IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 10, NO.

2, MARCH 2002

297

Network Design Consideration for Distributed Control Systems

Feng-Li Lian, James Moyne, and Dawn Tilbury

AbstractThis paper discusses the impact of network architecture on control performance in a class of distributed control systems called networked control systems (NCSs) and provides design
considerations related to control quality of performance as well as
network quality of service. The integrated network-control system
changes the characteristics of time delays between application devices. This study first identifies several key components of the time
delay through an analysis of network protocols and control dynamics. The analysis of network and control parameters is used to
determine an acceptable working range of sampling periods in an
NCS. A network-control simulator and an experimental networked
machine tool have been developed to help validate and demonstrate
the performance analysis results and identify the special performance characteristics in an NCS. These performance characteristics are useful guidelines for choosing the network and control
parameters when designing an NCS.
Index TermsDistributed control systems, networked control
systems (NCSs), network transmission delay, processing time
delay, sampling time selection.
Fig. 1. Performance comparison of continuous control, digital control, and
networked control cases.

I. INTRODUCTION

HE trend of modern industrial and commercial systems

is to integrate computing, communication and control
into different levels of factory operations and information processes [1]. For example, an integrated manufacturing system
might include computer-numerically controlled machines,
computer-aided design tools, supervisory controllers, and
intelligent monitoring devices. The introduction of control
network bus architectures can improve the efficiency, flexibility, and reliability of these integrated applications, reducing
installation, reconfiguration, and maintenance time and costs.
The change of communication architecture from point-to-point
to common-bus, however, introduces different forms of time
delay uncertainty between sensors, actuators, and controllers.
These time delays come from the time sharing of communication medium as well as additional functionality required for
physical signal coding and communication processing. The
characteristics of time delays could be constant, bounded, or
even random, depending on the network protocols adopted and
the chosen hardware. This type of time delay could potentially
degrade a systems performance and possibly cause system
instability.
Manuscript received November 30, 2000. Manuscript received in final form
October 26, 2001. Recommended by Associate Editor A. Ray. This work was
supported in part by the NSF Engineering Research Center for Reconfigurable
Machining Systems under Grant EEC95-92125.
F.-L. Lian and D. Tilbury are with the Department of Mechanical Engineering, University of Michigan, Ann Arbor, MI 48109-2125 USA (e-mail:
[email protected]; [email protected]).
J. Moyne is with the Department of Electrical Engineering and Computer
Science, University of Michigan, Ann Arbor, MI 48109-2122 USA (e-mail:
[email protected]).
Publisher Item Identifier S 1063-6536(02)01770-0.

Research in this class of systems, called networked control

systems (NCSs), has focused on two areas: communication
protocols and controller design. A proper message transmission
protocol is necessary to guarantee the network quality of
service (QoS), whereas advanced controller design is desirable
to guarantee the control quality of performance (QoP). There
are a wide variety of different commercially available control
networks such as ControlNet, DeviceNet, Ethernet, Profibus,
Sercos, WorldFIP, etc., for the implementation of a NCS
architecture. During the NCS design, a performance chart
can be derived as shown in Fig. 1. This performance chart
provides a clear way to choose the proper sampling period for
an NCS. Fig. 1 is the comparison of control performance versus
sampling period for continuous control, digital control and networked control. The worst, unacceptable, acceptable, and best
regions can be defined based on control system specifications
such as overshoot, steady-state error, and/or phase margin. The
performance axis in Fig. 1 could be chosen to reflect a subset
of these metrics. Since the performance of continuous control
is not a function of sampling period, the performance index is
constant for a fixed control law. For the digital control case, the
performance only depends on the sampling period assuming
no other uncertainties. The performance degradation point
of sampling period
in digital control could be estimated
based on the relationship between control system bandwidth
and sampling rate. For the networked control case, point
can be determined by further investigating the characteristics
and statistics of network-induced delays and device processing
time delays. As the sampling period gets smaller, the network
traffic load becomes heavier, the possibility of more contention

