IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 63, NO.

11, NOVEMBER 2014 2651

Fault Detection and Classification in Medium

Voltage DC Shipboard Power Systems With
Wavelets and Artificial Neural Networks
Weilin Li, Member, IEEE, Antonello Monti, Senior Member, IEEE,
and Ferdinanda Ponci, Senior Member, IEEE

Abstract— This paper proposes a fault detection and clas- converters is pervasive [1]. These power electronic devices can
sification method for medium voltage DC (MVDC) shipboard greatly simplify the system by providing more available space,
power systems (SPSs) by integrating wavelet transform (WT) possible cost reduction, higher efficiency, higher safety, and
multiresolution analysis (MRA) technique with artificial neural
networks (ANNs). The MVDC system under consideration for variable control to achieve self-healing and survivability [3].
future all-electric ships presents a range of new challenges, in However, the detection, location, identification and isolation
particular the fault detection and classification issues addressed in of various faults within the zonal distribution system remain
this paper. The WT-MRA and Parseval’s theorem are employed a challenge.
in this paper to extract the features of different faults. The There are already some established fault detection methods
energy variation of the fault signals at different resolution levels
are chosen as the feature vectors. As a result of analysis and for a high voltage DC (HVDC) power system [4]. However,
comparisons, the Daubechies 10 (db10) wavelet and scale 9 are the fault detection schemes for a MVDC system of naval ship-
the chosen wavelet function and decomposition level. Then, ANN board power system (SPS) are still in the developmental stages
is adopted to automatically classify the fault types according because of the smaller scale and higher safety requirement.
to the extracted features. Different fault types, such as short Most of all, while HVDC has been so far used in point to
circuit faults on both dc bus and ac side, as well as ground
fault, are analyzed and tested to verify the effectiveness of the point application, the SPS is by nature a more challenging
proposed method. These faults are simulated in real time with multiterminal dc network.
a digital simulator and the data are then initially analyzed with Some fault detection and classification issues in MVDC
MATLAB. The case study is a notional MVDC SPS model, and power systems have been addressed in [5]–[17]. In particular,
promising classification accuracy can be obtained according to a fault location technique based on active impedance esti-
simulation results. Finally, the proposed fault detection algorithm
is implemented and tested on a real-time platform, which enables mation (AIE) was proposed in [6]–[8], which used a power
it for future practical use. converter connected to the bus to evaluate the bus impedance
by injecting a very short current spike and analyzing the
Index Terms— Artificial neural networks (ANNs), fault
detection and classification, medium voltage dc (MVDC) system, associated current and voltage transient responses. However,
wavelet transform (WT)-based multiresolution analysis (MRA). obtaining the accurate system impedance in a short time during
the system operation is a challenge. Aurangzeb et al. [9]
I. I NTRODUCTION proposed an overcurrent protection method for detecting and

T HE development of integrated electrical propulsion, the

increase in high energy electrical loads and the wide-
spread use of power electronics are placing a heavy demand
locating the disturbances that may occur on voltage source
converter (VSC)-based multiterminal dc distribution systems.
Hard thresholds were adopted to detect the faults: a fault is
on the quality and survivability of shipboard power distribution assumed to have occurred if the monitored signal exceeds a
systems [1]–[3]. Medium voltage dc (MVDC) zonal distribu- given threshold for a given duration of time. However, the
tion architecture is proposed as a new distribution system for accuracy of this method turns out to be easily affected by
the all-electric ships (AES) in which the presence of power disturbances or noise. A more detailed review of the related
work in the field of fault detection and classification in MVDC
Manuscript received November 6, 2013; revised February 4, 2014; accepted
February 5, 2014. Date of publication April 9, 2014; date of current version power systems will be given in Section II.
October 7, 2014. This work was supported by the U.S. Office of Naval Recently, the use of high frequency components of the fault
Research under Grant N0014-08-1-0080. The Associate Editor coordinating generated transient signals was studied for fault protection and
the review process was Dr. Kurt Barbe.
W. Li is with the Department of Electrical Engineering, Northwestern some useful solutions were obtained [18], [19]. This method
Polytechnical University, Xi’an 710072, China, and also with the E.ON enables accurate fault detection unaffected by power swings.
Energy Research Centre, RWTH Aachen University, Aachen 52074, Germany However, it is only adopted for transmission line protection in
(e-mail: li.weilin@hotmail.com).
A. Monti and F. Ponci are with the E.ON Energy Research Centre, RWTH traditional ac power systems. Thus, it will be of great value to
Aachen University, Aachen 52074, Germany (e-mail: amonti@eonerc.rwth- develop a signal pattern analysis-based method for detection
aachen.de; fponci@eonerc.rwth-aachen.de). and classification of the faults in MVDC SPSs according to
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. fault generated transient signals. Based on the advanced signal
Digital Object Identifier 10.1109/TIM.2014.2313035 processing and artificial intelligence techniques, the obtained
0018-9456 © 2014 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
2652 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 63, NO. 11, NOVEMBER 2014

new method can avoid the problems present in traditional through detailed analysis and comparison, as presented in
schemes, and has a strong capability of generalization to Section IV. In this paper, the energy variations of the fault
different kinds of faults. signal at different resolution levels are adopted as feature
Fast Fourier transform (FFT) method, short time Fourier vectors. By employing the Parseval’s theorem, the information
transform (STFT) method, and wavelet transform (WT) are is condensed, thus reducing the data size and yielding a
among the signal processing methods for frequency and time- manageable feature vector. Then, an ANN is adopted to
frequency analysis used in power systems. The intrinsic draw- automatically classify the fault types according to the extracted
back of the traditional methods, such as FFT, is the lack features, which is suitable for RT operation of fault type
of information in the time domain [20], thus they are unfit classification problems. Numerical simulation has been used to
for detecting transients or short spikes in the signal. STFT verify the capability of the proposed fault detection algorithm
method is also limited due to the constraint on the window’s on a notional MVDC SPS. In order to verify the effectiveness
size: using a wide window results in good frequency but poor of the proposed approach, the SPS is first simulated in RT,
time resolution, while using a narrow window results in poor and results are then postprocessed with MATLAB. Simulation
frequency but good time resolution. The WT has the ability to results show that good classification accuracy can be obtained
decompose wideband and nonstationary signals into specific with the proposed method. Then, the proposed algorithm is
time-frequency resolutions. In particular, WT can be used to implemented and tested on a full RT simulation platform. The
extract the abrupt-changing features of signals (such as the proposed classifier can be used as a postfault diagnosis method
sharp increase of current during short circuit fault), which is to recognize, and identify the sources and causes of faults thus
the representation of the signal dynamics [21]. Comparative supporting the decision on mitigating actions.
studies of the aforementioned signal processing techniques Compared with the previous literature, the innovative
for feature extraction of different power quality events are contributions of this paper can be summarized as follows.
given in [22]. Wavelet analysis techniques have been proposed
1) This paper first applies WT-based MRA technique and
extensively for fault detection, localization, and classification
ANN-based algorithm for fault detection and classifi-
of different territorial power system transients [23]. Few papers
already investigated the possibility to apply WT-based method cation in MVDC SPSs, using the energy variation as
feature vectors. The proposed algorithm has been tested
for fault location and classification in dc power systems
[13], [14]. However, two practical problems must be overcome for short circuit faults and for ground faults. Compared
in the implementation of wavelets for fault detection in MVDC with the other methods, this method does not require
interzone communication and external signal injection.
power systems.
1) Adopting directly the WT coefficients for fault detection Moreover, its performance is not influenced by variation
of electrical parameters.
requires large memory space and computing time. Thus,
2) The process of selection of the suitable wavelet function
a feature vector with reduced quantities and without
losing the property of the original signal needs to be and of the optimal decomposition level is detailed, to
achieve a tradeoff between the accuracy and computing
chosen.
2) The wavelet function, as well as the decomposition level efficiency.
must be properly chosen to meet the tradeoff between the 3) The simulation results of the extracted features are based
on the analysis of RT simulated data obtained from real-
accuracy of detecting the fault type and the computing
efficiency. time digital simulator (RTDS) [28], [29], which sets the
conditions for a more realistic scenario and for easy
Regarding the selection of the pattern recognition method to
further transition, e.g., to hardware-in-the-loop (HiL)
automatically and accurately identify the fault types based on
the feature vectors, the methods of interest include artificial testing and communication assessment.
4) The proposed algorithm has been implemented and
intelligence and expert systems. However, expert systems
are expensive in their development and time consuming in tested in RT on a LabVIEW RT target [30] interfaced
their execution, hence not suited for real-time (RT) appli- with the RTDS, thus the obtained classifier has online
applicability for the purpose of practical use.
cations [24]. Artificial neural network (ANN) methodologies
appear to be well suit for these applications [25]. They are The rest of this paper is organized as follows. Section II gives
fit for RT use due to their relatively fast calculation time. a brief review and comparison of related work in MVDC
ANNs are capable of finding complex and hidden relationships system fault detection and classification. The notional diagram
among the presented data [26]. of the MVDC SPS and the general structure of the proposed
Li et al. [27] investigated the use of WT MRA technique algorithm are provided in Section III. Section IV introduces
to extract the features of different fault signals in MVDC the WT theory and the MRA technique in brief, as well as
SPSs. This paper is extended here by integrating WT-based the procedures to select the optimal wavelet function and
MRA technique with ANN to construct a classifier for auto- decomposition level. Section V presents the Parseval’s theorem
matic recognition of the fault types in future AES. First, the and how it is used to extract the feature vectors of the fault
WT-based MRA technique and the Parseval’s theorem are signal. Section VI introduces the structure and principles of the
employed to extract the feature vectors of different types of adopted ANNs, together with the classification results. The RT
fault. The db10 wavelet and scale 9 are chosen as appropri- implementation and test of the proposed algorithm are given
ate wavelet function and decomposition level, respectively, in Section VII. The conclusion is presented in Section VIII.
LI et al.: FAULT DETECTION AND CLASSIFICATION IN MVDC SPSs 2653

