Documente Academic
Documente Profesional
Documente Cultură
he defining characteris-
tic of a networked control
system (NCS) is having one
or more control loops
closed via a serial communi-
cation channel. Typically
when the words networking and control
are used together, the focus is on the
control of networks, but in this article
our intent is nearly inverse: not control
of networks but control through net-
Walsh (gwalsh@eng.umd.edu) and Ye are with the Department of Mechanical Engineering, University of Maryland, College Park, MD
20742, U.S.A.
0272-1708/01/$10.00©2001IEEE
February 2001 IEEE Control Systems Magazine 57
CAN devices range from stand-alone In summary, designers choosing to
CAN interface chips such as Intel’s 82527 use a networked control system archi-
(see Fig. 1 for the authors’ application), tecture are motivated not by perfor-
Microchip’s MCP2510, and Philips’ mance, but by cost, maintenance, and
SJA1000 to microcontrollers with an inte- reliability gains. The use of serial com-
grated on-chip CAN interface, such as munication networks in a control sys-
Motorola’s 68HC05, Siemens’ (or tem is such a clear win that it is now
Infineon’s) SAB-C167CR, and Texas In- widely supported from the device to
struments’ TMS320C24x series DSP. the system level. The next section out-
The serial communication channel, lines some application issues with
which multiplexes signals from the sen- NCSs and narrows the discussion to
sors to the controller and/or from the scheduling. Scheduling control net-
controller to the actuators, serves Figure 1. A networked control system work traffic is the focus of the following
many other uses besides control (see smart node designed by the authors circa section. The last section reports simu-
Fig. 2). Each of the system components 1996. The card is composed of a Motorola lation and physical experiments.
connected directly to the network is 68HC11 with an Intel 82527 CAN interface
denoted a physical node. Logical subdi- chip. Application Issues
vision is also common, so, for example, Control networks differ from data net-
although the high speed CAN network in an automobile might works in important ways, such as having short, frequent pack-
have 50 logical nodes, physically only a half a dozen devices ets with real-time requirements. Consider the generic NCS
may be connected. Control networks are typically local area problem in Fig. 2, where several sampled continuous- time
networks, and while hierarchical collections of networks are (vector-valued) outputs, data records, are sent through a sin-
found, control loops are closed locally. Important exceptions gle serial communications channel and reconstructed by a
to this rule exist. For example, in teleoperation systems, su- smart actuator, which then uses the resulting data image to
pervisory control of discrete-event systems, and sensor net- compute the control action. The computation may be carried
works control loops are often closed over wide area out remotely and the commands also sent over the network,
networks. A control network should have at least three although this is not shown in the figure. A control network im-
nodes; with two nodes there is no reason to use a network. pacts the closed-loop performance by creating differences be-
Plant outputs from different nodes are often coupled, and tween the data records and their associated remote images,
outputs on different time scales are also found. Congestion is and the performance of a well-designed NCS degrades grace-
a common issue. fully in the presence of congestion on the network. In a net-
To Unrelated Nodes
Smart Sensors with Network
To Unrelated Nodes
#1
y^1 y1
u1
Interface
Network
Controller Plant #2
y^2 um y2
#p
y^p yp
Serial Network
UN UN UN
Figure 2. The essential NCS problem. The data record is located at the smart sensor and the data image at the smart actuator. Although in
the figure only the measured plant outputs y( t ) are transmitted over the network, plant control signals u( t ) may also be so transmitted.
