Content uploaded by Thomas Lake
Author content
All content in this area was uploaded by Thomas Lake on Oct 03, 2019
Content may be subject to copyright.
Motion tracking of a free-yawing floating tidal
stream turbine platform
Thomas Lake, Alison Williams, and Ian Masters
Abstract—As part of ongoing work to develop and
validate a combined model for the behaviour and potential
power generation of a floating tidal energy converter, the
motion of a prototype device has been recorded over a two
month testing period. The inputs from several types of
sensor have been combined with GPS position information
to improve accuracy of the reported position, and the
distribution of this position and orientation under different
operating conditions is illustrated and discussed. The po-
sition distribution is consistent with other published work
describing the motion of this platform. Further work will
improve the processing of the underlying data and extend
the model to take advantage of the redundant information
available.
Index Terms—Tidal Energy, Motion tracking, Data acqui-
sition
I. INTRODUCTION
RECENT years have seen a range of floating tidal
energy converters (FTECs) deployed as part of
the still growing marine energy sector. These devices
are effectively a combination of a floating structure
(a subject on which there is a great deal of existing
expertise, regulation and understanding) and a tidal
energy converter - a category of device which is still
under rapid and active development, but for which
there is also a good deal of existing information.
Investigating and identifying the combined response
of these FTECs to varying and extreme environmental
loads can help determine some of the ways in which
both the capital and operating expenditure required
can be reduced. The response of the device under loads
can also inform predictions of how the device may
be affected by fatigue, with associated implications
for design life and therefore design criteria for new
devices.
Swansea University is part of the SURFTEC project
(Sustainability and Reliability of Floating Tidal Energy
Converters), investigating the measurement, prediction
and implications of the loads on this class of tidal
energy device. The first two work packages on this
project include updating and adapting an in-house
blade element momentum theory (BEMT) code and
coupling this to a floating body model, and collecting
and analysing data from PLAT-I - a prototype FTEC
Paper 1328, Tidal device development and testing
This work was supported in part by the EPSRC under grant
EP/N02057X/1
All authors are part of the Marine Energy Research Group, Col-
lege of Engineering, Swansea University, Bay Campus, Fabian Way,
Swansea, SA1 8EN.
The authors’ email addresses are t.lake@swansea.ac.uk,
alison.j.williams@swansea.ac.uk and i.masters@swansea.ac.uk
respectively.
designed and operated by Sustainable Marine Energy
(SME) [1].
The data collected and analysed from this platform
will then be used to validated the outputs of the
coupled BEMT and floating body model, before using
that model to investigate possible loadings on a generic
device under a range of environmental conditions and
operating scenarios.
This paper aims to illustrate how Kalman filtering
can be used to combine recorded position and sensor
data to reconstruct the motion and orientation of a
floating tidal device.
II. EQUIPMENT DESIGN AND SPECIFICATIONS
In order to validate the coupled BEMT and floating
body model, it is necessary to record the motion of the
platform alongside its operating condition, generated
power and the incident flow onto the turbines. The op-
erating condition, inlet flow and the device geometry
can be provided to the coupled model as input data,
with the output of the model compared to the recorded
motion and generated power results.
After some preliminary investigation and testing, it
was not possible to find a commercial off-the-shelf
(COTS) system that was able to record the desired data
while also meeting other design constraints, so a com-
bined system was designed based on a combination of
commercial and in-house components.
This system consists of a core data logging compo-
nent, two custom sensor inertial measurement units
and a single board computer providing additional
logging, supervision and remote access to the system.
A. Core logger
The core logging component of the SURFTEC system
is a Race Technology DL1 Pro, which incorporates a
GPS receiver (providing position and velocity data), 3-
axis accelerometer and supports logging external data
provided digitally via two serial inputs as well as
external analog inputs for voltage based sensors [2].
This unit can be monitored and controlled over a
USB connection, as well as logging directly to SD Card.
Due to the volume of data being recorded it was not
possible to monitor all data over USB, leading to two
separate streams of data – a high sample rate set stored
on the SD card and a lower sample rate set recorded
by the control computer over the USB connection. This
provided some level of redundancy in case of storage
media failure or if the SD card capacity was reached.
B. Inertial Measurement Unit (IMU)
Supplementing the accelerometers and GPS receiver
in the core logger, a pair of remote sensor units were
designed and built. Each unit is designed to be robust,
waterproof and with the option to reprogram in the
field if required.
