International Conference on Control, Automation and Systems 2008

Oct. 14-17, 2008 in COEX, Seoul, Korea

Applicability of ZigBee for Real-Time Networked Motor Control Systems

Ulugbek R. Umirov1, Seong-Hyun Jeong and Jung-Il Park2
1
Department of Electronic Engineering, Yeungnam University, Gyeongsan, Korea
(Tel : +82-53-810-3927; E-mail: [email protected])
2
(Tel : +82-53-810-2498; E-mail: [email protected])

Abstract: This paper discusses networked real-time control systems. The common network and ZigBee specific
problems are discussed and methods to overcome them are explained. Limitations of ZigBee networks, sources of delay
and benefits of broadcast mode over unicast mode for control loop time delay minimization are reviewed. It is
explained how play-back buffer, originally used in multimedia play-back, can help to eliminate variance of loop time
delay. To cope with achieved constant loop time delay the Smith predictor is used. Performance of control loop utilizing
conventional PID controller, Smith predictor and play-back buffer is verified via experiments and proved to be efficient.

Keywords: networked control, ZigBee, play-back buffer.

1. INTRODUCTION sensors are always assumed to be time-driven, that is,

the plant output is sampled periodically. Second,
A networked control system (NCS) is a feedback actuators are event-driven, and therefore apply control
control system where information from the sensors and signals as soon as they are received. The actuator also
the controllers is sent over an electronic communication holds the last received control signal until a new one is
network [1], as shown in Fig. 1. NCSs offer reduced received. Finally, the controller is also event-driven,
cost and relatively simple implementation, as well as meaning that the control signal is calculated as soon as a
greatly increased flexibility. sensor value is received.
In following sections we describe ZigBee standard,
its limitations and ways to overcome them, experimental
system, controller design and method of play-back
buffer, which removes variance in round-trip time (RTT)
delay, thus enabling us to use predictive control
designed for fixed time delay. Finally results of
Fig. 1 A block diagram of an NCS experiments and conclusion are presented.

NCSs are not without their drawbacks. At best, 2. ZIGBEE/IEEE 802.15.4

communication networks can introduce nontrivial
delays, but the network can also introduce ZigBee is the name of a specification for a suite of
nondeterministic elements such as time-varying random high level communication protocols using small,
delays and packet loss. Additionally, the use of an low-power digital radios based on the IEEE 802.15.4
electronic network requires the discretization of standard for wireless personal area networks (WPANs).
measurements as well as the control signal, and limited ZigBee is designed for low-rate applications. The
network capacity affects the ability of the system to use channel bandwidth in 2.4GHz radio band is 250kbps.
fast sampling frequencies. The claimed throughput is 120kbps, but during
In this paper the wireless ZigBee network is used for experiments it was discovered that maximum
real-time DC motor control. Going wireless has some throughput on given hardware is 80kbps, which is
obvious advantages; without cables we have a much achieved by unidirectional sending of packets with
greater freedom to physically distribute the nodes. It is maximum allowed payload without ACK
possible to place actuators and sensors without worrying (acknowledgment). However, for RTC (Real-Time
about the location of the control node. For example, it Control) purposes the packets should be small (less lag
allows us to place the sensor and actuator on a mobile due to smaller transmission time) and be sent bothway
object and still run the controller from a stand-still frequently. Experiments with two nodes sending packets
platform. However, going wireless will also result in of 5 and 4 bytes length to each other have shown up that
less dependable systems. The likelihood of a wireless packet interspace should be at least 25ms (control
link containing bit errors is several orders of magnitude frequency 40Hz), i.e. achievable throughput is only
larger than for a wired link. This means that the 3.2kbps. More nodes we have, the lower is achievable
possibility of a message being corrupt or delayed due to control frequency rate (10Hz in case of 1 master and 3
retransmission is much larger. slaves).
In [2], Nilsson analyzes several important features of ZigBee supports star, clustered tree and mesh
NCSs. Nilsson’s work standardized a few basic networks. It is obvious, that for RTC of multiple slaves
assumptions, which are used throughout this work. First, by single master only star network is appropriate,

2937
because each hop retransmission in clustered tree and 3.2. DC Motor Controller
mesh networks increases packet delay, negatively
The DC motor model is derived during experiment,
impacting on control performance. Average measured
and has the form of first-order transfer function, where
hop-to-hop time delay is 20ms, giving average RTT
input is torque, and output is angular velocity.
40ms. In fact, in not congested network the RTT varies
uniformly between 30 and 50ms. 1450
Gm ( s) = (1)
In RTC process packets are sent continually, thus s + 150
packet loss is acceptable and retransmissions are not Encoder is used in the role of sensor, thus the plant
required, moreover retransmissions are undesirable, has integrator block added to the motor block as shown
because they introduce additional lag. ZigBee allows us in Fig. 3.
to disable ACK mechanism for that purpose. By
disabling ACK we do not only minimize lag, but also
save channel throughput. However, some vendors’
T w a
ZigBee modules do not allow ACK disabling for unicast 1450 1
packets, then we can use broadcast packets, which do s + 150 s

not cause ACK by design.