2
must be larger than the average transmission time interval, 1.5
1
otherwise the finite sensor queue will overflow. The statically 0.5
scheduled policy employing a queue is compared to our 0
try-once-discard (TOD) policy in Fig. 3. The TOD policy, which –0.5
0 2 4 6 8 10 12 14 16 18 20
has no queue, discards data if the network is unavailable. Time (s)
(a)
The transport delays in an NCS are not physical: network
speeds are such that the data transmission time is very 1.5
small. Simply removing the queue, which is possible since
y ( ), y^(---)
1
the network is local, dramatically improves the congested
behavior of the networked control system. The example 0.5
serves to point out a fundamental difference between data
and control networks: in a control network, minimizing the 0
0 2 4 6 8 10 12 14 16 18 20
l ∞ norm of the difference between a data record and its as- Time (s)
sociated image is more important than exactly reproducing (b)
the data image at the controller node, as would be done for a
Figure 3. Comparison of the NCS with and without queue (same
normal data network using queues and retransmissions. random transfer interval distribution τ = 300 ms). Admitting packet
The problem of congestion immediately suggests the ap- loss yields better performance. (a) Queued NCS, τ = 0.30, Tp = 0.32,
plication of information theoretic tools, but closer examina- no overflow; (b) NCS without queue, τ = 0.30.
n$( t ). Such predicting and filtering will add extra states and
with state x c ∈ Rn c and output u ( t ) ∈ R q ; and the network
n
reported versions of u ( t ) and y( t ). Without loss of general- Without a network, e( t ) = 0, and hence the dynamics re-
duce to x&( t ) = A11 x ( t ). It is assumed that the controller has
ity, we have assumed Dp = 0. Outputs measured locally at an
been designed ignoring the network, hence A11 is Hurwitz.
actuator can be incorporated directly into the controller
Consequently, there exists a unique symmetric positive def-
and do not require treatment in our model. If such outputs
inite matrix P such that
are needed elsewhere, the actuator node can also be consid-
ered a smart sensor. Because of the network, only the re-
ported output y$( t ) is available to the controller and its AT11 P + PA11 = − I . (3)
prediction processes; similarly, only u$( t ) is available to the
actuators on the plant. Commonly used networks support Define the constants σ1 = λ min ( P ) and σ2 = λ max ( P ). Since we
broadcast, hence n$( t ) is globally known, and in such a case are modeling the network as a perturbation on the system,
the controller itself may be physically distributed. choosing the right-hand side of (8) equal to −I is desirable
Distribution of
Random Transfer Interval
6 80 0.4
y1
y2
Random Transfer Interval dt (s)
4 70 0.35
Number of Occurrences
y1 & y2
2 60 0.3
0 50 0.25
0 5 (a) 10 15
40 0.2
2 y1 30 0.15
y2
1.5
y1 & y2
20 0.1
1
10 0.05
0.5
0 0 0.02
0 5 10 15 0 0.1 0.2 0.3 0.4 0 50 100 150 200
Time (s) Time (s) Index Number
(b)
Figure 6. Comparison of different protocols under one run (with batch reactor as plant) with associated distribution of the random
transfer interval. (a) Step response with token-passing protocol (τ = .08 s); (b) step response with TOD protocol (τ = .08 s).
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0.035 0.04 0.045 0.05 0.055 0.06 0.065 0.07 0.075
τ(s)
then the networked control system is globally exponentially Batch Reactor Experiment
stable. The unstable batch reactor (see [24, p. 62]) is a coupled
Proof Sketch: Initially no assumption can be made about two-input, two-output NCS. Based on the linearized process
the initial conditions of the error e( t ). Using the initial condi- model (see (5)), a proportional-plus-integral controller (see
tion z ( t 0 ) and the Lipschitz constants of the system equa- (6)) is designed in advance to stabilize the feedback system
tions, a growth bound β can be calculated for the first pτ and achieve good performance.
seconds from the bound on the growth of z ( t 0 ). At transmis-
sion times, the error e( t ) can only decrease, so it may be
shown thatβ applies to the entire interval[t 0 , t 0 + pτ]. At time 138. −0.2077 6.715 −5.676 0 0
5.679 0
t 0 + pτ the lemmas can be employed to compute a smaller −0.5814 −4.29 0 0.675
&
x= x+ u,
bound on e( t ). Once it is verified that this bound holds be- 1067
. 4.273 −6.654 5.893 1136
. −3146
.