Each unit consists of an Arduino Micro and two
LSM9DS0 iNEMO inertial modules on a breakout
board. The inertial modules combine a 3-axis ac-
celerometer, 3-axis gyroscope and 3-axis magnetometer
into a single package, along with a temperature sensor.
The sensors can be configured to either return data
across a wide measurement range (at lower precision)
or across a smaller measurement range at greater pre-
cision. As an example, the accelerometer can be con-
figured to report values between ±2g with a precision
of 6.1×10−5g or between ±16g with a precision of
7.3×10−4g. Similar options exist for both the magne-
tometer and gyroscope portions of the sensor.
In each IMU unit, the two inertial modules are
configured with different measurement ranges to take
advantage of this option to allow a more precise
measurement of values in the general case while also
allowing larger peak values to be recorded in the case
of a more extreme event. The use of two sensors also
provide for a level of redundancy in case of failures
The two IMU units also include supporting circuitry
to allow the units to be powered from a range of input
voltages and to convert the output data signalling from
TTL logic level communication to RS-485 differential
serial signalling. This signalling standard was chosen
in order to provide more robust data transfer between
the sensor units and the core logger in the possibly
noisy electrical environment of a power generating
platform.
C. Control, supervision and remote access
One of the design requirements of the data logging
system is that it needed to be able to operate with as
little on site interaction as possible. This was imple-
mented by planning enough storage space to exceed
the planned deployment duration and placing a single
board computer within the equipment case to provide
remote access. The core logger was connected to this
single board computer, allowing data from the core
logger to be recorded and for the limited remote control
of the logger. The data streamed from the logger to the
PC stores data at a lower sample rate ( 25Hz) than the
data saved to the SD card in the core logger (which
records at 85Hz).
To allow remote access without requiring extensive
configuration changes to be made to the networking on
the platform, the computer established two SSH tun-
nels at boot, connecting to a server running on Amazon
Web Services and to a desktop machine at Swansea
University. This allowed two routes of access in case of
misconfiguration or network difficulty, while allowing
larger file transfers to avoid transiting through the
Amazon network (and subsequent bandwidth costs).
The control PC and core logger were contained in
a single equipment case along with DC-DC power
supplies to supply 5VDC and 12VDC to the PC, logger
and RS-485 interfaces. The RS-485 interface boards
converted the signal from the sensor units to standard
RS-232 TTL signal that could be decoded by the core
logger.
For the first deployment of the logging equipment,
this control computer was a Raspberry Pi 3B, with
additional flash memory to provide storage. The Rasp-
berry Pi was found to be operating very close to the
limit of its input/output capacity at various points dur-
ing this initial deployment, and has been replaced with
an ODROID XU-4 computer and solid state drive for
the second deployment (not presented here), providing
more capacity and computing power.
D. ADV Integration
The data logging system also incorporates data from
a Nortek Vector acoustic doppler velocimeter (ADV)
running in continuous sampling mode, with the data
recorded by the control PC. The maximum sample rate
of 64Hz was used. Due to the planned deployment
length the internal power supply and storage of the
ADV were not used. The recording process was au-
tomatically stopped and restarted briefly at midnight
UTC each day in order to resynchronise the ADV clock
with the control PC and minimise drift.
III. DEPLOYMENT
E. PLAT-I
PLAT-I is a floating tidal energy converter designed
and under test by Sustainable Marine Energy. The plat-
form is “a three-hulled tidal energy platform” [3], capa-
ble of freely rotating with the dominant flow direction.
The platform hosts 4 SCHOTTEL instream turbines
(SITs), for a total generating capacity of 280kW [3]. This
initial version of the platform is not grid-connected,
with generated energy being transferred into a load
bank immersed off one side of the central hull.
Fig. 1. PLAT-I on location at Connel (Photo by T.Lake)
The platform, shown in fig. 1, has three hulls joined
by a cross-deck structure towards the rear of the plat-
form. The locations of key equipment are indicated in
fig. 2. The two IMUs were installed under this cross
deck structure towards the outboard ends, with the
LAKE et al.: MOTION TRACKING OF A FREE-YAWING FLOATING TIDAL STREAM TURBINE PLATFORM
core logging equipment installed inside the shipping
container on the central hull. The GPS antenna was
attached to one of the standoffs securing the shipping
container to the central hull, at the position indicated
on the diagram. Other instrumentation installed by
SME is not included in the figure.
