Content uploaded by Peter Lindahl
Author content
All content in this area was uploaded by Peter Lindahl on Sep 06, 2020
Content may be subject to copyright.
Noncontact Sensors and Nonintrusive Load
Monitoring (NILM) Aboard the USCGC Spencer
Peter Lindahl & Steven Leeb
Massachusetts Institute of Technology
Cambridge, Massachusetts 02139
Email: lindahl@mit.edu
John Donnal
U.S. Naval Academy
Annapolis, MD, 21402
Lt. Greg Bredariol
U.S. Coast Guard
& Massachusetts Institute of Technology
Cambridge, Massachusetts 02139
Abstract—Modernization in the U.S. Navy and U.S. Coast
Guard includes an emphasis on automation systems to help
replace manual tasks and reduce crew sizes. This places a
high reliance on monitoring systems to ensure proper operation
of equipment and maintain safety at sea. Nonintrusive Load
Monitors (NILM) provide low-cost, rugged, and easily installed
options for electrical system monitoring. This paper describes
a real-world case study of newly developed noncontact NILM
sensors installed aboard the USCGC SPENCER, a Famous class
(270 ft) cutter. These sensors require no ohmic contacts for
voltage measurements and can measure individual currents inside
a multi-phase cable bundle. Aboard the SPENCER, these sensors
were used to investigate automated testing applications including
power system metric reporting, watchstander log generation, and
machinery condition monitoring.
I. INTRODUCTION
Modernization in the U.S. Navy and U.S. Coast Guard
includes an emphasis on automation systems to help replace
manual tasks and reduce crew sizes. This places a high reliance
on monitoring systems to ensure proper operation of equip-
ment and maintain safety at sea. Modern monitoring systems
generally rely on individual sensors at a machinery level.
These sensors require an extensive power and communication
network infrastructure making them expensive to install and
intensive to maintain.
The Nonintrusive Load Monitor (NILM) [1], [2] provides
a low-cost, rugged, and easily installed alternative for ma-
chinery monitoring. Installed at central points in the power
distribution network, e.g. the main switchboard of a Coast
Guard vessel, NILM systems record aggregate power demand
through voltage and current measurements. Automated com-
puter processing of this data allows for disaggregation and
characterization of appliance level transient events imparted
by each load as they change state, e.g. a motor turn on event
[3]. This disaggregated information can be integrated into
both manual and automated processes for improved equip-
ment usage tracking, energy management and efficiency, and
condition-based maintenance.
This paper presents a military application and case study
of newly developed noncontact NILM sensors [4] installed on
each electrical feeder-system of the USCG Famous Class (270
ft) Cutter SPENCER, i.e. both onboard generators and shore-
power distribution cables. These sensors require no ohmic
contact to the feeder cables and can be installed around
the multi-phase cable bundle. This makes them extremely
quick to install. Each noncontact NILM contains a capacitive-
electric field sensor and network of hall-effect magnetic-field
sensors for measuring linear combinations of phase voltages
and currents, respectively. Individual phase information is then
determined by calibrating these measurements with a custom
load.
The noncontact NILM sensor package provides “smart” ca-
pabilities via an on-board low-cost computer. These computers
run a custom database framework, NILM DB [5], that enables
efficient storage, retrieval, and processing of the electrical
measurements. Combined with a custom programming and
data management system, NILM Manager [6], the package
allows users to upload scripts or “apps” for localized data
compression and processing, which dramatically reduces the
“big data” bandwidth requirements of high-frequency energy
monitoring. Further, these systems allow users to quickly
visualize power consumption at any time-scale and generate
automated reports, e.g. a watchstander log, based on results of
the local data processing.
The noncontact NILM system was installed in December
of 2015 and has now collected data throughout two long
operational patrols both lasting approximately two months.
Data from the first patrol during which the ship operated both
at sea and in port, and with ship power sourced from shore,
either generator, or both generators, was collected from the
NILM computer. This operational data was then analyzed and
labelled by matching electrical transients to crew generated
machinery logs and previously installed subpanel meters [7].
This data set represents an extensive, sparsely labelled NILM
dataset unique in both application and scope compared with
current publicly available datasets, e.g. [8], [9]. For initial
deliverables, this dataset was used to investigate the use
of the noncontact NILM system for automated information
reporting of power system metrics, watchstander log events,
and machinery condition monitoring.
II. NONCONTACT NILM PACKAGE
The installation of a traditional power meters typically
requires an electrician to interrupt electrical service and po-
tentially expose him or herself to dangerously high voltages.
This is because standard voltage sensors require a direct
ohmic contact with conductors, and current sensors require
Fig. 1: The noncontact sensors installed on the outside of a
cable bundle.