yond the instant t 0 + pτ, a Lyapunov argument shows that, 1136 0
0.048 4.273 1343
. −2104
. .
for a t 1 > t 0 + pτ, z ( t 1 ) < ρ z ( t 0 ) for a ρ ∈[0 ,1). At time t 1 , the (4)
argument may be repeated, generating a smaller β and
hence bound on e( t ). By induction one concludes 1 0 1 −1
z ( mt 1 ) < ρ m z ( t 0 ) . y= x,
0 1 0 0 (5)
Simulations and Experiments
Two example systems, an unstable batch reactor and a dryer, 2s + 2
are used to support the analysis above and verify the reason- 0 s .
K (s) =
ableness of the assumptions. The basic configuration is −5 s − 8
0
shown in Fig. 5. The controller is located near the actuator, s (6)
10 8
u τ=0.08
9 y^ τ=0.15
7
8
6
7
5
6
u & y^
4
y^
4 3
3
2
2
1
1
0 0
0 2 4 6 8 10 12 14 16 18 20 0 2 4 6 8 10 12 14 16 18 20
Time (s) Time (s)
(a) (b)
Figure 9. Step response of the PT326 under a constant transfer interval of Ts = 80 ms and under a random transfer interval with plant
delay estimated at 400 ms using a CAN network. (a) PT326 with fixed transfer interval Ts = 0.08 and estimated delay = 4Ts ; (b) random
transfer interval and estimated delay = 5Ts .
cost, maintenance, and reliability. ator, and generate a control signal that
determines the input u.
The identified plant model with
throttle setting at 40% and sampling
One run of the system under the MEF-TOD protocol and period 80 ms is shown in the following:
the token-passing protocol is shown in Fig. 6 with the corre-
sponding random transfer interval distribution. With the 0.03837 + 0.05656 z −1 −4
Gp ( z ) = z .
same random interval distribution, the system with the . z −1 + 0.4516 z −2
1 − 137 (7)
MEF-TOD protocol performs better than the system with
the token-passing protocol. Since the plant has a transport delay of about 320 ms, the
The results of the batch reactor experiment over many control law was designed using the Smith predictor com-
runs are presented in Fig. 7. With the same confidence level bined with a PI controller. We treat the network as transpar-
(probability of passing the performance specifications set ent when designing the control law. For comparison, the
at 90% for a fixed average transfer interval time), the maxi- results of a step response test of the PT326 under constant
mum average transfer intervals for the system with the to- transfer intervals (control transmissions with fixed network
ken-passing protocol and with the MEF-TOD protocol are 43 access delays) and random transfer intervals (control
ms and 54 ms, respectively. transmissions with variable network access delays) are il-
lustrated in Fig. 9. For the constant transfer interval, the test
reveals the transport delay to be about 300 ms, similar to
the estimated delay. Under random transfer intervals, the
15
Thermister Output (V)
0.4 lay of 400 ms, the approximate sum of the network and plant
0.3 transport delay.
0.2 The results of the dryer experiment validate our former
0.1 theoretical analysis and assumptions. They also suggest
0
that including the average transport delay into the control-
0 5 10 15 20 25 30 ler design may improve controller performance. The sam-
Time (s) pled mean of the random transfer interval can be added to
(b) the estimated plant delay in the predictor; however, this
Figure 10. Experimental results using an IEEE 802.11 network. method is limited to small average delay intervals and varia-
The upper graph shows the commanded and reported step responses tions. Fig. 10 shows the step responses of two dryer plants
of both dryer plants with plant 2 signals shifted by 7 V. The lower sharing a wireless IEEE 802.11 network modified for net-
graph details the relevant network statistics during the course of the worked control.
run. (a) Plant commands and reported outputs, dryers 1 and 2; (b)
network transfer inrervals, mean, and standard deviation.