Anda di halaman 1dari 24

sensors

Article
A Novel Arc Fault Detector for Early Detection of
Electrical Fires
Kai Yang 1 , Rencheng Zhang 1, *, Jianhong Yang 1 , Canhua Liu 1 , Shouhong Chen 1 and
Fujiang Zhang 1,2
1 College of Mechanical Engineering and Automation, Huaqiao University, Xiamen 361021, China;
yangknh@126.com (K.Y.); yjhong@hqu.edu.cn (J.Y.); xsjtan@sina.com (C.L.); shouhong55@163.com (S.C.);
zfj@fjut.edu.cn (F.Z.)
2 School of Mechanical and Automotive Engineering, Fujian University of Technology, Fuzhou 350118, China
* Correspondence: phzzrc@hqu.edu.cn; Tel.: +86-592-6162-566

Academic Editor: Ingolf Willms


Received: 18 January 2016; Accepted: 5 April 2016; Published: 9 April 2016

Abstract: Arc faults can produce very high temperatures and can easily ignite combustible materials;
thus, they represent one of the most important causes of electrical fires. The application of arc fault
detection, as an emerging early fire detection technology, is required by the National Electrical Code
to reduce the occurrence of electrical fires. However, the concealment, randomness and diversity
of arc faults make them difficult to detect. To improve the accuracy of arc fault detection, a novel
arc fault detector (AFD) is developed in this study. First, an experimental arc fault platform is built
to study electrical fires. A high-frequency transducer and a current transducer are used to measure
typical load signals of arc faults and normal states. After the common features of these signals are
studied, high-frequency energy and current variations are extracted as an input eigenvector for
use by an arc fault detection algorithm. Then, the detection algorithm based on a weighted least
squares support vector machine is designed and successfully applied in a microprocessor. Finally,
an AFD is developed. The test results show that the AFD can detect arc faults in a timely manner
and interrupt the circuit power supply before electrical fires can occur. The AFD is not influenced
by cross talk or transient processes, and the detection accuracy is very high. Hence, the AFD can be
installed in low-voltage circuits to monitor circuit states in real-time to facilitate the early detection of
electrical fires.

Keywords: arc fault detector (AFD); electrical fire; high-frequency energy; current variation; weighted
least squares support vector machine; cross talk

1. Introduction
As reported by the United States Fire Administration (USFA), there were an estimated
372,900 residential building fires in the United States each year from 2011 to 2013, causing an estimated
13,125 injuries, 2530 deaths and $7 billion in property damage [1]. Fires represent a serious threat
to humans. Electrical fires are one of the most frequently occurring types of fires, representing on
average up to 30.2% of all fires and exceeding 50% of all large fires [2,3]. Statistical data from fire
administrations show that electrical fires are usually caused by arc faults, over currents, short circuits,
leakages [1–6], etc. After the USFA studied the sources of heat in residential building electrical fires,
they stated that arc fault accounted for 82% of the electrical fires [6]. Specifically, arc faults represent
one of the most important causes of electrical fires [1,5,6]. Traditional fire detection technologies
usually monitor smoke, gas, temperature, etc. Recently, several fire feature signals could be detected
by multisensors [7]. The multi-wavelength LiDAR technique was applied to study the smoke of a
fire [8]. A laser spectroscopic carbon monoxide sensor was used to detect fires [9]. Some wireless

Sensors 2016, 16, 500; doi:10.3390/s16040500 www.mdpi.com/journal/sensors


Sensors 2016, 16, 500 2 of 24
Sensors 2016, 16, 500 2 of 24

intelligent
sensors technologies
and intelligent were introduced
technologies for fire identification
were introduced [10–12]. These
for fire identification technologies
[10–12]. typically
These technologies
detectdetect
typically fires after
fires they
after have occurred.
they have Various
occurred. advanced
Various early detection
advanced methods
early detection may bemay
methods ablebetoable
effectively reduce fire losses and the number of false alarms. Early detection technologies
to effectively reduce fire losses and the number of false alarms. Early detection technologies can protect can protect
electrical equipment and humans before fires occur. Consequently, the development of advanced
electrical equipment and humans before fires occur. Consequently, the development of advanced early
early detection technologies for reducing the occurrence of electrical fires caused by arc faults
detection technologies for reducing the occurrence of electrical fires caused by arc faults represents an
represents an urgent task.
urgent task.
What is an arc fault? According to the Underwriters Laboratories (UL) Standard UL 1699, arcing
isWhat is anasarca fault?
defined According
luminous to the
electrical Underwriters
discharge across Laboratories
an insulating(UL) Standard
medium, and ULthe1699, arcing
partial
is defined as a luminous
volatilization electricalisdischarge
of the electrodes across an insulating
usually accompanied medium,
by the arcing andAn
process. thearc
partial volatilization
fault is defined
of the
as electrodes
an unintendedis usually
arcingaccompanied
situation inby a the arcing
circuit [13].process.
However,An visible
arc faultsparks,
is defined as annaturally
or arcs, unintended
arcing situation
generated in a circuit
in relays [13]. motors
and brush However,are visible
differentsparks,
from arcor faults.
arcs, naturally
Such sparks generated in relays and
are arcs generated
brush motors
under normal areoperation
differentand fromarearc faults. Such
considered to be sparks are arcsingenerated
non-hazardous terms of fireunder normal
detection. Arcoperation
faults
and are
are generated
considered in to
circuits for many reasons.
be non-hazardous Typical
in terms causes
of fire include Arc
detection. electrical
faultscords or wires having
are generated in circuits
loose connections;
for many reasons. Typical worn insulation on electrical
causes include cordscords
electrical or wires
or resulting from corrosion,
wires having age, bending
loose connections; worn
stress or heat; and damaged or misapplied electrical equipment. Figure 1
insulation on electrical cords or wires resulting from corrosion, age, bending stress or heat; and shows a process of an
electrical fire caused by an arc fault. The insulating material layer of the damaged electrical wire is
damaged or misapplied electrical equipment. Figure 1 shows a process of an electrical fire caused by
carbonized, and sporadic arcs are subsequently generated. Over time, the carbonized paths are
an arc fault. The insulating material layer of the damaged electrical wire is carbonized, and sporadic
further expanded, and severe arcs are formed. These arcs can release a tremendous amount of energy
arcs are subsequently generated. Over time, the carbonized paths are further expanded, and severe
and can reach temperatures of up to 13,000 K [14]. If combustible materials are nearby, a fire will
arcs easily
are formed. These[14,15].
be produced arcs can release a tremendous amount of energy and can reach temperatures of
up to 13,000 K [14]. If combustible materials are nearby, a fire will easily be produced [14,15].

Figure1.1.An
Figure Anelectrical
electrical fire
fire resulting
resultingfrom
fromananarc
arcfault.
fault.

Arc faults are categorized into three types: series arc faults, parallel arc faults and ground arc
Arc faults
faults. As a are categorized
result into three
of the existence types:
of fault series arc faults,
arc impedance, parallelofarc
the currents faults
series arcand ground
faults arc faults.
are usually
As alower
resultthan
of the
the existence of fault
rated currents arc impedance,
of loads, the currents
and the currents of parallelofand
series arc faults
ground are are
arc faults usually lower
usually
thanlower
the rated currents of loads, and the currents of parallel and ground arc faults
than those of short circuits. The arc fault currents are not sufficient to trigger traditional are usually lower
thanprotection
those of short circuits. The arc fault currents are not sufficient to trigger traditional
devices such as overcurrent protectors and ground fault circuit interrupters. Therefore, protection
devices
thesesuch as overcurrent
protection protectors
devices cannot and to
be used ground
protectfault circuit
circuits frominterrupters. Therefore,
arc faults [16]. Moreover,these
arcprotection
faults
are often
devices cannot latent, intermittent
be used andcircuits
to protect transient;
fromthus,
arcone should
faults [16].monitor
Moreover,circuit
arcstates
faultsinare
real-time to
often latent,
detect such faults to prevent electrical fires. Hence, an arc fault detector (AFD)
intermittent and transient; thus, one should monitor circuit states in real-time to detect such faultsrepresents an urgent
need in electrical
to prevent electrical fire detection.
fires. Hence, an arc fault detector (AFD) represents an urgent need in electrical
Arc
fire detection. fault circuit interrupters (AFCIs) represent an important AFD for preventing fires. The
president of the National Association of State Fire Marshals (NASFM), John C. Bean, stated that
Arc fault circuit interrupters (AFCIs) represent an important AFD for preventing fires. The
“NASFM strongly supports the broad adoption of AFCI technology through national, state, and local
president of the National Association of State Fire Marshals (NASFM), John C. Bean, stated that
building codes. AFCIs are the most welcome addition to fire prevention in decades. AFCIs promise
“NASFM
to savestrongly
hundreds supports
of lives the broad
every year”adoption
[17] (pageofnumber
AFCI technology through national,
of quote). Moreover, the USFAstate,
statedand
thatlocal
building codes. AFCIs
“The Consumer are Safety
Product the most welcome(CPSC)
Commission additionhasto fire prevention
identified in decades.
arc fault circuit AFCIs
interrupter promise
(AFCI)
to save
technology as an effective means of preventing fires caused by electrical wiring faults in homes” [17] that
hundreds of lives every year” [17] (page number of quote). Moreover, the USFA stated
“The(page
Consumer
number Product
of quote).Safety
UL Commission
has released the (CPSC) has identified
standard UL 1699: UL arc standard
fault circuitfor interrupter (AFCI)
safety arc-fault
technology as an effective
circuit-interrupters [13],means
and theofInternational
preventing fires caused by electrical
Electrotechnical Commission wiring
(IEC)faults in homes”
has released the [17]
(page number of quote). UL has released the standard UL 1699: UL standard for safety arc-fault
circuit-interrupters [13], and the International Electrotechnical Commission (IEC) has released the
standard IEC 62606: general requirements for arc fault detection devices [18]. Furthermore, the
National Electrical Code (NEC) requires the installation of AFCIs.
Sensors 2016, 16, 500 3 of 24