Fig. 2: Schematic for the Hall-effect current sensor.
separating conductors from bundles for individual conductor
encirclement. In contrast, noncontact power sensors can be
installed quickly and safely by an end user, and without service
interruption as they can be zip tied around the insulting jackets
of individual wires or cable bundles. Effectively, these sensors
give users the ability to nonintrusively install power meters
nearly anywhere in the electrical system electrical conductors
are accessible.
Fig. 1 shows an example installation. The sensor package
consists of multiple hall-effect sensors, which are strategically
placed around the bundle to measure the magnetic fields
induced by the currents through the conductors. Additionally,
a single capacitive-coupling sensor is used to measure the
electric field emanating from the voltages on the conductors.
The following subsections give high-level descriptions of these
sensors along with a calibration technique required for accu-
rate power measurements. Further details of their construction,
operation, and accuracy can be found in [4].
A. Current Sensors
Ampere’s law states that the integrated magnetic field of a
closed loop path in the space surrounding an electrical conduc-
tor is proportional to the current running through it. For this
reason, traditional current transducers for cable conductors are
designed to encircle the conductor of interest. Unfortunately,
for conductors in a cable bundle or tray, it may be impossible
or at least difficult to install such a transducer. The noncontact
sensors alleviate this difficulty by detecting the magnetic field
at a particular area in space around the conductor.
Fig. 3: Schematic for capacitive pick-up voltage sensor.
The schematic for each current sensor is shown in Fig. 2.
The sensor consists of an Allegro MicroSystem’s A1362
Hall Effect sensor chip buffered (the AD8676 op-amp) and
capacitively coupled to a double-gain stage amplifier (the two
AD8676 op-amps). The chip has a programmable gain of up
to 16 mV/G, sufficient for standard electrical system current
levels. The 2.2 µF capacitor and 49.9kΩinput resistance to
the gain stages acts to high-pass the measured signal at 1.5
Hz. This serves to reduce effects of low-frequency drift and
avoid signal saturation.
B. Voltage Sensors
The steady-state voltage amplitude in a power system is
typically known and well controlled (e.g. 120 V). While
electrical faults and strong load transients can cause short
fluctuations in the phase voltage amplitudes [10], the voltages
under normal operating conditions can be assumed to be
constant and balanced. Still, the phase relationships between
system voltages and currents do not remain constant and
determine the levels of real and reactive power.
To nonintrusively sense phase voltages, a differential
capacitive-pickup sensor was developed to measure the electric
field emanating from a conductor or conductor bundle. This
sensor has the schematic shown in Fig. 3. Here, the differential
inputs to the AD8421 instrumentation amplifier are connected
to copper plates on a PCB. Each plate creates a capacitive
coupling, C, with the conductor of interest. Along with the
R=1 MΩresistors providing input bias currents to the
amplifier, this capacitance creates an input stage with a transfer
function,
H(s) = sRC
sRC + 1 .(1)
While the 1 MΩbias resistors are large, the mutual capaci-
tance between the conductor and sensor are very small, on the
order of picofarads. As such, RC is small, and (1) reduces
to the transfer function of a differentiator, i.e. H(s)≈sRC.
Thus, the instrumentation amplifier is followed by an inte-
grator to compensate and then followed by an inverting op-
amp to provide additional gain. Also, an integrator circuit
provides feedback to the reference pin of the instrumentation
amplifier. This feedback path in effect subtracts offsets and
low-frequency drift errors from the instrumentation amplifier
output so as to not saturate the sensor output.
C. Power Calculations
Each NILM package records current and voltage mea-
surements at 3kHz. These measurements are then fed via
ribbon cables to a local microcontroller for analog to digital
processing, after which the digital representations are streamed
via USB to a local mini-computer for NILM preprocessing.
This preprocessing algorithm, the Sinefit Spectral Envelope
Preprocessor [11], compresses the high frequency current and
voltage streams into their spectral envelopes of fundamental
real (P1) and reactive power components (Q1) as well as
the real and reactive 3rd, 5th, and 7th current harmonics at
a rate equivalent to the line-frequency (60 Hz in the case
of the ship’s main distribution system). This combination of
high frequency “raw” data collection for capturing detailed
electrical information and the compressed real and reactive
streams for reduced data storage requirements affords a broad
feature space for detecting and identifying electrical signatures
of the ship’s loads while limiting the “big data” problem.
D. Calibration of Sensors
The drawback to using “point” reference sensors is that
the magnetic and electric fields are not necessarily uniform
around the conductor of interest and decrease in strength with
distance. Further if other conductors are nearby, as is neces-
sarily the case in a cable bundle, the magnetic and electric
fields in the space around the cable are linear combination of
fields due to each conductor’s current and voltage, respectively.
Thus, it is necessary to calibrate the sensors for disagreggating
individual phase power streams.