GPS Antenna
ADV
IMU A
IMU B
X
Y
Fig. 2. Simplified top view of PLAT-I showing key instrument
positions
F. Trial site: Connel, Scotland
The initial trial deployment was carried out south
west of the Falls of Lora at Connel, Argyll and Bute,
Scotland [4]. The site is relatively sheltered from waves,
but has a pronounced ebb-flood asymmetry. This asym-
metry is due to constrained flow of water entering and
leaving Loch Etive through a shallow, narrow neck
under Connel bridge. The shallow depth at the bridge
creates underwater overfalls on the ebb tide which lead
to a pronounced jet to the west of the bridge on the ebb
tide [3], [4]. This jet leads to a faster but more variable
flow past the platform on the ebb side of the tide.
Based on the ADV data recorded, the average flow
speed at the platform over a flood tide was 0.20ms−1
and 1.20ms−1during ebbs. The average of the peak
velocities for flood and ebb respectively were 0.58ms−1
and 2.37ms−1.
G. Recorded data
Testing took place between November 2017 and June
2018, with principal data collection for the SURFTEC
project taking place in December 2017 and January
2018. Due to some equipment failures during the de-
ployment, some data was lost or not recorded. In total,
744 hours of data were recorded at the highest sample
rate available, with a further 371 hours of data recorded
with full rate ADV data but only the lower sample rate
stream of motion data. There were 373 hours during
this period where no data was recorded or usable.
IV. PRO CE SS IN G AN D ANA LYSIS
A combined dataset was created using higher rate
IMU and GPS data where available, with gaps in the
higher rate record being filled with lower rate data. An
example of data from one of the sensors in this dataset
is shown in fig. 3.
The data available from the IMUs gives information
about the linear acceleration (from the accelerometers),
angular velocity (from the gyros) and orientation (from
the magnetometers). These sensors are imperfect, and
are subject to noise, bias and interference from external
sources. The GPS data returns position information,
but is also subject to error (although an estimate of
this error is also returned). It is, however, possible to
combine data from all of these sources to compute a
best estimate of the actual value of each parameter. This
can be carried out using Kalman filters [5], which are
frequently used to address this class of problem.
Fig. 3. Example timeseries showing x-axis acceleration from unit A,
sensor A for 1 second and 10 minute averages over a 24 hour period.
Axes as per fig. 2
H. Kalman filtering
Kalman filtering is an iterative process comparing
a modelled internal state of a system to a set of
measurements of that system. The process takes the
internal state at a given time, predicts the state at the
next timestep based on the internal physics imposed on
the model and compares this prediction to measured
values. Kalman filters can be applied to a range of lin-
ear problems, provided that the model used to predict
the behaviour of the system is sufficiently described.
Tracking the motion of a device by fusing together
information from a number of sensors is a popular use
of a Kalman filter [6].
A basic Kalman filter operates as follows [7]:
1) Initialise filter with an initial state ~x and ‘belief’
in that state, P
2) Make a prediction for the state of the system at
time t+δt using the transition matrix, F
•Because the prediction is uncertain, update P
to reflect this
3) Get a measurement of the system (~z and belief in
that measurement, R)
4) Compute the difference between the predicted
and measured states
•Use this to compute a scaling factor to select
whether the prediction or measurement is
most reliable
5) Set new state to a combination of the predicted
and measured values using scaling factor
6) Update the ‘belief’ in this new state using the
scaling factor
7) Go to 2 and make new prediction
In this document, the notation and symbols corre-
spond to those in [7]. The ‘belief’ in the state of the sys-
tem or in the accuracy of the measurements provided
is more mathematically described as the variance (and
co-variance) of the relevant variables. The predictions
of how the system moves from one state at tto another
at t+δt are encoded in the transition matrix F.
The uncertainty in the prediction in step 2 is repre-
sented mathematically by a Q, the ‘process noise’ [7]
or ‘covariance of random excitation’ [5]. This represents
any unmodelled effects that aren’t included in the tran-
sition matrix, or which aren’t represented in the state
variables. An example of this would be higher order
effects neglected in a simple model - e.g. accelerations
in a constant velocity model.