Because the detection method is the main component of an AFCI, many researchers have studied
arc fault detection technologies. Some researchers have developed mathematical arc fault models such
as an instantaneous arc model for a resistive-inductive system [19,20], a simplified arc fault model for
analyzing and determining basic parameters [21–23], and an arc fault meta model of conductance [24].
Many studies have shown that arc faults exhibit some abnormal behaviors in terms of arc radiation,
arc sounds, arc light, arc temperatures, arc currents and arc voltages in circuits [25–29]. When arc
faults occur in circuits, the radiated electromagnetic signals increase in intensity and can be collected
using a loop or stick antenna [25]; sounds and light from fault arcs appear and significantly increase in
intensity [26,27]; and arc temperatures quickly increase to up to 13,000 K, with a large amount of heat
being instantaneously released to the surrounding area [14]. Fault arc voltages in transmission lines
can be estimated through the least error squares method and can be used to identify arc faults [28,29].
However, mathematical arc fault models are usually used in theoretical analysis and simulations.
Arc radiation, arc sounds, arc light, arc temperatures and arc voltages are often characterized by
randomness because the times and locations of arc faults are always random. Therefore, arc fault
detection approaches based on such behaviors have certain limitations when used in practice.
Line currents, which are easily measured in circuits, have been used to study arc fault features
in many studies. Many high-frequency components have been found in the currents of arc faults,
the frequency bands of which are usually wide [30,31]. The spectral energy of currents has been
calculated to discriminate arc faults from load states [32]. Some arc fault detection algorithms have
been implemented in microcontrollers [33,34], and various patents for arc fault detection devices have
been approved [35–38]. However, the accuracies of many detection devices have not been high [39].
Because arc fault features are always hidden in the line currents, many advanced methods have
been used to extract them. A Kalman filtering algorithm, artificial neural networks and fuzzy logic
algorithms have been separately used to classify arc faults and normal states [40–42]. The harmonic
components of arc fault currents have been analyzed [43,44]. The reconstructed information entropy
of each current frequency band has been calculated to obtain the feature frequency bands of arc
faults [45]. Arc fault currents have been decomposed based on the discrete wavelet transform (DWT),
where the wavelet transform coefficients were used as fault features [46]. In one study, currents were
first de-noised by DWT to improve the signal-to-noise ratio, and the current eigenvectors were then
trained through wavelet networks to detect arc faults [47]. The Gabor Transform (GT) has been used to
obtain an optimal feature extraction; the extracted GT coefficients of currents were used to identify arc
faults [48].
However, arc fault currents are always strongly affected by different types of loads. Fault currents
produced by certain home appliances do not change noticeably, and arc fault features are usually
easily masked by load currents and background noise. It is difficult to find a general fault feature that
performs well for all loads. Moreover, the residential electrical standard for single-phase alternating
current (AC) is 120 V (60 Hz) in the United States, whereas it is 220–240 V (50–60 Hz) in other countries
including China, Germany, Switzerland and Korea. Higher voltages more readily produce electrical
fires caused by arc faults [49–52]. For example, the probability of fire ignition due to arc faults increases
from 3.5% for 120 V to 83% for 240 V, and their nominal current levels are both 15 A [49]. Compared
to 120 V, the higher voltages from 220 to 240 V are more likely to break down gaps and lead to more
arcs [50,51]. Arc currents are usually continuous in higher voltage systems, but they are sometimes
intermittent in 120 V systems [49,50]. Thus, higher voltage systems generate larger arc energy and thus
provide better conditions for the ignition of electrical fires [50,51].
Furthermore, some commercial AFCIs often fail to trip when necessary or trip when they should
not; the accuracy of arc fault detection was only approximately 50% in a previously published research
report [39]. Consequently, it remains difficult to accurately detect all arc faults in circuits, and certain
detection methods continue to require refinement, especially for 220–240 V operation.
In this study, to increase the accuracy of arc fault detection, a novel AFD will be developed for the
early detection of electrical fires. First, many currents and high-frequency signals from low-voltage
circuits will be acquired to study the general features of arc faults. Then, common features will
when addressing high dimensions or small samples [53–55]. A weighted least squares support vector
machine (WLS-SVM) algorithm will be introduced for arc fault identification in this study. Finally,
the arc fault detection algorithm will be applied to hardware toward developing an AFD.
The remainder of the paper is organized as follows. In Section 2, an experimental platform for
electrical fire research is built, and a substantial amount of experimental data is collected. In Section
3, different
Sensors 2016, 16, 500 types of load signals are analyzed to find the common features of arc faults, and high- 4 of 24
frequency energy and current variations are extracted as the eigenvector for arc fault detection.
Section 4 presents the arc fault detection algorithm based on WLS-SVM. In Section 5, the final design
be extracted
of the AFDandforused tothe
use in identify arc faults
early detection using anfires
of electrical advanced identification
is introduced, method. of
and the performance Compared
the
to general
AFD is identification
tested. Finally,algorithms, such as
Section 6 concludes thisneural
paper. network algorithms, support vector machine
(SVM)-based algorithms can obtain global optimums and have good generalization ability, especially
2. Experiments
when addressing high dimensions or small samples [53–55]. A weighted least squares support vector
The experimental
machine (WLS-SVM) arc fault
algorithm platform
will for electrical
be introduced for fire
arcresearch mainly consists
fault identification in of electrical
this study. loads,
Finally, the
an arc generator, transducers, and a data acquisition system. Figure
arc fault detection algorithm will be applied to hardware toward developing an AFD. 2 shows a schematic and the
actual experimental platform. The platform is built based on the relevant standards for arc faults in
The remainder of the paper is organized as follows. In Section 2, an experimental platform
low-voltage circuits. The standards include IEC 62606: 2013, UL 1699 and GB/T 14287.4-2014
for electrical fire research is built, and a substantial amount of experimental data is collected.
[13,18,56]. Typical electrical loads for the experiments are listed in Table 1.
In Section 3, different types of load signals are analyzed to find the common features of arc faults, and
high-frequency energy and current variations areexperimental
Table 1. Typical extracted as the eigenvector for arc fault detection.
loads.
Section 4 presents the arc fault detection algorithm based on WLS-SVM. In Section 5, the final design
Number Load Name Quantity Unit Power (W) Total Power (W)
of the AFD for use in the early detection of electrical fires is introduced, and the performance of the
1 Fluorescent lamp 2 40 80
AFD is tested. Finally, Section 6 concludes this paper.
2 Halogen lamp 6 50 300
3 Computer 1 350 350
2. Experiments
4 Electrical drill 1 750 750
The experimental
5 arc faultlamp
Dimming platform for electrical
5 fire research
200 mainly consists 1000
of electrical loads,
an arc generator,
6 transducers, and a data acquisition
Electric stove 1 system.1000Figure 2 shows a schematic
1000 and the
actual experimental
7 platform. The
Vacuum cleaner platform is
1 built based on the
1200 relevant standards for
1200 arc faults in
low-voltage8circuits. The standards
Air conditioner include IEC1 62606: 2013, UL 1699
1500 and GB/T 14287.4-2014
1500 [13,18,56].
Typical electrical
9 loadsAir
forcompressor
the experiments are1 listed in Table 22001. 2200

Arc generator

Transducers

Load 1 Socket 1
220V
sockets
Sensors 2016, 16, 500 Socket 25 of 24

(a) (b)
CT

HFT

PXI high- Arc generator


speed data
acquisition
system

Different
types
of loads

(c)
Figure 2. Experimental platform for electrical fire research: (a) schematic; (b,c) actual platform.
Figure 2. Experimental platform for electrical fire research: (a) schematic; (b,c) actual platform.
The arc generator includes a moving electrode and a stationary electrode, as shown in Figure 3.
One electrode is a carbon-graphite rod 6.4 mm in diameter, and the other electrode is a copper rod.
The arc generator is used to generate arc faults in circuits. The transducers consist of a high-frequency
transducer (HFT) and a current transducer (CT), as shown in Figure 2c. A schematic of the HFT is
shown in Figure 4. The input and output terminals of the transducer are connected to the power
supply terminals of the loads and data acquisition system, respectively. A transducer CT210 is
speed data
acquisition
system
Sensors 2016, 16, 500 5 of 24

Table 1. Typical experimental loads.

Number Load Name Quantity Unit Power (W) Total Power (W)
Different
1 Fluorescent lamp 2 40 80
2 types Halogen lamp 6 50 300
3 Computer 1 350 350
4 of loads Electrical drill 1 750 750
5 Dimming lamp 5 200 1000
6 Electric stove 1 1000 1000
7 Vacuum cleaner 1
(c) 1200 1200
8 Air conditioner 1 1500 1500
Figure 2. Experimental platform for electrical fire research: (a) schematic; (b,c) actual platform.
9 Air compressor 1 2200 2200

The arc generator includes a moving electrode and a stationary electrode, as shown in Figure 3.
The arc generator
One electrode includes a moving
is a carbon-graphite rod 6.4 electrode and a stationary
mm in diameter, and theelectrode, as shown
other electrode is aincopper
Figure rod.
3.
One electrode is a carbon-graphite rod 6.4 mm in diameter, and the other electrode
The arc generator is used to generate arc faults in circuits. The transducers consist of a high-frequency is a copper rod.
The arc generator
transducer (HFT) and is used to generate
a current arc faults
transducer in circuits.
(CT), as shownTheintransducers
Figure 2c.consist of a high-frequency
A schematic of the HFT is
transducer (HFT) and a current transducer (CT), as shown in
shown in Figure 4. The input and output terminals of the transducer are connected Figure 2c. A schematic ofto
thethe
HFT is
power
shown in Figure 4. The input and output terminals of the transducer are connected to the power supply
supply terminals of the loads and data acquisition system, respectively. A transducer CT210 is
terminals of the loads and data acquisition system, respectively. A transducer CT210 is selected as the
selected as the CT. A PXI-5124 digitizer fixed in PXIe-1071 is used for high-speed data acquisition,
CT. A PXI-5124 digitizer fixed in PXIe-1071 is used for high-speed data acquisition, and the sample
and the sample rate is set as 125 MS/s. As shown in Figure 4, the high-frequency signals in the power
rate is set as 125 MS/s. As shown in Figure 4, the high-frequency signals in the power supply circuit
supply circuit are coupled by a capacitor C and a transformer T. A high-pass filter composed of a
are coupled by a capacitor C and a transformer T. A high-pass filter composed of a capacitor C and the
capacitor C and the primary coil L1 of T is used. The power frequency signals at 50 Hz will be first
primary coil L1 of T is used. The power frequency signals at 50 Hz will be first filtered by the high pass
filtered by the high pass filter. According to the principle of electromagnetic induction, many high-
filter. According to the principle of electromagnetic induction, many high-frequency current signals
frequency current signals
will be generated will be generated
in the secondary in thethe
coil L2. Then, secondary coil L2.voltage
high-frequency Then, the high-frequency
signals voltage
will be generated
signals will be generated in a resistor R. Finally, they will be collected by the PXI-5124
in a resistor R. Finally, they will be collected by the PXI-5124 digitizer. The transformer T is used as digitizer. The
transformer
an isolatedTunitis used as anthe
between isolated
strongunit between
electricity the strong
signal and theelectricity signal and
weak electricity theBoth
signal. weakarc electricity
faults
signal. Both arc faults and normal states of typical loads can be realized using
and normal states of typical loads can be realized using the experimental platform. Large quantities of the experimental
platform. Large quantities
high-frequency signals and ofcurrents
high-frequency signals
in different and currents
load states in different
are collected to studyload states arefeatures
the common collected
to of
study the common
arc faults in the next features
section.of arc faults in the next section.

(a) (b)

Sensors 2016, 16, 500 Figure 3. The arc generator: (a) structure; (b) the actual device.
Figure 3. The arc generator: (a) structure; (b) the actual device. 6 of 24

Figure
Figure4.4.The
The HFT schematic.
HFT schematic.

3. Analysis of Experimental Results

3.1. High-Frequency Signal Analysis

3.1.1. High-Frequency Signals and Power Spectra


High-frequency signals were found to be a reflection of the dynamic arcing process in many arc
fault experiments. Upon the initial arcing, many air molecules between electrodes begin to be ionized;
Figure 4. The HFT schematic.