1) Current Measurement Calibration: The USCGC
SPENCER’s electrical distribution system is a delta-
configured 254/440 V system. Assuming no ground faults,
the phase current relationship for the distribution system is,
IA+IB+IC= 0.(2)
This means that the currents on the ship only span IR2as one
current can always be considered a linear combination of the
third, e.g. IC=−(IA+IB). Then, the output of an N-length
hall-effect sensor network monitoring the conductors depend
on only two phases, i.e.
S1
S2
.
.
.
SN
=
m1Am1B
m2Am2B
.
.
..
.
.
mNA mN B
IA
IB.(3)
Here, each mterm corresponds to the sensor output contri-
bution of the two independent phase currents. So long as the
sensor network spans the current space, i.e. the sensors are
located around the conductors such that they sufficiently mea-
sure the independent currents, then knowing these m-terms
allows the calculation of currents from sensor measurements
via (2) and
IA
IB=K
S1
.
.
.
SN
,(4)
where Kas the pseudoinverse of the m-terms matrix in (3).
One method to finding these mterms is to cycle a known
resistive load across two of the delta-leg connections and
measure the sensor outputs. For example, terms m1Athrough
mNA can be found by placing the load across phases A and
C so that IB= 0. Similarly, terms m1Bthrough mNB can be
found by placing the load across phases B and C so that IA
= 0.
2) Voltage Measurement Calibration: For a three-phase
electrical distribution system, the phase voltages are related
as,
VA(t) = Vcos (ωt)
VB(t) = Vcos ωt −2π
3(5)
VC(t) = Vcos ωt +2π
3.
Thus, the electric field sensor output is a linear combination
of voltages with the same frequency, and itself will be a single
sinusoid of the form,
Sv=Acos (ωt +φ).(6)
Here, φrepresents the phase-relationship between the mea-
sured sensor output and VA. When the known resistive load is
connected from VAto VC, then the load current has the form,
IL=√3V
Rcos ωt −π
6.(7)
where Ris the resistive value of the load.
With ILavailable from the calibrated current sensors, φcan
be calculated as φ=θ−π
6where θis the phase difference
between Svand IL.θcan be calculated from the output of
the preprocessor (Sect. II-C) during this calibration process.
That is, the power factor angle formed from P1and Q1prior
to voltage calibration is equal to −θ, i.e.
θ=−arctan Q1
P1.(8)
Collectively then, VA(t),VB(t), and VC(t)can then be
estimated by scaling Svby the factor V
Aand phase shifting
the waveforms in accordance with the phase relationships of
(5), (6), and (8).
E. User Interface
The spectral envelop data for each phase, i.e. P1,Q1, and
the current harmonics are stored locally on a host PC using
NILM DB [5], a time series database optimized for high band-
width datasets. This PC features a web-based interface, NILM
Manager [6], which can be securely accessed via a network
or directly on the computer, allows guardsmen, officers, or
researchers with access credentials to view and manipulate
this data using an intuitive “Google Earth” style interface.
That is, under different levels of zoom, the rendered plots
never show more information than is actually distinguishable
to a user. When viewing long duration time series data,
only a representative sample of the raw spectral envelope
Fig. 4: The NILM-sensor package features a “filter” environ-
ment for creating custom scripts for automated analysis of
power meter data.
is transmitted. As a user zooms in to shorter time scales,
progressively higher resolution data sets are returned until the
user selects a time range containing less data than a pre-set
amount, after which the raw stream data is returned. Figs. 6
and 11 of this paper show examples of this plotting feature.
In addition to the plotting feature, NILM Manager also
allows users to create custom scripts, or “filters”, using Python
and load them to the local computer of the NILM. This
environment is shown in Fig. 4. Each generated filter runs
locally against historical and/or real-time data and stores the
results as a new stream in NILM DB. This secondary stream
can be viewed using the same plotting interface as the spectral
envelope streams, or used to perform automated analyses
such as calculating and reporting operation metrics, generating
watchstander logs, and condition monitoring equipment.
III. INS TALL ATION ON THE USCGC SPENCER
The USCGC SPENCER is a 270ft (82 m) Medium En-
durance vessel commissioned in 1986 and currently based
out of Boston, MA. The on board electrical generation plant
consists of two V12 Caterpillar diesel generator sets rated
at 475 kW each. Additionally, the ship has two electrical
feeder systems for receiving power from shore when in port.
The SPENCER, which can sustain a crew of 100 personnel
for months at sea, has electrical loads ranging from large
three-phase motors for salt water cooling, HVAC systems, fire
protection, and hydraulic systems; to line-line loads including
lights, washers and driers, and ovens.
Fig. 5 shows a general installation schematic on the
USCGC SPENCER. To monitor electric power demanded
of the entire vessel at all times, noncontact sensors were
installed on each of the feeder cable systems that input to
the main switchboard, i.e. the two sets of cables from the
two onboard generators, and the two sets of cables available
for shore power. By monitoring these four systems, the full
aggregated power stream is available even at times when more
than one source is providing power, e.g. when both generators
are running. In addition to the noncontact NILM systems
which are the subject of this paper, two standard NILM
systems, i.e. ones using commercially available current and
voltage transducers, were previously installed on the USCGC
SPENCER to monitor 19 loads distributed over two subpanels
containing main machinery room gear [7]. These subpanel
NILMs aided in identifying and confirming transients of the
full aggregated noncontact NILM data streams.