I. A simplified motion model
To start the analysis of the motion data captured
on PLAT-I, a simplified model was constructed around
three degrees of freedom. The three degrees of freedom
represented are global position (E, N ) and heading
(ψ). The model also tracks the first and second time
derivatives of each of these degrees of freedom - i.e. ve-
locity and acceleration. The accelerations are assumed
to remain constant throughout each timestep (i.e. the
third time derivative of each degree of freedom is zero).
This model was implemented in Python, using the
filterpy library from [7] as the basis.
The position of the platform is represented as UTM
coordinates in zone 30V, with the heading represent-
ing rotation clockwise from north. The sensor values
returned from the device are all reported in a local
coordinate frame, and must be converted to equivalent
values in the global frame based on the current heading
of the device.
The transition between each timestep is calculated
as:
d=d+˙
dδt +1
2¨
dδt2(1)
˙
d=˙
d+¨
dδt (2)
¨
d=¨
d(3)
for each degree of freedom, d∈E, N, ψ.
These equations are written in matrix form for each
degree of freedom to create the transition marix, F
referenced above. Additionally, the value of ψat the
end of each timestep is checked, and if found to
lie outside the range [−π, +π]it is converted to an
equivalent value within the range.
The duration of each timestep (represented as δt
above) is determined from the provided input data.
The measurements provided to the filter at each
timestep include the reported GPS position and ac-
curacy, the magnetic heading (with some caveats dis-
cussed later) and reported angular velocity about the
vertical axis of the device - equivalent to the rate of
change of heading. In the results presented here, only
one sensor from one IMU is included.
The initial state of the filter is set to the first mea-
surement from the recorded data used, with relatively
large variances set in Pto reduce run-in time. For each
subsequent set of measurements input to the filter, the
associated belief Rin the measurements is updated
with the reported accuracy (for GPS position data), or
based on values in the sensor datasheet.
An example of the output of this model is shown
in fig. 4, showing the positions output by the filter
alongside the corresponding input GPS data for a
30 minute period with the platform operational. The
colour of each marker in fig. 4b is mapped to the sum
of the variances for the two coordinates and the GPS
data shown in fig. 4a is coloured by the square of the
reported accuracy, to allow comparison on the same
scale as the modelled data. The range of the colour
(a) GPS input (b) Model output
Fig. 4. GPS position data and model output, showing position over a 30 minute period. Points coloured by reported accuracy squared (GPS)
or variance (model) respectively
LAKE et al.: MOTION TRACKING OF A FREE-YAWING FLOATING TIDAL STREAM TURBINE PLATFORM
map has been clipped to highlight variations in a larger
selection of points.
Comparing the input GPS data shown in fig. 4a to
the output of the model in fig. 4b shows that the motion
here has been smoothed, with some of the horizontal
banding evident in the input GPS data removed.
J. Sensor drift, interference and calibration difficulties
One of the disadvantages of low cost MEMS devices
is that the outputs can be noisy, and prone to drift over
longer periods. Some of this is handled by the Kalman
filtering process - drift in the accelerometers can be
compensated for by the GPS position, and drift in the
gyros can be compensated for by the magnetometers.
Any change in the behaviour of the magnetometers,
however, cannot be corrected for in this manner.
This is compounded by the difficulty in calibrating
the magnetometers during the initial fitting. (While
rotating the box on all axes in close proximity to its
fitted position is a simple enough procedure in theory,
it is somewhat harder to do while providing reliable
power and data connections on a partially assembled
tidal platform!)
As described in [8], the magnitude of the external
magnetic field acting on the sensor at a given location
is effectively constant, allowing the magnetometer data
to be normalised by fitting the returned values to an
ellipsoid and using this fit to transform the data to lie
on a unit sphere centered at the origin. Orientation data
can then be obtained using trigonometry and then fed
into the Kalman process.
Fig. 6 shows both raw and corrected magnetometer
readings. As the figure illustrates, the rotation of PLAT-
I about it’s mooring ensures that the magnetometer
data covers enough of an ellipse in the x,y plane to
be fit reasonably. While this allows the device heading
to be recovered, the lack of significant pitch and roll
motion makes it harder to apply this method in those
axes - although the smaller motion in pitch and roll is
understood to be a good design feature in any other
respect!