3. Analysis of Experimental Results


Sensors 2016, 16, 500 6 of 24

3.1. High-Frequency Signal Analysis


3. Analysis of Experimental Results
3.1.1. High-Frequency Signals and Power Spectra
3.1. High-Frequency Signal Analysis
High-frequency signals were found to be a reflection of the dynamic arcing process in many arc
fault 3.1.1. High-Frequency
experiments. Upon the Signals and
initial Powermany
arcing, Spectraair molecules between electrodes begin to be ionized;
subsequently, the plasma
High-frequency motion
signals werequickly
found intensifies. Because
to be a reflection electrode
of the dynamicgaps,
arcingoxide layers,
process surface
in many
states,
arcadsorbed gases andUpon
fault experiments. dielectric materials
the initial arcing,are usually
many heterogeneous,
air molecules betweenthis dynamic
electrodes motion
begin to is
be ionized; subsequently, the plasma motion quickly intensifies. Because electrode
always uncertain. According to electromagnetic theory, many high-frequency signals can be easily gaps, oxide
layers, in
generated surface states,These
circuits. adsorbed gases will
signals and dielectric materials are
spread through usually heterogeneous,
conductors this dynamic
or radiate into the air. An
electromagnetic field E r is generated by the current i, as shown in Figure 5. The valuecan
motion is always uncertain. According to electromagnetic theory, many high-frequency signals of E r
be easily generated in circuits. These signals will spread through conductors or radiate into the air.
oriented along the x-axis can be expressed as:
An electromagnetic field Er is generated by the current i, as shown in Figure 5. The value of Er oriented
along the x-axis can be expressed as:
sin   di
2 0
Er  D , t   dx (1)
 0 Dc δ didt
4sinφ ż
Er pD, tq « dx (1)
4πε 0 Dc2 dt
where D is the distance from the component, t is time,0  is the angle with the y-axis,  0 is the
where D is the distance from the component, t is time, φ is the angle with the y-axis, ε 0 is the vacuum
vacuum permittivity, c is the speed of light, and  is the length of the current component [57]. For a
permittivity, c is the speed of light, and δ is the length of the current component [57]. For a spot arc, the
spot arc, the length  is negligible; therefore, Equation (1) can be simplified as:
length δ is negligible; therefore, Equation (1) can be simplified as:
dia . 
E  t   dia “
Ea ptq
a « γ
γi ia (2)
(2)
dtdt
a

where  isγaisconstant.
where a constant.

Figure 5. 5.
Figure The electromagnetic
The fieldalong
electromagnetic field alongthe
thex-axis.
x-axis.

The power spectrum


The power of of
spectrum thethe
arc
arcfault
faultcan
canbe
be expressed as:
expressed as:

γ2 2iai. a 0p0q 2ω2 2


ı2
 

ˇż 8  ˇ2
 jt ˇ2
Pa pωq   
Pa “ ˇ = EE
ˇ
0 a ptq
a 
te´e jωt dt
dtˇ “  ` (3) (3)
 
˘2
ˇ
0
ˇ
µ22 ´ω22 2`44ϕ22 ω2 2
a
where R is the arc resistance, L is the arc inductance, C is the arc capacitance, µ “ 1{ L C is the
where R a isa the arc resistance, La a is the arc inductance, C a is the arc capacitance, =1
a a a LaCa is
natural frequency of the arc fault, and ϕ “ R a {2L a is the arc inductive damping rate [58]. Considering
the natural frequency
that the fault of the can
arc inductance arc befault, andthe power
ignored, is theofarc
= Ra 2 Lspectrum
a the inductive damping
fault arc in the ratebe[58].
circuit can
approximated
Considering as: fault arc inductance can be ignored, the power spectrum of the fault arc in the
that the ”. ı2
circuit can be approximated as: γ2 i a p0q µ2 ω 2 R2a Ca2
Pa pωq « (4)
1 ` R2a Ca2 ω 2
An electrical drill’s high-frequency signals and corresponding power spectral density (PSD) are
shown in Figure 6. The high-frequency signals of arc faults are densely distributed and exhibit large
amplitudes. Various frequency bands with large amplitudes can also be found in the power spectrum.
 2 ia  0   22 Ra2Ca2
2

Pa    (4)
1  Ra2Ca22

An
Sensors electrical
2016, 16, 500 drill’s high-frequency signals and corresponding power spectral density (PSD) are
7 of 24
shown in Figure 6. The high-frequency signals of arc faults are densely distributed and exhibit large
amplitudes. Various frequency bands with large amplitudes can also be found in the power spectrum.
However, when
However, when anan electrical
electrical drill
drill is
is operated
operated inin the
the normal
normal state,
state, random
random interference
interference pulses
pulses are
are
generated in circuits. The pulses are caused by a brush motor in the electrical drill.
generated in circuits. The pulses are caused by a brush motor in the electrical drill. Electrical Electrical discharges
are usuallyare
discharges accompanied by friction between
usually accompanied the brush
by friction between andthe
rotor. Electrical
brush sparks
and rotor. are easily
Electrical observed
sparks are
from the electrical drill’s windows. These sparks are load arcs and are not hazardous.
easily observed from the electrical drill’s windows. These sparks are load arcs and are not hazardous. Interference
generated bygenerated
Interference these sparksby should be overcome
these sparks shouldinbearcovercome
fault detection.
in arcTherefore, the varied
fault detection. regularities
Therefore, the
of spectra over time will help facilitate further analysis.
varied regularities of spectra over time will help facilitate further analysis.

30
1.5
20
1

Voltage (V)
Voltage (V)

10
0.5
0
0
-10
-0.5
-20 -1
0.01 0.02 0.03 0.04 0.01 0.02 0.03 0.04
Time (s) Time (s)
(a) (b)

2
30

1.5
PSD(W/Hz)

PSD(W/Hz)

20
1

10
0.5

0 0
1 2 3 4 5 6 0 1 2 3 4 5 6
Frequency (Hz) 7 Frequency (Hz) 7
×
x 10 x× 10
(c) (d)
Figure
Figure 6. Signalanalysis
6. Signal analysisofof
anan electrical
electrical drill:drill: (a) high-frequency
(a) high-frequency signalssignals of arc(b)
of arc faults; faults; (b) high-
high-frequency
frequency
signals under normal operation; (c) PSD of arc fault signals; (d) PSD under normal operation.normal
signals under normal operation; (c) PSD of arc fault signals; (d) PSD under
operation.

3.1.2. Short-Time Fourier Transform (STFT) of High-Frequency Signals


3.1.2. Short-Time Fourier Transform (STFT) of High-Frequency Signals
Because arc faults are non-stationary, their frequency spectra and power spectra are usually time
Because arc faults are non-stationary, their frequency spectra and power spectra are usually time
variant. To further research the features of arc faults in different time and frequency bands, STFT is
variant. To further research the features of arc faults in different time and frequency bands, STFT is
introduced. The arc fault signals’ STFT for the electrical drill is shown in Figure 7. The high-frequency
introduced. The arc fault signals’ STFT for the electrical drill is shown in Figure 7. The high-frequency
signals of the arc faults can be found in both the line current and phase voltage, and they are generated
signals of the arc faults can be found in both the line current and phase voltage, and they are
periodically following the current cycles, as shown in Figure 7a,b. They clearly change in a wide
generated periodically following the current cycles, as shown in Figure 7a,b. They clearly change in
frequency band, especially under 30 MHz, as shown in Figure 7c.
a wide frequency band, especially under 30 MHz, as shown in Figure 7c.
Sensors 2016, 16, 500 8 of 24
Sensors 2016, 16, 500 8 of 24

(a)

(b)

(c)

Figure
Figure7.7.Signals’
Signals’STFT
STFTfor
forthe
theelectrical
electricaldrill.
drill.(a)
(a)line
linecurrent;
current;(b)
(b)high-frequency
high-frequencysignals;
signals;(c)
(c)signals’
signals’
STFT of arc faults.
STFT of arc faults.

To find the common features of arc faults, many high-frequency signals produced by different
To find the common features of arc faults, many high-frequency signals produced by different
typical loads are collected, as shown in Figures 8 and 9. Substantial power grid noise is generated by
typical loads are collected, as shown in Figures 8 and 9. Substantial power grid noise is generated
power lines, and some short pulses are initially generated when the vacuum cleaner operates in the
by power lines, and some short pulses are initially generated when the vacuum cleaner operates in
normal state, as shown in Figure 8b. The power grid noise mainly originates from power line carrier
the normal state, as shown in Figure 8b. The power grid noise mainly originates from power line
communication, switching power supplies, load faults, etc. According to the provisions from the State
carrier communication, switching power supplies, load faults, etc. According to the provisions from
Grid Corporation of China, the power line carrier frequency band is 10–500 kHz [59]. Some
the State Grid Corporation of China, the power line carrier frequency band is 10–500 kHz [59]. Some
frequencies of other strong noise sources in the power grid can be up to 1.5 MHz [59,60]. According
frequencies of other strong noise sources in the power grid can be up to 1.5 MHz [59,60]. According
to the signals’ STFT under normal operation, the large amplitude noise contains frequencies that are
to the signals’ STFT under normal operation, the large amplitude noise contains frequencies that are
also mainly under 1.5 MHz, as shown in Figure 8d. The switch transformers in the halogen lamps
also mainly under 1.5 MHz, as shown in Figure 8d. The switch transformers in the halogen lamps
often alternate between on and off states; they are controlled based on the pulse width modulation.
often alternate between on and off states; they are controlled based on the pulse width modulation.
Therefore, many high-frequency signals are also periodically generated when operating in the normal
Therefore, many high-frequency signals are also periodically generated when operating in the normal
states of the halogen lamps, some frequencies of which are up to hundreds of kHz [61]. The normal
states of the halogen lamps, some frequencies of which are up to hundreds of kHz [61]. The normal
signal frequencies with large amplitudes are mostly less than 1.5 MHz, as shown in Figure 9d.
signal frequencies with large amplitudes are mostly less than 1.5 MHz, as shown in Figure 9d.
Broadband frequency components from arc faults can be clearly found in Figures 7c, 8c and 9c.
The amplitudes decrease with increasing frequency. The frequency components are not the same at
different times. They are also not the same for different types of loads. Noticeable spectrum
signatures appear near the zero-crossing regions of arc fault currents, and currents may disappear in
a short period of time, as shown in Figure 7a. The zero-crossing regions are also reflected in the
can be collected by an HFT, as shown in Figures 7b, 8a and 9a. In the practical engineering, according
to the propagation characteristics of electromagnetic interference, conduction propagation is the
main propagation mode of electromagnetic interference below 30 MHz [62]. Moreover, according to
the signals’ STFT, the high-frequency signals of arc faults under typical loads concentrating in the
1.5–30 MHz frequency band increase more noticeably than those under normal operation, as shown
in Figures
Sensors 2016, 16, 7c,
5008c,d and 9c,d; moreover, the signals may have greater energy. Therefore, 1.5–30 MHz9 of 24
can be regarded as a feature frequency band.