Nonintrusive Calibration Procedure
While the calibration of the sensors as described in the
previous section is simple and effective, it is unrealistic and
against the philosophy of the noncontact sensors to require
turning off all non-calibration loads during the calibration
process. Instead, to calibrate the sensors without interrupting
service, the calibration load is pulse-width modulated (PWM)
to create a discernible signal in the measured currents amongst
the intrinsic ship loads which are treated as background noise.
In this way, the calibration load can be differentiated from the
background loads using spectrum analysis.
Using a microcontroller and a solid state relay (SSR), the
calibration load, a 2.5 kW heater, is pulse-width modulated
with a duty cycle of 33% at a frequency of 0.33Hz. This
oscillating load is connected across phases B-C for 90 sec-
onds, and then A-C for 90 seconds. During this process,
the NILM software takes the fast Fourier transform (FFT)
of the preprocessor output and finds the fundamental and
harmonic components corresponding to the PWM calibration
load. These components are then matched against the Fourier
Series coefficients of a PWM signal with a duty cycle of 33%
and magnitude equal to that of the 2.5 kW load. Generalized
but detailed descriptions of this automated process are further
explained in [4] and [12].
IV. USCGC SPENCER DATASE T
Data collection began aboard the USCGC SPENCER in
December of 2015 and continues today. At the time of this
writing, the noncontact NILM systems have collected data
throughout two long operational patrols, both lasting approx-
imately two months, as well as several months with the ship
stationed in port in Boston, MA. Over the course of this time
period, the ship has operated for long periods both at sea and
Fig. 5: General layout of the four noncontact and two standard
subpanel NILM installations on the USCGC SPENCER.
Fig. 6: Nine day period of three-phase power measurements
from the three power sources utilized: generator 1 (red),
generator 2 (green), and shore power (blue).
in port, and with the ship powered from shore or either one
or both generators.
Fig. 6 shows an approximately nine day period of aggregate
real power data collected while the vessel was underway
and returned to port. In this figure, the red, green, and blue
traces indicate the three-phase power supplied by generator
1, generator 2, and shore power, respectively. While the
ship is underway, from Dec. 24th through Dec. 30th, each
generator alternates supplying power to the ship interspersed
with periods of both generators supplying power. Early on
Dec. 31st, the ship pulled into port, connected to shore power,
and shut down the two diesel generators. In these traces,
the darker solid line signifies a decimated moving average
of the recorded power, while the lighter shaded region gives
an indication of the local variance of the power values about
the mean trace. This plot shows that the average power draw
during this period was approximately 200 kW.
NILM methods for disaggregating power streams typically
fall into two categories, event-based or non-event based, which
generally utilize supervised and unsupervised machine learn-
ing techniques, respectively [13]. For supervised techniques,
algorithms require labelled data for training the algorithm,
while unsupervised techniques do not. Even so, labelled data
aids in evaluating the performance of unsupervised techniques.
To begin compiling a dataset useful for both categories of load
disaggregation, events are manually labelled by correlating
electrical transients in the data streams to previously measured
load “exemplar” waveforms around times when crew logs
specifically recorded events, e.g. energizing a fire pump.
These exemplar waveforms are short data streams captured in
coordination with the crew, who energized and de-energized
specified loads while we watched and marked the transients
in the data streams in near real-time. As a further aid to
labelling, the previously installed submeter NILM sensors also
help by narrowing the set of potential loads corresponding to
a transient when that transient appears in their data streams as
well.
Correlation-Based Event Identification
A correlation-based event identification algorithm, Trainola
[5], is built into the NILM Manager and NILM DB system.
In this algorithm, a measure of the correlation between an
exemplar waveform and a window of electrical data equal
in length to the exemplar waveform is calculated as the
window moves across a data stream. When this measure peaks
within a defined range, the algorithm marks the time instance
corresponding to the beginning of the moving window as that
of an event corresponding to the exemplar.
Consider an exemplar, g, of length N, and the equal-length
windowed portion of a data stream, f. The sum of squared
errors between these two waveforms is defined as,
E=
N
X
n=1
(f[n]−g[n])2.(9)
This error term can be expanded and ultimately rewritten in
terms of dot products, i.e.
