Quarterly Report
On the Contributions from
MIT
To the
Electric Ship Research and Development Consortium
Sponsored by the Office of Naval Research
Under Award No: N00018 -1-0080
and Coordinated by Florida State University
Period Covered: October 1, 2008- December 31, 2008
Submitted By:
Chryssostomos Chryssostomidis
Franz Hover
George Karniadakis
James Kirtley
Steven Leeb
Michael Triantafyllou
Table of Contents
1.
EXECUTIVE SUMMARY................................................................................................................................5
2.
TECHNICAL DETAILS BY TASK OR PROJECT: .....................................................................................8
3.
PUBLICATIONS AND REPORTS RESULTING FROM ONR SUPPORT .............................................41
4.
PERSONNEL STATISTICS DURING THIS REPORTING PERIOD......................................................44
4.1.
Total No of Co-PI ......................................................................................................... 44
4.2.
Total No of students working on the grant ................................................................... 44
2
Table of Figures
Key accomplishments
HOVER:
Figure 1. (untitled)
Technical detail
CHRYSSOSTOMIDIS/KARNIADAKIS:
Figure 1. Sketch of DDG 51 Flight I (left) and IPS components (right).
Figure 2. Torque and angular velocity variation (top row), and thrust loss and ship speed (lower
row).
Figure 3. Motor (upper) and rectifier (lower) current variations due to propeller emergence.
KIRTLEY:
Table 1. Comparison between reported values and computer model of high power induction
motors
Table 2. Comparison of calculated size, weight, and full load performance of three designs based
on the existing 19 MW motor
Table 3. Change in predicted full load performance with speed for the 20 MW motor design
LEEB:
Figure 1: Program flow diagram. Shows detect-classify-verify logic used for incoming power.
Figure 2: Steady-state changes for turn-on and turn-off of a computer and a lamp (see text).
Figure 3: USCGC Escanaba CHT GUI
Figure 4: Real-time NILM monitoring CHT on the USCGC Escanaba.
Figure 5: Finite state model for the CHT system.
TRIANTAFYLLOU:
Figure 1. Schematic of three phase AC powered propulsion system.
Figure 2. Controlled gas turbine model.
3
Figure 3. Equivalent circuit for coupled electric machinery.
Figure 4. Traditional rectifier-inverter.
Figure 5. Black box equivalent of the traditional rectifier-inverter.
Figure 6. An example of an induction motor torque-speed curve, for a constant voltage, Vm.
Figure 7. Volts-per-Hertz speed control for induction motor.
Figure 8. Energy conversion efficiency between gas turbine and synchronous generator.
Figure 9. Energy conversion and storage in power electronics.
Figure 10. Energy conversion efficiency across induction motor.
Figure 11. Energy conversion efficiency between propeller and translational ship motion.
4
1. EXECUTIVE SUMMARY
CHRYSSOSTOMIDIS/KARNIADAKIS:
We have developed a new computational framework consisted of detailed models for the two
primary components of the All-Electric Ship (AES), namely Integrated Power System (IPS) and
Hydrodynamics. This is the first time that all components of each system have been modeled and
coupled together at this level. Simulating such a coupled system on today's computers is feasible,
and results can be obtained in a matter of minutes even for long-time integration of the system to
include the characteristic electric, mechanical and hydrodynamic scales. We note that these
characteristic time scales, even of the small subcomponents, are not input to the problem but they
are part of the solution, unlike other models where simplified lumped parameterizations are
employed. This new computational framework is applied to a model of the DDG-51 destroyer
that involves a 19 MW 15-phase induction machine (IM) and an indirect field oriented controller
(IFOC). In particular, we simulate the extreme event of propeller emergence.
HOVER:
We investigate the robustness of integrated power system layouts from a network theoretic
perspective. We find that an abstracted electric ship model based on a notional cruiser line
diagram behaves more like an Erdos-Renyi random network than a scale-free network, and
furthermore that scale-free networks are significantly less robust than random and notational
cruiser-type networks, under the loading and failure conditions considered. A second area of
research has been particle filtering. Specifically, we apply stochastic collocation, a tool we
previously used for uncertainty analysis in the AES, to make sampling from proposal
distributions more efficient, and hence increase the real-time applicability of particle filters.
KIRTLEY:
Our efforts this past quarter focused on the design and evaluation of motors to match reported
specifications and performance for two different Alstom machines, a 19 MW, up to 150 RPM,
motor and a 20 MW, up to 180 RPM, design. An IEEE paper, “The Advanced Induction Motor,”
by Clive Lewis from the 2002 IEEE Power Engineering Society Summer Meeting describes both
5
machines. As a physically realizable design, previously built and tested, the 19 MW machine
provides a baseline for verifying the MIT motor evaluation software on a low speed, multimegawatt motor as well as a realistic framework for examining design modifications. The initial
designs were also altered for comparison with two different cases, first doubling the initial motor
power output, speed, and frequency, and second doubling the speed and frequency while
maintaining the original power output but with a reduced motor size and weight. After designing
around the full load operating points, changes in motor efficiency with varying speed and
corresponding load were also compared between designs.
LEEB:
Power system monitoring is an exciting approach for creating an inexpensive, highly capable
“black-box” for monitoring the performance of critical shipboard systems. With remarkably
little installation effort or expense, we have fielded a non-intrusive load monitor (NILM) that can
reliably monitor and track diagnostic conditions for multiple devices. During the past quarter, we
have worked to automate the recognition of load operation and diagnostic monitoring to make
results available to the crew in real-time. To do so, we have developed a software package
known as ginzu that eliminates the need for off-line analysis by a skilled observer. Initial
tracking of load operation and diagnostic condition are now provided automatically by the NILM
on-board ship. Furthermore, our field-tested systems have been installed at a central point that
allows them to monitor multiple loads simultaneously. The ginzu software application
implements a detect-classify-verify loop that locates electrical load transients, identifies them
using a decision-tree-based expert classifier, and then generates event files that contain relevant
information. Additionally, the ginzu application provides streaming data to a graphical user
interface known as the Ginzu Graphical User Interface (GinzUI)
TRIANTAFYLLOU:
We developed robust models for ship hydrodynamics, including forward propulsion and
maneuvering, as well as models for the propeller and the linear and nonlinear loads in a random
sea-state; models for the gas turbine, generator, electric system, and motor. The models were
developed at various levels of accuracy so as to assess the principal time constants and
6
parameters of the system, and thorough testing of the power flow and transient dynamics was
made. Controllers for the gas turbine, generator and motor were developed and tested. This led
to a system that can be parameterized in terms of the significant parameters so as to perform
sensitivity analysis and stochastic simulation. Simulations that involve random loads from the
sea, random events such as propeller emergence, and maneuvering motions in a storm are
possible with the models developed. Sensitivity tables exhibiting the parameter to parameter
finite-amplitude sensitivity have been developed that allow time-varying assessment of the
critical parameters.
