Sicon/05 – Sensors for Industry Conference
Houston, Texas, USA, 8-10 February 2005
Daniel Sexton Michael Mahony Michael Lapinski
Phone 518-387-4121 Phone 518-387-5485 Phone 518-387-6690
Email: sextonda@crd.ge.com Email: michael.mahony@crd.ge.com Email: lapinski@crd.ge.com
GE Global Research
One Research Circle
Niskayuna, NY 12309
Jay Werb
Sensicast System
220-3 Reservoir Street
Needham, MA 02494
Phone 781-453-2555 x203, Email: jwerb@sensicast.com
Abstract –
Wireless Mesh Sensor Networks are being deployed today
in various monitoring and control applications. Some radio
network designs, such as ZigBee, presume that radio connectivity is
reasonably consistent over time. Others take the opposite approach
of presuming that links are entirely unreliable, and build large
degrees of physical redundancy into the network in the hope that a
collection of redundant but unreliable individual links will result in
a reliable overall system. Surprisingly little work has been done in
the middle ground, endeavoring to understand the root cause of
link failure in real-world factory environments and applying this
knowledge in the design of protocols that adaptively detect and use
workable radio channels.
In collaboration under a Department of Energy grant for the
Industries of the Future, General Electric and Sensicast Systems
have studied theoretical and actual performance of 2.4 GHz IEEE
802.15.4 radio transceivers on the lab bench and on the factory
floor, with particular attention to jamming from 802.11 and
multipath fading. Temporal and frequency variations in link
quality are explored. The implications for network reliability and
protocol design are discussed.
I. INTRODUCTION
Wireless sensors for industrial applications are expected to
open large opportunities for data collection where it has
traditionally been considered technically impossible or cost
prohibitive. To overcome installation and acceptance barriers
a wide variety of requirements must be satisfied. Some of
these barriers include cost and reliability. Short-range wireless
technologies such as IEEE 802.15.4 [1] combined with mesh
networking techniques are being widely considered as the
answer to both cost and reliability in industrial settings.
However RF communications, particularly indoors, is well
known to be unpredictable.
Proper understanding of the channel characteristics is
needed in order to determine adequate design margins to
minimize the installation effort or the amount of physical
network reconfiguration required as the environment around
the network changes. One approach would simply be to
over-configure the network by increasing the node density
with additional mesh routing nodes. However this can cause
issues with additional installation cost, network maintenance
and decreased network capacity. A better approach would be
to just slightly over-configure the network by understanding
what the appropriate required design margins. Of course the
success in this approach requires collection and analysis of a
statistically relevant and representative set of RF channels and
environments.
In our work on the Department of Energy grant we have
attempted to better understand what effects are present in the
RF environment in industrial facilities. As part of earlier work
we have decided to focus our attention on the performance of
2.4 GHz IEEE 802.15.4 physical radios in this environment.
With this information we hope to gain insight into the
characteristics required of a mesh network as well as the
suitable design margins required.
In our paper we present results of several types of
measurements taken in several industrial facilities.
II. OVERVIEW
In our project several important technology and design
decisions needed to be made early in our program so that we
could launch critical design activities. One of these decisions
was the choice of a physical radio technology. There were
several options available to us. To make an initial decision we
first took measurements of a readily accessible and
representative harsh indoor environment. We took in initial
channel measurement using a network analyzer and two
antennas spaced 25 feet apart. An image of the facility is
shown in Figure 1 and the schematic of the test equipment in
shown in Figure 2. Initial measurements of this environment as
shown in Figure 3 illustrates that in the 2.4GHz ISM band
there is both significant frequency selective fading as well as
flat fading depending on the area of interest. The dark points
represent the actual measurement whereas the lighter points
represent the fading as averaged over a 2MHz bandwidth.
Clearly this initial measurement illustrates the need to utilize
RF transmission techniques that can tolerate both frequency
selective as well as flat fading.
Figure 1: Initial Test Environment
Figure 2: Initial Test Configuration
At the time of this experiment the first IEEE 802.15.4
Direct Sequence Spread Spectrum radios were being
introduced and were an obvious consideration for this project.
We evaluated the ChipCon CC2420 [3] and decided to
adopt this device for this project. With the transceiver output
power of of 0dBm and the receiver sensitivity of –94dBm the
RF link budget could then be evaluated. However additional
channel measurement data would be required to get a better
understanding of the link margin requirements given a desired
maximum range requirement of 100m both line of sight and
non line of sight. Figure 4 shows a plot of the first estimate of
path loss using equation 1, based on our initial data and a
simple break point model we assume free space loss for the
first d=5m and n=3.5 thereafter [13]. Note that for 0dBm
output power we fall below the receiver sensitivity at about a
65m distance using this path loss model.