1063-6536/02$17.00 2002 IEEE

298

IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 10, NO. 2, MARCH 2002

time or data loss increases in a bandwidth-limited network and

longer time delays result. This situation causes the existence of
point in an NCS. The analysis in this paper will show how
these points can be computed.
The goals of this paper are to study the key components of
time delays and to provide guidelines for obtaining the optimal
working range of sampling times. The characteristics of time delays are analyzed as a combination of data processing and message transmission delays; data collected from actual traffic on an
industrial control network validates the analysis. The data analysis results provide several time delay models for performance
evaluation purposes and the usage of control applications. Based
on digital control theory, the impact of sampling period selection is then analyzed. A network-control simulator that takes as
input actual device performance profile data, is used to simulate
the network traffic with different network protocols, evaluate
the control performance of an NCS and validate the phenomena
found in the experimental study.
This paper is organized as follows. Section II provides
detailed background and related work associated with this
study. Section III addresses the key components of time delay
between two devices and provides several scenarios to identify
these components. Section IV presents the network and control
performance analysis in terms of network QoS and control QoP.
Section V details the simulation and experimental case study of a
networked machine tool. This paper concludes with Section VI.
II. BACKGROUND
There are now a large number of networks available for applications at the information level as well as at the control level.
The goal of a control network is to provide a guaranteed quality
of service such as deterministic time delays and maximum
throughput for real-time control applications. These networks
target various types of industrial automation and processing
applications and are distinguished through static parameters
such as data rate, message size, medium length, supported
topology, number of nodes, and dynamic parameters such as
medium access control mechanism, message connection type,
interoperability, and interchangeability. The detailed comparison of these control networks in terms of the above-mentioned
parameters can be found in [2][4].
Much of the research on the network architecture in control applications has been focused on the network QoS of existing network solutions and then on modifications of the message scheduling algorithm or medium access control to guarantee a certain level of network QoS. The network QoS includes
predicting time delays, improving throughput, utilization, and
efficiency, minimizing lost data rate and calculating message
schedulability. Comparisons of network QoS metrics such as
network utilization and efficiency versus static network parameters such as message size and data rate were considered in [6]
for several popular control networks. The results provided a first
understanding of the NCS time delay without considering other
network dynamic parameters which may cause time delay variance. Data latency, or time delay, was analyzed by Ray and
Halevi on the medium access control (MAC) mechanism for
centralized controlled access, distributed controlled access and

random access [7], [8]. Statistical models of network-induced

delays of sensor-controller and controller-actuator links were
also studied in terms of message size, period, and station response time. This statistical information provided a foundation
for further analysis of control system stability and performance
in a stochastic process setting.
When the time delays do not meet the control system requirements, several traffic smoothers, for example, were
adopted to adjust the message generation rate [9]. These traffic
smoothers were installed between the UDP or TCP/IP layer and
the MAC layer and gave higher priority to real-time packets over
nonreal-time ones. The adaptive traffic smoother harmonically
increased and multiplicatively decreased the input limit of
real-time or nonreal-time packets to avoid network congestion
and improve throughput. Other approaches such as assigning
different data sampling periods for different control loops and
redesigning network topology into a multihop network were
studied by Hong [10] and Tovar et al. [11], respectively. The
sampling period of every individual control loop was determined
by the limitation of maximum loop delays and the availability of
network bandwidth, so as to meet both the control stability condition and the network schedulability condition [10]. Multihop
topology, on the other hand, provided one quick solution for
maintaining the timing requirement [11]. However, additional
hardware and software were needed to implement the message
transmission between hops, increasing the system cost.
When encountering the degradation of control QoP of an
NCS, most studies assumed that, given the information of
network-induced delays, robust or optimal controllers could be
designed using discrete-time or continuous-time models. Ray et
al. used an augmented state to model the network-induced delays
within one sampling period and provided design considerations
on the stability issues of different time skews and sampling rates
in [12] and [13]. Moreover, a stochastic regulator solution that
uses dynamic programming and the optimality principle was
studied in the presence of randomly time-varying delays [14]. On
the other hand, Krtolica et al. expanded the number of states to
include the delayed signals, assuming the time delays are equal
to multiples of one sampling period [15]. For continuous-time
models, studies focused on finding the maximum allowable
delay bound based on a Lyapunov stability approach [16]. Based
on this bound, they also provided some scheduling algorithms or
protocols to guarantee this requirement.
III. TIMING COMPONENTS
The important time delays that should be considered in a
distributed control system analysis are the sensor to controller
and controller to actuator end-to-end delays. In an NCS, message
transmission delay can be broken into two parts: device delay
and network delay. The device delay1 includes the time delays at
the source and destination nodes. The time delay at the source
and the waiting
node includes the preprocessing time,
. The time delay at the destination node is only the
time,
. The network time delay includes the
postprocessing time,
1Each device requires processing time for both network communication and
data acquisition. There may be one processor performing both function or multiple processors. In this paper, we define the total elapsed time required as the
pre- or post-processing time.

IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 10, NO. 2, MARCH 2002

299

Fig. 2. A timing diagram showing time spent sending a message from a source node to a destination node.