7
Technical Details by Task or Project:
•
Key Accomplishments
HOVER:
In particle filtering, we have applied stochastic collocation to allow faster sampling of proposal
distributions and thus enhanced real-time applicability. Particle filters, also known as bootstrap
filters, sequential Monte Carlo, and survival of the fittest, are a general solution to the recursive
Bayesian estimation problem. Particle filters differ fundamentally from Kalman filters in how
they represent the distribution of the state. Kalman filters propagate only the mean and
covariance of the state because Gaussian distributions are uniquely determined by their first two
moments. In contrast, particle filters approximate arbitrary distributions with sets of weighted
points, i.e. particles. A significant shortcoming is computational cost, however; particle filters
sacrifice efficiency for generality. Usually a large number of particles must be carried to obtain
an accurate representation of the distribution. The proposal distribution is a probability
distribution (typically derived from Bayes rule and importance sampling) which when sampled
from is a mapping from the state space to the state space at the next time step. This proposal
distribution can significantly affect the efficiency of the particle filter – well designed proposals
require significantly fewer particles. However, better proposals are often more expensive to
sample from. Here we have applied stochastic collocation to the proposal distribution, which in
our examples include simulation of the system and Kalman filter-like calculations. We achieve
two fold computational savings on a standard, easy to evaluate one-dimensional benchmark
system, and order of magnitude savings on the significantly more expensive Kraichnan-Orszag
system.
We have applied statistical network theory to assessing the robustness of electric ship models.
We consider the following two specific scenarios: an intentional random attack in which high
degree nodes are removed with greater probability than lower degree nodes, and a geographic
attack in which a large cluster of physically neighboring nodes is removed. The distribution of
power to loads is computed by a simplified network flow, large-scale linear programming
optimization, and nodes linked to failed nodes are then failed at successive iterations with
probability proportional to the flow in the connecting link. We find that Erdos-Renyi networks
8
are significantly more robust in this scenario than scale-free networks and slightly more so than
the notational cruiser type networks. Shown in the figure below is a histogram from Monte Carlo
trials of the ratio of load served after (LC) and before (LN) the intentional random attack failure
scenario – this is a measure of how much functionality has been lost. It is clear that the random
network is the most robust in this situation than the notional cruiser (labeled “CGX” in the
figure).
Figure 1. (Hover)
KIRTLEY:
We have created a computer model of the existing 19 MW Alstom induction motor that has been
corrected to match total machine weight and outer envelope dimensions and closely agrees with
the reported machine performance. This initial baseline machine provides both a check on our
induction machine evaluation software for this type of low speed, high power motor as well as a
9
realistic starting point for design modifications. After developing the initial baseline machine, the
motor speed and frequency were doubled to examine the possibilities of either doubling the
power output using nearly the same machine design (with half the stator turns for consistent air
gap magnetic fields), or maintaining the same power output but at a reduced size and weight. An
additional 20 MW, up to 180 RPM design was also evaluated and compared to a 40 MW, up to
360 RPM, version. Results for this second generation, smaller, higher power density machine
again seem consistent with the reported values of motor size and performance. However, it is
unknown if this machine has been built and tested or only simulated, making these designs less
reliable until a more detailed examination of motor temperature rise and cooling system is
included in our evaluation software.
All of the initial designs focused on the full load steady state motor operation, but the actual
operation of motor and propeller will vary over a wide range of speed and load. The maximum
efficiency for the three machines based on the existing, 19 MW motor were also evaluated over a
range of speeds from 20% up to full speed, with each speed corresponding to a particular motor
load. Assuming a ship driven by two propellers of equal speed, each with a maximum load of 38
MW at full speed, higher efficiency and lower losses were predicted at every load point using the
single motor with twice the original motor speed and power output. Using dual 19 MW motors
on each shaft, equally dividing the total load between the two machines gave higher efficiency
than more fully loading only one of the machines for loads less than 50%. However, the even
more lightly loaded 40 MW machine consistently outperformed either operating strategy of the
19 MW motors.
LEEB:
We have demonstrated the first real-time NILM that has provided diagnostic information in near
real-time to a serving military crew. The commissioning process for the NILM requires system
knowledge, but is not onerous. It has been shown that the NILM is capable of performing as a
stand-alone diagnostic tool. The performance on the USCGC Escanaba indicates that the NILM
is a maturing and capable technology that could work well supporting ICAS and other conditionbased maintenance efforts in the USN and USCGC.
10
•
Technical Detail
CHRYSSOSTOMIDIS/KARNIADAKIS:
1. SYSTEM MODELING
To investigate the influence of the propeller loading and the pulsed power load on the AC
propulsion and power generation in the AES, modeling of both power system and
hydrodynamics of ship and propeller must be introduced. In our study, the main AC subsystem
of the IPS consists of a 21 MW, 3-phase, 60Hz Synchronous Machine (SM) as a generator, a
close-loop drive using the indirect rotor field oriented control of 19 MW, 15-phase Induction
Machine (IM) as a propulsor, and a pulsed power charger for an EM gun. The generator supplies
the voltage to the 3-phase 4160V AC distribution bus, which must be filtered by a harmonic
filter before connecting to the induction motor drive. A power converter and the Indirect Field
Oriented Control (IFOC) using the sine-triangle modulation technique for a current regulator are
incorporated into the propulsion system to control the torque of the induction machine. The IPS
configuration is similar to a split plant arrangement shown in Figure 1.
The propeller model is based on a five-blade, fixed-pitch, highly skewed propeller with a
maximum skew angle of 32 degrees. Propeller emergence can abruptly reduce the propeller
thrust because of a loss in the propeller effective disc area. In particular, when the ship operates
in rough sea conditions, the propeller emergence can occur intermittently, resulting in ship speed
reduction, sudden increase in shaft angular speed, and rapid decrease in motors' current.
Propeller thrust and torque loss can be represented by the appropriate factors available in the
literature. To model the wave effects, first, we assume that nonlinear and viscous effects are
small compared to wave inertia for a ship motion in the sea. Moreover, we assume that deep
water random waves are modeled through the one-parameter Pierson and Moskowitz spectrum,
derived on the basis of North Atlantic data. To model ship hydrodynamics, we must consider the
interaction between the propeller and ship hull because the wake created by the hull modifies the
propeller advance velocity. Moreover, the presence of the propeller at the stern also increases
the drag force on the vessel; thus, the propeller thrust must be decreased by an appropriate factor.
11
In the equation of motion we allow only surge in the present study but we include both calm
water and added resistance (including second-order effects). Moreover, we assume that the
electric ship is driven by two propellers connected to two identical IPSs, in head sea. We note
that the total propulsion system of DDG-51 consists of four identical IPS and achieves 31 knots;
with two IPS it achieves a maximum speed of about 14 knots. The ship's motions in other
directions i.e., sway, heave, roll, pitch, and yaw are assumed to be small.
Vw : Wave
x
Vs : Forward
λ
Wave
IPS
Prim
S
DC
Zone
Prim
S
s
Y
Driv
Driv
I
I
Figure 1: Sketch of DDG 51 Flight I (left) and
IPS components (right).