>
+
+ ≤
= n d d m
d d m
Lo 5
5
54 10 log
40 20log( ) 5
(1)
Figure 3: Channel Fading – Line of sight channel
(1)
Figure 4: Breakpoint model for path loss
III. CHANNEL MEASUREMENT STRATEGY
With the radio selection made we took a focused strategy
on channel measurements. We decided on a two-pronged
approach to the collection of channel data. One approach used
classical techniques for channel sounding. The other used the
actual radios themselves to gather performance data. Relevant
channel measurements made at ranges beyond 25 ft were not
practical with out initial measurement technique. We therefore
choose to use the Spread Spectrum Sliding Correlator
approach [2] for channel sounding. The test equipment we
chose for these measurements was the Berkeley Varitronics
Systems Raptor, which is well suited for indoor channel
measurements [4]. The Raptor consists of a stand-alone direct
sequence spread spectrum transmitter that has field-adjustable
RF output power and selectable transmit frequency, and a
mobile receiver that captures data using a PDA. In addition to
the Raptor data, we characterized local noise sources in the
2.4GHz ISM band using a spectrum analyzer set for max hold
to measure aggregate interference in the 83.5 MHz-wide band,
and time-based, zero-span measurements of each IEEE
802.15.4 channel (see Figure 5)
Figure 5: Spectrum Analyzer
The second measurement approach we used for channel
characterization primarily consisted of a spectrum analyzer set
for zero span and a CW RF signal source. . This measurement
enabled us to extract the coherence time for the channel as well
as calculate the fading statistics.
The third measurement approach used was to experiment
with the performance of IEEE 802.15.4 radios in the actual
industrial environment. To perform this test we constructed six
units each of which housed a radio module and a single board
computer with local data storage – Figure 6. These units were
placed at locations in industrial facilities where wireless
sensors for equipment monitoring would typically be placed.
The boxes then recorded the transmission performance for
every packet sent on every channel so that a history of path
performance could be determined. The data was then extracted
from the units and stored in a database where it could be
processed and sorted as desired.
The indoor channels that we measured displayed various
amounts of interference from other radio systems. Although
interference is not a topic for this paper it had a substantial
effect on our measurements and had to be dealt with as the data
was processed and analyzed.
The outcome of the coherence time measurements is an
important discussion topic because it directly affects the
channel stability and fading over time. In the channels we
measured, channel coherence time averaged approximately
0.1sec – which represented people moving around within the
channel environment, Figure 7. For many of the installations,
the channel is relatively static; we are not considering mobile
RF devices for our analysis at this time.
The bit rate for IEEE 802.15.4 operating at 2.4GHz is
250kbps and in our system typical messages are in the range of
46 bytes including overhead, for a message duration of about
1.5 msec. With channel coherence time measured at about 100
msec, the channel can assumed to be stable over a single
message and in fact can assumed to be stable over multiple
transmissions of a message in the case of immediate retries.
Figure 6: Radio Performance Tester
IV. MEASUREMENTS
Off the shelf IEEE 802.15.4 radios currently do not have
equalization built in. In industrial facilities with large metal
surfaces the multipath delay spread may represent a problem.
Andersen [5] reports that RMS delay spreads of 300ns might
be expected and Ganesh [6] reports mean RMS delay spreads
ranging as high as 73nsec (max of 150 nsec). This amount of
delay spread could cause a problem without equalization. In
an IEEE 802.15.4 system with a chip rate of 2Mc/s [1], an
RMS delay spread of greater than 50nsec would be cause for
concern. In 2002 we took extensive measurements of RMS
delay spread and path loss for large industrial facilities that
were housed in metal and concrete structures – in one
example: a power generation facility with four GE Frame 7
turbines, two steam turbines and associated boilers. we
calculated mean RMS delay spreads of 34.4 nsec using the
direct RF pulse method [2] which are consistent with the
findings of Ganesh [6]. In the metallurgy facility at GE Global
Research (a large metal and brick building) we measured
RMS delay spreads of 10 to 200nsec (with excess delay in the
order of microsecnds) using the Raptor which is consistent
with Andersen[5]. A recent presentation by Gorday [11]
simulates IEEE 802.15.4 with various RMS delay spreads up
to 500 nsec. This analysis shows that for an RMS delay spread
of 400 nsec a degradation of about 17dB might be
experienced. We do not need to add additional margin for
RMS delay spread on top of margin for fading as they are
related. We do need to verify that we use the worst-case effect
between the two. We needed to verify the fading
characteristics of the channel to validate assumptions on
fading and RMS delay spread.