E=Xf2[n] + Xg2[n]−2Xf[n]g[n]
E= (f·f)+(g·g)−2 (f·g)
E=|f|2+|g|2−2 (f·g).(10)
If the waveforms are identical, the sum of squared errors
term is E= 0, and (10) can be reformed as,
f·g=|f|2+|g|2
2.(11)
Further, if the magnitudes of the exemplar and the measured
waveform are the same, i.e. |g|=|f|, then (11) reduces to,
f·g
|g|2= 1.(12)
Thus, we can define a function,
M[k] = f·g
|g|2,(13)
that generates a local maximum with a value of 1 when the
windowed section, f, of a data stream starting at time instance,
k, matches in amplitude and shape to exemplar, g.
The loads with at least one identified exemplar are listed
in Table I. These loads are primarily pumps ranging in size
from 3 kW to 56 kW with power factors ranging from
approximately 0.8 to 0.9. All except for the gray water pump
are loads whose operations are recorded manually by the
watchstander. Thus, for these loads we are able to use the
correlation technique described above (and backed by our
own transient recognition abilities), to find and label further
transients. The number of on and off events confirmed with a
high level of confidence are shown in the “Labelled” (3rd and
4th) columns of the table. The 5th and 6th columns, entitled
“Minimum Unlabelled”, correspond to events indicated in the
machinery logs, but whose corresponding electrical transients
were not identified with extreme confidence. This could be due
to mislabelled or mistimed entries in the logs, or the event
transient might be masked by other simultaneous events, or
both. Still, it remains likely that in many cases the recorded
event occurred around the time logged. Thus, that an unla-
belled transient occurred can still be useful information in
developing disaggregation algorithms.
The grey water tanks happen to be in close proximity to
the installation locations of the noncontact NILMs. As such,
TABLE I: Initial set of confirmed load exemplars from the
first monitored operational patrol.
Load Electrical
Logged Operational Events
Name Specifications Labelled Minimum
Unlabelled
On Off On Off
Watchstander Logged Loads
Fire 56 kW 8 8 0 0
Pump 1 0.88 PF
Fire 56 kW 3 3 0 0
Pump 2 0.88 PF
Aft Steering 22 kW 6 6 0 0
Pump 1 0.85 PF
Aft Steering 22 kW 6 6 0 0
Pump 2 0.85 PF
CPP 7.5 kW 11 8 9 10
Pumps 0.80 PF
Inport ASW 7.5 kW 2 2 1 1
Pump .79 PF
Underway 11 kW 2 2 1 1
ASW Pump .79 PF
Diesel Engine Auxiliary Loads
Jacket 9 kW 13 13 34 34
Water Heater 1 PF
Lube Oil 12 kW 13 13 34 34
Heater 1 PF
Pre-lube 3 kW 13 13 34 34
Pump 0.8 PF
Other Loads
Gray 3.7 kW 7 7 N/A N/A
Water Pump 0.79 PF
it was easy to note the timing of the tank pump run events
during inport testing of the noncontact NILM systems. Thus,
several exemplar waveforms were identified for use with the
Trainola correlation feature, however no estimate of the total
number of events is available.
Fig. 7 shows an overlay of the 8 labelled fire pump #1
energized (turn on) events, with (a) giving the fundamental real
power component and (b) the reactive. When the fire pump
is energized, the large starting currents of the motor cause
massive spikes in the power draw which last for approximately
1.5 seconds. Following this spike, the real and reactive power
consumed by the pump steadies out at levels in line with its
nameplate ratings, i.e. 56 kW and 30 kVAR (for a 0.88 PF),
respectively.
Fig. 8 helps to understand the performance of the Trainola
algorithm. The top plot shows the reactive power transient
during the 3rd fire pump energized event within the context of
the actual data stream. This even occurs at approximately the
15 second mark. Using the first fire pump energized event as
the “exemplar”, the Trainola algorithm metric for the reactive
power stream, MQ1, results in a positive peak with a value
of 0.92, very close to the ideal value of 1 if both events were
identical. Of note, similar peaks also occur in the correlation
metric for the real power stream, MP1, however these peaks
tend to vary more in amplitude due to the smaller peak in P1
compared to that in Q1, and because P1is a “busier” data
streams. Thus, MP1provides less certainty in fire pump #1
event identification than MQ1.
(a) Fundamental Real Power
(b) Fundamental Reactive Power
Fig. 7: Overlay of several fire pump #1 energized transients.
(a) Data Stream
(b) Correlation Measurement
Fig. 8: Trainola performance using the first fire pump #1
transient as the exemplar around the time of the third fire
pump #1 transient.
V. AUT OM ATED TE ST IN G APPLICATIONS
Using the dataset described above, and the baseline transient
detection algorithm, Trainola, several automated testing appli-
cations were investigated including measuring and reporting
power system metrics, generating watchstander log reports,
and condition monitoring cyclic loads.
(a) Dec. 27, 2015, 1140:50 - Load Balance (b) Dec. 28, 2015, 1342:32 - Load Balance
(c) Dec. 27, 2015, 1140:50 - Power Factors (d) Dec. 28, 2015, 1342:32 - Power Factors
Fig. 9: Generator operation metrics showing load balancing between phases and each generator, and power factor for each
phase and each generator.