The first major component of IPS is the synchronous generator or machine (SM), which supplies
electric power to AC power distribution and propulsion system. To maintain the bus voltage at a
specified level, an exciter is needed to feedback the bus voltage to the generator. The
propulsion system of IPS is driven by a 19 MW, 15-phase (squirrel cage) IM. The larger number
of phases in the induction motor is, the larger the electromagnetic torque is and hence a more
reliable motor operation can be achieved. In this study, the model of the 15-phase induction
machine is represented by three 5-phase equivalent circuits. Control of the 15-phase IM drive
requires an input torque command and a desired rotor flux command and outputs semiconductor
signals, driving the H-bridge type inverters. Because of the multiple phases in the induction
motor, the power converter topology must be separated into three five-phase parallel rectifier-dc
link-inverter paths. As a result, each of the five phases of the induction machine can be
independently controlled using the paralleling control. In this study, maintaining the IM's torque
at a specified level is the main objective of IFOC.
12
2. RESULTS
The interaction between the ship speed, propulsion load, and electric machines' states are
examined for sea state 6 (very rough). Two identical IPS, as shown in Figure 1, with the torque
control using the IFOC, directly connected to the propeller, drive the 1DOF ship motion in head
seas. We note that this is a simplified model and does not correspond to the full DDG-51
configuration which requires 4 IPS to reach 31 knots. Here we further assume that initially two
generators operate near their steady-state conditions and then after 2 seconds the two propulsion
drives are turned on while the ship moves forward at about 20 knots. The mean added resistance
and the slowly-varying added resistance of the scaled electric ship are computed using highorder models, hence, the propellers driven by IPS experiences the slow-varying force from the
added resistance.
Next, we examine the specific effects of propeller emergence that causes a load torque reduction
on the IM drive. The propeller emergence leads to an increase in motor's or propeller's angular
speed by about 7% and a minor decrease in ship speed, as shown in Figure 2; these results
indicate the effectiveness of the IFOC. We also plot in Figure 3 typical currents for the IM,
which at first decreases and subsequently increases while the propeller is out of the water.
However, the rectifier’s phase current, also shown in Figure 3, increases slightly to compensate
for the abrupt increase in IM's speed.
Figure 2: Torque and angular velocity variation (top), and thrust loss and ship speed (lower row).
13
Figure 3: Motor (upper) and rectifier
(lower) current variations due to
propeller emergence.
KIRTLEY:
Table 1. Comparison between reported values and computer model of high power induction
motors
Rated Power [MW]
Approx. Full Load Torque [kN m]
Airgap Shear Stress [kN/m^2]
Rotor R^2 L Volume [m^3]
Envelope Length [m]
Envelope Width [m]
Envelope Height [m]
Total Volume [m^3]
Total Weight [kg]
19 MW, up to 150 RPM
Estimated(1)
Published
19
19
1210
1210
76
76
2.53
2.53
4.8
4.81
4.5
4.5
4
3.99
86.4
86.25
117e3
117e3
20 MW, up to 180 RPM
Published
Estimated
20
20
1061
1061
100
101
1.71
1.72
3.3
4.7
3.6
3.7
3
3.27
35.64
56.96
89e3
84.5e3
(1) Envelope dimensions and weight for the 19MW motor are corrected to match the reported values. The
same correction factors are applied to the 20MW motors.
The calculated motor performance matches the expected values with efficiencies in the mid to
upper 90s and power factor around 0.85. The motor models are 12 pole machines and include the
14
reported 8 mm air gap required for shock resistance. Initial calculations focused on the
electromagnetic design of the motor itself while the outer envelope dimensions and total machine
weight must also account for the motor shaft, rotor spider, bearings, coupling, heat exchangers,
reinforced housing, and foundation. These additional requirements contribute significant size and
weight to the total motor package. Correction factors of about 1.3 and 1.15 increased the overall
stator outer diameter to match the reported machine total width and height. The overall length
was increased more significantly, by a factor of 2.25. This is more extreme, but published photos
and finite element motor models suggest that the motor stack length is less than half of the outer
envelope length (and show significant heat exchanger volume and frame reinforcement). Plus,
the known rotor volume from the reported air gap shear stress gives a good approximation of the
stack length for reasonable rotor designs. After including estimates for the weight of the shaft,
rotor spider, and frame, a correction factor of about 1.20 increased the calculated weight to
match the total reported weight, accounting for the heat exchanges, bearing, and additional
frame. The same process and correction factors applied to the 20 MW design produce a good
approximation of the machine, except for the overall length. The reported 3.3 m “frame length”
probably does not include the additional space required for the section of shaft, bearing,
coupling, heat exchangers, and housing included in the 19 MW design.
Table 2. Comparison of calculated size, weight, and full load performance of three designs based
on the existing 19 MW motor
Full Load Efficiency (2)
Full Load Power Factor
Rotor Outer Diameter [m]
Stator Outer Diameter [m]
Motor Stack Length [m]
Overall Motor Length [m]
Motor Active Material Weight [kg]
Corrected Weight Estimate [kg]
Corrected Length [m]
Corrected Width [m]
Corrected Height [m]
Total Corrected Volume [m^3]
19 MW, 150 RPM
0.965
0.857
1.45
3.5
1.41
2.17
51e3
117e3
4.81
4.5
3.99
86.25
19 MW, 300 RPM
0.973
0.851
1.45
3.5
0.6
1.46
26.75e3
75e3
3.25
4.5
3.99
58.24
38 MW, 300 RPM
0.981
0.856
1.45
3.5
1.41
2.17
51e3
117e3
4.81
4.5
3.99
86.25
(2) Estimated core, bearing friction, and windage losses have not been verified against test data for this size and type
of machine
15
The 38 MW, up to 300 RPM, design is very nearly identical to the 19 MW, up to 150 RPM,
design except with double the operating frequency and output power along with half of the
number of turns in each stator winding coil to give consistent air gap magnetic flux density. The
19 MW, up to 300 RPM, design has half of the original active motor length, but the diameters
were unchanged in order to avoid increased heating or decreased power factor. The calculated
motor efficiency increases slightly with increased speed and power output.
Table 3. Change in predicted full load performance with speed for the 20 MW motor design
(2)
Full Load Efficiency
Full Load Power Factor
20 MW, 180 RPM
0.947
0.846
1.15
2.85
1.52
2.12
39e3
84.5e3
4.7
3.7
3.27
56.96
Rotor Outer Diameter [m]
Stator Outer Diameter [m]
Motor Stack Length [m]
Overall Motor Length [m]
Motor Active Material Weight [kg]
Corrected Weight Estimate [kg]
Corrected Length [m]
Corrected Width [m]
Corrected Height [m]
Total Corrected Volume [m^3]
40 MW, 300 RPM
0.969
0.846
1.15
2.85
1.52
2.12
39e3
84.5e3
4.7
3.7
3.27
56.96
(2) Estimated core, bearing friction, and windage losses have not been verified against test
data for this size and type of machine
Again, the calculated motor efficiency improves with the increased speed and power output.