3
5
2

1
Voltage (V)

Voltage (V)
0 0

-1

-2
-5
0 0.02 0.04 0.06 0.08 -3
0 0.02 0.04 0.06 0.08
Time (s) Time (s)
(a) (b)

×× ×

(c) (d)
-4 -4
x 10
× x× 10
5 1

4 0.8
Energy (J)
Energy (J)

3 0.6

2 0.4

1 0.2

0 0
0.02 0.04 0.06 0.08 0.02 0.04 0.06 0.08
Time (s) Time (s)
(e) (f)
Figure
Figure 8. 8. Signal
Signalanalysis of aofvacuum
analysis cleaner:
a vacuum (a) high-frequency
cleaner: signals of arc
(a) high-frequency faults;of(b)arc
signals high-
faults;
frequency signals under normal operation; (c) signals’ STFT of arc fault signals; (d) signals’ STFT
(b) high-frequency signals under normal operation; (c) signals’ STFT of arc fault signals; (d) signals’
under normal operation; (e) high-frequency energy of arc faults; (f) high-frequency energy under
STFT under normal operation; (e) high-frequency energy of arc faults; (f) high-frequency energy under
normal operation.
normal operation.

Broadband frequency components from arc faults can be clearly found in Figures 7c, 8c and 9c.
The amplitudes decrease with increasing frequency. The frequency components are not the same at
different times. They are also not the same for different types of loads. Noticeable spectrum signatures
appear near the zero-crossing regions of arc fault currents, and currents may disappear in a short period
of time, as shown in Figure 7a. The zero-crossing regions are also reflected in the figures showing the
high-frequency signals and their spectra, as shown in Figures 7b,c, 8a,c and 9a,c. Near the zero-crossing
regions, arcs may extinguish themselves. However, they will re-ignite if there are sufficient voltages
to reestablish the arcs between the electrodes. As a result of the abrupt changes of the arc currents,
many high-frequency signals with large amplitudes are rapidly generated. They can be collected by an
HFT, as shown in Figures 7b, 8a and 9a. In the practical engineering, according to the propagation
characteristics of electromagnetic interference, conduction propagation is the main propagation mode
Sensors 2016, 16, 500 10 of 24

of electromagnetic interference below 30 MHz [62]. Moreover, according to the signals’ STFT, the
high-frequency signals of arc faults under typical loads concentrating in the 1.5–30 MHz frequency
band increase more noticeably than those under normal operation, as shown in Figures 7c, 8c,d and 9c,d;
moreover, the signals may have greater energy. Therefore, 1.5–30 MHz can be regarded as a feature
frequency band.
Sensors 2016, 16, 500 10 of 24

0.6
3
2 0.4

1 0.2
Voltage (V)

Voltage (V)
0 0
-1 -0.2
-2
-0.4
-3
0.02 0.04 0.06 0.08 0 0.02 0.04 0.06 0.08
Time (s) Time (s)
(a) (b)

× ×

(c) (d)
× ×

(e) (f)
Figure 9. Signal analysis of halogen lamps: (a) high-frequency signals of arc faults; (b) high-frequency
Figure 9. Signal analysis of halogen lamps: (a) high-frequency signals of arc faults; (b) high-frequency
signals under normal operation; (c) signals’ STFT of arc fault signals; (d) signals’ STFT under normal
signals under normal operation; (c) signals’ STFT of arc fault signals; (d) signals’ STFT under normal
operation; (e) high-frequency energy of arc faults; (f) high-frequency energy under normal operation.
operation; (e) high-frequency energy of arc faults; (f) high-frequency energy under normal operation.
3.1.3. The Periodic Energy of High-Frequency Signals
3.1.3. The Periodic Energy of High-Frequency Signals
The original signals from the HFT are filtered to obtain the high-frequency signals in the 1.5–30
The frequency
MHz original signals from
band. The the HFT
periodic areEhffiltered
energy to obtain
(k ) of these the high-frequency
high-frequency signals can besignals in the
calculated
as: MHz frequency band. The periodic energy Eh f pkq of these high-frequency signals can be
1.5–30
calculated as: ż kT kT
( k1)T ss2(ptqdt
E (k)=AC t)dt ( k  1,2, ,K)
2 AC

Eh f pkq “hf pk “ 1, 2, ¨ ¨ ¨ , Kq (5) (5)


AC
pk´1qTAC
where T = 10 ms is half the AC current period for a 50 Hz power grid, also being half an fault arc
where TAC =AC10 ms is half the AC current period for a 50 Hz power grid, also being half an fault arc
period,
period, andand
s(t)s(t) represents
represents thethe high-frequencysignals.
high-frequency signals.
The vacuum cleaner’s high-frequency energy Ehf (k ) is calculated as shown in Figure 8e,f, and
the halogen lamps’ Ehf (k) is calculated as shown in Figure 9e,f. Based on many calculations of Ehf (k ) ,
the Ehf (k) of arc faults is larger than that under normal operating conditions for a load. Nevertheless,
Sensors 2016, 16, 500 11 of 24

The vacuum cleaner’s high-frequency energy Eh f pkq is calculated as shown in Figure 8e,f, and
Sensors 2016, 16,lamps’
the halogen 500Eh f pkq is calculated as shown in Figure 9e,f. Based on many calculations of E11h fofpkq, 24

the Eh f pkq of arc faults is larger than that under normal operating conditions for a load. Nevertheless,
the Ehf (k) of different loads often varies. It is difficult to select a reasonable constant threshold to
the Eh f pkq of different loads often varies. It is difficult to select a reasonable constant threshold to
discriminate arc faults from the load states states for
for all
all loads.
loads. Furthermore, high-frequency signals are
easily influenced
influenced by by cross
crosstalk,
talk,asasshown
shownininFigure
Figure10.10.The
The high-frequency
high-frequency signals
signals in in branch
branch B1 B1
cancan
be
be detected
detected by by
thethe
HFT HFT in AFD
in AFD 2, which
2, which is an
is in in an adjacent
adjacent circuit.
circuit. Hence,
Hence, this
this will
will lead
lead to to false
false alarms.
alarms.

Figure
Figure 10.
10. Cross
Cross talk.
talk.

The high-frequency signals can propagate via radio frequency and conduction. Cross talk is
The high-frequency signals can propagate via radio frequency and conduction. Cross talk is
formed by electromagnetic coupling. It can be divided into capacitive coupling cross talk and
formed by electromagnetic coupling. It can be divided into capacitive coupling cross talk and inductive
inductive cross talk. Due to the capacitive coupling, the induced current is
cross talk. Due to the capacitive coupling, the induced current is
du
iM 1 =C M 1 du N 1 (6)
i M1 “ C M1 dtN1 (6)
dt
where CM1 is the coupling capacitance between wires M1 and N1, and uN 1 is the voltage of the
where C M1 is the coupling capacitance between wires M1 and N1, and u N1 is the voltage of the wire
wire N1. If uN 1 changes quickly, iM1 will become large. Thus, the high-frequency signals can
N1. If u N1 changes quickly, i M1 will become large. Thus, the high-frequency signals can generate a
generate
large i M1 .aDue
largeto ithe. Due to the inductive coupling, the induced voltage is
M1 inductive coupling, the induced voltage is

diN 1
u M1u“ =LM 1 di N1
M 1 L M1
(7)
(7)
dt
where LLM1
where is the coupling inductance between wires M1 and N1, and iN1 isisthe
M 1 is the coupling inductance between wires M1 and N1, and i N1
the current
current ofof the
the wire
wire
Similarly, the
N1. Similarly, the high-frequency
high-frequencysignals
signalscancangenerate large uuMM1
generateaalarge 1
[62]. Fortunately,
Fortunately, the
the voltage
voltage u M1
M1
induced by power frequency currents at 50 Hz is small and is easily masked in the power
induced by power frequency currents at 50 Hz is small and is easily masked in the power frequency frequency
voltages. In other
voltages. In other words,
words, power
power frequency
frequency currents
currents at
at 50
50 Hz
Hz are
are not
not easily
easily influenced
influenced byby cross
cross talk.
talk.
To overcome the influence of cross talk, the features of currents will be extracted and added
To overcome the influence of cross talk, the features of currents will be extracted and added for for analysis
in the next
analysis in section.
the next section.
3.2. Current Analysis
3.2. Current Analysis
3.2.1. Typical Load Currents
3.2.1. Typical Load Currents
Line currents of various typical loads exhibit various characteristics. When loads operate in
Line currents of various typical loads exhibit various characteristics. When loads operate in
stable states, as in the case of “normal” conditions, as shown in Figure 11, currents in circuits remain
stable states, as in the case of “normal” conditions, as shown in Figure 11, currents in circuits remain
relatively stable and exhibit good symmetrical characteristics and certain periodicities. When loads
relatively stable and exhibit good symmetrical characteristics and certain periodicities. When loads
can be characterized as transient processes, such as starting loads, dimming lamps, regulating speeds
can be characterized as transient processes, such as starting loads, dimming lamps, regulating speeds
and plugging in or unplugging, currents rapidly change monotonically at first but then return to stable
and plugging in or unplugging, currents rapidly change monotonically at first but then return to
states, as shown in Figure 12. Nevertheless, currents will become random or noticeably distorted, their
stable states, as shown in Figure 12. Nevertheless, currents will become random or noticeably
periodicities will be lost, and their amplitudes will become erratic if there are arc faults in circuits, such
distorted, their periodicities will be lost, and their amplitudes will become erratic if there are arc faults
as under “arc fault” conditions, as shown in Figure 11. Moreover, current amplitudes may become
in circuits, such as under “arc fault” conditions, as shown in Figure 11. Moreover, current amplitudes
zero at times. This is because the inductive and capacitive reactances of fault arcs are usually unstable,
may become zero at times. This is because the inductive and capacitive reactances of fault arcs are
and they may change sporadically as a result of external conditions. The entire varied regularities of
usually unstable, and they may change sporadically as a result of external conditions. The entire
currents from normal conditions to arc faults under different typical loads can be observed in Figure 11.
varied regularities of currents from normal conditions to arc faults under different typical loads can
be observed in Figure 11. Because current features can be used to overcome the influence of cross talk,
current variations should be quantitatively described to accurately extract arc fault features.
Sensors 2016, 16, 500 12 of 24

Sensors 2016, 16, 500 12 of 24


Sensors 2016, 16, 500 12 of 24
Because current features can be used to overcome the influence of cross talk, current variations should
overcome the influence of cross talk, current variations should be quantitatively described to
be quantitatively described of
overcome the influence to accurately
cross talk, extract
current arc fault features.
variations should be quantitatively described to
accurately extract arc fault features.
accurately extract arc fault features.
40
40 normal arc fault normal arc fault
normal arc fault 10 normal arc fault
10
20
5

(A)(A)
20
(A)(A)

Current
Current

Current
0 0
Current

0 0
-5
-20 -5
-20 -10
0 0.2 0.4 0.6 0.8 -100 0.2 0.4 0.6 0.8
0 0.2 Time
0.4 (s) 0.6 0.8 0 0.2 Time
0.4(s) 0.6 0.8
(a) (s)
Time Time
(b) (s)
6 (a) normal (b) arc fault
6 normal arc fault
4 normal arc fault
normal arc fault 6
4
2 (A)(A) 46
(A)(A)