II. R ELATED W ORK

As indicated previously, the fault detection schemes in
MVDC SPSs are still in the early developmental stage, due
to its smaller scale compared with territorial power systems,
higher safety requirement, and the more complex multitermi-
nal distribution network. This section will give a more detailed
review of the state of the art methods that have been applied
for fault detection and classification in MVDC power systems.
Table IV summarizes the related work in this research field
(Appendix).
Among the solutions proposed in literature, the current
differential protection (CDP) scheme may be the simplest [5].
The principle of CDP scheme is based on Kirchhoff’s current
law that states that the sum of the currents entering a node
must be zero. In the current differential scheme, all the currents
entering a piece of equipment are monitored and if their sum is Fig. 1. MVDC SPS notional scheme.
nonzero it is assumed there is an internal fault. This protection
scheme is very precise and fast, with a high sensitivity. The
structure of ANN is very complex and does not support on-line
main drawbacks are: a pilot wire or communication link
implementation.
between all the device terminals is needed; only the zone
The discrete WT (DWT) has been applied to detect any
between the current transformers is protected, it does not
surges in dc third rail current waveform of dc transit systems
provide backup protection to other parts of the system, and
in [13] and [14]. In the event of a surge, the WT extracts
it is susceptible to current transformer errors and pilot wire
a feature vector from the current waveform and feeds it to
asymmetries.
a self-organizing neural network. The neural network then
AIE fault location technique uses a power converter con-
determines whether the feature vector belongs to a fault current
nected to the bus to estimate the impedance of the bus to
surge due to short circuit or to a normal surge due to the
source [6]–[8]. To do this, it injects a very short current spike
starting of the chopper train. The only disadvantage of this
onto the system (via a coupling inductance) at the point of
method is that, all the coefficients at certain decomposition
measurement (POM). The injected current, and the resultant
level are selected as a feature vector, which leads to very slow
bus voltage transient are recorded and used to calculate the
training process.
system impedance. The system impedance measured at POM
Reference [15] investigates the main challenges of detecting
is expected to contain the information on both the fault
and localizing a fault, and interrupting it as quickly as pos-
location and the fault severity. However, challenges still lie in
sible in a VSC-based multiterminal dc system. It illustrates
the appropriate choice of frequency under which the measured
that protection can be simplified using the power-electronic
impedance can provide a sufficiently high resolution.
converters’ ability to act as fault current-limiting circuit
Traveling wave fault location is a well known and estab-
breakers (CBs). It is shown that by associating relays with
lished technique for accurately locating faults in transmission
the converters and by adopting overcurrent-based protection
lines and to a more limited extent in distribution lines [9]–[11].
schemes for these relays, the faults on the system can be
When a fault occurs in a transmission line there is a sudden
detected and localized very fast. Hard thresholds were adopted
drop in voltage at the fault location. This abrupt change in
to detect the faults: a fault is assumed to have occurred if
operating conditions causes high frequency travelling waves
the monitored signal exceeds a given threshold for a given
to propagate in both directions away from the fault location.
duration of time. Thus, the accuracy of this method turns out
Traveling wave fault location can be performed by making
to be easily affected by disturbances or noise. Moreover, due
time synchronized measurements at both ends of the trans-
to the small line inductance and resistance of the compact
mission line (double-ended) and time-tagging the arrival of
power network, and the variation of fault resistance, similar
the initial travelling waves usually using a Global Positioning
response may present for different fault locations, leading to
System timing signal. However, the MVDC power system is
selectivity issues.
of small scale, making it difficult to measure with enough
accuracy the time difference. Also, real effect, such as cable
III. MVDC SPS S AND THE G ENERAL S TRUCTURE
terminations, junctions, and terminal connections may impact
OF THE P ROPOSED A LGORITHM
the performance of this method.
Chanda and Fu [12] propose an ANN-based method for A. MVDC SPSs
the fault classification and location in the MVDC SPSs using The notional diagram of the MVDC SPS considered here
the transient information in the fault voltage and current is shown in Fig. 1. The system is organized in five zones
waveforms. It is found to be efficient in detecting the type (separated by dashed lines and numbered one to five in figure).
and location of the dc cable faults. However, they directly The prototype architecture considered here is a revised version
use the sampled data as the input of ANN, which leads to a of the system considered by the U.S. Navy, and it represents a
very long training and calculation time. Moreover, the adopted generic multiterminal dc distribution system [31]. Zones 3–5
2654 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 63, NO. 11, NOVEMBER 2014

are connected to the primary dc buses (Starboard and Portside

bus) via power converters based on power electronic building
blocks [(PEBBs), labeled PEBB_zoneNR_NR]. The system
also has two ac sources (generators) in Zones 1 and 2 feeding
the primary dc buses through rectifying converters. Vital loads,
generic Load 2 in Zone 3 and induction motor drive in Zone 5,
can be supplied from either the Starboard bus or the Portside
bus but not by both buses at the same time. Auctioneering
diodes are adopted to realize this function. The secondary
dc buses at the same zone are kept at different voltages (800
and 750 V), so that if a contingency occurs on one bus
that causes the voltage to drop, then the auctioneering diodes
allows the load to be picked up by the other bus. During
normal operation, the primary dc buses kept at 5-kV dc, in
particular the positive pole has +2.5 kV and the negative
pole −2.5 kV. Each bus feeds two kinds of loads: vital loads Fig. 2. Simplified schematic of the notional MVDC SPS.
(100 kW) and nonvital loads (50 kW).
To verify the proposed fault detection approach in this paper,
a set of data obtained from RT simulation is analyzed. For
this purpose, the notional MVDC SPS model is implemented
in RTDS. In particular, the small time step model is built to
represent the circuit dynamics with time steps in the range
1–4 μs. It is to be noted that although the model can execute
at microseconds level, the controller may not be so fast
due to the fact that it needs time for calculation and the
limitation from the switching capability of the converters.
In this case, the controller is executed at milliseconds level.
The detailed description of the control methods applied to
PEBB-based converters in MVDC power systems can be found
in [32] and [33].