The following plots show the motor efficiency and total losses as the speed varies from 20% to
full speed and corresponding motor load vary from less than one percent to full load. For these
initial plots, the shaft and ship speed are assumed to be linearly related, and the turbine and
motor output power is assumed to vary with the turbine or ship speed cubed. The required range
of output speed and power is achieved by varying the input electrical frequency and voltage.
Operating points for maximum efficiency were determined over the operating range while
avoiding machine saturation by limiting the terminal voltage to no more than the rated value of
16
volts per hertz. Substantially lower voltage improved efficiency (and power factor) for the lower
power points below about 15% of the full load.
At every operating point the calculated efficiency of the single 38 MW motor at up to twice the
baseline speed exceeds both cases of dual, 19 MW motors, up to either the baseline or twice the
baseline speed with reduced motor dimensions. The higher speed, higher power machine
improves performance, even when running at low power levels.
LEEB:
In general, NILM classification methods have focused on identifying system-specific events
based on signal characteristics. The methods implemented in the ginzu classification software
compare the shape characteristics of a transient to shape characteristics of known events.
Specifically, the shape characteristics are defined as (1) the relative steady state power change
across the transient event index and (2) the shape of the spectral envelope during the transient.
The comparisons are aided by continuously tracking the state of the system (i.e. the running
status of the known motors and other electric components in the system) and limiting the
classification decisions to only those permitted by the associated finite state diagram of possible
operating conditions.
17
Figure 1 shows a simplified flow diagram of ginzu’s program logic. The algorithm initializes by
loading a 10 second data window consisting of relevant spectral envelopes, e.g., corresponding
to in-phase and quadrature (“real” or “P” and “reactive” or “Q”) components of current. This
window is passed to a detection algorithm that locates vector indices where rapid changes in the
envelopes have occurred. These indexes represent system transients and are candidates for
classification.
Start
Change-of-Mean
Filter
Is an
Yes
Classify
event
No
d t t d?
Verify
Event
State
Create
Event
End
Figure 1: Program flow diagram. Shows detect-classify-verify logic used for incoming power.
Once an event has been detected, the classifier may be called. The classifier implements a
hierarchy of classification decisions to make a ‘best guess’ based on the relative power levels
around the event, the state of the system prior to the event, and if possible, the correlation
between the shape of the power signal during the event and the shape of a known library event.
On the other hand, if a rapid power change is not detected, a state verification and correction
function is called. This function attempts to verify that the current power levels are consistent
18
with what the ginzu algorithm anticipates them to be based on current system state. The
algorithm then reads one additional period of data from the input buffer; this data is inserted into
the P/Q buffer and the old data is discarded. This new P/Q window is then passed to the detectclassify-verify loop, and the cycle is repeated.
The following sections provide an overview of the main components of the ginzu software along
with a brief description of the GinzUI application. For in-depth discussion refer to Proper
(2008).
EVENT DETECTION
The preprocessor located upstream in the program flow of the ginzu software provides spectral
envelopes for fundamental and higher harmonic content at a sample rate of 120 Hz. Therefore,
the ten second data windows form several 1200 index arrays. For example, one array contains
“real” power and another “reactive” power. The 1200 index power array is passed to the
detection algorithm where rapid power changes are located. This is accomplished by using a
change-of-mean filter that calculates the difference between the original power signal and the
output of a low pass filter. The result is a processed signal that only contains rapid power
changes. Ultimately, these power changes are compared to pre-determined detection thresholds
(based on the monitored system’s characteristics). The output of the comparator is an index of
‘Event Detections’.
CLASSIFICATION TECHNIQUES
The ginzu software algorithm recognizes events by examining changes in both steady-state
consumption levels and also transient shape.
STEADY-STATE POWER CHANGE
When individual loads are cycled within the system (i.e. pump on/off), they produce a
corresponding change in the real envelope (∆P) and reactive envelope (∆Q) and possibly other
spectral envelopes. These changes can be used as a simple classifier to identify loads.
19
It is advantageous to look for changes in steady-state levels in as many spectral envelopes as
contain useful information. This is illustrated in Figure 2, which shows steady-state power levels
after turn-on and turn-off of two loads, a computer and a lamp. The top plot in Figure 2 shows
changes in Q versus P. The middle plot shows the change in third harmonic content versus P.
The bottom plot shows change in Q versus third harmonic content. Notice that in the P-Q space
(top graph), the two loads are essentially indistinguishable. They both turn on and consume
approximately 150 watts in steady state, with essentially no reactive power. Similarly, they turn
off with a -150 watt steady-state change as expected. Observations summarized in the top graph
Δ 3rd Harm
(In-Phase) ΔQ (VAR)
alone could not be used to differentiate the operation of the two loads.
100
0
-100
-200
-150
-100
-50
-150
-100
-50
-150
-100
150
50
100
150
200
0
50
100
150
200
0
50
100
150
200
0
-150
-200
ΔQ (VAR)
0
ΔP (W)
100
ΔP (W)
0
-100
-200
-50
Δ3rd Harm (In-Phase)
(W)
Computer On
Incandescent On
Computer Off
Incandescent Off
Figure 2: Steady-state changes for turn-on and turn-off of a computer and a lamp (see text).
The computer, however, draws a third harmonic current, distinguishing it from the lamp. The
middle and bottom traces in Figure 2 show that this difference makes it easy to detect the
operation of the computer with respect to the lamp in an information space that includes as many
useful spectral envelopes as can be reliably recorded.
20
TRANSIENTS
The ginzu algorithm also classifies individual loads based on distinctive load transient shapes.
Overall transient profiles tend to be preserved even in loads that use active waveshaping or
power factor correction. Most loads observed in the field have repeatable transient profiles, or at
least sections of the transient profile that are repeatable.
Transient-based recognition permits near-real-time identification of load operation, especially
turn-on events. Transients are identified by matching events in the incoming aggregate power
stream to previously defined transient signatures, or “exemplars.” Exemplars can be determined,
for example, by a one-time direct observation of the device in question, or by previous training in
the laboratory. Pre-training has proven to be a reasonable approach for very repeatable loads that
show up in large quantities, such as fluorescent lamp ballasts. The exemplar may be comprised
of multiple parts for loads whose transients have a number of distinct sections. Each section of
the exemplar can be shifted in time and offset to match incoming transient data. In addition, an
overall gain may be applied to all sections of the exemplar to achieve a better fit. Each event
detected is compared to the full set of exemplars by using a least squares criterion to select the
appropriate shifts and gains. The match with the lowest residual norm per number of points is
then compared to a threshold. If the fit is good enough, the event is classified as a match to the
exemplar. If not, the event is left unclassified.
Correct classification of overlapping transients is possible using properly designed exemplars.