2
0 24
Current
Current

0 02
Current
Current

-2
-2 0
-2
-4 -2
-4
-4
-6 -4
-6
-6 -6
-8
0 0.2 0.4 0.6 0.8 0 0.2 0.4 0.6 0.8
-8 Time
0 0.2 Time
0.4(s) 0.6 0.8 0 0.2 0.4(s) 0.6 0.8
(c) (s)
Time (d) (s)
Time
(c) (d)
Figure 11. Currents: (a) an air conditioner; (b) dimming lamps; (c) a computer; (d) an electric stove.
Figure 11. Currents: (a) an air conditioner; (b) dimming lamps; (c) a computer; (d) an electric stove.
Figure 11. Currents: (a) an air conditioner; (b) dimming lamps; (c) a computer; (d) an electric stove.
6 8
6 68
4
4 46
2
(A) (A)
(A) (A)

24
2
Current

02
Current

0
Current
Current

0 -20
-2
-4
-2
-2
-4 -4
-6
-4 -8
-6
-6
0 0.2 0.4 0.6 0.8 -80 0.2 0.4 0.6 0.8
-6 Time Time
0.4(s)
0 0.2 0.4 (s) 0.6 0.8 0 0.2 0.6 0.8
(a) (s)
Time (b) (s)
Time
10
(a) 3 (b)
10
23
5
12
(A) (A)

(A) (A)

5
Current

01
Current

0
Current

Current

0 -10
-5 -1
-2
-5
-2
-3
-10
0 0.2 0.4 0.6 0.8 -30 0.2 0.4 0.6 0.8
-10 Time (s) Time (s)
0 0.2 0.4 0.6 0.8 0 0.2 0.4 0.6 0.8
(c) (s)
Time (d) (s)
Time
(c)
Figure 12. Currents of transient (d) (b) dimming lamps; (c)
processes: (a) starting an electrical drill;
regulating
Figure 12. the speed of
Currents of atransient
vacuum processes:
cleaner; (d)(a)
plugging
startinginan
a halogen lamp.
electrical drill; (b) dimming lamps; (c)
Figure 12. Currents of transient processes: (a) starting an electrical drill; (b) dimming lamps;
regulating the speed of a vacuum cleaner; (d) plugging in a halogen lamp.
(c) regulating the speed of a vacuum cleaner; (d) plugging in a halogen lamp.
Sensors 2016, 16, 500 13 of 24

3.2.2. Current Feature Extraction


The
Sensors current
2016, 16, 500 integral can be used to observe changes in currents. This can be expressed as 13 of 24
ż kTAC
J pkq “ J (k )= kTAC pk “ 1, 2, ¨ ¨ ¨ , Kq (8)

|iptq|dt
i(t )dt ( k  1, 2, , K)
pk´1(qkTAC
1)TAC
(8)

where
where TT AC
AC
= =1010msmsisishalf
halfthe
theACACcurrent
currentperiod
periodandandi(t)
i(t)represents
representsthetheline
linecurrent.
current.An Anarc
arcfault
faultisis
clearly
clearly aa stochastic process.The
stochastic process. Theline
line currents
currents of
of arc
arc faults
faults are
are often
often seriously
seriously distorted
distorted or or change
change inin
uncertain
uncertain manners, as shown in Figures 13a, 14a and 15a. Current periodic integrals of arc faultsoften
manners, as shown in Figures 13a, 14a and 15a. Current periodic integrals of arc faults often
fluctuate
fluctuate and
and become
become erratic,
erratic, as
as shown
shown inin Figures
Figures13b,13b,14b
14bandand15b.
15b. Nevertheless,
Nevertheless, the the changes
changes inin
current periodic integrals can also be observed when loads operate under
current periodic integrals can also be observed when loads operate under normal conditions,normal conditions, especially
during transient
especially duringprocesses.
transient The electrical
processes. drill
The currentdrill
electrical during startup
current and its
during periodic
startup andintegrals are
its periodic
shown
integralsinare
Figure
shown 16a,b, respectively.
in Figure The vacuum
16a,b, respectively. Thecleaner
vacuum current
cleanerwhen regulating
current speed and
when regulating its
speed
periodic integrals are shown in Figure 17a,b, respectively. The current periodic integrals
and its periodic integrals are shown in Figure 17a,b, respectively. The current periodic integrals also also change
during
changethese
during transient processes;
these transient they change
processes; they in termsinofterms
change regularity and remain
of regularity monotonic
and remain over a
monotonic
certain periodperiod
over a certain of time. Hence,
of time. therethere
Hence, are are
some limitations
some limitations in in
identifying
identifying arc
arcfaults
faultsfrom
fromtransient
transient
processes only via current periodic integrals.
processes only via current periodic integrals.

4 normal arc fault normal arc fault


Current Integral (A·ms)

15
2
Current (A)

10
0

5
-2

0
0 0.2 0.4 0.6 0.8 0.2 0.4 0.6 0.8
Time (s) Time (s)
(a) (b)
25 normal arc fault
Current variation (A·ms)

20

15

10

0
0.2 0.4 0.6 0.8
Time (s)
(c)
Figure 13. Arc fault current analysis of the electrical drill: (a) current waves; (b) current periodic
Figure 13. Arc fault current analysis of the electrical drill: (a) current waves; (b) current periodic
integrals; (c) current variations.
integrals; (c) current variations.
Sensors 2016, 16, 500 14 of 24
Sensors 2016, 16, 500 14 of 24
Sensors 2016, 16, 500 14 of 24
15 50
arc fault normal arc fault normal
15

(A·ms)
50
10 arc fault normal arc fault normal
40

(A·ms)
10 40
(A)(A)

Integral
5 30

Integral
Current

5 30
0 20
Current

Current
0 20

Current
-5 10
-5 10
-10 0
0 0.2 0.4 0.6 0.8 0 0.2 0.4 0.6 0.8
-10 0 Time
0 0.2 Time
0.4(s) 0.6 0.8 0 0.2 0.4(s) 0.6 0.8
Time
(a) (s) Time
(b) (s)
(a) arc fault normal (b)
(A·ms)

arc fault normal


(A·ms)

100
100
variation
variation

50
Current

50
Current

0
0.4 0.6 0.2
0.8
Time 0
0.4 (s) 0.6 0.2
0.8
Time
(c) (s)
(c)
Figure 14. Arc fault current analysis of the vacuum cleaner: (a) current waves; (b) current periodic
Figure 14. Arc fault current analysis of the vacuum cleaner: (a) current waves; (b) current periodic
Figure 14.(c)
integrals; Arc fault variations.
current current analysis of the vacuum cleaner: (a) current waves; (b) current periodic
integrals; (c) current variations.
integrals; (c) current variations.

10 arc fault normal arc fault normal


(A·ms)

10 arc fault normal arc fault normal


(A·ms)

5 10
(A)(A)

10
Integral

5
Current

Integral

0
Current

0 5
Current

-5 5
Current

-5
-10
0
-100 0.2 0.4 0.6 0.8 0 0.2 0.4 0.6 0.8
Time 0
0 0.2 0.4(s) 0.6 0.8 0 0.2 Time
0.4(s) 0.6 0.8
Time
(a) (s) Time
(b) (s)
(a) 30 (b)
arc fault normal
(A·ms)

30 arc fault normal


(A·ms)

20
variation

20
variation

10
Current

10
Current

0
0.2 0.4 0.6 0.8
0 Time
0.2 0.4 (s) 0.6 0.8
Time
(c) (s)
(c)
Figure 15. Arc fault current analysis of the halogen lamps: (a) current waves; (b) current periodic
Figure 15.(c)Arc
integrals; fault variations.
current current analysis of the halogen lamps: (a) current waves; (b) current periodic
Figure 15. Arc fault current analysis of the halogen lamps: (a) current waves; (b) current periodic
integrals; (c) current variations.
integrals; (c) current variations.
Sensors 2016, 16, 500 15 of 24
Sensors 2016, 16, 500 15 of 24
Sensors 2016, 16, 500 15 of 24
30

(A·ms)
5 30

(A·ms)
(A)(A) 5
20

Integral
20
Current

Integral
0
Current

0
10

Current
10

Current
-5
-5
0
0 0.2 0.4 0.6 0.8 00 0.2 0.4 0.6 0.8
0 0.2 Time
0.4(s) 0.6 0.8 0 0.2 Time
0.4(s) 0.6 0.8
Time
(a) (s) Time
(b) (s)
(a) 15 (b)
15
(A·ms)
(A·ms)

10
variation

10
variation

5
Current

5
Current

0
0.4 0.6
0 0.2 0.8
Time
0.4 (s) 0.6 0.2 0.8
Time
(c) (s)
(c)
Figure 16. Transient process analysis of the electrical drill: (a) current waves; (b) current periodic
Figure 16. Transient process analysis of the electrical drill: (a) current waves; (b) current periodic
Figure 16.(c)
integrals; Transient process analysis of the electrical drill: (a) current waves; (b) current periodic
current variations.
integrals; (c) current variations.
integrals; (c) current variations.

50
(A·ms)

10
50
(A·ms)

10 45
5
(A)(A)

45
Integral

5 40
Current

0
Integral

40
Current

0 35
-5 35
Current

30
-5
Current

-10 30
25
-10 25
0 0.2 0.4 0.6 0.8 0 0.2 0.4 0.6 0.8
0 0.2 Time
0.4(s) 0.6 0.8 0 0.2 Time
0.4(s) 0.6 0.8
Time
(a) (s) Time
(b) (s)
(a) (b)
(A·ms)

15
(A·ms)

15
variation

10
variation

10
Current

5
Current

0
0.2 0.4 0.6 0.8
0 Time
0.2 0.4 (s) 0.6 0.8
Time
(c) (s)
(c)
Figure 17. Transient process analysis of the vacuum cleaner: (a) current waves; (b) current periodic
Figure 17.(c)
integrals; Transient process analysis of the vacuum cleaner: (a) current waves; (b) current periodic
current variations.
Figure 17. Transient process analysis of the vacuum cleaner: (a) current waves; (b) current periodic
integrals; (c) current variations.
integrals; (c) current variations.
Sensors 2016, 16, 500 16 of 24

To avoid incorrect determinations during transient processes, the current variations in several
current periods are introduced. According to the geometric meaning of definite integrals, the definite
integral can be used to represent the total area of a graph. This area contains positive and negative
areas such as the current wave area over a current period. The current period is 20 ms. Load current
waves of arc faults are often asymmetrical, whereas those of normal states are symmetrical. Therefore,
the definite integrals in AC current periods can be used to represent current variations. The current
variations can be calculated as the summation of several current periodic integrals over a certain period
of time expressed as: ˇ ˇ
K ˇż pk`1qTAC
ÿ ˇ
Jsum pkq “ i ptqdtˇ pk “ 1, 2, ¨ ¨ ¨ , Kq (9)
ˇ ˇ
ˇ
ˇ pk´1qTAC ˇ
k “1

where TAC = 10 ms is half the AC current period and K = 8; therefore, the total time is 80 ms. The current
variations of the electrical drill, vacuum cleaner and halogen lamps are shown in Figures 13c, 14c
and 15c, respectively. From the analysis of many current variations, we find that the current variation
thresholds between arc faults and normal states are unequal under different loads. Taking start-up
states and arc faults as examples, some current variations in the electrical drill startup state are similar
to those of the arc faults of halogen lamps, as shown in Figures 15c and 16c. This may hence lead to
false identification.
In conclusion, the high-frequency energy of arc faults clearly increases, and the current periodic
integrals of arc faults randomly change. Nevertheless, the high-frequency signals are easily influenced
by cross talk, and the use of only current variations may result in false identification under different
loads. To accurately detect arc faults, several features, including the high-frequency energy and
current variations, should be used simultaneously. Furthermore, the development of an intelligent
detection algorithm also represents a very important and urgent need. The detection algorithm will be
introduced and detailed in the next section.