B. Short Circuit Fault

Fig. 3. Current waveforms of primary dc bus when short circuit faults occur.
Physical system faults of SPSs can be categorized as
phase-to-ground fault (ground fault), and phase-to-phase (short The following cases are simulated in RTDS (as shown in
circuit fault). Fig. 2, which is a simplified schematic of the notional MVDC
The potential short circuit faults on board of a ship are SPS).
shown in Fig. 1, labeled as F1 on the portside bus, F2 on the 1) Short circuit fault on primary dc bus.
secondary bus in Zone 5, and F3 on the load side. In case of a 2) Short circuit fault on ac side (phase-to-phase fault), as
short circuit fault at location F3, the induction motor drive this kind of fault will also cause a dramatic change of
(vital load) loses power. However, assuming the converter the voltage and current in primary dc bus, and should
modules (PEBB5_3) can operate in current limiting mode [32], be distinguished.
the current then decreases slightly to zero until the converters In actual MVDC power systems, the PEBB-based converters
are shut down. The fault will therefore not affect the remainder and disconnectors will be used to interrupt the fault current
of the system except for a drop in load. In the case of a short when short circuit fault occurs. However, in this paper, we did
circuit fault at location F2, the Portside converter module limits not consider the protection action of the system after the fault
the current and the Starboard converter takes over the power has been detected and classified due to the following reasons:
supply to induction machine through the auctioneering diodes. the fault can be detected before the protection actions will be
The auctioneering diodes ensure that the transfer of power taken (for example, before the CBs are open); the duration of
from main bus to alternate bus is seamless and without inter- the fault will not affect the performance of the fault detection
ruption. The system is thus designed to be robust with respect method due to the fact that high frequency components of
to these faults. However, a short circuit fault at location F1, is the fault generated transient signals was used.
more detrimental to the whole power system than the previous The current of the primary dc bus (measurement point_SC
two kinds of fault. In this case, in fact, the entire Portside bus as shown in Fig. 2) is monitored for detection of short circuit
is temporarily lost. Thus, the detection of this kind of fault faults. The waveforms, when different short circuit faults
becomes the main focus in this case study. Although we only occur, are shown in Fig. 3 [27]. It is noticed that both short
focus on the most serious fault, the obtained fault detection circuit faults will lead to a significant increase in the dc bus
algorithm can also be applied to other fault types. current.
LI et al.: FAULT DETECTION AND CLASSIFICATION IN MVDC SPSs 2655

Fig. 4. Noise signal of measurement point when negative line to ground

fault occurs in Zone 3 at 0.04 s.

C. Ground Fault
Fig. 5. Flowchart of the proposed algorithm.
High-resistance grounding (or ungrounded) is implemented
in the MVDC SPSs to permit continuity of service during network for classification. The outputs of the neural networks
a ground fault [35]. To prevent subsequent phase-to-ground indicate the fault types (normal operation, dc bus short circuit
faults on another phase from causing a disruptive phase-to- fault, disturbance caused by ac side short circuit fault, or
phase fault, it is essential that fault identification and location ground fault, those are the most likely faults that may occur
methods are accurate and robust. However, locating ground on board of a ship according to [9]).
faults in an ungrounded shipboard power distribution system is This method reduces the quantities needed for training with
inherently difficult because it has to be done while the system the use of feature vectors and supports RT online implementa-
is still in operation (which is a salient feature of ungrounded tion. Moreover, the performance of this method is not affected
power system), and without interrupting the service to vital by variations in electrical parameters like fault resistance and
loads. Furthermore, such faults do not provide considerable loads. The objective is to obtain a practical classifier that can
fault current for tracing the fault location. be used as a fault diagnosis method to recognize, and classify
In this paper, the ground fault detection and identification the sources and causes of faults before appropriate mitigating
is based on the pattern analysis of inherent high frequency actions can be taken.
background noise, which is introduced by repetitive switching
events of power electronic converters interacting with system
IV. WAVELET-BASED MRA
parasitic elements (such as cable insulation capacitance and
stray inductance) [35]. In particular, as shown in Fig. 2, the In the following sections, the WT theory, the multiresolution
midpoint voltage of the dc source capacitor (measurement analysis (MRA) technique, the selection of appropriate wavelet
point_G as indicated) was measured to obtain the noise signal function, and the calculation of optimal decomposition level
since a significant change in it will not only indicate the will be described in detail.
occurrence of a fault, but also directly identify general fault
types. A. WT Theory
In this paper, the ground fault at Zone 3 has been used Based on the characteristics of the bandpass filters, the WT
as case study to validate the proposed algorithm. The signals has been proposed to investigate the transient phenomena of
of the measurement point_G under both normal condition and power signals from different scales of the WT coefficients.
ground fault occurrence in Zone 3 are shown in Fig. 4. Bigger Particularly, the transient behaviors with high-frequency com-
oscillations are noticed when ground fault occurs. ponents can be analyzed.
The definition of continuous WT (CWT) for a given signal
D. General Structure of the Proposed Algorithm x(t) with respect to a mother wavelet ψ(t) is
The proposed fault diagnostic algorithm is based on the  ∞  
signature analysis of the monitored signals, performed with 1 t −b
CWT(a, b) = √ x(t)ψ dt (1)
WT-based MRA techniques for feature extraction and ANN for a −∞ a
feature classification. The flowchart of the proposed algorithm where a is the scale factor and b is the translation factor. For
is shown in Fig. 5. At the data acquisition stage, the monitored CWT, t, a, and b are all continuous.
signal selected for fault diagnostic is sampled. Then, the The application of WT in engineering areas usually requires
WT-based MRA technique is used for decomposing each a DWT, which implies the discrete form of t, a, and b in (1).
signal into low frequency approximation and high frequency The representation of DWT can be written as
details. Parseval’s theory is applied to extract the feature  
1  k − nb0 a0m
vectors of each sampled signal. The extracted feature vectors DWT(m, n) =  m x(k)ψ (2)
are then used as the inputs of a three-layer feedforward neural a0 a0m
k
2656 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 63, NO. 11, NOVEMBER 2014

Fig. 6. Decomposition of WT-based MRA.

where the original a and b parameters in (1) are changed to Fig. 7. Frequency response of the filters based on different wavelets.
be the functions of integers m, n, and k is an integer variable
and it refers to a sample number in an input signal.
1) The wavelet function should have enough number of
B. Multiresolution Analysis vanishing moments to represent the salient features of
the disturbances.
The MRA is a framework for hierarchical representation of
2) The wavelet should provide sharp cutoff frequencies to
a signal at its different scales. The goal of WT-based MRA is
reduce the amount of leakage energy to the adjacent
to develop representations of a signal f (t) in terms of both
resolution levels.
the scaling and the wavelet functions, in the following terms
3) The wavelet should obtain higher total wavelet energy
[36], [37]:
of the same feature of the same signal.
 