Fingerprint traces provide positive identification of specific events occurring during system
operation. By comparing the shape of the transients to known system events, a numerical score
can be assigned to grade the degree of similarity of the two signals; this score is known as the
correlation score. It is derived using the method of least squares in the ginzu algorithm. This
method is discussed in detail in Lee (2003) and Proper (2008).
STATE VERIFICATION
If no transients are detected within a given window, the classifier does not need to be called.
Ginzu uses this opportunity to verify the current state of operation. This is accomplished by
21
calculating the average power level for a ten second window and its standard deviation. These
values are used in a state verification function to perform various checks and correct the state
status if needed.
GinzUI
The GinzUI application provides the interface between the event file and the NILM user. Figure
3 illustrates the front-end display used on board the USCGC Escanaba. The primary functions of
GinzUI are:
- To continuously check the user interface directory for newly created event files.
- To read event file contents and move the event files to an archive directory.
- To peform diagnostics on event file data and alert the user if a diagnostic has failed.
- To allow the user to graphically view event file contents.
Figure 3: USCGC Escanaba CHT GUI
CHT TEST SYSTEM
To demonstrate the ability of ginzu to monitor multiple loads, we installed a NILM at the service
entry to the power panel supplying the Shipboard Waste Collection and Disposal System (CHT)
onboard the USCGC Escanaba. The CHT system represents a common shipboard auxiliary
system used to transfer sewage from installed heads to a sanitary tank where it is pumped
overboard.
22
NILM has been monitoring the CHT system since 2003 and various problems have been detected
and classified through the application of NILM signal analysis. The CHT system operation and
performance has been detailed in previous research (Mosman 2006, Piber 2007).
Figure 4 shows the prototype installation aboard the Escanaba. The NILM system with ginzu
software runs completely on the touch tablet computer shown in Figure 4, installed next to the
power panel serving the CHT system. The crew can interact with the NILM through the touch
screen, both through diagnostic reports and also by requesting additional data and analysis from
the NILM.
Figure 4: Real-time NILM monitoring CHT on the USCGC Escanaba.
The CHT system consists of a 360 gallon sewage collection tank, which collects drains from
eighteen vacuum toilets, two urinal lift valves, one urinal non-lift valve, and the ship’s garbage
grinder. There are four pumps associated with the system including two vacuum pumps and two
discharge pumps. A number of other ancillary single-phase loads (i.e. lamps and small motors)
are also installed.
The vacuum pumps are each rated at 1.5 horsepower and connected upstream to the top of the
collection tank and downstream to the vacuum seal tank. Their function is to maintain the
necessary vacuum on the system for proper operation. If pressure falls below 14 in-Hg, one of
23
the pumps automatically turns on to increase vacuum within the tank. The pumps alternate
operation in order to minimize wear. If pressure falls below 12 in-Hg, both pumps are energized
to restore proper vacuum pressure. Pumps are automatically secured when pressure reaches 18
in-Hg.
CHT SYSTEM STATE RECOGNITION
As previously mentioned, the system states can be defined by measuring the real power usage.
This approach was applied to the CHT system. The final allowable states that were defined for
the CHT system are shown in Figure 12. By identifying the most likely state transitions and
tracking these states, the classifier algorithm can be tuned so that the most likely transitions are
given additional consideration.
One additional note on Figure 5 is that it includes the most common transitions from state to
state. In other words, if an ON event is detected while both vacuum pumps are already running,
the event cannot be a vacuum pump ON. As stated in the previous section, the state information
can be combined with the power change information to create accurate classifiers. Consider an
event where the post-event power was zero. Any loads known to be operating before the change
should now be classified as OFF.
2 Vacuum Pumps On
0 Discharge Pumps On
2 Vacuum Pumps On
1 Discharge Pump On
1 Vacuum Pump On
0 Discharge Pumps On
1 Vacuum Pump On
1 Discharge Pump On
All Pumps Off
0 Vacuum Pumps On
1 Discharge Pump On
Figure 5: Finite state model for the CHT system.
24
CHT DIAGNOSTIC PACKAGE
A rudimentary diagnostic package was included in GinzUI to provide real-time detection of CHT
system faults that had been observed by Mosman (2006) and Piber (2007). These failures
include clogs in gauge lines and/or priming orifices, tank level probe failures, and system leaks.
When a system fault is detected, a comment is printed to the diagnostic log and the log turns
yellow to indicate an abnormal condition. Additionally, a line is printed to a text file that
contains the description of the failure and the time of the detection.
TRIANTAFYLLOU:
MODEL DESCRIPTION
In this section, the all-electric ship model that we use for further research into propeller design is
presented. This model is intentionally kept simple and low-order to be conducive to fast
simulation and post-processing/analysis. However, the model preserves the fully-coupled and
complex nature of the actual all-electric ship. The model considers the following subsystems: a
prime mover (gas turbine), an AC synchronous generator, an AC induction motor, a power
electronic control system, and the hydrodynamic systems including the propeller and ship hull
surface. A general schematic for the system is shown in Figure 1; the circuitry is shown in detail
in the following subsections. Because the electrical states have time constants orders of
magnitude less than the mechanical states, a steady-state assumption is used to reduce the state
space to turbine shaft speed, propeller shaft speed, ship speed, and field current,
X1=(n_t,n_p,u,i_f). The control inputs for the system are the turbine fuel flow rate and the
generator excitement voltage U1=(f,v_f). With the addition of power electronic control, a fifth
state is added to account for the voltage across the added power storage element,
X2=(n_t,n_p,u,i_f,V_c). This allows for actuation of the induction motor applied voltage and
excitement frequency, U2=(f,v_f,V_m,ω), discussed in more detail below. The primary
disturbance considered in this paper is random torque loading, Qprop, which gets added to the
predicted hydrodynamic propeller load, Qpsp.
25
The target steady-state values for a 10,000 tonne ship are given in Table 1. These values are used
to hone the flexible electrical/mechanical parameters.
GAS TURBINE
Many all-electric ship models assume a constant input speed to the electric generator obviating
the need to model the prime mover. One of the incorporations here is simple gas turbine model
to drive the synchronous generator. This way, we can inspect the effect of propeller load changes
on the gas turbine.
We use a GE LM2500 gas turbine with rated maximum power output of 21.5 MW, a full-locking
torque of approximately 123 kN-m (at full throttle), and a no-load speed of 6875 rpm (at full
throttle). Torque input of gas turbine is linearly interpolated with the input shaft speed and fuel
flow rate
where f/fmax is the throttle ratio.
26
To maintain steady-input power in an efficient range, nt=3600 rpm, discrete PI feedback control
of the turbine shaft speed is used actuate the throttle. The gain values of proportional and integral
control can be optimized for prescribed step changes in shaft speed, but here this is not done;
gain values have been selected to achieve less than 10% overshoot, less than 5% steady-state
error, and less than a 300 s settling time from start-up. A saturation limit, 0 < f < 1, and a slew
rate limitation, |df/dt| < go, are imposed on the fuel flow rate to represent physical bounds and
prevent engine surge due to rapid changes in fuel flow rate. The entire subsystem model is
shown in Figure 2; the emf torque from the generator, Qq, is detailed in the next section.
