7078 IEEE JOURNAL OF EMERGING AND SELECTED TOPICS IN POWER ELECTRONICS, VOL. 9, NO.

6, DECEMBER 2021

DC Fault Detection and Pulsed Load Monitoring

Using Wavelet Transform-Fed LSTM Autoencoders
Yue Ma , Student Member, IEEE, Damian Oslebo, Member, IEEE, Atif Maqsood , Member, IEEE,
and Keith Corzine , Fellow, IEEE

Abstract— The extensive deployment of power electronics loads Pulsed dc loads especially provide a challenge for traditional
in naval ship power systems indicates the ship electrification load monitoring and fault detection techniques. For example,
is inevitable in future trends. Next-generation warships require more electronic loads, such as electromagnetic rail guns and
high-power density weapons drawing pulse power from the
dc grid. A particularly concerning issue is that these pulse advanced radar on an electric ship power system, are irregular
loads draw large currents in short periods of time, similar pulsed-power load in nature, drawing pulsating current from
to fault behavior, and maybe indiscernible from a fault. This the grid. This presents a challenge for conventional fault
article proposed a novel machine learning-based algorithm—long monitoring devices, as many fault currents are also pulsating
short-term memory (LSTM) recurrent neural network (RNN)- in nature and can be difficult to differentiate from a desirable
based autoencoder (AE) networks to detect dc faults and monitor
load conditions applied to naval pulse loads. The novel load pulse event.
monitoring solution presented herein can be applied to any A new load monitoring approach is urgent to be pro-
load profile that exhibits repetitive transients during normal posed for naval pulse loads. The feature extraction is the
operation. The frequency-domain features of the load current first step, in which certain attributes about the load profile
are extracted under wavelet transform for the network training characteristics are extracted and stored and can be used in the
to set the network weights and biases. Once the network training
is completed, the LSTM RNN-based AE will produce both signal following classification step to identify the current load profile.
classification and signal reconstruction of the pulse load based on A comprehensive summary of the feature extraction methods is
wavelet features of input current. Any faults should yield large presented in [11]. These feature extractions can be either time
reconstruction errors for protective action. Finally, the method domain, such as active load profile components [12], [13],
is demonstrated in experimental results. or frequency domain, such as harmonic content [14], [15].
Index Terms— Autoencoder (AE), dc fault detection, electric Some researchers have also proposed a combination of both
ship, load monitoring, long short-term memory (LSTM), pulsed time- and frequency-domain methods for feature extractions
dc load, recurrent neural network (RNN), wavelet transform. [16]. Recent research has shown that wavelet analysis applied
to both ac and dc systems is an effective means to study
I. I NTRODUCTION
the transient behavior of load shifts and disturbances [17].

N AVAL shipboard power systems are moving toward

larger incorporation of power electronic-based loads
assisted by enabling technologies, including the wide bandgap
Some recent studies have also utilized wavelet analysis to also
identify nascent arcing faults, another crucial aspect for load
monitoring [18], [19].
devices, medium-voltage dc (MVDC), integrated power and The recent machine learning renaissance that was brought
energy system (IPES), and so on [1]–[6]. An MVDC system by innovations in parallel computing offers new opportunities
offers benefits, including less power conversion modules, for neural network applications in load monitoring and fault
enhanced power density, and efficiency, that make an MVDC diagnostics. A significant advantage of the neural network
distribution a strong contender for implementation in future architecture branch of machine learning applications is that
ship power systems [7]–[9]. It is critical to implement a these architectures can model nonlinearities when multiple
reliable monitoring and protection scheme in an MVDC sys- layers of nodes are applied in succession during operation
tem that can detect and react to any incipient faults [10]. in both ac and dc distribution system transient analysis [20],
Manuscript received May 30, 2020; revised July 22, 2020; accepted [21]. A recurrent neural network (RNN) and its LSTM variant
August 3, 2020. Date of publication August 25, 2020; date of current provide for time-domain signal classification by accounting
version December 1, 2021. Recommended for publication by Associate Editor for previous inputs to influence the classifier output at the
Frederick P. Page. (Corresponding author: Yue Ma.)
Yue Ma, Atif Maqsood, and Keith Corzine are with the Department of current time. LSTM RNNs more aptly model the causal
Electrical and Computer Engineering, University of California at Santa Cruz, dynamics of MVDC distributions compared to other tradi-
Santa Cruz, CA 95064 USA (e-mail: [email protected]; [email protected]; tional neural network architectures because RNNs retain the
[email protected]).
Damian Oslebo is with the Department of Electrical and Computer Engi- previous system state. Moreover, artificial neural network or
neering, Naval Postgraduate School, Monterey, CA 93943 USA (e-mail: convolutional neural network layers can also be leveraged
[email protected]). simultaneously with an LSTM RNN for feature extraction for
Color versions of one or more of the figures in this article are available
online at https://ieeexplore.ieee.org. better predictor accuracy [22]. RNNs have been applied suc-
Digital Object Identifier 10.1109/JESTPE.2020.3019382 cessively in harmonics analysis, natural language processing,

2168-6777 © 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: Pontificia Universidad Javeriana. Downloaded on October 23,2023 at 03:41:30 UTC from IEEE Xplore. Restrictions apply.
MA et al.: DC FAULT DETECTION AND PULSED LOAD MONITORING 7079

electrocardiography, and stock market prediction [23]–[26].