J 
4) The wavelet basis should be orthonormal.
f (t) = c J,k φ J,k (t) + d j,k ψ j,k (t) There are several orthogonal wavelet families, which are the
k j =1 k Daubechies family (dbN), Symlets family (symN), Coiflets

c J,k = f (t)φ̄ J,k (t) family (coifN), and Meyer family (meyr) [39], with N being
k the order of the wavelet function. The problem of selecting

d j,k = f (t)ψ̄ j,k (t). (3) an optimum wavelet function can be defined as selecting
k
orthonormal basis function. The minimum description length
(MDL) data criterion is applied in this case to select the
In (3), φ̄(t) and ψ̄(t) are conjugate of scaling function φ(t) optimum wavelet function, which states that the best model
and wavelet function ψ(t), respectively. within the group of models has the shortest description of
The procedure of WT-based MRA is shown in Fig. 6 [38], data and model itself. The available orthogonal wavelets
where x[n] is the sampled signal of f (t). MRA is performed are tested by the MDL data criterion to select an optimum
by passing the signal x[n] through a high-pass filter h[n] wavelet function in [38], and the evaluation results show that
and a low-pass filter g[n], and this is followed by a down the Daubechies wavelet functions have the smallest MDL
sampling by two to compute both the scaling coefficients indices. Thus, only Daubechies wavelets are considered as
[(SCs), also called approximation] and wavelet coefficients they enable fast computation speed and are therefore suit-
[(WCs), also called detail]. SCs represent the low frequency able for RT application [13]. In this paper, db2–db10 are
components, while WCs represent the high frequency com- chosen as mother wavelet candidates because, according to
ponents. Assuming the original sampling frequency is f s , the the literature, they show excellent performance in analyz-
frequency bands of different decomposition levels are shown ing dynamic signals with discontinuity and abrupt changes
in Fig. 6. By transforming a signal, the features of the signal [39], [41], and also higher order wavelets lead to longer
can be extracted from SCs and WCs. In the case of disturbance computation time.
signals in power systems, WCs are more important because the Fig. 7 shows the frequency response of the filters based on
transient features of the signals are mostly represented by the different mother wavelets: db2, db10. The frequency responses
high-frequency components [38]. of the low-pass filter (scaling function) are represented by
the dashed lines, and the high-pass filters (wavelet functions)
C. Selection of Mother Wavelet are represented by the solid lines. The frequency response
Before WT-based MRA is applied for feature extraction, the of db10 shows a dramatic reduction in the overlapping
appropriate wavelet function and decomposition scales need region and sharper cutoff frequency, compared with that
to be carefully selected considering that they will impact the using db2. Minimum leakage energy is thus expected during
accuracy and processing time of the proposed fault detection the wavelet decomposition process with smaller overlapping
method. The criteria adopted in this paper for selecting the region between the two filters and more information will
mother wavelet are summarized here [40]. be remained after decomposition for feature extraction and
LI et al.: FAULT DETECTION AND CLASSIFICATION IN MVDC SPSs 2657

classification. Based on the above analysis, db10 wavelet is A. Parseval’s Theorem

chosen in this case. The comparison of these two wavelet If the selected scaling function and the wavelet function
functions in extracting the features are presented more in detail form an orthonormal basis, then Parseval’s theorem relates the
in Section V with simulation results. energy of the fault signal to the energy of the coefficients.
Hence, referring to the wavelet decomposition, this means
D. Selection of Decomposition Level that the energy (E Signal) of fault signal can be expressed
It is preferable to use a higher decomposition level in order mathematically in terms of the expansion coefficients [43],
not to miss the features. However, as the decomposition level [44] as
increases, the computational time becomes significantly larger.  ∞ 
 ∞
Therefore, it is crucial to select a suitable decomposition level E Signal = |c J (k)| + 2
|d j (k)|2 . (6)
that makes a good tradeoff between the number of candidate k=−∞ j =1 k=−∞
features and computing time. Equation (6) could also be written in the following form:
Once a mother wavelet is selected, the data-independent
selection (DIS) method is considered to determine the optimal 
J

levels of decomposition in this paper [42]. E Signal = E c J + Ed j (7)

The DIS approach is based on the following steps. j =1
1) The number of decomposition levels (n Ls ) to be adopted with
depends on the sampling frequency f s of the signal qj
being analyzed. It has to be chosen to allow the high- 
qJ 
Ec J = |c J i |2 , Ed j = |d j i |2 (8)
level signals (approximation and details) to cover all
i=1 i=1
the range of frequencies along which the sideband is
localized. The sideband components are known as the where subscript j stands for the scale, and the maximum
harmonic components that appear around (left and right) number of scale is J ; subscript i stands for the point in WCs;
the fundamental frequency components. and q j is the length of WCs at scale j .
2) The minimum number of decomposition levels that is This energy reflects the energy concentration at each scale.
necessary for obtaining an approximation signal so that The larger the wavelet energy, the more relevant the infor-
the upper limit of its associated frequency band is mation contained in the decomposition terms [38]. In this
under the fundamental frequency f is described by the paper, the difference of energy (E) contained in the WCs at
following condition [42]: each scale between faulted signal (E Fault ) and normal signal
(E Normal ) is adopted as the vector for feature extraction.
2−(n Ls +1) fs < f. (4) The proposed feature vector x can then be mathematically
From this condition, the decomposition level of the approx- represented as
imation signal is the integer n Ls , given by ⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎡ ⎤
x1 E Fault.d1 E Normal.d1 E d1
  ⎢ .. ⎥ ⎢ .. ⎥ ⎢ .. ⎥ ⎢ .. ⎥
log( fs / f ) ⎢. ⎥ ⎢. ⎥ ⎢. ⎥ ⎢. ⎥
n Ls = floor (5) ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥
log(2) x =⎢ x ⎥ = ⎢ E ⎥ − ⎢ E
⎢ j ⎥ ⎢ Fault.d j ⎥ ⎢ Normal.d j ⎥ ⎢
⎥ = ⎢ E dj ⎥ ⎥
where floor() means to take the integral part of the calculation ⎢ .. ⎥ ⎢ . ⎥ ⎢. ⎥ ⎢. ⎥
⎣ . ⎦ ⎣ .. ⎦ ⎣ .. ⎦ ⎣ .. ⎦
result.
For our application, further decomposition of this signal has xJ E Fault.d J E Normal.d J E d J .
to be done so that the frequency band [0, f ] will be decom- (9)
posed into more bands. Usually, two additional decomposition
levels, (i.e., n Ls + 2) would be adequate for the analysis [42]. B. Feature Extraction Results
After calculation according to (5), scale 9 is selected as the
1) Features of Short Circuit Faults: By applying the
decomposition level in this paper. The frequency bands of
WT-based MRA method with db10, and the Parseval’s theory,
different decomposition levels for fault analysis are provided
the detailed energy distributions of the decomposition levels of
in the Appendix.
the dc bus current signals under different operating conditions
(normal and faults) are shown in Figs. 8–10, respectively.
V. F EATURE E XTRACTION The x-axis is the decomposition level and the y-axis is the
The detail coefficients and approximation coefficients energy. In this case, the sampling frequency is 5 kHz, and
obtained from wavelet MRA should not directly be used the frequency bands of the decomposition levels are provided
as the classifier inputs. To reduce the feature dimension, in the Appendix, in Table II. The results clearly show an
the feature extraction methods are generally implemented increased energy at the decomposition levels from 5 to 8 of the
to these coefficients at each decomposition level. Common fault signals, compared with the normal operating condition.
statistical approaches, such as mean, standard deviation, rms, This is because these four decomposition levels cover the
and Shannon-entropy, are not robust with noise [22]. Thus, frequency range from 9.265 to 156.25 Hz, which comprises
in this paper, we proposed to use energy-based technique for the fundamental frequencies (60 Hz, due to the presence of
feature extraction, which uses Parseval’s theorem. the fundamental frequency of the ac side of the MVDC power
2658 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 63, NO. 11, NOVEMBER 2014

Fig. 8. Detailed energy distribution of each decomposition level of dc bus Fig. 11. Extracted feature vectors of short circuit fault with db10 wavelet.
current under normal operation condition.

