constant R Hz,
Rav = R.
Commonly the optimization problem in NCS can be divided into two classes of constrained optimization problems
as follows:
Given a network traffic R, minimize the control error
or control cost E.
(6)
Given a control error E, minimize the network traffic
load R.
and
.
Vk = V (x(tk+1 ))V (x(tk )) (kx(tk )k),
(7)
where e is the error between the actual and reference trajectories. t0 and tf are the initial time and final times of the
(10)
= (A BK)x(t) + Br(t),
(11)
0.7
control performance metric: IAE
80
70
60
50
40
30
20
0
0.1
0.2
0.3
threshold H
0.4
0.5
0.6
0.5
0.4
0.3
0.2
0
0.1
0.2
threshold H
0.3
0.4
0.5
uk = |u(tk+1 ) u(tk )|
Fig. 3. Impact of varying threshold on network traffic and
control performance.
(12)
+ K(I
k ) K r(tk )|,
R
= min{h1 , h2 }.
= eAh ,
= h eAs B ds, and h
where
0
Here we calculate the next input increment under the assumption that the next sampling interval is the smallest one in set
H, and compare it with threshold H, so that we could guarantee that, the actual input increment of two successive sam is always greater than threshold
pling instants with interval h
H. Our first regulation rule is:
Regulation Rule 1 At time tk ,
0.7
control performance metric: IAE
= |[r(tk+1 ) Kx(tk+1 )]
+ [r(tk ) Kx(tk )]|
Kx(tk + h)]
= |[r(tk + h)
0.6
constant sampling rate
0.5
0.4
0.3
variable sampling rate
0.2
20
30
40
50
60
average feedback rate R
70
80
av
A=
25
1
0
0
;B =
32
0
; C = [ 0 32.51 ] (13)
0.6
Regulation Rule 3: At time tk , if the packet which contains sensed full state x(tk ) is lost during transmission, sen where h
= min{h1 , h2 };
sor will sample data at tk + h,
otherwise the sensor will follow the Regulation Rule 2.
...
0.5
h = 1/25 s; h = 1/75 s
1
h = 1/40 s; h = 1/100s
1
0.4
h = 1/50 s; h = 1/120 s
1
...
0.3
0.2
20
40
60
80
average feedback rate: R
100
120
av
Fig. 5. Relationship between feedback rate and control performance at different values of h1 and h2
We select these points in all obtained curve. Working at
these points imply that, the network can be most efficiently
utilized: Under the same maximum network traffic constraint,
they can obtain best control performance, while under the
same maximum control error constraint, these points have
minimal network traffic utilization. Because with higher feedback rate the system behaves more like continuous system,
and the increase of feedback rate has smaller impact on the
control error, the difference between our approach and constant sampling rate will be smaller with the increase of average feedback rate.
Now we use the Lagrangian cost function (10) to select
1
. Here we
the optimal best operating point. We set = 500
dont describe how to select in detail, just assume that we
have it from practice in advance. The system performance
1
metric is J = E + 500
R. We can first use this cost function to
select the optimal constant sampling rate for system (13). The
optimal constant sampling rate is R = 62 Hz or packets/s, and
Jmin = 0.3811. For the above simulated threshold and sampling interval range, by using our approach, Jmin = 0.32572
1
1
s and h2 = 125
s.
with threshold H = 0.0112 and h1 = 25
The resulted average feedback rate is Rav = 45.7 packets/s.
5. OVER DROPOUT NETWORK
Now we implement our approach in a dropout network. We
use the simplest fixed average model for packet dropout. P is
defined as the possibility of packet dropout. If the feedback
packet from sensor is lost during transmission, new control
signal wont be generated and sent to plant, neither the signal for next sampling instant of sensor. The sensor has two
choices to determine the next sampling time: it will sample
with the same interval as last one or with the smaller interval
(in our case h2 ).
We have two corresponding new regulation rules as follows:
Regulation Rule 2: At time tk , if the packet which contains sensed full state x(tk ) is lost during transmission, sen-
0.7
0.8
0.6
0.2
30
40
50
60
70
80
90
average feedback rate Rav
100
110
120
Fig. 6. Relationship between feedback rate and control performance over dropout network.
In Fig.6, the R-E curves of VR are still lower the R-E
curve of CR. The resultant R-E curve by using Regulation
Rule 3 is a bit lower than that by using Regulation Rule 2,
because Regulation Rule 3 will use small sample interval to
compensate the lost information caused by packet dropout. If
the the threshold is great enough, the system will work only
with h2 regardless of applied regulation rules, so that both RE curves of VR will approach the CRs curve with increase of
Rav , and end at same point. If threshold is zero, with Regulation Rule 2 the system will work only with sampling interval
h1 , so the left edge points of R-E curve with Regulation Rule
2 and CR are overlapped. But with Regulation Rule 3, if H is
zero, the sample interval will be switched to h2 , so the resultant average feedback rate Rav > h11 . We could use the same
method as Chapter 4 to determine the best operating points,
and optimal best operating point by minimizing Lagrangian
cost function 10. We omit this part here.
In Fig.7 we illustrate how packet dropout impacts our variable sampling rate approach. The parameters of regulation
1
1
rule is: H = {h1 = 25
, h2 = 125
}, H = 0.0012. Higher
packet dropout possibility leads to heavier network traffic and
greater control error for both of regulation rules. The result-
P = 50%
0.9
0.8
regulation rule 2
P = 40%
0.7
system intensity index should be used. We have used heuristic method in our study. Most of our results was obtained by
simulation and some ideas are still waiting for more strongly
theoretic support. Analytical method should be developed.
The most possible way to analytical method is use optimal
control theory and we should convert the optimization into
numerical solvable question.
In future we could implement our approach in other network mode like two state Markov chain or network simulator,
so we could more characteristics about our approach. Some
interesting aspects should be taken into account, e.g impact of
sampling rate over packet dropout [10, 11]. Implementation
in real environment is also expected.
0.6
P = 30%
0.5
0.4
References
P = 10% P = 20%
P=0
0.3
0.2
45
regulation rule 3
50
55
60
average feedback rate: R
65
70
av