Figure 7: Channel Coherence Time Measurement
From measured data at various industrial locations,
calculations of channel fading statistics varied substantially
from location to location. Most of the measurements exhibited
Nakagami-m characteristics while some exhibited Lognormal
and Rayleigh characteristics. At this stage in the data
collection it is unclear whether or not a single pdf can be used
for the purposes of characterizing the channel. In general we
have seen a large variance in the m- factor that represents the
severity of fading, ranging from 0.9 to 55.9 (Figure 8 and
Figure 9).
We are still in the process of collecting and characterizing
channel data and cannot draw any definitive conclusions with
the limited data set that we currently have. We will report our
findings at a later date when the analysis is completed.
Even with channel fading data collected, having actual
radio performance data will be useful to validate the
characterization of the channels. At this point data collected
from the actual radios using the units illustrated in Figure 5 has
yielded the most interesting information. Again, we have not
completely harvested all the information in this data but we
have gathered some interesting results. In one experiment the
boxes were distributed in a machine room (Figure 1) with a
floor plan illustrated in Figure 10. Each of the locations for the
units is marked 1 through 6. The test was allowed to run for 4
hours. There is little to no motion in this installation as it is in
an isolated area. During the experiment we collected lost
packets, RSSI (received signal strength indicator) and LQI
(link quality indicator) for each packet sent. Since the units are
synchronized in time there can be no collisions. At any point in
time only one unit was transmitting while the remainder were
listening and recording. Transmission was then cycled to
another unit until all had a time to transmit. This sequence
occurred over and over again. The results of each transmission
were recorded. Figure 12 plots the packet loss rate versus path
and channel. The nomenclature for path is as follows: Path12
represents the packet loss information for the path from unit 1
as the transmitter to unit 2 as the receiver and Path21 represents
the reverse path. From this experiment we not only studied the
performance of the radios but also the symmetry of the
channel. In this experiment paths were both line of sight (LOS)
as well as non line of sight (NLOS).
Figure: 8 Field data demonstrating a channel with Nakagami-m
statistics showing a low degree of
fading
Figure 9: Field data demonstrating a channel with Nakagami-m
statistics showing a high degree of fading
Figure 13 plots the same experiment in a different location
– this data was collected in an industrial gas compression
facility. Both data sets show rather vividly the effect of
frequency selective fading. In this case of Figure 12 there was
no channel that allowed for reliable communications over all
paths for all units throughout the entire test period. In figure 13
only channel 15 was clear for all paths. None of the paths were
very symmetric for all channels. The results of these
experiments clearly show that a frequency agile approach or a
multi-channel approach might be more robust than a single
channel approach and is consistent with the initial
measurements – Figure 3.
Figure 10: Machine room floor plan
Figure 11: Compressor house floor plan
Figure 12: Packet loss versus path and channel – Machine Room
Figure 13: Packet loss versus path and channel – Compressor House
V. CHANNEL FADE MARGIN
Ultimately we need to determine the required fade margin
or excess power required to obtain the desired reliability
(99%).
Free space channel path loss (dB) is given by equation 2
Π
=
O
O d
L n
4
10 log λ
(2)
Where λ is the wavelength of the RF signal and dO is the
path distance and n is the path loss exponent (n=2 for ideal free
space). From line of sight data measured in industrial facilities
we estimated that n=1.6 gives a close approximation of the
average attenuation. Our value of n was derived from averaged
measurements from industrial facilities and is consistent with
measurements reported by Andersen [5]. For non line of sight
n=3.73 is consistent as determined in reference [7].
Although these estimates for path loss give a good starting
point for calculating the mean loss, the variance around that
mean must also be considered. Link margin is required to
overcome this variance and maintain a certain confidence level
for the channel. Link margin can be gained by using multiple
strategies. The three we choose to investigate are power
amplification, spatial (path) diversity and frequency diversity.
Both spatial and frequency diversity will provide a diversity
gain that will allow us to reduce additional output power while
maintaining our range.
In general we need to design our system for NLOS with
multipath. Using a Rayleigh distribution as our model for the
average receive power we will have the calculated mean
power as defined above. To establish a goal for probability of
success we also assume very slow fading or a static channel.
We can assume that the distribution around the mean follows
the Rayleigh pdf.