Fig. 12. Extracted feature vectors short circuit fault with db2 wavelet.

Fig. 9. Detailed energy distribution of each decomposition level of dc bus

current under short circuit fault on dc bus. The short circuit fault can be detected by comparing the vector
x 6 [ which equals to E d6 as given in (9)] with a properly
predefined threshold value: a fault is confirmed only if the
feature vector x 6 exceeds the threshold value. The capability
of the proposed method to distinguish the short circuit fault
on dc bus from short circuit fault on generator ac side could
also be realized by comparing the feature vector of the faulted
signal, the short circuit fault on primary dc bus will lead to
significantly higher values.
To show that the db10 wavelet performs better as expected,
the feature vectors extracted via db2 wavelet is shown in
Fig. 12. Comparing the results in Figs. 11 and 12, it is noticed
that when the order of the mother wavelet is increased, the
values in the feature vectors are getting more concentrated
to vector x 6 , and the differences of the value of vector x 6
becomes larger between the feature vectors of different faults.
This shows good consistency with the analysis in Section IV,
Fig. 10. Detailed energy distribution of each decomposition level of dc bus Part C. As the frequency response of the mother wavelet
current under short circuit fault on ac side. function with higher order has a dramatic reduction in the
overlapping region and sharper cutoff frequency. Minimum
system), the second harmonic (120 Hz), as well as some low leakage energy is thus expected during the wavelet decompo-
frequency components. sition process.
Substituting the energies of each decomposition level 2) Features of Ground Faults: Fig. 13 shows the extracted
into (9), we can get the feature vectors for different short features of the ground fault occurred in Zone 3. The
circuit fault locations (as shown in Fig. 2), as shown in Fig. 11. x-axis is the decomposition level and the y-axis is the energy
LI et al.: FAULT DETECTION AND CLASSIFICATION IN MVDC SPSs 2659

Fig. 14. Structure of a three layer feedforward neural network.

The three layer feedforward neural network has been

Fig. 13. Extracted features of negative line to ground fault in Zone 3 with
db10 wavelet.
designed in this paper for nine inputs and two outputs. In the
proposed method, the inputs are the extracted feature vectors
variation. In this case, the sampling frequency is 40 kHz due to through WT-based MRA technique, as described in Section IV.
the change of monitored signal, and the frequency bands of the The outputs are mathematically shown in (11) to indicate
decomposition levels are provided in the Appendix, Table III. whether the monitored part is under normal condition or a
As we can see, there is a significant increase in vector fault condition, respectively
x 1 and x 2 , which reflects the high frequency component of
background noise signal, as analyzed in Section II. Moreover, φ : R n → {0, 1}m
the values of vector x 6 to x 9 also slightly increase, because the φ(x) = (c1 , c2 , . . . , ci , . . . , cm ) (10)
bandwidth of these decomposition levels covers the switching
frequency of pervasive buck converters (600 Hz). with

Through the above analysis, we can see that, the application 1 fault
of WT-based MRA method and Parseval’s theorem, yields ci = (11)
0 normal
a compact feature of the fault signal. Thus, less memory
space and computing time are required for implementation where i = 1, 2, . . . , m. m is the number of output neurons, in
compared with the fault detection methods that adopt directly this case m equals to 2. n is the number of inputs and equals
the sampled raw data or the WT coefficients. The extracted to 9.
features may be affected by the noise at the input signals, As stated before, the number of neurons in the hidden layer
however, this can potentially be reduced by denoising the input has been selected by trial and error with the goal of reaching
signals by using a wavelet-based denoising method presented better stability and high convergence rate. Different numbers
in [45]. The obtained feature vectors with db10 wavelet are of neurons in the hidden layer have been tried and it has been
then used as the inputs of ANN for pattern recognition, and found that the feedforward networks with eight or more hidden
this process is detailed in the following section. neurons are sufficient for this particular application, where
the training error is set to be lower than 0.001. However, to
VI. C LASSIFICATION W ITH ANN reduce the computational burden of the proposed algorithm for
ANN has the capability of recognizing underlying relation- the purpose of the RT implementation, a feedforward network
ship between the input and output patterns. It can consider sev- with eight hidden neurons is chosen. It is to be noted that the
eral features of the input signals simultaneously and compare selected number of hidden layer may not be the optimal, the
the patterns according to their mutual similarity instead of the number is chosen only if it satisfied the preset training error;
hard thresholds. Of the different topologies, the feedforward the optimal issue is out of the scope of this paper.
network is probably the most popular, and it constitutes more The log-sigmoid [49], [50] activation function is used in
than 80% of the current neural network applications for fault both hidden and output layers of the network, because we want
diagnostic [46]. to use binary outputs to indicate the fault types. The network
The basic multilayer feedforward network contains one is trained in a supervised manner with the back propagation
input layer, one output layer, and any number of hidden function traingdm [49]–[51]. The weight and bias values of the
layers in between [47], [48]. Currently, the design of network hidden and output layers are updated according to the gradient
configurations to solve a particular engineering problem is still descent with momentum. The actual output is compared with
a trial and error process [46]. In this paper, the three layer the desired or target output. Depending on the error between
feedforward neural networks has been adopted for implement- actual and target outputs, the weight of each connection and
ing the proposed fault detection and classification method for the bias of each processing unit of the network are updated.
MVDC SPSs. The specific structure of the adopted three layer The training process of the network continues until the average
feedforward neural networks is shown in Fig. 14. error of the entire training set reaches a defined error goal.
2660 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 63, NO. 11, NOVEMBER 2014

TABLE I
C LASSIFICATION R ESULTS

Fig. 15. Accuracy of the methods with different SNRs.