4. Arc Fault Detection Algorithm


Compared to general identification algorithms, such as neural network algorithms, SVM-based
algorithms can avoid falling into local minima, over learning, the dimensional disaster [63,64], etc.
Such algorithms can obtain a global optimum and have good generalization ability, especially when
addressing high dimensions or small samples [53–55]. Based on the traditional SVM, Suykens
and Vandewalle presented the LS-SVM algorithm, which can ensure satisfactory identification
accuracy while reducing the computational complexity [53]. To more accurately detect arc faults, the
high-frequency energy and current variations are selected as the eigenvector. They can be calculated
using Equations (5) and (9), respectively. Based on the input eigenvector y, the WLS-SVM algorithm
will be applied to arc fault detection. The WLS-SVM algorithm has emerged as an improved algorithm
for enhancing the sparseness and robustness of LS-SVM [65].
A nonlinear function f (y) is used to map the eigenvector y from the observation space to a higher
dimensional feature space. The optimal identification between arc faults and normal states can be
obtained in the higher dimensional feature space. The constrained optimization problem can be
described as [53,65]:
N
$
g ř
& min W pU, ζ n q “ 12 UT U ` 2 bn ζ n2
U,ζ n ,q n “1 (10)
s.t.zn UT f yn ` q “ 1 ´ ζ n , pn “ 1, . . . , Nq
% “ ` ˘ ‰

where U is the weight vector, N is the sample number, g is the penalty parameter, bn pn “ 1, . . . , Nq are
the weighting
# factors, q is the bias term, ζ n pn “ 1, . . . , Nq are the error variables, and the output data
1, yn P arc faults
are zn “ .
0, yn P normal states
Sensors 2016, 16, 500 17 of 24

To solve Equation (10), Lagrange multipliers β n are used. The Lagrangian becomes

N
ÿ ! ” ı )
β n z n UT f y n ` q ´ 1 ` ζ n
` ˘
L pU, ζ n , β n , qq “ W pU, ζ n q ´ (11)
n “1

Based on the Karush–Kuhn–Tucker condition, the partial derivatives of U, ζ n , β n , q in Equation (11)


are equal to 0. One obtains $
řN ` ˘
U“ β n f yn






’ n “1
& β n “ gbn ζ n
“ T ` ˘ ‰ (12)
’ z n U f yn ` q ´ 1 ` ζ n “ 0



’ N
’ ř

% β n zn “ 0 pn “ 1, . . . , Nq
n “1

The solution can be given as


˜ ¸˜ ¸ ˜ ¸
V ` Bg z β I
“ (13)
zT 0 q 0

where V “ f T pxn q f px p q, I “ p1, ¨ ¨ ¨ , 1qT , β “ pβ 1 , ¨ ¨ ¨ , β N qT , z “ pz1 , ¨ ¨ ¨ , z N qT , n “ 1, . . . , N, p “


1, . . . , N, and the diagonal matrix Bg is
ˆ ˙
1 1
Bg “ diag , ¨¨¨ , (14)
gb1 gb N

According to the Mercer condition, the kernel function is Fpyn , y p q “ f T pyn q f py p q. Here, the
radial basis function (RBF) is chosen as the kernel function
´ ¯
Fpyn , y p q “ exp ´||yn ´ y p ||2 {σ2 , n “ 1, . . . , N, p “ 1, . . . , N (15)

where σ2 is the RBF kernel parameter. The optimal parameters σ2 and g are obtained through the
ten-fold cross-validation method [66]. The parameters β and q can be solved using the least square
method [65].
Based on the above discussion, Equations (13) and (15) are used to solve Equation (10), and the
identification results can be expressed as:
« ff
N
ÿ ` ˘
˚
z pyq “ sgn β n zn F y, yn ` q (16)
n“1

Considering all the derivations and discussions above, a flow chart of the arc fault detection
algorithm is designed, as shown in Figure 18. It will be implemented in the following.
The experimental data for nine typical loads under different states are selected as the sample set
(y, z). The input eigenvector y contains the current variation y1 and the high-frequency energy y2 .
The vector z represents the actual classification results, which were denoted by the authors. It is noted
that “0” represents the normal state and “1” represents the arc fault state. The sample set includes
260 samples, as listed in Table 2, 200 of which are used as the training samples; the remaining samples
are used as the test samples.
Sensors 2016, 16, 500 18 of 24
Sensors 2016, 16, 500 18 of 24

Start

Preprocess high-frequency signals and currents

Calculate high-frequency energy and current variations

Determine the training samples and test samples

Choose RBF kernel function, obtain 2 and g

Train samples and obtain WLS-SVM model

Assess the generalization ability of the arc fault recognizer

End

Figure 18. The


Figure 18. The flow
flow chart
chart of
of the
the arc
arc fault
fault detection
detection algorithm.
algorithm.

In the arc fault detection algorithm, the arc fault recognizer is developed based on the training
and2.gThe
Table
samples. The parameters σ2 = 0.095 sample
= 3.26 set of typical
are obtained vialoads.
the ten-fold cross-validation method.
The test samples are used to test the arc fault recognizer. The calculated classification results are
Sample Number 1 2 3 ... 260
compared with the actual classification results z n . The generalization ability of the recognizer can
y1 (A¨ ms) 1.5513 22.1634 8.6517 ... 40.6125
be assessed through the error rate
y (mJ) 0.9726 11.7258 1.4209 ... 14.2903
2
z 0 M 1 0 ... 1
z *
 zn
n
(17)
e M  n =1  100%
In the arc fault detection algorithm, the arc fault M recognizer is developed based on the training
samples.M The
is theparameters σ2number and g =
= 0.095 and 3.26 are obtained via the ten-fold cross-validation method.
where test sample z *  n  1, , M  are the calculative classification results. The
The test samples are used to test the arc nfault recognizer. The calculated classification results are
identification
compared with results for the
the actual load statesresults
classification are shown in generalization
zn . The Figure 19. The ability
symbolof“□” represents can
the recognizer an arc
be
fault, and “*” represents the
assessed through the error rate normal load state. The feature vector space is used to identify arc faults
from the test samples. The optimal plane of M classification between normal states and arc faults is
ř ˚´z |
found in the high-dimensional feature space.|zThese n n identification results for load states are the
n “1
e “
globally optimal solution based on theMobtained M RBF kernel ˆ 100%
parameters. If the test sample appears(17)in
the arc fault feature space as shown in Figure 19, then an arc fault has occurred in the circuit;
where M is the test sample number and z˚n pn “ 1, . . . , Mq are the calculative classification results.
otherwise, there is no arc fault.
The identification results for the load states are shown in Figure 19. The symbol “˝” represents an
The confusion matrix of the arc fault detection algorithm after all the sample tests are completed
arc fault, and “*” represents the normal load state. The feature vector space is used to identify arc
is given in Table 3. The results show that the average error rate of the arc fault detection algorithm is
faults from the test samples. The optimal plane of classification between normal states and arc faults is
3.33%. Therefore, the arc fault identification rate is 96.67%. The developed algorithm demonstrates
found in the high-dimensional feature space. These identification results for load states are the globally
good generalization ability under different loads.
optimal solution based on the obtained RBF kernel parameters. If the test sample appears in the arc
fault feature space as shown in Figure 19, then an arc fault has occurred in the circuit; otherwise, there
Table 3. Confusion matrix for arc fault detection algorithm.
is no arc fault.
Calculative Classifications
Actual Classifications
Normal Arc Fault
Normal 27 1
Arc Fault 1 31
Sensors 2016, 16, 500 19 of 24
Sensors 2016, 16, 500 19 of 24

Feature Space
25
Normal
Arc fault
20

15
y (mJ)
2

10

5
1 1
1

20 40 60 80 100
y (A·ms)
1

Figure 19. Identification results of load states.

5. Designing AFD for


The confusion the of
matrix Early
the Detection of Electrical
arc fault detection Fires after all the sample tests are completed
algorithm
is given
TheinAFD
Tableis3. mainly
The results show that
composed of the average error
a transformer rate of the
module, arc fault
a signal detection algorithm
conditioning module, isa
3.33%. Therefore, the arc fault identification rate is 96.67%. The developed algorithm
microprocessor module, an alarm module, and a power module. The AFD schematic is shown in demonstrates
good
Figuregeneralization ability under
19a. The transducer module different
containsloads.
an HFT and a CT, which were introduced in Section 2.
These components are used to obtain the high-frequency signals and line currents, respectively, in
Table 3. Confusion matrix for arc fault detection algorithm.
low-voltage circuits. The obtained signals are filtered, amplified and integrated in the signal
conditioning module. Then, the signals are sent to the microprocessor module for further processing.
Calculative Classifications
The arc fault detection algorithm is successfully applied in the STM8S103 module, which is selected
Actual Classifications
as a microprocessor. The STM8S103 has 8 Kbytes Normal
of flash programmable Arc memory,
Fault 640 bytes of true
data Electrically Erasable Normal
Programmable Read-Only 27 Memory (EEPROM)1 and 1 Kbyte of Random
Access Memory (RAM); Arcits
Fault
maximum master clock 1 frequency is 16 MHz.31In the alarm module, the
electrical relay is controlled by a silicon controlled rectifier. When arc faults occur, the microprocessor
5. Designing
generates AFD for
a tripping the Early
signal; then,Detection
the earlyof Electrical
alarm Fireslights up, and the relay interrupts the
indicator
circuit power
The AFDsupply
is mainly to prevent an electrical
composed fire. A voltage
of a transformer of 220 V
module, a (50 Hz) conditioning
signal is directly supplied
module, to
the power module. The AC-to-DC (direct current) convertor based
a microprocessor module, an alarm module, and a power module. The AFD schematic is shown on the LNK302 chip can output
±12Figure
in V and20a.
±5 V fortransducer
The modules in the AFD.
module All modules
contains an HFT areandintegrated
a CT, which into
werea complete
introduced AFD. An AFD
in Section 2.
prototype has been developed, as shown in Figure 19b. The prototype can
These components are used to obtain the high-frequency signals and line currents, respectively, be directly installed in a
low-voltage circuit.
in low-voltage circuits. The obtained signals are filtered, amplified and integrated in the signal
According
conditioning to various
module. Then,standards,
the signalsincluding
are sent IEC
to the62606: 2013 and GB/T
microprocessor module31143-2014, numerous
for further test
processing.
experiments
The were performed
arc fault detection algorithm to is
verify the detection
successfully applied effects
in theunder different
STM8S103 loads.
module, The is
which schematic
selected asof
the tripping test is shown in Figure 20a. The fault arc voltage, the line current
a microprocessor. The STM8S103 has 8 Kbytes of flash programmable memory, 640 bytes of true data and the tripping signal
from the AFD
Electrically were observed
Erasable Programmable using aRead-Only
Tektronix Memory
DPO3014(EEPROM)
oscilloscope,andas1 shown
Kbyte of in Random
Figure 20b, and
Access
the tripping time of the AFD was 84 ms. The experimental schematic of
Memory (RAM); its maximum master clock frequency is 16 MHz. In the alarm module, the electricalcross talk is shown in Figure
10. When
relay arc faultsby
is controlled occurred
a siliconincontrolled
branch B1, AFD 1 When
rectifier. could detect them
arc faults in a the
occur, timely manner and generates
microprocessor interrupt
athe circuit signal;
tripping power then,
supply before
the early an electrical
alarm fire lights
indicator was produced to ensure
up, and the the prevention
relay interrupts of electrical
the circuit power
fires. Moreover, AFD 2 did not alarm or trip. This is because the current variation
supply to prevent an electrical fire. A voltage of 220 V (50 Hz) is directly supplied to the power module. was added as an
The AC-to-DC (direct current) convertor based on the LNK302 chip can output ˘12 V and ˘5 V not
important feature in the arc fault detection algorithm. The power frequency currents at 50 Hz are for
easily influenced by cross talk. Therefore, the developed AFD can overcome
modules in the AFD. All modules are integrated into a complete AFD. An AFD prototype has been the influences of cross
talk and transient
developed, as shown processes.
in Figure 20b. The prototype can be directly installed in a low-voltage circuit.
Sensors 2016, 16, 500 20 of 24
Sensors 2016, 16, 500 20 of 24