2
2
2
( ) 2 σ
σ
r
f r = r e r ≥ 0. (3)
If we assume that for any single transaction (which may
involve multiple transmissions each of which is less than
4msec) the channel is fixed then we can calculate the excess
power required to guarantee with a 99% certainty that we will
be at or above the expected power level from the following
equation.
2
2
( ) ( ) 2σ
R
R
P R f r dr e
∞ −
= ∫ = (4)
Where σ2 is the time-averaged power received signal and
P(R) is the probability of the received signal equal to or
exceeding power level R2. If R2 is the power level we need for
reliable transmission, then we can solve for σ2 as the average
power level, or we can solve for 2
2
R
σ
as the static link margin
LM
L 10log( 2ln(P(R))) m = − − (5)
If we chose a 99% probability that we will obtain a
particular power level, we will need to provide a 17dB margin
in our link budget (10dB for 95%). This is consistent with the
margin estimated for RMS delay spread [11] and with data
measured in reference [7]. The link budget equation then
becomes:
Me(dB) = Tp – Lo –Lm - Rs+Gta+Gra (6)
In our case we are using an omni-directional patch antennas
that provide approximately 1dBi. Our transmit power Tp is
+15dBm and our receiver sensitivity is –92dBm (typical). We
use a slightly lower receiver sensitivity because of the losses
associated with adding a power amplifier – these are measured
values. If we calculate the link margin at 100m NLOS (slow
Rayleigh fading where the channel is considered stable and
n=3.73) our link margin requires 23dB additional gain.
Additional processing gain can be achieved through diversity
as described in [12].
Refs [8] [9] [10] analyze the use of spatial diversity and
frequency diversity techniques in terms of processing gain.
Todd [8] evaluated both spatial and frequency diversity at
1.7GHz and compared this to results obtain for 900MHz in the
same environment. With spatial separations of 1 wavelength a
diversity gain of 10.8dB could be achieved with a frequency
separation of 10 MHz (two 802.15.4 channels). Combining
these two diversity schemes, a total diversity gain of 16dB was
achieved [8] assuming a link availability of 99%. One
wavelength spatial diversity requires a spacing of 12cm at
2.4GHz, which in our application would make the overall
package size too great. Large spatial diversity can be achieved
by the use of multiple routing nodes either of which can
receive and relay a message. Although this is not the same
because of the significant path loss differences, it can be used
to achieve additional diversity. For spatial diversity we can
assume a worst-case analysis such that both receiving nodes
are at the maximum distance of 100m. The result of ref [8]
also effectively assumes that both receivers are at the same
statistically average power point or at the same range and
somewhat correlated.
In our system we will have a much higher diversity factor
than two and our spatial separation is much greater. Until we
collect more data we assume we can use uncorrelated
channels. We therefore performed a simulation assuming
totally independent uncorrelated channels for each level of
diversity. If we consider our system to operate as a simple
form of selection diversity [12] in a Rayleigh channel the
probabilities multiply as the order of diversity N increases:
N
R
N P (1 e ) 2
2
2σ
−
= − (7)
The results of this simulation are illustrated in Figure 14
and agree very closely to measured data as shown in Table 1.
Since we have 16 channels to choose from and we assume a
minimum of two routing nodes available to each sensor node
we have 32 degrees of freedom. Our simulation shows that in
the best case we might expect a 23dB gain from diversity with
99% availability in a pure Rayleigh channel; however it is
doubtful we will ever achieve this amount of gain in practice.
Coincidently to meet out link budget we need 23dB.
In this analysis we assume that when changing channels as
well as changing paths we move to a totally independent yet
statistically similar channel. In typical mesh routing networks,
changing frequencies is much less costly than changing paths,
as fewer routing nodes are needed. In addition, with frequency
diversity all the routing tables remain stable and new routes do
not need to be established. However in our network design we
do keep a single alternate route in the routing tables if that
route can be established.
No margin was added for interference due to coexistence
issues, we are studying this and will be a subject for another
paper. We do know that frequency diversity can be an
effective tool in circumventing coexistence issues.
We do know from the data we have collected that we are
getting significant advantage due to frequency diversity as
shown in Figures 12 and 13, but we have not been able to
extract the data to accurately quantify the diversity gains we
are achieving. RSSI might be one valuable measure of this but
we have no RSSI values for packets we did not receive. We do
have RSSI as well as LQI values for both good and bad
packets received and we will attempt to normalize those into
information that will relate to diversity gain.