Fig. 3. Waiting time diagram.

total transmission time of a message and the propagation delay

of the network. The total time delay can be expressed by the
following equation:
(1)
The key components of each time delay are shown in Fig. 2 and
will be discussed in the following sections.
A. Preprocessing Time at Source Nodes
The preprocessing time at the source node is the time needed
to acquire data from the external environment and encode it into
the appropriate network data format. This time depends on the
device software and hardware characteristics. In many cases,
it may be assumed that the preprocessing time is constant or
negligible. However, we will show in Section III-E that this assumption is not true in general; in fact, there may be noticeable
differences in processing time characteristics between similar
devices and these delays may be significant. We will show how
experimentally-identified models of the processing time can be
used to characterize this delay.
B. Waiting Time at Source Nodes
A message may spend time waiting in the queue at the senders
buffer and could be blocked from transmitting by other messages
on the network. Depending on the amount of data the source node
must send and the traffic on the network, the waiting time may be
significant. The main factors affecting waiting time are network
protocol, message connection type and network traffic load.2 For
example, consider a strobe message connection in Fig. 3. If Slave
2In a practical control application, most message transmissions are periodic
and only the most recent data are used. Hence, in this paper, we do not consider
the case where the sender transmits stale messages.

Fig. 4. Nine identical devices with strobed message connection.

1 is sending a message, the other eight devices must wait until the
network medium is free. In a CAN3 -based DeviceNet4 network,
it can be expected that Slave 9 will encounter the most waiting
time because it has a lower priority on this priority based network.
However, in any network, there will be a nontrivial waiting time
after a strobe, depending on the number of devices that will respond to the strobe.
Fig. 4 shows experimental data of the waiting time of nine
identical devices on a DeviceNet network. These devices have
a very low variance of processing time. We collected 200 pairs
of messages (request and response). Each symbol denotes the
mean and the distance between the upper and lower bars equals
3CAN stands for controller area network which is a bit-synchronized control
network that utilizes a nondestructive collision resolution scheme through message priority specified in the message arbitration field.
4DeviceNet implements the physical layer and data link layer of a standard
CAN protocol and defines its own node and message priority. Specifically,
DeviceNet further utilizes the message arbitration field to define different
classes of message and node priorities. The waiting time in a DeviceNet network
is a function of the message or node number, i.e., priority.

300

IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 10, NO. 2, MARCH 2002

Fig. 5. Processing time histogram of six typical DeviceNet devices.

two standard deviations. If these bars are over the limit (maximum or minimum), then the value of limit is used instead. It
can be seen in Fig. 4 that the average waiting time is proportional to the node number (i.e., priority). Also, the first few devices have a larger variance than the others, because the variance
of processing time occasionally allows a lower-priority device
to access the idle network before a higher-priority one.
C. Transmission Time on Network Channel
The transmission time is the most deterministic parameter
in a network system because it only depends on the data rate,
the message size, and the distance between two nodes. The formula for transmission time can be described as follows:
, where is the message size in terms of bits,
is the bit time and
is the propagation time between
any two devices. Since the typical transmission speed in a communication medium is 2 10 m/s, the propagation time
is negligible in a small scale (100 m or shorter) control network.
D. Postprocessing Time at Destination Nodes
The postprocessing time at the destination node is the time
taken to decode the network data into the physical data format
and output to the external environment. Some experimental
analysis of postprocessing time along with preprocessing times
will be discussed in the following section.
E. Experimental Investigation of Timing Components
In practical applications, it is very difficult to identify each
individual timing component discussed above. Instead, by mon-

itoring the time-stamped traffic of the request-response messaging on a DeviceNet network, we will show the characteristics
of processing times, i.e., the sum of preprocessing and postprocessing times of one device.
In the experimental setup, there is only one master and one
slave connected to the network and the master continuously
polls this slave. Refer to Fig. 2 and let Node A be the master
and Node B be the slave. Here, there is no other network traffic
other than the request-response messages between the master
and the request-response frequency
and slave, i.e.,
is set low enough such that no messages are queued up at the
sender buffer. By monitoring the message traffic on the network medium and time-stamping each message, we can further calculate the processing time of each request-response, i.e.,
, after subtracting the transmission time.
Fig. 5 shows the histogram of 400 samples of six typical
DeviceNet device processing times. The (right) solid and (left)
dashed lines are the maximum and minimum values of processing times, respectively. The histogram plots indicate the
nondeterministic processing times of different network devices
and their variance. Devices 1 and 5 have similar functionality of
discrete inputs/outputs, but different numbers of inputoutput
modules. Device 5 provides several augmentable modules and
hence has more processing units and computation load. Device
1, on the other hand, has only one unit. Devices 2, 3, and 4 have
fairly consistent processing times, i.e., low variance. Note that
the smallest time that can be recorded is 1 s. The uniform distribution of processing time at Device 6 is due to the fact that
it has an internal sampling time which is mismatched with the

IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 10, NO. 2, MARCH 2002

301

ability. Here we classify Devices 1, 2, 3, and 5 as model (a),

Device 6 as model (b), and Device 4 as model (c).
IV. NETWORK AND CONTROL PERFORMANCE ANALYSIS
Networks are used to transmit data among control system devices. Hence, the network QoS should be analyzed before implementing control systems with network architectures. On the
other hand, the control QoP should be specified to help evaluate
the control system performance. The main evaluation measures
of the network QoS are time delay statistics, network efficiency,
network utilization, and the number of lost or unsent messages
and have been studied in [6]. These measures are used to determine the capability of network medium and provide information
to specify control parameters such as sampling period.

(a)

A. Control Quality of Performance

In this section we discuss several performance criteria such as
IAE/ITAE and control system specifications and their relation
to sampling periods. Simulation results relating different time
delays and system dynamics are also provided in this section to
help provide an understanding of the impact of time delay on
control performance.
Two criteria, IAE and ITAE, are generally used to evaluate
control system design and performance. IAE is the integral of
the absolute value of the error and ITAE is the integral of the
time multiplied by the absolute value of the error [18]. Their
mathematical formulas are as follows:

(b)

(c)

or

IAE

and

Fig. 6. Three models of processing time.

request frequency. Hence, the processing time recorded here is

the sum of the actual processing time and the waiting time inside
the device. Device 6 also provides more complex functionality
and has a longer processing time than the others.
In an effort to classify device processing time, we propose
three device processing time models as shown in Fig. 6, based on
the histograms shown in Fig. 5. The classification of device processing times is based on a heuristic grouping of experimental
data. The first model is similar to the network delay model in
s, , and , we can
[17]. By using these parameters, i.e., s,
specify the probability distribution of the processing time. In
model (a), and are the minimum and maximum processing
is the processing times with higher probability.
times and
and are the probability of the two portions. That is

Prob

(2)

. Model (b) is the uniform distriwhere

bution and can expressed by the following form:

or

ITAE

and
are the initial and final times of the
where
evaluation period in continuous (discrete) time and is the error
between the actual and reference trajectories. ITAE weights
later errors heavier and discounts the transient response,
whereas IAE weights all errors equally.
When sensors, actuators and controllers are interconnected by
one common-bus network, all devices need to share the transmission medium. In addition, application signals are discretized.
Hence, it is natural to analyze these types of systems using a
digital control approach. In order to guarantee system stability
and control performance, two control measures can be used to
determine the best sampling period: phase margin and control
system bandwidth.
The phase margin is the amount by which the phase of an
open-loop system exceeds 180 when the magnitude equals
one. The primary effect of the sampling time delay is additional
and time
phase lag. The phase lags due to discretization,
, are summarized as follows [19], [20]:
delay,
and

Prob

(3)

. Model (c) is the general case; the mulwhere

tiple majority processing times are assumed to have same prob-

(4)

(5)

is the sampling period and

where is the system frequency
is the time delay.
is defined as the maxThe control system bandwidth
imum frequency at which the output of a system will track an

302

IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 10, NO. 2, MARCH 2002

same sampling period. In addition, the case with uniformly distributed time delays shows a further degradation of system performance because the time-delay variation causes uncertainty at
actuators actions.
Fig. 7 demonstrates how point in Fig. 1 can be determined
using (5)(6). If the statistics of the additional time delay are
can be determined as follows. Using (5), the
known, point
total phase lags due to digital control without time delay,
and digital control with time delay
can be further expressed
as follows:
and
(8)

Fig. 7.

Impact of time delay mean and variance on control performance.

input sinusoid in a satisfactory manner. In order to guarantee the

control QoP, the rule of thumb for selecting sampling periods
)
in digital control is that the desired sampling multiple (
is [19]:
(6)

Suppose both digital control and digital control with delay have
. Hence, we have
same phase lags, i.e.,
or
. Furthermore, (6) can be used
in terms of control system bandwidth, i.e.,
to estimate
. So, the sampling period of point can be described as
follows:
(9)
Therefore, using (9) and assuming a 2ms delay,
for the system discussed is estimated as follows:
ms.
B. Network Bandwidth Versus Control Performance