the sampled raw data as the inputs for ANN for classification.
A. Classification Results
Method 2 (M2) adopts all the DWT coefficients of the second
After training with specific momentum constant and adap- decomposition level as the inputs for a self-organizing maps.
tive learning rate, the proposed technique is tested offline Method 3 (M3) is the method proposed in this paper, which
in the MATLAB software package using appropriate set of feeds ANN with the features extracted through DWT MRA
weights and biases, and the data collected from RT simulation and Parseval’s theorem.
in RTDS. The training and testing cases include different fault Fig. 15 shows the classification accuracies of different
types (short circuit fault on dc bus, disturbance caused by methods under various signal-to-noise ratio (SNR) values
short circuit fault on ac side, ground fault in Zone 3, and (from 20 to 50 dB). The value of SNR is defined as follows:
normal operations), and normal operations with disturbances  
Ps
(such as load switching, capacitor switching). Forty cases SNR(dB) = 10 log (13)
are used for training of each scenario with the proposed Pn
algorithm, 240 cases are used for testing. These training and where Ps is the power of the original signal and Pn is that of
testing cases are obtained randomly with varying time of the noise. A typical SNR value of 30 dB is equivalent to a
occurrence, fault location (in step of 0.1 p.u), fault resistances peak noise magnitude of nearly 3.5% of the original signal.
(in step of 0.01 p.u), and the load in the zone (in step of As can be noted from Fig. 15, the proposed method in this
1 ohm). The randomly generated scenarios may consist of paper, M3, shows the best robustness under different levels of
some extreme cases and thus can demonstrate an overall SNRs. The overall accuracy can still reach 97.08% even under
statistical performance of the proposed algorithm. the noise level of 20dB. The performance of M1 is easily
Since the outputs of a multilayer neural network rarely affected by the noises, due to the directly use of sampled raw
give exactly the target of zero or one on each output neuron, data as the inputs for classification. Moreover, it has the most
the pattern was deemed to have been correctly classified if complex structures and highest computation cost for training.
the error on each output neuron is within 0.3 [52]. If any of the Although M2 takes the DWT coefficients of the second
outputs takes a value between 0.3 and 0.7, this is considered decomposition level as input vectors, its performance also
as error in the detection of the pattern. decreases a lot with the noise levels. Because the use of only
Accuracy defined as in (13) is used as the metric for deter- one decomposition level may not reflect the characteristics of
mining the performance of ANN, which can be mathematically the original signal, and some of the properties of the original
defined as follows: signal will be lost.
  A comparative study of the classification accuracies with the
Ne
A = 100 · 1 − (12) features obtained from Parseval’s-based energy technique and
N
other statistical extraction techniques (such as, mean, variance,
where A represents the accuracy, Ne and N represent the rms, and entropy) has been done, and the results are shown in
number of classification errors and the total number of tested Fig. 16. It can be noticed that, the proposed energy technique
examples, respectively. has the best accuracy and is robust with noise.
The results of the test cases are summarized in Table I. Although only the occurrence of time, fault location, fault
Notice that, the algorithm is able to correctly identify the cases resistances, and the loads are considered during the perfor-
with the accuracy of 99.58%, which shows the effectiveness mance studies of the classifier, the obtained scheme has the
of the proposed algorithm to correctly classify the faults on generality for different fault and operating conditions. The real
MVDC system. system may have different operating conditions and different
The performance of the proposed method has also been system configurations from the cases that we have studied
compared with other signal processing-based methods pro- here, in these cases, higher classification performance could be
posed for fault detection and classification in MVDC power guaranteed only by retraining the neural network with newly
systems, as shown in Table V. Method 1 (M1) directly uses obtained features.
LI et al.: FAULT DETECTION AND CLASSIFICATION IN MVDC SPSs 2661

Fig. 16. Accuracies of different feature extraction techniques.

Fig. 18. Real view of experiment configuration in ACS lab.

Fig. 17. Schematic structure of the RT testing platform.

VII. RT V ERIFICATION
The algorithm of WT-based MRA and ANN-based fault
diagnosis algorithm can be implemented for online applica-
tions at an affordable cost [53], [54]. Thus, it is feasible
to embed the proposed fault diagnosis algorithm into local
monitoring devices. It is to be noted that, the neural network
is trained offline with the obtained data from RTDS and then
tested online. This section illustrates in detail the RT imple-
mentation and testing of the proposed algorithm.
The schematic structure of the RT platform for online Fig. 19. Correct recognition of short circuit fault on primary dc bus.
testing of the proposed fault detection algorithm is shown in
Fig. 17, and a picture of the experimental setup is shown in
Fig. 18. The platform is a HIL setup based on RTDS [28] compiled in Microsoft Visual C++ to generate Model.dll,
and LabVIEW RT target [30]. The notional MVDC SPS is which is executed on the LabVIEW RT target. When the
implemented in RTDS and executed in RT. The monitored process is started, the samples of the monitored signals are
signals for fault detection are accessed via the analog output sent to the memory of the RT target, where they are processed.
board of RTDS, and fed to the RT target, where the proposed Fault detection and diagnosis are achieved by output the
fault detection and classification algorithm has been fully indicators in binary values from RT target.
implemented. Data acquisition in RT target is realized with One of the main design considerations during hardware
the field-programmable gate array card, a detailed description implementation is to minimize the cause of high frequency
of this module can be found in [30]. noise. Low-pass R-C antialiasing filter has been used at the
The C code of the proposed fault diagnosis algorithm input side (in the interface board, as shown in Fig. 18) to
that integrates WT-based MRA and ANN is generated from remove the high frequency component of the signal outside
MATLAB RT Workshop. The generated C code is then the sampling bandwidth. The cut off frequencies of the filters
2662 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 63, NO. 11, NOVEMBER 2014

is solved by discarding the highest frequency detailed WCs,

which ultimately results in an effective sampling frequency
equal to half of the original sampling frequency. For example,
if the sampling frequency is 20 kHz, then the wavelet energies
are calculated using levels 2–10 WCs. This is equivalent to the
use of levels 1–9 components with a sampling frequency of
10 kHz.
The dc bus current and the indicator signals for the case of
short circuit fault on primary dc bus are shown in Fig. 19. It
clearly shows that the indicator signal (c1 ) changed its status
from low to high after the fault occurrence indicating that a
fault has been detected for the case of dc bus short circuit fault
(as illustrated in Fig. 14). In addition, the algorithm generated
the high indicator signal (c2 ) after the fault occurrence for the
case of disturbance caused by short circuit fault on ac side,
as shown in Fig. 20. Similarly, the corrected recognition of
ground fault has been shown in Fig. 21, where both of the
indicator signals are high after the occurrence of the fault.
VIII. C ONCLUSION
This paper proposes a method that integrates WT-based
MRA technique with ANN for fault detection and classifi-
cation in MVDC SPSs. The application issues of WT-based
MRA, such as selection of proper wavelet function and optimal
decomposition scales, are discussed and supported with sim-
ulation analysis. Taking WCs energy to generate the feature
Fig. 20. Correct recognition of disturbance caused by short circuit fault on
ac side. vector, the db10 wavelet and scale 9 are chosen as mother
wavelet and the decomposition level of interest, respectively,
according to the criteria discussed in this paper. ANN is then
adopted to classify these extracted features to identify the fault
types automatically, with its strong capability of generalization
and training mechanism. Numerical simulation has been per-
formed to verify the capability of the proposed algorithm for
fault detection and the results show that promising accuracy
can be obtained for distinguishing different types of faults.
Finally, the proposed algorithm is implemented in LabVIEW
RT target for the purpose of online practical use. However,
more adjustments are needed with this obtained classifier to
make it be able to cope with specific system structures. To be
more precise, we need to train the ANN classifier first with
the data obtained from the new system.
Compared with the other methods, the adopted method
in this paper does not require interzone communication and
external signal injection. The main properties of the proposed
method are summarized as follows.
1) The WT-based MRA method provides an efficient way
to extract signal components at different frequency
bands.
2) The neural network provides an intelligent method and
a soft criterion for feature comparison.
3) The proposed algorithm decreases the computational
burden due to less input data (only nine values of the
Fig. 21. Correct recognition of ground fault in Zone 3. feature vector), thus it supports RT online implementa-
tion.
are set to half the sampling frequency. However, they also 4) The algorithm has a high accuracy and is robust to
introduce errors into the high frequency components just inside various faults, its performance will not be affected by
the sampling bandwidth, which may affect the magnitudes of changes in electrical parameters like fault resistance and
the high frequency wavelet decompositions [55]. This problem loads.
LI et al.: FAULT DETECTION AND CLASSIFICATION IN MVDC SPSs 2663