In addition to the effect of the metal surrounding
the IMUs and the drift of sensors over time, additional
interference was observed in the magnetometer read-
ings due to the use of a winch located near to the IMU
Fig. 6. Plots of the X,Y magnetometer readings pre- and post-
rescaling
units. The brief periods where this equipment appears
to have been in use have been excluded from analysis,
with the magnetometer rescaled using the ellipse fitting
method described above after the excluded data.
V. RE SU LTS
During the testing conducted at Connel, PLAT-I
can be described as being in one of three states, as
discussed in [4]. These states are:
a) Operational - generating power
b) Rotors parked - turbines in the water, but station-
ary
c) SDMs1raised - turbines raised out of the water
The recorded data has been classified into each of
these three states, and each contiguous period filtered
using the Kalman filtering process described abov, us-
ing the one second mean values of the underlying data
as input. The one second mean values were used in this
instance to ensure a consistent timestep, and to reduce
the computational cost of this initial investigation.
1SDM: SIT deployment module - a moveable arm connecting each
SIT to the platform
Fig. 5. Distribution of modelled position for operating, rotors parked and SDMs raised conditions
Fig. 7. Platform position and orientation, based on two minute mean values
The distribution of positions for each state are shown
in fig. 5, represented as hexagonal bins coloured by
the number of samples in each bin. As the total time
the platform was recorded in each of the three opera-
tional states varies (with the operating state being the
smallest dataset), the colour range in each plot has been
normalised against the duration of the input data.
It should be noted that for both the raw GPS data
and the model results, the reported position is the lo-
cation of the SURFTEC GPS antenna, shown relative to
the rest of the device in fig. 2. This is why these results
differ from the positions reported in [4], although the
trends illustrated are similar.
K. Operational
The distribution of positions for the turbines operat-
ing state is the smallest of the three cases, as generation
only occurred on the ebb tide. The distribution shown
in fig. 5 is based on generating periods from the ebb
tide on three successive days. The resulting position
data is spread around a central dominant location,
with most points confined within the region extending
approximately 15m north and 15m south east of the
mean. This would be consistent with PLAT-I rotating
about the mooring turret during operation, but primar-
ily located around the mean flow direction as discussed
in [4].
This is also reflected in the yaw data in the model,
with the reported heading varying over a 160◦range,
with a standard deviation of 8.6◦based on the 1Hz
input data. This distribution reduces to a range of 51◦
and a standard deviation of 6.7◦when averaged over
two minute intervals. The two minute average values
of platform position have been used to generate the
plots shown in fig. 7, with the head of the arrows
representing the approximate location of the bow of
the platform.
L. Rotors Parked
Comparing the distribution of position with the ro-
tors parked (fig. 5) to the distribution for the opera-
tional state, the most immediate difference is that the
distribution forms an ellipse around a central point -
this shows the rotation of the platform about the turret,
with an approximately 14m radius between the turret
and the GPS antenna position being reported here.
The distribution has two distinct higher density re-
gions - one to the south west in a similar location to
the operational case, and one to the north east, almost
but not quite diametrically opposite. These regions
represent the dominant position during ebb and flood
phases of the tide respectively. The distribution shows
more occupied bins to the north than the south, which
is consistent with the platform rotating more often to
the north of its moorings when transitioning during the
changing tides. The higher density region associated
with ebb tides is slightly larger than the corresponding
region in the operating case - it is likely that the lower
thrust in the parked state allows the platform to more
easily move in the lateral direction.
The motion model estimates position and heading,
but there is no assumption imposed that these are
correlated. However, we observe that the platform
heading does appear to be correlated with position
in a manner consistent with the both position and
heading depending on the dominant flow direction
and mooring design. As could be expected from the
position plots, the range of headings recorded with the
rotors parked covers the full range from 0◦to 360◦.
The region to the south west of the reported po-
sitions shown for parked rotors in fig 7 shows a
predominantly easterly heading, with the region to the
north east showing a predominantly westerly heading,
reflecting the approximate flow directions for the ebb
and flood respectively. It can be seen that a cluster of
values toward the north west of the rotors parked plot
in fig. 7 appear to be pointing away from centre - this
seems improbable based on the platform position and
mooring design and appears to be an artefact of the
averaging process. This requires further investigation.