Sensors 2016, 16, 500 20 of 24

(a) (b)
Figure
Figure 19. AFD:
20. AFD: (a) schematic;
(a) schematic; (b) prototype.
(b) prototype.

According to various standards, including IEC 62606: 2013 and GB/T 31143-2014, numerous test
experiments were performed to verify the detection effects under different loads. The schematic of
the tripping test is shown in Figure 21a. The fault arc voltage, the line current and the tripping signal
from the AFD were observed using a Tektronix DPO3014 oscilloscope, as shown in Figure 21b, and the
tripping time of the AFD was 84 ms. The experimental schematic of cross talk is shown in Figure 10.
When arc faults occurred in branch B1, AFD 1 could detect them in a timely manner and interrupt
the circuit power supply before an electrical fire was produced to ensure the prevention of electrical
fires. Moreover, AFD 2 did not alarm or trip. This is because the current variation was added as an
important feature in the arc fault
(a) detection algorithm. The power frequency currents(b) at 50 Hz are not
easily influenced by cross talk. Therefore, the developed AFD can overcome the influences of cross
talk and transient processes. Figure 19. AFD: (a) schematic; (b) prototype.
Figure 20. Tripping test: (a) schematic; (b) results.

Table 4. Test results under typical loads.

Start or Stop Arc Fault Detection


Number Load Name
Condition Accuracy
1 Computer without tripping 98%
2 Halogen lamp without tripping 100%
3 Electrical drill without tripping 96%
4 Dimming lamp without tripping 100%
5 Vacuum cleaner without tripping 100%
6 Air conditioner without tripping 100%
7 Figure 20. Tripping
Air compressor test: (a) schematic;
without(b) tripping
results. 100%
Figure 21. Tripping test: (a) schematic; (b) results.
8 Fluorescent lamp and electric stove without tripping 100%
Table 4. Test results under typical loads.
Test results under typical loads are shown in Table 4. The accuracy of the arc fault detection is
very high,Test
andresults under
the AFD typical
exhibits loads
good are shown in Table
generalizability use4.or
Start
for The accuracy
Stop
under ofloads.
Arc
different the arcBecause
Fault fault detection
some is
Detection
Number
very high, and the Load
AFD Name good generalizability for use under different loads. Because some
exhibits
weak arc faults produced by electrical drills and computers were used in the tests, the
Condition features of arc
Accuracy
weak arc faults produced by electrical drills and computers were used in the tests, the features of arc
1 Computer without tripping 98%
faults were very similar to those of normal states. Hence, the AFD sometimes failed to detect these
2 Halogen lamp without tripping 100%
3 Electrical drill without tripping 96%
4 Dimming lamp without tripping 100%
Sensors 2016, 16, 500 21 of 24

faults were very similar to those of normal states. Hence, the AFD sometimes failed to detect these arc
faults, as reflected under certain loads such as those of electrical drills and computers. This limitation
will be improved in future work.

Table 4. Test results under typical loads.

Number Load Name Start or Stop Condition Arc Fault Detection Accuracy
1 Computer without tripping 98%
2 Halogen lamp without tripping 100%
3 Electrical drill without tripping 96%
4 Dimming lamp without tripping 100%
5 Vacuum cleaner without tripping 100%
6 Air conditioner without tripping 100%
7 Air compressor without tripping 100%
8 Fluorescent lamp and electric stove without tripping 100%

6. Conclusions
Arc faults occur in many electrical circuits and represent one of the most important causes of
electrical fires. AFDs are urgently needed for the early detection of electrical fires. In this study,
common arc fault features under different loads are extracted for the study of arc fault detection.
The main conclusions are as follows:

1. High-frequency signals collected by an HFT noticeably increase when arc faults occur in circuits.
However, they are strongly affected by cross talk.
2. The line currents of arc faults are often distorted or become unstable; however, current periodic
integrals also clearly change under normal states, especially during transient processes.
3. High-frequency signal energy and current variations were extracted as the input eigenvector of
the arc fault detection algorithm. An algorithm based on WLS-SVM was designed to identify arc
faults from load states.
4. The arc fault detection algorithm was successfully applied in a STM8S103 microprocessor, and an
AFD prototype was developed. The AFD was able to detect arc faults in a timely manner and
interrupt the circuit power supply before electrical fires could be produced during testing. The
prototype was not influenced by cross talk or transient processes. The detection accuracy was
very high, and the AFD exhibited good generalizability under different loads.

Consequently, the AFD can be installed in low-voltage circuits to monitor circuit states in real-time
to facilitate the early detection of electrical fires. The accuracy of the arc fault detection technique used
by the prototype will be further enhanced in future research and development. Moreover, the AFD,
the overcurrent protector, the ground fault circuit interrupter and other protection devices will be
integrated to develop a combination circuit fault detection device. The combination device will be
used to prevent all electrical fires.

Acknowledgments: This research has been funded by the National Natural Science Foundation (No. 51506059),
the Fujian Industry-University Cooperation Science & Technology Major Project (No. 2016H6014), the Quanzhou
Science & Technology Project (No. 2014Z114) and the Huaqiao University Scientific Research Foundation of China
(No. 14BS305).
Author Contributions: Kai Yang designed the arc fault detection algorithm and wrote the entire manuscript.
Rencheng Zhang managed the research and supervised the revision of the manuscript. Jianhong Yang participated
in the design of the detection algorithm. Canhua Liu, Shouhong Chen and Fujiang Zhang participated in the
experiments and data processing.
Conflicts of Interest: The authors declare no conflict of interest.
Sensors 2016, 16, 500 22 of 24

References
1. United States Fire Administration National Fire Data Center. Residential building fires (2011–2013). Top. Fire
Rep. Ser. 2015, 16, 1–15.
2. Fire Department of Ministry of Public Security. China Fire Services (2014); Yunnan People’s Publishing House:
Kunming, China, 2014; pp. 396–407. (In Chinese)
3. Si, G. Analysis on China’s electrical fire situation and feature from 2008 to 2012. Fire Sci. Technol. 2014, 33,
569–572. (In Chinese).
4. Vadoli, C.N. Lessons learned from electrical circuit analysis in support of a fire probabilistic risk assessment.
In Proceedings of the International Topical Meeting on Probabilistic Safety Assessment and Analysis,
Wilmington, NC, USA, 13–17 March 2011; pp. 711–720.
5. Lee, D.A.; Trotta, A.M.; King, W.H., Jr. New Technology for preventing residential electrical fires: Arc-fault
circuit interrupters (AFCIs). Fire Technol. 2000, 36, 146–162. [CrossRef]
6. United States Fire Administration National Fire Data Center. Residential building fires (2009–2011). Top. Fire
Rep. Ser. 2014, 14, 1–11.
7. Wang, X.G.; Lo, S.M.; Zhang, H.P. Exploring the role of feature subset combinations on performance of
multisensor fire detection. J. Appl. Fire Sci. 2013, 23, 179–192. [CrossRef]
8. Veselovskii, I.; Whiteman, D.N.; Korenskiy, M.; Suvorina, A.; Kolgotin, A.; Lyapustin, A.; Wang, Y.; Chin, M.;
Bian, H.; Kucsera, T.L.; et al. Characterization of forest fire smoke event near Washington, DC in summer
2013 with multi-wavelength lidar. Atmos. Chem. Phys. 2015, 15, 1647–1660. [CrossRef]
9. Hangauer, A.; Chen, J.; Strzoda, R.; Fleischer, M.; Amann, M.C. Performance of a fire detector based on
a compact laser spectroscopic carbon monoxide sensor. Opt. Express 2014, 22, 13680–13690. [CrossRef]
[PubMed]
10. Luis, J.A.; Galán, J.A.G.; Espigado, J.A. Low power wireless smoke alarm system in home fires. Sensors 2015,
15, 20717–20729. [CrossRef] [PubMed]
11. Yang, J.; Zhou, J.; Lv, Z.; Wei, W.; Song, H. A real-time monitoring system of industry carbon monoxide
based on wireless sensor networks. Sensors 2015, 15, 29535–29546. [CrossRef] [PubMed]
12. Jee, S.W. Distinction of fire source from smoke using discrete probability distribution and neural networks.
Fire Technol. 2014, 51, 887–904. [CrossRef]
13. Underwriters Laboratories Inc. UL Standard for Safety Arc-Fault Circuit-Interrupters, 2nd ed.; Underwriters
Laboratories Inc.: New York, NY, USA, 2011.
14. Lee, R.H. The other electrical hazard: Electric arc flash. IEEE Trans. Ind. Appl. 1982, IA-18, 246–251. [CrossRef]
15. Fuller, J.F.; Hanna, W.J.; Kallenbach, G.A. Arcing faults in metallic conduit at 120 and 240 V. IEEE Trans.
Ind. Appl. 1985, IA-21, 820–825. [CrossRef]
16. Gregory, G.D.; Wong, K.; Dvorak, R. More about arc-fault circuit interrupters. IEEE Trans. Ind. Appl. 2004, 40,
1006–1011. [CrossRef]
17. What Others Are Saying. Available online: http://www.afcisafety.org/endorsements.html (accessed on
8 May 2013).
18. International Electrotechnical Commission. General Requirements for Arc Fault Detection Devices; IEC 62606;
International Electrotechnical Commission: Geneva, Switzerland; 9; July; 2013.
19. Gammon, T.; Matthews, J. Arcing-fault models for low-voltage power systems. In Proceedings of the IEEE
Industrial and Commercial Power Systems Technical Conference, Clearwater Beach, FL, USA, 7–11 May 2000;
pp. 119–126.
20. Wu, H.R.; Li, X.H.; Stade, D.; Schau, H. Arc fault model for low voltage AC systems. IEEE Trans. Power Deliv.
2005, 20, 1204–1205. [CrossRef]
21. Kizilcay, M.; Pniok, T. Digital simulation of fault arcs in power systems. Eur. Trans. Electr. Power Eng. 1991, 1,
55–60. [CrossRef]
22. Song, Y.H.; Aggarwal, R.K.; Johns, A.T. Digital simulation of fault arcs on long-distance compensated
transmission systems with particular reference to adaptive autoreclosure. Eur. Trans. Electr. Power 1995, 5,
315–324. [CrossRef]
23. Parise, G.; Martirano, L.; Laurini, M. Simplified arc-fault model: The reduction factor of the arc current.
IEEE Trans. Ind. Appl. 2013, 49, 1703–1710. [CrossRef]
Sensors 2016, 16, 500 23 of 24