In order to visualize the impact of sampling and time delays on control QoP, a second-order system is considered. Its
open-loop transfer function is
(7)
rad/s (asThe phase margin of this system is 40 at
suming unity feedback). Hence, the maximum sampling period,
ms. If considering an additional time
based on (5), is
delay of 2 ms, the phase lag due to time delay is
using (5). This additional time delay will further reduce the maximum sampling period.
Fig. 7 shows the simulation study of the impact of sampling
effect and time delay on control QoP. A closed-loop system with
control system bandwidth of 2.5 Hz is studied and the results are
shown in Fig. 7. The horizontal axis is sampling period and the
vertical axis is ITAE. The (red) solid lines with are the result of only considering the sampling effect. The (blue) dashed
lines with are the result of considering the sampling effect
and a constant time delay of 2 ms. The (green) dashdot lines
with are the result of the sampling effect and a uniformly
distributed time delay between 0 and 4 ms. These delays are
experienced at both the sensing and actuation nodes. The result of random delay case is the average value of three identical
simulation runs. Based on the sampling rule of thumb in (6),
the suggested maximum sampling period is 20 ms. For the case
without any additional time delay, i.e., only the sampling effect,
the control performance gets worse when the sampling periods
are larger than the maximum values. When additional time delays are also considered, the point of degradation moves to the
right. That is, the control performance becomes worse with the

For smaller sampling periods, the network traffic is saturated

and longer time delays will adversely impact the control performance. The result of smaller sampling periods and high network
traffic load describes the location of point in Fig. 1. Therecan be estimated by the total transmission time
fore,
of all cyclic messages in a control application. However, using
means that the network is saturated. Hence, a better
estimation suggested by the sufficient schedulability condition
in [25] is as follows:
(10)
where 0.69 is the maximum ratio of utilization to meet the sufficient schedulability condition for infinite messages. Furtherequals
for strobe or
for poll,
more,
is the transmission
where is the number of devices and
time of each message. When considering the device processing
, will increase and can
time, the minimum sampling period,
be further modified as follows:
(11)
is the maximum of all device processing time for
where
strobe, or the sum of all device processing times for poll.
C. Sampling Period Selection
Due to the interaction of the network and control requirements, the selection of the best sampling period is a
compromise. Based on the previous analyzes, smaller sampling periods guarantee a better control QoP. However, in a
bandwidth-limited control network, smaller sampling periods
introduce high-frequency communications and may degrade

IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 10, NO. 2, MARCH 2002

Fig. 8.

303

Simulation model of the networked Robotool in Matlab/Simulink.

the network QoS. The degradation of network QoS could

further worsen the control QoP due to longer time delays
when network is near saturation. These issues will be further
illustrated in the following section by the simulation and
experimental study of a networked machine tool.
V. CASE STUDY: A NETWORKED MACHINE TOOL
In this section, we provide a case study of network configuration on a machine tool. We investigate the interaction between
network and control system parameters through simulation and
experiment.
A. System Setup of the Machine Tool
The machine tool studied has three axes: , , and . Each
axis moves on a linear slide and is driven through a ball screw
by a dc motor with a tachometer, which provides an angular
velocity measurement. The dc motor is driven by a PWM drive
with control input between zero and 255. Each axis also has
a linear encoder that provides linear position measurement.
Therefore, both position and velocity feedback are available.
The three axes operate independently. The mathematical model
of each axis between the PWM input ( ) and the position
output ( ) is described by a second-order linear transfer
function:
(12)
The time constants (s) for each axis are 0.055 ( ), 0.056 ( )
and 0.040 ( ) and the overall gains ((mm/s)/PWM) are 8.346
( ), 8.956 ( ) and 1.606 ( ), respectively. In the past, several
authors used the machine tool as a testbed for research on open
architecture control systems [21], supervisory control [22] and
cross-coupling control [23]. All previous work used a traditional
point-to-point communication architecture. We have installed a
DeviceNet network system on the machine tool to study the advantages of a distributed architecture and the impact of network
induced delays on control performance.

B. Simulation Model
A network-control simulation model has been implemented
in Matlab/Simulink as shown in Fig. 8. The simulation model
has two parts: a network simulator and a control simulator.
By specifying the MAC mechanism of different protocols and
network parameters such as node numbers, data rates, data
sizes and message periods, the network simulator produces
the network induced delays between different nodes with
experimentally determined device performance profiles and
provides statistics of these time delays and related network QoS
evaluation of the network configuration. Three types of network
protocols, Ethernet, ControlNet (token-passing) and DeviceNet
(CAN-based) are implemented in the network simulator.5 The
time delay data can be fed into the control simulator to simulate
the dynamic response of the closed-loop system with network
induced delays. A PID controller is used on each axis of the
machine tool; gains are calculated based on a discrete-time
model of the system and a set of desired closed-loop poles.
The control simulator also provides the control QoP in terms
of ITAE and IAE. The detailed description of this simulation
software package can be found in [24].
C. Simulation Study With Network Delay
In this section, we consider the cases with network-induced
delays only. Device processing time delay will be discussed in
the next section. In this study we consider two fundamental message connections: strobe and poll. In a strobe connection, all
sensors are asked for new information at the same time and the
response messages are received by the controller one at a time
based on the different network protocols. The controller then
calculates the actuation command based on the control law and
sends the command to the actuators through the network. For
the case of a poll connection, sensors respond with new information after they have received poll requests. Because sensing
messages arrive at the controller at different time instants, it is
possible for the controller to update the actuation command immediately to obtain better control performance. However, this
5The detailed description and comparison of Ethernet, ControlNet, and
DeviceNet can be found in [6]