A. Generator Operation Metrics
In addition to performing load-disaggregation based moni-
toring tasks, the NILM sensors and data management systems
also afford opportunities for aggregate local power system
monitoring. On the USCGC SPENCER, noncontact NILM
sensors are installed on the feeder cables connecting the two
diesel generators to the main switchboard. Thus, the NILMs
provide an ability to analyze the preprocessed power streams
in near real-time for metrics concerning the health of the
generation and distribution system. These metrics can then
be reported via a local network to crew members instigating
further investigation if required.
Two mechanisms for load induced failure in diesel genera-
tors such as those aboard the SPENCER are imbalanced loads
and low power factor loads. Imbalanced current demands on
the three phases of the generator can cause poor efficiencies,
thermal and mechanical stresses [14], [15], and vibrations
that increase the noise signature of the ship [16]. Low power
factor loads require increased currents for a given power
demand. Thus, low power factors require derating the delivery
capabilities of the generator to avoid overheating.
The NILM sensors can be programmed via the filter en-
vironment to generate metrics of phase balance and power
factor and report them to the crew in near real-time. Fig. 9
shows an example of such metrics. The figure features bar
plots reporting each generator’s loading by phase as well
as a comparison of overall loading between generators on
Dec. 27th 2015 at 1140:50 (Fig. 9a) and Dec. 28th at 1342:32
(Fig. 9b). From Fig 9a, crew operators can immediately note
that while Generator 1 is well balanced, Generator 2 is not.
Instead, Phase B is roughly 40% more loaded than Phase
A and Generator 2 is about 50% more loaded overall than
Fig. 10: Example image of the manually generated watch-
stander log.
TABLE II: List of acronyms used in the watchstander log of
Fig. 10.
Acronym Description
BMDE both main diesel engines
STBY standby
FP fire pump
E/R engine room
RMD restricted maneuver doctrine
SSDG ship service diesel generator
F/O fuel oil
IAW in accordance with
CPP controllable pitch propeller
Generator 1. At 1342:32 on Dec. 28th, the plot shows that
now only generator 2 is operating, and while it is slightly
more balanced, the generator is nearing its max capacity.
Operating at this max capacity however requires an operat-
ing power factor of at least 0.8. The power factors for each
generator can be similarly reported, again by phase and by
the overall generator power factor, Fig. 9c and Fig. 9d. Here,
Fig. 9c show that each phase of both generators are operating
at or above their rated power factor levels on Dec. 27th at
1140:15. Similarly, generator 2 is operating above the rated
power factor level on Dec. 28th at 1342:32 (Fig. 9d). Thus,
the real power demanded of the generator at this time is still
in compliance with the generator’s power delivery capabilities.
B. Automated Watchstander Log
As described in [17], the machinery log is an official
document, and in the event of an accident caused by machinery
malfunction or operator error, the document becomes a legal
reference in court. Thus, it is important that they be as
accurate as possible. Further, automating log generation helps
to modernize the marine environment by reducing crew time
commitments dedicated to low-level manual tasks. The non-
contact NILM system can aid in both objectives. Specifically,
given a crew generated log with rough timings of events, the
NILM can identify exact times of transients corresponding to
the log events, thus improving the log accuracy. Ultimately,
as improved event detection and identification algorithms are
applied, the need for a crew generated log disappears entirely.
TABLE III: Comparison of manually generated log of Fig. 10
to NILM detected events.
Time Event Manual NILM Notes
Log Log
1505 Secured both X X NILM does not
MPDE distinguish engines
1509 Energized #1 X X NILM does not
fire pump distinguish pumps
1515 Secured #1 X X NILM does not
fire pump distinguish pumps
1520 Secured both X X NILM does not
steering pumps distinguish pumps
1530 Round of XNo exemplar
E/R transient for NILM
1540 Secured from XNo exemplar
Special Sea transient for NILM
1540 Secured from X X Recorded at
RMD 1541 by NILM
1541 Secured both XNot recorded
CPP “C” pumps by watchstander
1602 Swapped potable XNo exemplar
water suction transient for NILM
1655 Singled electrical XWatchstander log
load on Gen. #2 indicates incorrect Gen.
1707 Secured X X Recorded at
Gen. #1 1653 by NILM
1843 Started both XNot recorded
MPDE by watchstander
1900 Secured both X X NILM does not
MPDE distinguish engines
1903 Commenced fuel X X Recorded at
transfer 1912 by NILM
The NILM uses the Trainola algorithm to identify transient
events and then logs the timestamp of the transient event.
Using the filter interface, the NILM can be automated to create
a document containing these timestamps and a description of
the event similar to that of the manual watchstander log. An
example excerpt from this log for the afternoon of Dec. 24th,
2015 is shown in Fig 10. A list of acronyms used in this report
are given in Table II. Of note, this document was redacted
to remove sensitive information, e.g. the ship’s location and
the name of the watchstander officer, but the timing and
descriptions of the recorded events remain unchanged.