24. Andrea, J.; Schweitzer, P.; Martel, J.M. Arc fault model of conductance application to the UL 1699 tests
modeling. In Proceedings of the 57th IEEE Holm Conference on Electrical Contacts, Minneapolis, MN, USA,
11–14 September 2011; pp. 37–42.
25. Charles, J.K. Electromagnetic radiation behavior of low-voltage arcing fault. IEEE Trans. Power Deliv. 2009,
24, 416–423.
26. Yang, J.H.; Zhang, R.C.; Du, J.H. Study on application of arcing protective system faults based on multiple
information fusion. High Volt. Appar. 2007, 43, 194–196. (In Chinese).
27. Zhang, R.C.; Yang, J.H.; Du, J.H. Study on in-process detection and diagnosis of faults arc based on early
sounds signature and intermittent chaos. Key Eng. Mat. 2008, 381, 611–614. [CrossRef]
28. Lin, Y.H.; Liu, C.W.; Chen, C.S. A new PMU-based fault detection location technique for transmission lines
with consideration of arcing fault discrimination-part I theory and algorithms. IEEE Trans. Power Deliv. 2004,
19, 1587–1593. [CrossRef]
29. Meyar-Naimi, H.; Hasanzadeh, S.; Sanaye-Pasand, M. Discrimination of arcing faults on overhead
transmission lines for single-pole auto-reclosure. Int. Trans. Elecr. Energy Sys. 2013, 23, 1523–1535. [CrossRef]
30. Boksiner, J.; Silverman, E.J. Electrical Cable Arcing Fault Detection by Monitoring Power Spectrum in
Distribution Line. U.S. Patent 5,359,293, 25 October 1994.
31. Li, J.C.; Kohler, J.L. New insight into the detection of high-impedance arcing faults on DC trolley systems.
IEEE Trans. Ind. Appl. 1999, 35, 1169–1173.
32. Wang, S.C.; Wu, C.J.; Wang, Y.J. Detection of arc fault on low voltage power circuits in time and frequency
domain approach. IEEE Trans. Ind. Appl. 2012, 6, 324–331.
33. Koziy, K.; Gou, B.; Aslakson, J. A low-cost power-quality meter with series arc-fault detection capability for
smart. IEEE Trans. Power Deliv. 2013, 28, 1548–1591. [CrossRef]
34. Lezama, J.; Schweitzer, P.; Tisserand, E.; Humbert, J.B.; Weber, S.; Joyeux, P. An embedded system for AC
series arc detection by inter-period correlations of current. Electr. Power Syst. Res. 2015, 129, 227–234.
[CrossRef]
35. Haun, A.A.; Coats, A.G.; Wong, K.B.; Dvorak, R.F.; Scott, G.W. Arc Fault Circuit Interrupter. U.S. Patent
6,414,829, 2 July 2002.
36. Dvorak, R.F.; Wong, K.B. Arc Fault Circuit Interrupter System. U.S. Patent 7,151,656 B2, 19 December 2006.
37. Li, C.L.; Yue, G.L. Arc Fault Circuit Interrupter. U.S. Patent 20150155698, 4 June 2015.
38. Nicolescu, A.D.; Orr, R.K. Arc Fault Detection and Extinguishing. European Patent EP2,960,947,
30 December 2015.
39. Su, W.Y.; Wu, C.J. The Assessment of Introduction of the Arc Fault Circuit Interrupters; IOSH98-S302; Report
prepared for Institute of Occupational Safety and Health: Taipei, Taiwan, 2010.
40. Zhang, S.; Zhang, F.; Liu, P.; Han, Z. Identification of low voltage AC series arc faults by using Kalman
filtering algorithm. Elektronika ir Elektrotechnika 2014, 20, 51–56. [CrossRef]
41. Sidhu, T.S.; Singh, G.; Sachdev, M.S. Arcing fault detection using artificial neural networks. Neurocomputing
1998, 23, 225–241. [CrossRef]
42. Rivers, C., Jr. Arc Fault Detection Using Fuzzy Logic. U.S. Patent 8,054,592 B2, 8 November 2011.
43. Zhang, R.C.; Yang, J.H.; Li, X.H. Low-Voltage Arc Fault Detection Apparatus. Available online: http://www.
pss-system.gov.cn/sipopublicsearch/search/search/showViewList.shtml (accessed on 16 January 2013).
44. Cho, Y.; Lim, J.; Seo, H.; Bang, S.B.; Choe, G.H. A series arc fault detection strategy for single-phase boost
PFC rectifiers. J. Power Electr. 2015, 15, 1664–1672. [CrossRef]
45. Liu, Y.L.; Guo, F.Y.; Wang, Z.Y.; Chen, C.K.; Li, Y. Research on the Spectral Characteristics of Series Arc Fault
Based on Information entropy. Trans. China Electrotech. Soc. 2015, 30, 488–495.
46. Kim, C.H.; Kim, H.; Ko, Y.H.; Byun, S.H.; Aggarwal, R.K.; Johns, A.T. A novel fault-detection technique of
high-impedance arcing faults in transmission lines using the wavelet transform. IEEE Trans. Power Deliv.
2002, 17, 921–929. [CrossRef]
47. Jamali, S.; Ghaffarzadeh, N. A new method for arcing fault location using discrete wavelet transform and
wavelet networks. Eur. Trans. Electr. Power 2012, 22, 601–615. [CrossRef]
48. Kawady, T.A.; Elkalashy, N.I.; Ibrahim, A.E.; Taalab, A.M.I. Arcing fault identification using combined
Gabor Transform-neural network for transmission lines. Int. J. Electr. Power Energy Syst. 2014, 61, 248–258.
[CrossRef]
Sensors 2016, 16, 500 24 of 24

49. Kolker, D.; Campolo, S.; DiSalvo, N. A study of time/current characteristics of the ignition processes
in cellulosic material caused by electrical arcing for application in 240 V arc-fault circuit interrupters.
In Proceedings of the 53th IEEE Holm Conference on Electrical Contacts, Pittsburgh, PA, USA,
16–19 September 2007; pp. 105–114.
50. Shea, J.J. Comparing 240 Vrms to 120 Vrms series arcing faults in residential wire. In Proceedings of the 54th
IEEE Holm Conference on Electrical Contacts, Orlando, FL, USA, 27–29 October 2008; pp. 218–224.
51. Carvou, E.; Ben Jemaa, N.; Tian, S.; Belhaja, Z.; Jusselin, B. Electrical arc characterization for ac-arc fault
applications. In Proceedings of the 55th IEEE Holm Conference on Electrical Contacts, Vancouver, BC,
Canada, 14–16 September 2009; pp. 21–26.
52. Ko, W.S.; Moon, W.S.; Bang, S.B.; Kim, J.C. Analysis of ignition time/current characteristics and energy when
series arc-fault occurs at rated 220 V. Trans. Korean Inst. Electr. Eng. 2013, 62, 1184–1191. [CrossRef]
53. Suykens, J.A.K.; Vandewalle, J. Least squares support vector machine classifiers. Neural Process. Lett. 1999, 9,
293–300. [CrossRef]
54. Cheng, Q.; Tezcan, J.; Cheng, J. Confidence and prediction intervals for semiparametric mixed-effect least
squares support vector machine. Pattern Recogn. Lett. 2014, 40, 88–95. [CrossRef]
55. Si, L.; Wang, Z.B.; Liu, X.H.; Tan, C.; Liu, Z.; Xu, J. Identification of shearer cutting patterns using vibration
signals based on a least squares support vector machine with an improved fruit fly optimization algorithm.
Sensors 2016, 16, 90. [CrossRef] [PubMed]
56. General Administration of Quality Supervision, Inspection and Quarantine of the People’s Republic of
China; Standardization Administration of the People’s Republic of China. Electrical Fire Monitoring System
–Part 4: Arcing Fault Detectors; Standards Press of China: Beijing, China, 2014.
57. Bondiou, A.; Labaune, G.; Marque, J.P. Electromagnetic radiation associated with the formation of an electric
breakdown in air at atmospheric pressure. J. Appl. Phys. 1987, 61, 503–509. [CrossRef]
58. Damas, M.C.; Robiscoe, R.T. Detection of radio-frequency signals emitted by an arc discharge. J. Appl. Phys.
1988, 64, 566–574. [CrossRef]
59. Zhang, W.X.; Zhang, J.; Xu, X.Y. Analysis on line noise characteristics of low voltage grid. North China
Electr. Power 2012, 4, 27–30.
60. Zhu, Z.S.; Zheng, Y.S. The frequency range of switching power supply. TV Eng. 1994, 6, 43–49.
61. Cheng, W.B.; Zhong, Y.R.; Jin, S. Study of spread spectrum in suppressing acoustic resonance. Trans. China
Electrotech. Soc. 2006, 21, 83–88.
62. Wang, Q.M. A Study of Conducted Interference Testing and Corrective Measures for House Appliances;
Suzhou University: Suzhou, China, 2012.
63. Pontil, M.; Verri, A. Properties of support vector machines. Neural Comput. 1998, 10, 955–974. [CrossRef]
[PubMed]
64. Yang, X.W.; Tan, L.J.; He, L.F. A robust least squares support vector machine for regression and classification
with noise. Neurocomputing 2014, 140, 41–52. [CrossRef]
65. Suykens, J.A.K.; De Brabanter, J.; Lukas, L.; Vandewalle, J. Weighted least squares support vector machines
robustness and sparse approximation. Neurocomputing 2002, 48, 85–105. [CrossRef]
66. Han, X.H.; Du, S.H.; Su, J.; Guan, H.O.; Shao, L.M. Determination method of electric shock current based on
parameter-optimized least squares support vector machine. Trans. Chin. Soc. Agric. Eng. 2014, 30, 238–245.
(In Chinese).

© 2016 by the authors; licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC-BY) license (http://creativecommons.org/licenses/by/4.0/).