304

IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 10, NO. 2, MARCH 2002

high-frequency updating will generate more network traffic and

further increase the time delay. Hence, we only consider the case
of a single actuation during each sampling period. Also, if the
network is overloaded, the sending nodes only transmit the most
current data when the network is available, i.e., stale backlogged
data are discarded.
The reference trajectory considered in this simulation study
is
for the and axes, where
is the unit step function. The
reference trajectory of the axis is equal to half of that of the
and axes since the axis has a shorter working range in
the experimental setup. ITAE is used to evaluate the control performance. The ITAE results, shown in Fig. 9, illustrate the simulated control performance of each axis of the networked machine tool using Ethernet (at 10 Mb/s), ControlNet (at 5 Mb/s),
and DeviceNet (at 500 Kb/s), respectively, and strobe and poll
message connections. As expected from classical digital control
theory, the ITAE value becomes smaller as the sampling rate increases and the system performance becomes worse when time
delays are longer.
When the time delay varies due to network characteristics,
the ITAE values first decrease and then increase as the sampling
rate increases. The decrease occurs when the network has a relatively low traffic load or is unsaturated; thus, the performance
is similar to the constant time delay cases. However, as the sampling rate increases, a large number of messages (i.e., higher
frequency messages) are contending for the network medium
and the time delay of each message becomes larger when total
number of messages that need to be sent requires more network
bandwidth than available from the transmission medium. Therefore, the ITAE values increase above a critical period. These
critical periods are a function of the transmission time to finish
transmitting all messages. This is the occurrence of point in
Fig. 1. Because the node numbering has little effect in Ethernet
and ControlNet, the performance of the three axes are similar in
Fig. 9(a) and (b). However, for DeviceNet in Fig. 9(c), the node
numbering has a significant effect because nodes are prioritized.
In this simulation we set the -axis nodes as high priority and
the -axis nodes as low priority, thus the performance of the
-axis subsystem worsens first.
Although we only show the comparison of three different network protocols, the simulation tool can be programmed for other
network solutions as well. The type of simulation study is useful
for the first stage of NCS design where the physical network
and control system are not available yet. Hence, by simulating
network and control models, an NCS designer can compare the
performance and decide on the best network protocol for this
specific control application. In addition to the network transmission delays, device processing times can also be included in
the simulation tool [24].
D. Simulation Study With Network and Device Delays
The simulation study in previous subsection provides information on how the delays associated with different network protocols impact control performance. However, in practical applications, device delays such as software and hardware processing
times are also important as discussed in Section III. In this section, a set of timestamped network traffic data is collected from

(a)

(b)

(c)
Fig. 9. Control performance (ITAE) versus message period of three control
networks and two messaging connections. (a) Ethernet. (b) ControlNet.
(c) DeviceNet.

a DeviceNet network. There is one controller node and nine

nodes of inputoutput modules which are used as the interface

IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 10, NO. 2, MARCH 2002

Fig. 10.

305

Simulation and experimental results of control QoP: centralized and networked cases. Top row: (a), (b), (c). Bottom row: (d), (e), (f).

between physical sensing/actuating devices and the DeviceNet

communication medium. By analyzing the timestamped network traffic, the actual time delays between the sensors/actuators and controllers can be obtained. These time delays can be
further fed into the network-control simulator.
Fig. 10(a)(c) show the simulation results with actual network data. In each plot, the horizontal axis is the sampling period (s) and the vertical axis is the performance index, i.e., trajectory error in terms of ITAE (ms). The (red) dashed lines with
are the results of digital sampling effect only and the (blue) solid
lines with are the results with actual network traffic data
which includes network and device delays. In this example, the
axis has a higher priority than the and axes. Hence, with
smaller sampling periods, the network traffic load is heavy and
some of the - or -axis messages do not get through. Therefore, the control performance of and axes degrades first.
Note that the and axes have similar dynamic characteristics and reference trajectories, but the axis is slightly different,
as described in (12). Fig. 10(a)(c) demonstrate the same time
delay effect discussed in Section IV and shown in Fig. 7. However, Fig. 10(a)(c) also indicate that as the sampling period decreases, the control performance gets worse when the network
load is near saturated. Therefore, in this example, at a sampling
period of about 7 ms the axis is still in control, but the and
axes are out of control.