Each event in Fig. 10 is listed by the time of the event
in bold followed by the event description, e.g. “1505 Se-
cured BMDE’s placed in 30 min STBY.” The ability of the
NILM system to generate an automated watchstander log
is contrasted against this manual log in Table III. Here,
each event recorded by the watchstander is listed with the
time in the first column and a shortened description in the
second. Additionally, typically recorded events which were
identified by the NILM but not recorded by the watchstander
are also shown, e.g. 1541 - “Secured both CPP “C” Pumps”.
Check marks in columns three and four indicate if the event
was recorded in the manually generated log and the NILM
generated log, respectively. The final column gives information
regarding the logging of each event.
Of the 12 events listed by the officer, the NILM is able
to automatically record 9, though with some caveats. At
the present time, the NILM cannot well distinguish between
identical loads, e.g. fire pumps #1 and #2 as their transients are
extremely similar and thus well correlated. Some events, e.g.
“round of E/R”, which means the watchstander inspected the
equipment in the engine room, provide no electrical transient
for the NILM. The NILM however did indicate that the
watchstander missed recording events typically included in
the log and incorrectly recorded other events. Notably, at
1655 the watchstander recorded that all electrical loads were
singled on Generator #1, though the NILM data streams at
that time show Generator #2 singling the loads. Regarding
the events detected by the NILM but not recorded by the
watchstander, its certainly possible that these events are false
positive detections. However, at least in the case of the
“Secured both CPP “C” pumps” event at 1541, being secured
from RMD almost always corresponds to also securing the
CPP “C” pumps thus giving a high level of confidence that
the event did indeed occur.
C. Cyclic Load Condition Monitoring
A final application for the noncontact NILMs is condition
monitoring. By detecting and assigning transients to specific
loads or simply grouping together transients with similar traits,
the NILM system can be automated to detect changes in event
timing patterns, which can be indicative of machinery faults
[18]. This is particularly important in closed-loop controlled
loads, e.g. cyclic loads such as gray water tank pumps, as
pre-catastrophic faults can easily go unnoticed.
The gray water waste system aboard the SPENCER collects
the waste water from non-sewage and non-recycling water
loads, e.g. sinks, showers, and washers. This waste is collected
in a holding tank before being filtered and pumped overboard
or combined with sewage depending on ship’s location and
local pollution regulations. Tank level indicators (TLIs) pro-
vide feedback to the pump’s control system indicating when
the water level is high and needs to be pumped down.
Using exemplar waveforms of such pump runs, the NILM
Trainola function reported very high frequency pump opera-
tion but with very short run times. Fig. 11 depicts this short
cycling operation. Here, the phase A real and reactive power
streams are plotted over a 7 minute period on Dec. 7, 2015.
During this period, 7 transients indicative of pump runs were
identified by the NILM system, but these transients only lasted
a few seconds.
According to the ship’s crew, a common failure mode for the
gray water pumps are TLI sensor faults. One fault mechanism
occurs when the sensors become clogged with debris and oil
leading to premature “full tank” indications and causing the
pumps to shut off only a few seconds after switching on.
Eventually this leads to pump failure by overworking the pump
and working it with little medium. On the SPENCER, this
fault was the root cause of the pump’s short cycle operation.
However, the fault had gone unnoticed by the crew as the
control system still moves grey water on from the tank and
full failure had not yet happened. Thus, the NILM system
can detect such faults before they cause failure and can be
made automated to do so by monitoring and comparing the
Fig. 11: Real and reactive fundamental power streams showing
the detected shot-cycling gray water pump.
statistical distributions of the pump’s run frequency and run
length to that when the system is in good condition [18].
VI. CONCLUSION
The noncontact, nonintrusive load monitoring (NILM) sen-
sors described in this paper represent next-generation sensor
technology for NILM systems. As shown via their application
aboard the U.S. Coast Guard SPENCER, these sensors when
combined with signal processing and signal disaggregation
techniques create a powerful tool for automated testing of
electrical systems and loads. These sensors are easy to install
and uninstall making them useful tools for crew technicians
needing to perform on-the-go diagnostic tests of subsystems
or individual pieces of equipment. Further, custom analyzer
scripts can be uploaded to the NILM computer for targeted
data processing thus reducing the required bandwidth as well
as improving data security when transmitting automated test
results via a local network.
The data collected from the USCGC SPENCER has been
matched against manually generated machinery logs in or-
der to build a dataset for testing disaggregation algorithms.
With the sensors still installed aboard the ship, more data is
continuously being collected and new instances of machinery
events identified. Moving forward, we plan to develop and
apply more advanced disaggregation, condition monitoring,
and energy score keeping algorithms to further improve the
autonomy, accuracy, and ultimately automated testing abilities
of the noncontact NILM system.