ELECTRIC MACHINERY
First, the physical models for the electric machinery are presented. The power electronic control
for this machinery will be introduced in the next section. The equivalent circuit for the coupled
AC synchronous generator and the P-pole AC induction motor is shown in Figure 3. Recall that a
steady-state assumption is used for the AC electrical states to allow for the sinusoidal steadystate analysis with complex impedances. For simulations requiring analysis of transience on the
order of tenths of a second, a non-averaged model should be used.
27
The synchronous generator is modeled as an input voltage source, Ea, a stator resistance, Rsg, a
stator winding inductance Lsg, and the separate excitement. The separate excitement is modeled
as a controllable DC voltage source, vf, a field resistance, Rfg, and a field winding inductance,
Lfg. The state equation for the excitement current is
The RMS voltage inputted by a synchronous generator is widely defined as Ea=Mg ωe if/sqrt(2),
where Mg is the mutual inductance of the generator windings, and ωe is the electrical excitement
frequency applied by the turbine on the generator in rad/s (in our case this is equivalent to the
mechanical frequency of turbine shaft because our synchronous generator has a single pole pair).
The synchronous generator produces torque opposing the shaft angular velocity both from
friction and the electromotive force produced by the windings. The friction force, physically
induced in the generator bearings, is modeled to act between the turbine and the generator,
Qfric= μ1 + μ2 ωe so that the power conversation is described Pturb=Pl,fric+Pgen. The
electromotive torque created by the generator must conserve power in the transition from
mechanical to electrical domains,
28
where the numerator represents the real (in-phase) power produced by all three phases of the AC
synchronous generator. The imaginary (out-of-phase) counterpart of Ea Is is the reactive power
in the electrical system; power that flows back and forth in the system.
At start up, the excitement voltage of the generator, vf, is ramped up from zero to avoid large
torque impulses on the turbine shaft.
The induction motor is modeled as a power conservation device instead of explicitly deriving
torque and magnetic flux. The inputs to the system are applied voltage, Vm, and excitement
frequency, ωe. The output of the system is rotor mechanical speed ωp, or rotor electrical speed,
ωr = Pm ωp, where Pm is the number of pole pairs. The equivalent circuit combines stator
resistance, Rsm, and winding inductance, Lsm, with a air gap leakage inductance, Lm, in parallel
with a rotor resistance, Rrm, a rotor winding inductance, Lrm, and a electromechanical
"resistor," Rrm (1-s)/s, where s is induction motor slip and is defined as the normalized
difference between excitement frequency and rotor electrical speed, s= (ωe - ωr)/ωe. The
"losses" in the electromechanical resistor are converted to mechanical power
The current computation is done with sinusoidal steady-state analysis with general impedances
defined Z=R+jωL, where R is a general resistance, L is a general inductance, ω is the frequency
across the inductor, and j is imaginary domain delimiter. The currents are thus defined
29
Finally, the state equations for the shaft angular velocities are derived from Newton's 2nd law for
rotational bodies
where It and Ip are the turbine and propeller shaft moments of inertia, respectively, and Qpsp is
the hydrodynamic torque discussed in the subsequent section.
POWER ELECTRONIC CONTROL
To attain the target mechanical outputs, as outlined in Table 1, with external disturbances or
unforeseen internal parameters, a control system must be added before the induction motor. To
meet this end, we introduce a power electronics rectifier-inverter between the generator and
motor to actuate both the applied voltage and excitement frequency. Traditionally, the full-wave
rectifier is modeled with an arrangement of diodes (uncontrolled) or SCRs (controlled) and the
inverter is modeled as an H-bridge with ideal switches (approximately attained with IGBTs) as
shown in Figure 4.
Simulating this model results in several extra states and certainly more complexity. To maintain
the simplicity of our model while maintaining a high level of fidelity, we use black-boxes for
both the rectifier and the inverter. The inputs and outputs of the black-boxes are accurate in the
30
sense of power conversion, and they approximate what one would expect from the intricate
power electronic devices. Our model, with all important parameters labeled, is shown in Figure
5. Note that generator inductance and resistance is neglected and a resistance, Rp, is added to the
power electronics to have the effect of an inductor acting on dIgDC/dt due to notching. The
convention we use in the below equations is that Ea is RMS and line-to-line.
The first thing to note is the addition of a state: the voltage across the energy storage element,
Vc. The state equation is derived from the voltage-current relationship of an ideal capacitor:
The DC voltage out of the rectifier is VDC=(3/π) sqrt(6) Ea. The sqrt(2) is to scale the RMS
voltage to voltage amplitude. The sqrt(3) is to scale the line-to-line voltage to line-to-neutral. The
(3/π) is a value slightly less than unity to account for non-perfect switching in the rectifier. The
DC current out of the rectifier, IgDC, is found by applying KVL to the loop containing the
rectifier and the capacitor:
31
The DC current into the inverter, ImDC, can be found first by assuming that the voltage across
the H-bridge, on the DC side, is equal to the voltage across the capacitor - a result of KVL. Then
we assert that power is conserved across the inverter resulting in
where the AC current out of the inverter, to the induction motor, is found as in a subsequent
section, Iin=Vm/Ztot. The conversion to mechanical power is computed as in Eqs. 4-6.
Finally, to achieve conservation of power across the aggregate electrical system, the
electromotive torque from the AC synchronous generator is found by assuming power
conservation between the turbine shaft and the rectifier, Qg=VDC IgDC/ωe. In summary, the
mechanical power into the electrical domain becomes DC electric power. This is stored in the
capacitor and eventually all but a finite amount of energy is converted back to AC electric power
and expired in resistances or turned back to mechanical power by the induction motor.
The inter-workings of the power electronics actuation mechanisms are outside of the scope of
this paper but will likely be a challenge handled in the future with the aid of a consult. That said,
the general control strategy is introduced here. We employ a Volts-per-Hertz strategy; that is the
actuation of the excitement frequency is directly influenced by the control system and the applied
voltage is then actuated to achieve a constant ratio Vm/ωe.
In greater detail, the slip speed to the induction motor, ωs= ωe-/ωr, is actuated with PI control of
propeller shaft angular velocity error, e(ωr)= ωr-ωrh (nominally, feedback is attained from a
tachometer mounted in series with the propeller shaft). Dynamic saturation is added to prevent
the controlled slip speed, ωs*, from becoming greater than ωs,max as labeled on the induction
motor curve shown in Figure 6, ωs*<sratedωe. This way the PI control works in a monotonically
increasing region of the torque-speed curve.
32
The applied voltage is actuated to relate to the excitement frequency linearly, Vm=ωe /ωrated
Vrated, where ωrated and Vrated are parameters of the induction motor. This protocol ensures a
maximum magnetic flux and results in an identical torque-speed curve for any excitement
frequency; essentially when smax is not equal to f(ωe). The latter attribute obviates the needs for
an optimizer in the control loop to solve for smax. The schematic in Figure 7 describes the entire
protocol.