The trend of the simulation result with actual network and

device delays here agrees with the DeviceNet case in the previous section, except for the exact sampling periods where the
system performance degrades. In the previous section, only network transmission time was considered. The simulation scenario here includes both network and device delays and the simulation result suggests that the device delays contribute more to
the performance degradation than the network delays. Hence,
in addition to using a high-speed network and a proper MAC
protocol for control applications, networked devices with small
processing delays must be chosen.
E. Experimental Study
Fig. 10(d)(f) illustrate the experimental results of the corresponding simulation results shown in Fig. 10(a)(c). The sampling periods considered range from five to 500 ms. The message transmission is controlled by the polling rate of the master
device in a DeviceNet network. Because of the variance in processing times, the sampling periods are not identical but have
a small variance. This further degrades the performance when
compared to the digital control case.
The experimental results in Fig. 10(d)(f) show that the
axis has wider feasible range of sampling periods than the
and axes because the
axis has the highest priority. When

306

IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 10, NO. 2, MARCH 2002

the network traffic load gets heavy, only the messages associated
with the -axis can get through. However, all three axes have
ms in
the same upper limit of sampling period, i.e.,
the example, since they have similar closed-loop dynamic characteristics. The feasible range of sampling periods for the ,
and axes are (7, 23), (13, 23), and (13, 23) ms, respectively.
The results are similar to the simulation results shown in the top
row of Fig. 10, although the performance is worse in the cases
of longer sampling periods.

VI. CONCLUSION
This paper addressed the design issues of network architecture in a type of distributed control system where sensors,
actuators, and controllers are interconnected by a common-bus
network. Design considerations include network parameters,
control parameters and networked control system performance
as the design guideline. Along with the theoretical analysis, a
numerical computer network-and-control simulation tool was
developed to assist in NCS design. The design procedures
are summarized as follows. At the first stage of a design
process (assuming actual network and control systems are not
available), an NCS designer can utilize the network and control
performance analysis in Section IV to investigate the feasibility
of this set of system parameters. Also incorporating the processing time model and other timing parameters described in
Section III, an integrated computer network-and-control simulator can be used to help understand the network availability
and control performance. Similar comparative analysis can be
applied to different network protocols and system dynamics.
Hence, the analysis result can be a guideline for choosing the
most suitable network system.
If a practical network is available at the next design stage, actual device processing times and network traffic can be collected
and used to simulate the control system model. At this stage,
different combinations of network parameters can be studied to
optimize the control performance. In order to guarantee the best
control performance, different controller designs can be tested
to verify their stability and performance.
During the design procedure, an NCS performance design
chart as shown in Fig. 1 can be derived. This paper investigated
the existence and location of the performance degradation points
with the aid of simulation and experimental results. The performance degradation point in digital control can be estimated
by using the relationship between control system bandwidth and
sampling rate. For the networked control case, point can be
determined by further investigating the characteristics and statistics of network-induced delays along with other processing
time delays. Messages with smaller sampling periods also generate high network traffic load. The high network traffic load
could increase the possibility of data loss or the waiting time
for message contention and induce longer time delays. Therefore, network saturation results in the existence of point in
networked control.
The improvement of NCS performance can be divided into
two areas. One is the minimization of device processing time
and the improvement of network protocols to further guarantee