ACKNOWLEDGMENT
The authors would like to thank Lt. William Cotta of the
U.S. Coast Guard for his installations of the contact NILM
sensors aboard the USCGC SPENCER and the crew of the
U.S. Coast Guard Cutter SPENCER for their generous ac-
commodations and help in understanding the ship systems and
operations. Additionally, the authors thank the Office of Naval
Research NEPTUNE Program and the Grainger Foundation
for their financial and technical support.
REFERENCES
[1] G. W. Hart, “Nonintrusive appliance load monitoring,” Proceedings of
the IEEE, vol. 80, no. 12, pp. 1870–1891, Dec 1992.
[2] S. R. Shaw, S. B. Leeb, L. K. Norford, and R. W. Cox, “Nonintrusive
load monitoring and diagnostics in power systems,” IEEE Transactions
on Instrumentation and Measurement, vol. 57, no. 7, pp. 1445–1454,
July 2008.
[3] C. Laughman, K. Lee, R. Cox, S. Shaw, S. Leeb, L. Norford, and
P. Armstrong, “Power signature analysis,” IEEE Power and Energy
Magazine, vol. 1, no. 2, pp. 56–63, Mar 2003.
[4] J. S. Donnal and S. B. Leeb, “Noncontact power meter,” IEEE Sensors
Journal, vol. 15, no. 2, pp. 1161–1169, Feb 2015.
[5] J. Paris, J. S. Donnal, and S. B. Leeb, “Nilmdb: The non-intrusive load
monitor database,” IEEE Transactions on Smart Grid, vol. 5, no. 5, pp.
2459–2467, Sept 2014.
[6] J. Donnal, J. Paris, and S. B. B. Leeb, “Energy applications for an energy
box,” IEEE Internet of Things Journal, vol. PP, no. 99, pp. 1–1, 2016.
[7] W. J. Cotta, “Machinery diagnostics and characterization through elec-
trical sensing,” Master’s thesis, Massachusetts Institute of Technology,
2015.
[8] S. Makonin, B. Ellert, I. V. Bajic, and F. Popowich, “Electricity, water,
and natural gas consumption of a residential house in Canada from 2012
to 2014,” Scientific Data, vol. 3, no. 160037, pp. 1–12, 2016.
[9] J. Z. Kolter and M. J. Johnson, “Redd: A public data set for energy
disaggregation research,” in In proceedings of the SustKDD workshop
on Data Mining Applications in Sustainability, 2011.
[10] J. Prousalidis, E. Styvaktakis, E. Sofras, I. K. Hatzilau, and D. Muthu-
muni, “Voltage dips in ship systems,” in 2007 IEEE Electric Ship
Technologies Symposium, May 2007, pp. 309–314.
[11] J. Paris, J. S. Donnal, Z. Remscrim, S. B. Leeb, and S. R. Shaw, “The
sinefit spectral envelope preprocessor,” IEEE Sensors Journal, vol. 14,
no. 12, pp. 4385–4394, Dec 2014.
[12] J. S. Donnal, “Home nilm: a comprehensive energy monitoring toolkit,”
Master’s thesis, Massachusetts Institute of Technology, 2013.
[13] A. Zoha, A. Gluhak, M. A. Imran, and S. Rajasegarar, “Non-intrusive
load monitoring approaches for disaggregated energy sensing: A survey,”
Sensors, vol. 12, no. 12, pp. 16 838–16 866, 2012.
[14] T. F. Chan and L. L. Lai, “Steady-state analysis and performance of
a stand-alone three-phase induction generator with asymmetrically con-
nected load impedances and excitation capacitances,” IEEE Transactions
on Energy Conversion, vol. 16, no. 4, pp. 327–333, Dec 2001.
[15] E. Muljadi, R. Schiferl, and T. A. Lipo, “Induction machine phase bal-
ancing by unsymmetrical thyristor voltage control,” IEEE Transactions
on Industry Applications, vol. IA-21, no. 3, pp. 669–678, May 1985.
[16] R. Zachar, P. Lindahl, J. Donnal, W. Cotta, C. Schantz, and S. B.
Leeb, “Utilizing spin-down transients for vibration-based diagnostics of
resiliently mounted machines,” IEEE Transactions on Instrumentation
and Measurement, vol. 65, no. 7, pp. 1641–1650, July 2016.
[17] B. P. Murphy, “Machinery monitoring technology design methodology
for determining the information and sensors required for reduced man-
ning of ships,” Master’s thesis, Massachusetts Institute of Technology,
2000.
[18] J. Paris, J. S. Donnal, R. Cox, and S. Leeb, “Hunting cyclic energy
wasters,” IEEE Transactions on Smart Grid, vol. 5, no. 6, pp. 2777–
2786, Nov 2014.