HYDRODYNAMIC MODELING
The models for determining the hydrodynamic efforts are adapted from Triantafyllou & Hover
(2003). These efforts include self-propelled propeller thrust, Tsp, self-propelled propeller
torque, Qsp, and hull and prop fluid drag, Rsp. Here, we will assume that the ship is in forward
motion and the propeller is rotating to create forward thrust.
Propeller thrust and torque are modeled as
33
where np is the propeller shaft angular velocity [Hz]. ηR < 1 is the empirically-derived relative
rotative efficiency of the propeller due to variations of the wake distribution and turbulence
caused by the ship hull. The coefficients kT and kQ are the thrust and torque coefficients
(accounting for losses) which can be uniquely interpolated given the pitch ratio of the propeller
and the dimensionless advance coefficient
where w<1 is the experimentally derived wake fraction used to account for differences in flow at
the front of the vessel and in its wake.
Finally, the hull/prop resistance, or fluid drag, has been studied to follow the relationship
where Cr is an empirical parameter of the ship hull which remains approximately constant at the
high value of Reynolds number common to ship travel, Aw is the wetted area of the ship, and t is
the experimentally derived thrust deduction. Thrust deduction is a value usually less than unity
and effectively accounts for the increased drag due to the propeller. With resistance computed
we can give our final state equation derived from Newton's second law
where m is the ship mass and ma is the mass of the entrained water.
CONSERVATION OF POWER
Here, we discuss the power flow of the system (without power electronic control). The
extraction of power from gas in turbine is omitted. Instead, empirical "torque vs. speed/fuel rate"
34
interpolation is used allowing us to accurately initiate our simulation at mechanical power in,
Pin=Qt ωe. Some power is lost to friction in the generator, the rest is converted to AC electrical
power $Pgen=3 R Ea Is*. Some of this power is lost to resistances in the generator stator and
induction motor, the rest is transferred to mechanical power, Pm=3 |Ir|2 Rrm (1-s)/s = Qm wp.
This rotational power is either converted to translational power, Pout=Tsp u, or lost to fluidic
dissipation/cavitation effects.
The plots below show an energy conversion analysis for a single simulation with steady-state
targets given in Table 1.
35
36
37
•
Plans for the Next Quarter
CHRYSSOSTOMIDIS/KARNIADAKIS:
In future work will extend the new framework to include stochastic modeling of the coupled IPShydrodynamics sub-systems to allow global sensitivity studies – a very useful tool in the early
design phase.
HOVER:
We will consider further the following novel approach to optimizing robustness of the AES
power system: optimizing the degree distribution to sample a network from, i.e. optimizing the
probabilities of having certain degrees over a particular failure scenario. Knowledge of this
probability mass function can then be an indicator of robustness for deterministic designs, which
can guide their structural layout. A second and related area of research is multi-armed bandits.
This is an optimization framework with application to sensor management and agent task
assignment. We will explore its application to agent-based reconfiguration
KIRTLEY:
Plans for the next quarter include improving the heat transfer description of the motor with the
addition of realistic first order approximations of air flow and temperature rise to provide a better
idea of thermal limits to motor improvements in size, weight, power density. Our existing
induction motor evaluation software assumes any design can be cooled without exceeding a
given steady state temperature rise. Increased machine temperature will reduce the efficiency and
lifespan of the motor and, in more extreme cases, can make designs impractical to build and
cool. A more detailed thermal analysis of the motor will help to predict approximate temperature
rise, establish limits for feasible motor size, losses, and current and power densities, and would
also help add approximate power losses from the motor cooling system. This is also a first step
towards modeling water cooled motors by establishing the baseline air cooled case for later
comparison to water cooled designs and determination of any changes in motor size, weight, or
efficiency.
38
LEEB:
During the next quarter, we expect to have field results demonstrating that the crew of the
USCGC ESCANABA has found the Ginzu software and NILM installation to be of high utility
in maintaining the operational readiness of shipboard systems. We also expect to have
characterized the design of a testbed under construction at MIT for exploring the utility of a
NILM in a zonal electrical distribution system (ZED). .
TRIANTAFYLLOU:
Over the next quarter, we plan to develop further our models and employ them to study the
response of a ship in a seaway to allow us to:
- Study the transient response of a maneuvering all-electric ship in a seaway. When
maneuvering in a storm, the ship is subject to random environmental loads and severe drag loads
that can cause large speed reduction, up to 50% or more. When combined with second order
wave forces, providing adequate propulsive and steering force is challenging, especially when
other functions must be served simultaneously, while several parameters are known only within
certain ranges.
- Study the sensitivity of the maneuvering ship to various parametric variations and assess the
finite sensitivity as function of time, accounting for the (fast) electrical and (slower) mechanical
and hydrodynamic time constants.
We will focus on this study because of the impact it will have on the propeller/motor design,
since during severe maneuvering the loads on the propeller (axial and transverse) become very
significant and transient.
REFERENCES:
KIRTLEY
Clive Lewis, “The Advanced Induction Motor,” IEEE 2002 Power Engineering Society Summer
Meeting, pp.250-253, IEEE, 2002.
39
LEEB
Cox, R., J. Mosman, T. McKay, S. Leeb, and T. McCoy. 2006. Diagnostic indicators for
shipboard cycling systems using non-intrusive load monitoring. In Proc. ASNE Day 2006, June,
Arlington, VA.
Cox, R., G. Mitchell, P. Bennet, M. Piber, J. Paris, W Wichakool, and S. Leeb. 2007.
Improving shipboard maintenance practices using non-intrusive load monitoring. In Proc.
ASNE Intelligent Ships Symposium VII, May, Philadelphia, PA.
Lee, K. 2003. Electric load information system based on non-intrusive power monitoring. Ph.D.
diss., Massachusetts Institute of Technology, Cambridge.
Leeb, S., S. Shaw, and J. Kirtley. 1995. Transient event detection in spectral envelope estimates
for nonintrusive load monitoring. IEEE Trans. on Power Delivery 10: 1200–1210.
Mitchell, G., R. Cox, J. Paris, and S. Leeb. 2007. Shipboard fluid system diagnostic indicators
using non-intrusive load monitoring. In Proc. ASNE Day 2007, June, Arlington, VA.
Proper, E. 2008. Automated Classification of Power Signals. S.M. thesis, Massachusetts
Institute of Technology, Cambridge.
Shaw, S., S. Leeb, L. Norford, and R. Cox. 2008. Nonitrusive load monitoring and diagnostics in
power systems. IEEE Trans. on Instrumentation and Measurement. 57(7): 1445–1454.
40
2. PUBLICATIONS AND REPORTS RESULTING FROM ONR SUPPORT
•
Journals
HOVER:
Josh A. Taylor and Franz S. Hover. “Efficient particle filtering using interpolation,” IEEE
Trans. on Sig. Proc., submitted, 2009.
Prempraneerach P., F.S. Hover, M.S. Triantafyllou, T.J. McCoy, C. Chryssostomidis, and G.E.