Main goal of ZigBee is low-power consumption. For
lag minimization, sleeping has to be turned off on all Fig. 3 Plant model: T – torque, ω – angular velocity,
nodes. α – output angle.
ZigBee defines GTS (Guaranteed Time Slot) feature
To cope with loop delay Smith predictor [3] is used,
for real-time purposes, but most of vendors do not
which structure is shown in Fig. 4. The quick, easy way
implement this feature, as in case of our hardware.
to understand the controller is to assume that the plant
model is perfect, i.e., G’(s)=G(s), and ignore the
3. DESIGN OF NCS
feedback from the plant and the feedback from the
delayed plant model. Then, we can see that the control
3.1 Setup of NCS
signal, u, will be the same control signal sent to the
The distributed control configuration is shown in plant as if it were controlled in a closed loop without
Fig. 1. The feedback loop consists of three parts: the delay. This means that the controller can be designed for
sensor, the controller and the actuator. The sensor node the delay-free system, and therefore can have a much
is time-driven while the controller and actuator are more aggressive response. However, because the plant
event-driven. The sensor samples the process model is not always perfect, the error signal must also
periodically and sends the measurement values to the include feedback from the plant. The feedback from the
controller. Upon receipt, the controller calculates a new plant must be compared to the output of the model, and
control signal and sends it to actuator node which because the plant has delay, the feedback from the
outputs the value. In our setup, the sensor node and model must also include that delay.
actuator node are located in the same hardware unit,
called the Remote I/O. Gc (s )
u
G (s )
y

The implementation structure of networked control

system is shown in Fig. 2. MaxStream XBee/XBee 2
Gm (s)
ZigBee modules are used and connected to PC and
Remote I/O via RS-232 link at 115kbps baudrate. API
(Application Programming Interface) frames are used to
control configuration and communication of modules.
Remote I/O is based on AVR ATMega128 MCU with
PWM controlled DC motor driver. Fig. 4 Smith predictor

For NCSs, while a good plant model may be

available, the delays in the loop are time-varying and
random. This presents a significant problem for the
Smith predictor, as differences between the delay in the
plant loop and the model of the delay in the controller
can degrade performance and destabilize the system.

3.3. Coping with Varying Delays

Control schemes designed to compensate for a
constant loop delay are easily destabilized by loop
delays which are not equal to the modelled delay, and
therefore can perform very poorly when the loop delay
is time varying. Of course, if the delays are known in
Fig. 2 NCS implementation structure
advance, the predictive control scheme can adjust

2938
accordingly, but in NCSs, the delays are typically 4. EXPERIMENTS
random. A play-back buffer at the actuator can be
designed to hold a control signal and apply it only after In order to verify effectiveness of Smith predictor
a specified amount of time has passed. Thus, the delay and play-back buffer we performed several experiments
in the loop is reshaped to be more deterministic, but the on system using single DC motor and 3 DC motors. As
delay must be increased for this to happen, a potentially controller the conventional PI controller is used. Smith
costly drawback. predictor’s prediction time is set to 50ms, same as
Play-back buffers were originally designed for play-back buffer. All experiments use step-function for
multimedia play-back [4]. In [5], Liberatore proposed reference angle. The step is equal to π/2 rad.
an algorithm to integrate a play-back buffer with
networked control for the control algorithm, actuator 4.1 Single Remote I/O
and sensor. The main feature is a play-back actuator
As we mentioned before, in case of single Remote
which delays the application of a control signal until a
I/O (2 nodes in network) the maximum achievable
specified play-back time is reached. If control signal is
control frequency is 40Hz (Ts=25ms). PI controller uses
arrived after specified play-back time, it is dropped, thus
Kp=10 and Ki=0.1 gains. It is shown in Fig. 7, that
play-back time has to be chosen carefully.
utilization of only PI controller causes oscillations of
The structure of system with play-back buffering is
motor shaft. Fig. 8 represents results of experiment
represented in Fig. 5.
utilizing predictor and play-back buffer. As we can see,
the oscillations are eliminated in that case.