Availability Diversity Todd[8]
Maximum
Simulation
90% 1 5.7dB 5.58dB
99% 1 10.8dB 10.62dB
99% 2 16.1dB 13.9dB
99.9% 1 16.9dB 14dB
Table 1: Diversity Gain Comparison
Figure 14: Estimated Diversity Gain
In our system we also require that a message is
acknowledged and that the acknowledge is received to
complete a transaction. (In a typical scenario, lost
acknowledgements cause unnecessary retries, wasting power
and bandwidth but not compromising reliability.) Currently
the acknowledge must occur on the same channel and the
reciprocal path as the original message. If there is minimal
reciprocity in the paths then the diversity gain could be
compromised. Our test data does not collect reciprocal path
data within the coherence times we measured but we ran a
correlation between the data to see if there was reciprocity
anyway. For the machine room data in Figure 12 the
correlation for reciprocity was 0.91 and for the compressor
house data in Figure 13 the correlation for reciprocity was 0.87
both with a high confidence factor. With this level of
correlation we can accept the diversity gains.
As previously mentioned, we have also seen Lognormal
and Nakagami fading statistics. In line of sight channels we
have measured exponential path loss factor of n=1.6. In these
channels we expect to have an excess margin of 35dB before
fading effects are considered. This is 52dB above our starting
point for NLOS channels so we do not anticipate a problem.
We will be analyzing the other statistic channel models in a
similar way to determine the overall link margin choosing a
conservative distribution.
We will be gathering additional data to validate the
assumptions in this paper but the data we have already
collected leads us to believe that we should have a high degree
of confidence that we can maintain an adequate link margin to
maintain our performance goals in all but the absolute worst
case scenarios (all path distances = 100m NLOS).
VI. CONCLUSIONS
IEEE 802.15.4 at 2.4 GHz appears to be a suitable physical
layer protocol for use in industrial environments; however
much more testing and experience is needed. Power
amplification beyond 15 dBm may be required to achieve the
desired range of 100 m NLOS, overcome the path loss
uncertainty and keep network density low. Diversity schemes
such as spatial diversity through mesh routing, and frequency
diversity will significantly help to increase the reliability of the
network by providing additional diversity gain. Channel
reciprocity is also important in systems that require packet
acknowledgements so that the full diversity gain can be
achieved. In our network design we attempt frequency
diversity before path diversity because of the lower cost.
REFERENCES
[1] IEEE Standard 802.15.4 – 2003, Standards for Telecommunications and
Information Exchange Between Systems – Local Area and Metropolitan
Area Networks - Specific Requirements Part 15.4: Wireless Media
Access Control (MAC) and Physical Layer (Phy) Specifications for Low
Rate Wireless Personal Area Networks (WPAN), Standard
[2] T. Rappaport. “Wireless Communications Principles and Practices,”
Book, 1996 by Prentice-Hill, Inc.
[3] ChipCon CC2420 Datasheet
[4] Berkeley Varitronics The Raptor Datasheet
[5] J. Andersen, T. Rappaport, “Propagation Measurements and Models for
Wireless Communications Channels,” IEEE Communications
Magazine, January 1995.
[6] R. Ganesh, K. Pahlavan, “On The Modeling of Fading Multipath Indoor
Radio Channels,” IEEE GlobeCom 1989, Volume 3, November 1989,
pages 1346 – 1349.
[7] D. Cheung, C. Prettie, A Path Loss Comparison Between the 5GHz
UNII Band (802.11a) and the 2.4GHz ISM Band (802.11b), Intel Labs
Intel Corporation
[8] S. Todd, M. El-Tanny, S. Mahmoud, “Space and Frequency Diversity
Measurements of the 1.7GHz Indoor Radio Channel Using a
Four-Branch Receiver,” IEEE Transactions on Vehicular Technology,
Vol. 41, No 3. August 1992
[9] J. Wintersm J. Salz, R. Gitlin, “The Impact of Antenna Diverstiy on the
Capacity of Wireless Communication Systems,” IEEE Transactions on
Communications, Vol 42. No. 2/3/4, February/March/April, 1994
[10] J. Winters, “The Diversity Gain of Transmit Diversity in Wireless
Systems with Raleigh Fading,” IEEE Transactions on Vehicular
Technology, Vol 47, No 1, February 1998
[11] P. Gorday, 802.15.4 Multipath, doc: IEEE 802.15-04-0337-00-004b,
July 2004
[12] D.G. Brennan, “Linear diversity combining techniques,” Proc. IRE, vol.
47, pp. 1075-1102, June 1959
[13] IEEE 802.11.b Radio Interface, Radionet presentation.
ليست هناك تعليقات:
إرسال تعليق