A PPENDIX
TABLE II TABLE III
F REQUENCY BANDS OF D IFFERENT D ECOMPOSITION L EVELS F REQUENCY BANDS OF D IFFERENT D ECOMPOSITION L EVELS
FOR A NALYSIS OF G ROUND FAULTS FOR A NALYSIS OF S HORT C IRCUIT FAULTS

TABLE IV
C OMPARISON OF R ELATED FAULT D ETECTION M ETHODS FOR MVDC P OWER S YSTEMS

TABLE V
S IGNAL P ROCESSING -BASED M ETHOD FOR MVDC S YSTEM FAULT D ETECTION AND P ROTECTION
2664 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 63, NO. 11, NOVEMBER 2014

R EFERENCES [25] I. Monedero, C. Leon, J. Ropero, A. Garcia, J. M. Elena, and

J. C. Montano, “Classification of electrical disturbances in real time
[1] M. E. Baran and N. Mahajan, “System reconfiguration on shipboard DC using neural networks,” IEEE Trans. Power Del., vol. 22, no. 3,
zonal electrical system,” in Proc. IEEE, Electric Ship Technol. Symp., pp. 1288–1296, Jul. 2007.
Jul. 2005, pp. 86–92. [26] A. Sadeghian, Z. Ye, and B. Wu, “Online detection of broken rotor
[2] A. Emadi, M. Ehsani, and J. M. Miller, Vehicular Electric Power bars in induction motors by wavelet packet decomposition and arti-
Systems. New York, NY, USA: Marcel Dekker, 2004. ficial neural networks,” IEEE Trans. Instrum. Meas., vol. 58, no. 7,
[3] T. Ericsen, N. Hingorani, and Y. Khersonsky, “Power electronics and pp. 2253–2263, Jul. 2009.
future marine electrical systems,” IEEE Trans. Ind. Appl., vol. 4, no. 1, [27] W. Li, M. Lio, A. Monti, and F. Ponci, “Wavelet based method for fault
pp. 155–163, Jan. 2006. detection in medium voltage DC shipboard power systems,” in Proc.
[4] N. Flourentzou, V. G. Agelidis, and G. D. Demetriades, “VSC-based IEEE I2MTC, May 2012, pp. 2155–2160.
HVDC power transmission systems: An overview,” IEEE Trans. Power [28] Real Time Digital Simulator RTDS, Calgary, AB, Canada. (2011,
Electron., vol. 24, no. 3, pp. 592–602, Mar. 2009. Feb. 20). RTDS Simulator [Online]. Available: http://www.rtds.com
[5] C. Christopoulos and A. Wright, Electrical Power System Protection, [29] A. Monti, F. Ponci, and A. Benigni, “Real time simulation and hardware
2nd ed. Dordrecht, The Netherlands: Kluwer Academic, 1999. in the loop as enablers for sustainable energy solutions,” in Proc. ICCEP,
[6] E. Christopher, M. Sumner, D. Thomas, X. Wang, and F. de Wildt, “Fault Jun. 2011, pp. 21–27.
location in a zonal DC marine power system using active impedance [30] National Instruments, Austin, TX, USA. (2012, Jan. 18). NI LabVIEW
estimation,” in Proc. IEEE ECCE, Sep. 2010, pp. 3050–3054. [Online] Available: http://www.ni.com/products/
[7] D. Thomas, M. Sumner, D. Coggins, X. Wang, J. Wang, and [31] H. Li, W. Li, M. Luo, A. Monti, and F. Ponci, “Design of smart MVDC
R. Geertsma, “Fault location for DC marine power systems,” in Proc. power grid protection,” IEEE Trans. Instrum. Meas., vol. 60, no. 9,
IEEE ESTS, Apr. 2009, pp. 456–460. pp. 3035–3046, Sep. 2011.
[8] X. Feng, Z. Ye, C. Liu, R. Zhang, F. C. Lee, and D. Boroyevich, [32] H. Xin, “Modeling and control of the multi-terminal MVDC system,”
“Fault detection in DC distributed power systems based on impedance M.S. thesis, Dept. Electr. Eng. Inf. Technol., RWTH Aachen Univ.,
characteristics of modules,” in Proc. IEEE Ind. Appl. Conf., Oct. 2000,
Aachen, Germany, Sep. 2012.
pp. 2455–2462.
[33] V. Arcidiacono, A. Monti, and G. Sulligoi, “Generation control system
[9] M. Aurangzeb, P. A. Crossley, and P. Gale, “Fault location using
for improving design and stability of medium-voltage DC power systems
the high frequency travelling waves measured at a single location on
on ships,” IET Electr. Syst. Transp., vol. 2, no. 3, pp. 158–167, Sep. 2012.
a transmission line,” in Proc. 7th Int. Conf. Develop. Power Syst.
[34] W. Li and A. Monti, “Integrated simulation with VTB and
Protection, 2001, pp. 403–406.
OPNET for networked control and protection in power sys-
[10] D. W. P. Thomas, R. J. O. Corvalho, and E. L. Pereira, “Fault location
tems,” in Proc. IMSC SummerSim, Symp. GCMS, Jul. 2010,
in distribution systems based on travelling waves,” in Proc. Power Tech.
pp. 386–391.
Conf., 2003, pp. 5–9.
[35] Y. Pan, M. Steurer, and T. L. Baldwin, “Ground fault location testing
[11] Z. Q. Bo, G. Weller, and M. A. Redfern, “Accurate fault location
of a noise-pattern-based approach on an ungrounded DC system,” IEEE
technique for distribution system using fault-generated high-frequency
Trans. Ind. Appl., vol. 47, no. 2, pp. 996–1002, Mar. 2011.
transient voltage signals,” IEE Generat. Transmiss. Distrib. Proc. C,
vol. 146, no. 1, pp. 73–79, Jan. 1999. [36] S. Mallat, “A theory for multiresolution signal decomposition: The
[12] N. K. Chanda and Y. Fu, “ANN-based fault classification and location in wavelet representation,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 11,
MVDC shipboard power systems,” in Proc. NAPS, Aug. 2011, pp. 1–7. no. 7, pp. 674–693, Jul. 1989.
[13] C. S. Chang, S. Kumar, B. Liu, and A. Khambadkone, “Real-time [37] S. Santoso, E. J. Powers, and W. M. Grady, “Power quality disturbance
detection using wavelet transform and neural network of short-circuit data compression using wavelet transform methods,” IEEE Trans. Power
faults within a train in DC transit systems,” IEE Electric Power Appl. Del., vol. 12, no. 3, pp. 1250–1257, Jul. 1997.
Proc., vol. 148, no. 3, pp. 251–256, May 2001. [38] J. Ning, “Wide-area monitoring and recognition for power system distur-
[14] C. S. Chang, Z. Xu, and A. Khambadkone, “Enhancement and laboratory bances using data mining and knowledge discovery theory,” Ph.D. dis-
implementation of neural network detection of short circuit faults in sertation, Dept. Comput. Sci., Tennessee Technol. Univ., Cookeville, TN,
DC transit system,” IEE Electric Power Appl. Proc., vol. 150, no. 3, USA, Aug. 2010.
pp. 344–350, May 2003. [39] W. Gao and J. Ning, “Wavelet-based disturbance analysis for power
[15] M. E. Barana and N. R. Mahajan, “Overcurrent protection on voltage- system wide-area monitoring,” IEEE Trans. Smart Grid, vol. 2, no. 1,
source-converter-based multiterminal DC distribution system,” IEEE pp. 121–130, Mar. 2011.
Trans. Power Del., vol. 22, no. 1, pp. 406–412, Jan. 2007. [40] V. Perrier, T. Philipovitch, and C. Basdevant, “Wavelet spectra compared
[16] L. Tang and B.-T. Ooi, “Locating and isolating DC faults in multi- to Fourier spectra,” J. Math. Phys., vol. 36, no. 3, pp. 1506–1519,
terminal DC systems,” IEEE Trans. Power Del., vol. 22, no. 3, Mar. 1995.
pp. 1877–1884, Jul. 2007. [41] L. Zhang and P. Bao, “Edge detection by scale multiplication in wavelet
[17] M. Naidu, T. J. Schoepf, and S. Gopalakrishnan, “Arc fault detection domain,” Pattern Recognit. Lett., vol. 23, no. 14, pp. 1771–1784, 2002.
scheme for 42-V automotive DC networks using current shunt,” IEEE [42] J. A. Daviu, M. Riera-Guasp, J. Roger-Folch, F. Martínez-Giménez, and
Trans. Power Electron., vol. 21, no. 3, pp. 633–639, May 2006. A. Peris, “Application and optimization of the discrete wavelet transform
[18] N. Zhang and M. Kezunovic, “Transmission line boundary protection for the detection of broken rotor bars in induction machines,” Appl.
using wavelet transform and neural network,” IEEE Trans. Power Del., Comput. Harmon. Anal., vol. 21, no. 2, pp. 268–279, Sep. 2006.
vol. 22, no. 2, pp. 859–869, Apr. 2007. [43] C.-C. Liao and H.-T. Yang, “Recognizing noise-influenced power quality
[19] V. Pathirana and P. G. McLaren, “A hybrid algorithm for high speed events with integrated feature extraction and neuro-fuzzy network,” IEEE
transmission line protection,” IEEE Trans. Power Del., vol. 20, no. 4, Trans. Power Del., vol. 24, no. 4, pp. 2132–2141, Oct. 2009.
pp. 2422–2428, Oct. 2005. [44] F. H. Magnago and A. Abur, “Fault location using traveling waves
[20] Y. H. Gu and M. H. J. Bollen, “Time-frequency and time-scale domain and wavelet transform,” in Proc. North Amer. Power Symp., Oct. 1997,
analysis of voltage disturbances,” IEEE Trans. Power Del., vol. 15, no. 4, pp. 1–3.
pp. 1279–1284, Oct. 2000. [45] H.-T. Yang and C.-C. Liao, “A de-noising scheme for enhancing wavelet-
[21] S. Mallat, A Wavelet Tour of Signal Processing. New York, NY, USA: based power quality monitoring system,” IEEE Trans. Power Del.,
Academic, 1999. vol. 16, no. 3, pp. 353–360, Jul. 2001.
[22] H. Eristi, A. Ucar, and Y. Demir, “Wavelet-based feature extraction and [46] M.-Y. Chow, R. N. Sharpe, and J. C. Hung, “On the application and
selection for classification of power system disturbances using support design of artificial neural networks for motor fault detection—Part I,”
vector machines,” Electr. Power Syst. Res., vol. 80, no. 7, pp. 743–752, IEEE Trans. Ind. Electron., vol. 40, no. 2, pp. 181–188, Apr. 1993.
Jul. 2010. [47] D. E. Rummelhart and J. L. McClelland, Parallel Distributed Processing.
[23] A. M. Gaouda, E. F. El-Saadany, M. M. A. Salama, V. K. Sood, Cambridge, MA, USA: MIT, 1986.
and A. Y. Chikhani, “Monitoring HVDC systems using wavelet [48] R. P. Lippmann, “An introduction to computing with neural nets,” IEEE
multi-resolution analysis,” IEEE Trans. Power Syst., vol. 16, no. 4, ASSP Mag., vol. 4, no. 2, pp. 4–22, Apr. 1987.
pp. 662–670, Nov. 2001. [49] H. Haykin, Neural Networks: A Comprehensive Foundation. Piscataway,
[24] W. R. A. Ibrahima and M. M. Morcos, “Artificial intelligence and NJ, USA: IEEE Press, 1994.
advanced mathematical tools for power quality applications: A survey,” [50] M. Works, MATLAB: Neural Network Tool Box. Natick, MA, USA:
IEEE Trans. Power Del., vol. 17, no. 2, pp. 668–673, Apr. 2002. MathWorks Inc., 2006.
LI et al.: FAULT DETECTION AND CLASSIFICATION IN MVDC SPSs 2665