1.8

1.3
angle (rad)

0.8 angle
reference
0.3
Fig. 5 Remote I/O with play-back buffer
-0.2 0 1 2 3 4 5 6 7 8 9 10
For an example of the effects of play-back buffering,
see Fig. 6. The f(t) is function sampled by level and sent time (s)

over the network, thus giving us fd(t) function, which is

delayed version of original one.
Fig. 7 Single Remote I/O, Ts=25ms, Only PI Controller

1.8

1.3
angle (rad)

0.8 angle
f (t ) f (t )
f d (t ) f d (t ) reference
0.3
t t
-0.2 0 1 2 3 4 5 6 7 8 9 10
Fig. 6 An example of the effects of play-back buffering. time (s)

In Fig. 6 the continuous-time signal to be sampled is

solid line. The left plot shows the delayed signal without Fig. 8 Single Remote I/O, Ts=25ms, PI Controller +
play-back buffering. The plot on the right shows the Smith predictor + Play-back buffer
delayed signal using play-back buffering, removing the
uncertainty in the delay. 4.2 Multiple Remote I/O
The play-back time is chosen to remove much of the
variation in the loop time delay, which makes the A number of experiments have been done on system
application of the control signal more predictable. As with 3 Remote I/O. In case of 3 Remote I/O the
we mentioned before, in non-congested ZigBee network achievable control frequency is only 10Hz (Ts=100ms),
the RTT varies between 30 and 50 ms, thus we select otherwise the network is congested and most of packets
50 ms as play-back time for all experiments described in are lost. In first experiment we used the same gains for
next section. PI controller as we used for experiment with Ts=25ms,
i.e. Kp=10 and Ki=0.1. The results are shown in Fig. 9

2939
and Fig. 10, which demonstrate identical behavior to 5. CONCLUSION
Fig. 7 and Fig. 8.
The problems of common and ZigBee networks and
their solution are discussed. Also the common problem
1.8 of variable time delay in networked control loop is
discussed in this paper and play-back buffer solution is
1.3 verified. It is verified that play-back buffering and
angle (rad)

model predictive control for constant time delay

0.8 angle together give satisfactory results.
reference However ZigBee’s original purpose is low-rate
0.3 low-power applications, it was not designed with
real-time control in mind. The satisfactory results of
-0.2 0 1 2 3 4 5 6 7 8 9 10 experiments were achieved by aggressive optimization
time (s) of packets structure, decreasing their size to minimum
reachable level – 4 bytes for control packet and 5 bytes
for information packet. Even in case of such extreme
Fig. 9 Multiple Remote I/O, Ts=100ms, Only PI optimization the only reachable control frequency is
Controller 40Hz, which is not appropriate for many RTC problems.
As far as only star network is appropriate for wireless
control purposes, i.e. mesh networking capabilities of
1.8 ZigBee standard are not used, utilization of pure
802.15.4 standard could be the better approach.
1.3
angle (rad)

ACKNOWLEDGEMENT
0.8 angle
This research was supported by the Center for
reference
0.3 Embedded Software Technology research grants in
2007.
-0.2 0 1 2 3 4 5 6 7 8 9 10
time (s)
REFERENCES
[1] D. Hristu-Varsakelis and W. S. Levine, editors.
Fig. 10 Multiple Remote I/O, Ts=100ms, PI Controller + Handbook of Networked and Embedded Control
Smith predictor + Play-back buffer Systems. Springer Verilag, 2005.
[2] J. Nillson, “Real-Time Control Systems with
Another experiment has been done for only PI Delays”. PhD thesis, Department of Automatic
controlled system with lower Kp=2 gain, to verify Control. Lund Institute of Technology, 1998.
whether it is possible to eliminate oscillations in that [3] O. J. M. Smith. “Closer control of loops with
case. As Fig. 11 demonstrates, the oscillations are deadtime,” Chemical Engineering Progress, 1957.
eliminated; however the raising time is obviously [4] L. L. Peterson and B. S. Davie, Computer
increased, what is undesirable. Networks, Morgan Kaufman, 2000.
[5] V. Liberatore, “Integrated play-back, sensing, and
networked control,” IEEE Infocom, 2006.
1.8

1.3
angle (rad)

0.8 angle
reference
0.3

-0.2 0 1 2 3 4 5 6 7 8 9 10
time (s)

Fig. 11 Multiple Remote I/O, Ts=100ms, Only PI

Controller, Kp=2, Ki=0.1

From results of experiments we can say definitely

that Smith predictor together with play-back buffer
efficiently copes with variable network delay.

2940