M. SDMs Raised
The last case represents a maintenance state, with
the turbines raised out of the water. This reduces the
drag on the platform, and examination of the position
data presented in [4] shows larger deviations from
LAKE et al.: MOTION TRACKING OF A FREE-YAWING FLOATING TIDAL STREAM TURBINE PLATFORM
the mooring centre in this case due to this reduction.
Comparing the higher density region associated with
ebb flows to the equivalent region in the operating and
parked cases, is slightly further north, with a narrower
east-west distribution and larger north-south spread.
This would be consistent with the increased range
of motion discussed in [4]. The whole distribution is
also narrow in the east-west axis, which would be
consistent with the reduced drag placing less load on
the moorings and allowing the platform to sit closer to
the centre.
The platform orientation data for this case provides
a very similar plot to that for the rotor parked case.
Numerically, the range of headings reported based on
two minute average values is smaller than in the rotors
parked case (295◦compared to 352◦)- largely due to the
reduced presence of the platform towards the south of
the area.
VI. CONCLUSIONS
The work shown here illustrates that a Kalman filter
can be used to combine recorded position and motion
data and reconstruct the motion and orientation of
a floating tidal device. Some of the challenges that
can arise when dealing with sensors over a longer
duration deployment and methods for recalibrating
magnetometer data in post processing have also been
shown. Despite this, the results presented are consis-
tent with existing published work and the expected
behaviour of the platform and provide a solid basis
for future development and analysis.
Following from this work, it is intended to extend the
filter to account for full three dimensional movement.
This will require the magnetometer data to be fit and
transformed onto a sphere rather than a circle (to give
a pitch and roll reference), and for the remaining data
in other sensor axes to be incorporated.
The model should also be extended to incorporate
the remaining sensors and take advantage of the re-
dundancy this gives. There are a number of approaches
that can be taken here, but they can include either the
use of multiple Kalman filters to process each input to
the model (i.e. Filter all x-axis acceleration to give a
single reading input into the motion model) or to run
the full motion model in parallel with some filtering of
the resultant data.
Since collecting the data used in this paper, PLAT-I
has been re-deployed at Grand Passage, Nova Scotia.
Further work will also include applying the model
presented here to the data collected in this second
deployment.
ACKNOWLEDGEMENT
The authors would like to thank Sustainable Marine
Energy for allowing the use of their platform to collect
the data presented here, and their ongoing support
with future work.
REFERENCES
[1] Sustainable Marine Energy. [Online]. Available: https:
//sustainablemarine.com/
[2] Race Technology Ltd. (2015) DL1 PRO/WP and DL2
Introduction. [Online]. Available: https://www.race-technology.
com/wiki/index.php/DL1PRO-WP/Introduction
[3] R. Starzmann, I. Goebel, and P. Jeffcoate, “Field performance
testing of a floating tidal energy platform - part 1: Power
performance,” in Asian Wave and Tidal Energy Conference, 2018.
[Online]. Available: https://www.researchgate.net/publication/
326572960 Field Performance Testing of a Floating Tidal
Energy Platform-Part 1 Power Performance
[4] P. Jeffcoate and N. Cresswell, “Field performance testing
of a floating tidal energy platform - part 2: Load
performance,” in Asian Wave and Tidal Energy Conference, 2018.
[Online]. Available: https://www.researchgate.net/publication/
326573036 Field Performance Testing of a Floating Tidal
Energy Platform-Part 2 Load Performance
[5] R. E. Kalman, “A new approach to linear filtering and prediction
problems,” Journal of Basic Engineering, vol. 82, no. 1, p. 35, 1960.
[6] L. Drolet, F. Michaud, and J. Cote, “Adaptable sensor fusion
using multiple kalman filters,” in Proceedings. 2000 IEEE/RSJ
International Conference on Intelligent Robots and Systems (IROS
2000) (Cat. No.00CH37113). IEEE, 2000.
[7] R. Labbe. (2018) Kalman and Bayesian Filters
in Python. [Online]. Available: https://github.
com/rlabbe/Kalman-and-Bayesian-Filters-in-Python/tree/
e84f8018366438c87189ccad40a56bf506f81ffc
[8] M. Kok, J. D. Hol, T. B. Sch¨
on, F. Gustafsson, and H. Luinge,
“Calibration of a magnetometer in combination with inertial
sensors,” in 15th International Conference on Information Fusion, pp.
787–793.