[51] M. Khan and M. A. Rahman, “Development and implementation of a Antonello Monti (M’94–SM’02) received the M.Sc.
novel fault diagnostic and protection technique for IPM motor drives,” (summa cum laude) and Ph.D. degrees in electrical
IEEE Trans. Ind. Electron., vol. 56, no. 1, pp. 85–92, Jan. 2009. engineering from Politecnico di Milano, Milan, Italy,
[52] J. Jin, C. S. Chang, C. Chang, T. Hoshino, M. Hanai, and N. Kobayashi, in 1989 and 1994, respectively.
“Classification of partial discharge events in gas-insulated substations He started his career with Ansaldo Industria, Abu
using wavelet packet transform and neural network approaches,” IEE Dhabi, UAE, and then with Politecnico di Milano
Proc. Sci. Meas. Technol., vol. 153, no. 2, pp. 55–63, Mar. 2006. as an Assistant Professor. In 2000, he joined the
[53] M. Forghani and S. Afsharnia, “Online wavelet transform-based control Department of Electrical Engineering, University of
strategy for UPQC control system,” IEEE Trans. Power Del., vol. 22, South Carolina, Columbia, SC, USA, as an Asso-
no. 1, pp. 481–491, Jan. 2007. ciate and then a Full Professor. Since 2008, he has
[54] O. A. S. Youssef, “Combined fuzzy-logic wavelet-based fault classifi- been the Director with the Institute for Automation
cation technique for power system relaying,” IEEE Trans. Power Del., of Complex Power System, E.ON Energy Research Center, RWTH Aachen
vol. 19, no. 2, pp. 582–589, Apr. 2004. University, Aachen, Germany. He has authored and co-authored more than
[55] N. Perera and A. D. Rajapakse, “Development and hardware implemen- 200 peer-reviewed papers published in international journals and in the
tation of a fault transients recognition system,” IEEE Trans. Power Del., proceedings of international conferences.
vol. 27, no. 1, pp. 40–52, Jan. 2012. Dr. Monti is an Associate Editor of the IEEE S YSTEMS J OURNAL.

Ferdinanda Ponci (M’00–SM’08) received the M.S.

Weilin Li (S’10–M’13) received the B.S., M.S., and and Ph.D. degrees in electrical engineering from
Ph.D. (Dr. Ing.) degrees in electrical engineering Politecnico di Milano, Milan, Italy, in 1998 and
from Northwestern Polytechnical University, Xi’an, 2002, respectively.
China, and the E.ON Energy Research Center, Insti- She joined the Department of Electrical Engi-
tute for Automation of Complex Power Systems, neering, University of South Carolina, Columbia,
RWTH Aachen University, Aachen, Germany, in SC, USA, in 2003, as an Assistant Professor, with
2007, 2009, and 2013, respectively. the Power and Energy Research Group, and was
He is currently an Assistant Professor with the promoted to Associate Professor in 2008. In 2009,
Department of Electrical Engineering, Northwest- she was with the Institute for Automation of Com-
ern Polytechnical University. His current research plex Power Systems, E.ON Research Center, RWTH
interests include medium voltage dc power system, Aachen University, Aachen, Germany, where she is currently an Apl. Profes-
integrated co-simulation, and power electronic applications in smart grid. sor.