the determinism of transmission time as well as reduce the

end-to-end delays. The designer should then pick a sampling
period between points
and
based on other necessary
design constraints. The other direction for improvement is
advanced optimal or robust controller design which can overcome the uncertainty in an NCS and achieve the best control
performance. These improvements can make the performance
curve of networked control move toward that of digital control.
Our future work will focus on the discrete-time modeling of
NCSs which will utilize the network and control parameters
identified in this paper. Based on the proposed model, an
advanced controller design that consists of both control and
network parameters will be implemented to improve the overall
NCS performance.
REFERENCES
[1] H. A. Abdullah and C. R. Chatwin, Distributed C environment for
small to medium-sized enterprises, Integrated Manufacturing Syst.,
vol. 5, no. 3, pp. 2028, 1994.
[2] P. Pleinevaux and J.-D. Decotignie, Time critical communication networks: Field buses, IEEE Network, vol. 2, no. 3, pp. 5563, May 1988.
[3] R. S. Raji, Smart networks for control, IEEE Spectrum, vol. 31, no. 6,
pp. 4955, June 1994.
[4] G. Schickhuber and O. McCarthy, Distributed fieldbus and control network systems, Comput. Contr. Eng., vol. 8, no. 1, pp. 2132, Feb. 1997.
[5] G. Cena, C. Demartini, and A. Valenzano, On the performances of
two popular fieldbuses, in Proc. IEEE Int. Workshop Factory Commun.
Syst., Oct. 1997, pp. 177186.
[6] F.-L. Lian, J. R. Moyne, and D. M. Tilbury, Performance evaluation
of control networks: Ethernet, ControlNet and DeviceNet, IEEE Contr.
Syst. Mag., vol. 21, no. 1, pp. 6683, Feb. 2001.
[7] A. Ray, Performance evaluation of medium access control protocols for
distributed digital avionics, ASME J. Dyn. Syst., Measurement, Contr.,
vol. 109, no. 4, pp. 370377, Dec. 1987.
[8] Y. Halevi and A. Ray, Performance analysis of integrated communication and control system networks, ASME J. Dyn. Syst., Measurement,
Contr., vol. 112, no. 3, pp. 365371, Sept. 1990.
[9] S.-K. Kweon, K. G. Shin, and G. Workman, Achieving real-time
communication over ethernet with adaptive traffic smoothing, in Proc.
IEEE Real-Time Tech. Appl. Symp., June 1999, pp. 90100.
[10] S. H. Hong, Scheduling algorithm of data sampling times in the integrated communication and control systems, IEEE Trans. Contr. Syst.
Technol., vol. 3, pp. 225230, June 1995.
[11] E. Tovar, F. Vasques, and A. Burns, Supporting real-time distributed
computer-controlled systems with multihop P-NET networks, Contr.
Eng. Practice, vol. 7, no. 8, pp. 10151025, Aug. 1999.
[12] Y. Halevi and A. Ray, Integrated communication and control systems:
Part Ianalysis, ASME J. Dyn. Syst., Measurement, Contr., vol. 110,
no. 4, pp. 367373, Dec. 1988.
[13] A. Ray and Y. Halevi, Integrated communication and control systems:
Part IIdesign considerations, ASME J. Dyn. Syst., Measurement,
Contr., vol. 110, no. 4, pp. 374381, Dec. 1988.
[14] L.-W. Liou and A. Ray, A stochastic regulator for integrated communication and control systems: Part Iformulation of control law, ASME
J. Dyn. Syst., Measurement, Contr., vol. 113, no. 4, pp. 604611, Dec.
1991.
[15] R. Krtolica, U. Ozguner, H. Chan, H. Goktas, J. Winkelman, and M.
Liubakka, Stability of linear feedback systems with random communication delays, Int. J. Contr., vol. 59, no. 4, pp. 925953, Oct. 1994.
[16] G. C. Walsh, H. Ye, and L. Bushnell, Stability analysis of networked control systems, in Proc. Amer. Contr. Conf., June 1999, pp.
28762889.
[17] J. Nilsson, Real-Time Control Systems With Delays, Ph.D. dissertation, Lund Inst. Technol., Lund, Sweden, 1998.
[18] G. F. Franklin, J. D. Powell, and A. Emami-Naeini, Feedback Control of
Dynamic Systems, 3rd ed. Reading, MA: Addison-Wesley, 1994.
[19] G. F. Franklin, J. D. Powell, and M. L. Workman, Digital Control of
Dynamic Systems, 3rd ed. Reading, MA: Addison-Wesley, 1998.

IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 10, NO. 2, MARCH 2002

[20] M. Malek-Zavarei and M. Jamshidi, Time-Delay Systems: Analysis,

Optimization and Applications. Amsterdam, The Netherlands:
North-Holland, 1987.
[21] Y. Koren, Z. J. Pasek, A. G. Ulsoy, and U. Benchetrit, Real-time open
control architectures and system performance, CIRP Ann., vol. 45, no.
1, pp. 377380, 1996.
[22] R. G. Landers and A. G. Ulsoy, Supervisory machining control: Design
approach and experiments, CIRP Ann., vol. 47, no. 1, pp. 301306,
Aug. 1998.

307

[23] B.-C. Chen, D. M. Tilbury, and A. G. Ulsoy, Modular control for

machine tools: Cross-coupling control with friction compensation, in
Proc. ASME DSC, vol. DSC-64, Nov. 1998, pp. 455462.
[24] F.-L. Lian, J. R. Moyne, and D. M. Tilbury, Networked Control Systems Toolkit: A Simulation Package for Analysis and Design of Control
Systems With Network Communications,, Tech. Rep. UM-ME-01-04,
2001.
[25] C. M. Krishna and K. G. Shin, Real-Time Systems. New York: McGraw-Hill, 1997.