Signal Processing - September 2016 - 152

Cruise
Ship

log-var of
Wavelet Detail Signal 3

log-var of
Wavelet Detail Signal 1

Lebanon
100
50
Quebec
0
-50
100

Texas

U.S. West

U.S. East

Wa
0
vel log-v
et
De ar of -50
tail
-100
Sig
na
l2

(a)

50
0
f
5
-50
o
nal
var
-100
log- tail Sig
t De
vele
Wa

100

China

50
0

Tenerife

-50

India

-100
100

Turkey

50
0
Wa
vel log-va -50
et D
r of
eta
-100
il S
ign
al 4

Ireland

-100

-95
Mean

-60
-70
-80
ent
Segm
f ENF

(b)

FigurE 3. An example of features derived from ENF signals for differentiating the originating power grids.

nominal ENF, 3) audio signals with
50-Hz nominal ENF, and 4) audio signals with 60-Hz nominal ENF. While
each team had their own design for the
learning and classification system,
their solutions all consisted of four
subsystems, one dedicated to each data
type. Some of these subsystems had
certain variations in the way the ENF
signals were extracted or utilized different discriminating features, depending on which of the four types to which
the signal belonged.
The first-place winning team, Team
Resonance_1011 from Bangladesh,
employed two-stage support vector
machine (SVM) classifiers for each data
type. The first one is a one-class SVM
classifier that decides if the signal belongs
to a grid seen in training, and the second
SVM classifier narrows down the list of
possible grids on the basis of different discriminating features extracted from the
embedded ENF signals. The last stage of

F-400 mA R1
TX1
11 kΩ
Power
Source

the classification is where the testing signal is passed to a "pole-matching" classifier to reach a final decision on the grid
of origin based on the minimum distance
between the estimated poles of the training and testing grids.
Team Hammer Down from the United
States, which took second place, carefully
studied the characteristics of the signals
for this problem and proposed a different
method from what is available in the
recent ENF literature. The team's novel
method utilized a multiharmonic histogram to analyze ENF signals for identifying power grids. Team members
computed ENF signals at multiple harmonic locations and extracted histograms
of 1) ENF magnitudes, 2) the signal
power around the ENF, 3) the noise
power around the ENF, and 4) the signalto-noise ratio of the ENF. At the classification stage, they proposed a histogram
matching method with a multilayer decision rule for identification.

R4-200 kΩ

0.1 µC
R2
1.5 kΩ

R3
50 kΩ

C2
Output
0.01 µC

220/28 V

FigurE 4. A circuit diagram from Team UpatrasECE from the University of Upatras, Greece.
152

IEEE Signal Processing Magazine

September 2016

Team Vidyut from India, which took
third place, used a multistage SVM system where, for each input test recording,
five predictions from five respective classifiers were received. To compute the
confidence of a prediction, team members
proposed an entropy-based measure of
confidence where lower entropy means
higher confidence. For audio signals, they
used an additional subband classifier that
captures the relative strength among the
different ENF harmonic components captured by a recording. The team used
custom-designed detectors to resolve
conflicts in decision between their SVM
and subband classifiers.

Sensing circuits
For the hardware and signal acquisition
tasks in the SP Cup that aimed at bringing
out the synergy of sensing, processing,
and learning, nearly all participating teams
built their own circuit for collecting reference power recordings from the team's
respective locations. Useful references
include an article in IEEE Signal Processing Magazine on observing the grid [7] as
well as ENF-related literature [3], [5], [8].
Among the submissions received, some
teams adopted the reference design, while
other teams made improvements, incorporated innovative features, or employed different implementations.
Given that this is the first time the SP
Cup competitions included sensing hardware design and implementation in addition to algorithmic tasks, the judging

Table of Contents for the Digital Edition of Signal Processing - September 2016