Karniadakis, “Stochastic Sensitivity Methods and Application to the Shipboard Power System,”
IEEE Transactions on Power Systems Engineering, submitted, 2009
•
Book or Chapters
•
Technical Report / Presentations
LEEB:
Ramsey, J.S., S.B. Leeb, T. Denucci, J. Paris, M. Obar, R. Cox, C. Laughman, T. McCoy,
“Shipboard Applications of Non-Intrusive Load Monitoring,” American Society for Naval
Engineers Reconfiguration and Survivability Symposium, Orlando, FL, February 2005. **
Greene, W.C., J.S. Ramsey, S.B. Leeb, T. Denucci, J. Paris, M. Obar, R. Cox, C. Laughman, T.
McCoy, “Non-Intrusive Monitoring for Condition-Based Maintenance,” American Society for
Naval Engineers Reconfiguration and Survivability Symposium, Orlando, FL, February 2005. **
DeNucci, T, R. Cox, S.B. Leeb, J. Paris, T.J. McCoy, C.R. Laughman, W.C. Greene, “Diagnostic
Indicators for Shipboard Systems using Non-Intrusive Load Monitoring,” IEEE Electric Ship
Technologies Symposium, Philadelphia, PA, July 2005.
Cox, R.W., S.B. Leeb, S.R. Shaw, L.K. Norford, “Transient Event Detection for Nonintrusive
Load Monitoring and Demand Side Management using Voltage Distortion,” Applied Power
Electronics Conference, Dallas, TX, March 2006. **
41
Cox, R.W., J.P. Mosman, D. McKay, S.B. Leeb, T. McCoy, “Diagnostic Indicators for
Shipboard Cycling Systems Using Non-Intrusive Load Monitoring,” American Society for Naval
Engineers Day 2006, Arlington, VA, June 2006. **
Cox, R., P. Bennett, D. McKay, J. Paris, S.B. Leeb, “Using the Non-Intrusive Load Monitor for
Shipboard Supervisory Control,” IEEE Electric Ship Technologies Symposium, Arlington, VA,
May 2007.
Cox, R.W., M. Piber, G. Mitchell, P. Bennett, J. Paris, W. Wichakool, S.B. Leeb, “Improving
Shipboard Maintenance Practices Using Non-Intrusive Load Monitoring,” ASNE Intelligent
Ships Symposium VII, Philadelphia, PA, May 2007. **
Wichakool, W., A. Avestruz, R.W. Cox, S.B. Leeb, “Resolving Power Consumption of Variable
Power Electronic Loads Using Nonintrusive Monitoring,” IEEE Power Electronics Specialists
Conference, Orlando, FL, June 2007.
G. Mitchell, R.W. Cox, M. Piber, , P. Bennett, J. Paris, W. Wichakool, S.B. Leeb, “Shipboard
Fluid System Diagnostic Indicators Using Nonintrusive Load Monitoring,” ASNE Day 2007,
Arlington, VA June, 2007. **
Laughman, C.R., R. LaFoy, W. Wichakool, P. Armstrong, S.B. Leeb, et.al., “Electrical and
Mechanical Methods for Detecting Liquid Slugging in Reciprocating Compressors,”
International Compressor Engineering Conference, Purdue University, West Lafayette, IN,
C1437, July 14-17 2008. **
Proper, E., Cox, R., Leeb, S., Douglas, K., Paris, J., Wichakool, W., Foulks, L., Jones, R.,
Branch, P., Fuller, A., Leghorn, J., Elkins, G., “Field Demonstration of a Real-Time NonIntrusive Monitoring System for Condition-Based Maintenance,” Electric Ship Design
Symposium, National Harbor, Maryland, February 2009.
42
HOVER:
Taylor, J.A., and F.S. Hover, “Cubature Using Sparse Grids with Shifted Indexing,” MIT Center
for Ocean Engineering Technical Report COE-031708, March 2008.
Langston, J., J. Taylor, F. Hover, J. Simpson, M. Steurer, and T. Baldwin, “Uncertainty Analysis
for a Large-Scale Transient Simulation of a Notional All-Electric Ship Pulse Load Charging
Scenario,” 10th International Conference on Probabilistic Methods Applied to Power Systems,
Cancun, Mexico, May 2008.
Taylor, J.A., F.S. Hover, and A. Ouroua, Uncertainty Analysis of Large-scale Power Systems
Using Collocation,” Grand Challenges in Modeling and Simulation, Edinburgh, Scotland, June
2008.
Langston, J., M. Steurer, T. Baldwin, J. Taylor, F. Hover, J. Simpson, “Considering Uncertainty
in Assessment of Impact of Pulse Load Charging Event on Shipboard Power System,” Grand
Challenges in Modeling and Simulation, Edinburgh, Scotland, June 2008.
Prempraneerach, P., F.S. Hover, M.S. Triantafyllou, C. Chryssostomidis and G.E. Karniadakis,
“Sensitivity Analysis and Low-Dimensional Stochastic Modeling of Shipboard Integrated Power
Systems,” 39th IEEE Power Electronics Specialists Conference, Rhodes, Greece, June 2008, pp
1999-2003.
Greytak, M., and F.S. Hover, “Robust Motion Planning for Marine Vehicle Navigation,”
Proceedings of the 18th International Offshore and Polar Engineering Conference, Vancouver,
BC, July 2008.
Greytak, M., and F.S. Hover, “Underactuated Point Stabilization Using Predictive Models with
Application to Marine Vehicles,” Intelligent Robots and Systems Conference, Nice, France,
September 2008.
43
Schmitt, K.P., F.S. Hover, and M.S. Triantafyllou, “System-Wide Power Bus Perturbations Due
to Propeller Load Changes in the All Electric Ship,” Electric Ship Technology Symposium, to
appear, 2009.
Hover, F.S., and J.A. Taylor, “Statistically Robust Design for the All-Electric Ship from a
Network Theoretic Perspective,” Electric Ship Technology Symposium, to appear, 2009.
Greytak, M., and F.S. Hover, “Analytic Error Variance Predictions for Planar Vehicles,”
International Conference on Robotics and Automation, to appear, 2009.
Greytak, M. and F.S. Hover, “Planning to Learn: Integrating Model Learning into a Trajectory
Planner for Mobile Robots,” International Conference on Information and Automation,
submitted Jan 2009.
** Outgrowth of supervised student research
•
Inventions
•
Honors
3. PERSONNEL STATISTICS DURING THIS REPORTING PERIOD
3.1. Total No of Co-PI: 7
•
No of Woman Co-PI: 1
•
No of Minority Co-PI
3.2. Total No of students working on the grant
•
Total Grad: 5
o Woman Grad Students: 1
o Minority Grad Students: 1
44
•
Total Under Grad
o Woman Under Grad Students: 1
o Minority Under Grad Students
•
Total Post Docs: 1
o Woman Post Docs: 1
o Minority Post Docs
•
No of degrees Granted during this report Period: 1
45