LSTM networks have been used in a few anomaly detection
applications in other research fields. Usually, these networks
require a labeled data set to train the model to look for
anomalous conditions [27]–[29]. LSTM AEs, which do not
require a labeled set, have also been used as fault detectors as
well [30]–[32]. Fig. 1. Three-stage DWT implementation [35].
This article proposed a machine learning-based approach to
pulsed dc load monitoring and fault detection applied to naval
pulse loads. Section II provides the description of the proposed
monitoring solution: a stationary wavelet transform (SWT)-
fed long short-term memory (LSTM) RNN-based autoencoder
(AE). Section III describes the operating principles and pro-
vides the parameters used in pulse load prototypes for training
and testing. The experimental results of the pulse load with Fig. 2. Three-stage SWT implementation [36].
load monitoring and fault detection are also discussed in this
section.
the signal. Fig. 1 shows the DWT process for three levels
of decomposition. The g[n] functions represent the low-pass
II. FAULT D ETECTION AND L OAD M ONITORING filter to produce an approximation coefficient of the signal
N ETWORK at the next larger dyadic scale. The h[n] functions represent
This section begins by describing the building block com- the high-pass filter to produce the detail coefficients that
ponents of the proposed load monitoring system in Sections correspond to the CWT values at each dyadic scale. Each
II-A–II-C. The final proposed load monitoring system and fault stage output is then downsampled by a factor of 2 to ensure
detector conclude in Sections II-D and II-E. the output signal size is the same as the input signal size after
all the stage outputs are concatenated. A key issue with the
DWT during event analysis is that the implementation reduces
A. Wavelet Transform Feature Extraction the output length of the coefficients for each dyadic scale. This
The proposed feature extraction implementation performs a means that a load shift occurring at one time in a sample block
preprocessing SWT step in the load monitoring system. The will appear differently than if the same load shift occurred
SWT was selected for its shift-invariance analysis and com- at another place in the sample block as successive filtering
putational advantages. This step conditions the input signal stages are applied. DWT, therefore, loses shift invariance—an
into the LSTM AE-based load monitoring system. The final important attribute of CWT.
SWT recursive implementation was demonstrated in sample The SWT is performed much like the DWT, except it keeps
frequency rates in excess of 20 kHz in our control hardware shift invariance by upsampling the high-pass and low-pass fil-
described later. ters after each filter stage, and there are no output signal down
Wavelet transformations are relatively nascent when com- sample operations. Fig. 2 shows the SWT implementation over
pared with their Fourier transform (FT) counterpart [33]. three stages. The number of stages of decomposition is limited
Alfréd Haar proposed the Haar wavelet when reiterating work by the relationship of 2N less than or equal to the sample block
on Fourier. Unlike the sine wave used to decompose a signal to size, where N is the number of stages. This allows transient
fundamental components in the FT, wavelets have a start and information to align by the correct time index to produce
an end defined by their supported mini wave-like signal, such the signal approximations that are passed to subsequent filter
as square waves, used by Haar wavelet. While much like the stages. The result of this process is a time-scale decomposition
FT, the continuous wavelet transform (CWT) uses the integral of transient information that is consistent without regard to the
inner product of the original signal with a continuous function. transient’s position in the sample block.
A time-shifted and scaled wavelet function is the basis for The final implementation of the real-time load monitoring
the wavelet transform shown in (1), where  represents the system recursively computes the Haar wavelet transformation.
wavelet function that is shifted in time by parameter b and For wavelet basis selection, the Haar wavelet was chosen for
scaled by parameter a the SWT because it is computed from two-point differences
 ∞   and sums. This means its simplicity in implementation and
1 ∗ t −b
CWT(a, b) = √ x(t)ψ dt. (1) in wavelet filter size, making it well-suited for employment
a −∞ a on resource-constrained microcontrollers. The recursive form
However the CWT is not feasible due to the computational of Haar SWT is presented in (2) and (3). The recursive
load for real-time applications, such as load monitoring. The wavelet filter approach uses previously calculated filter outputs
discrete wavelet transform (DWT) can be used to overcome along with “future” inputs to calculate the filter coefficients
this limitation by computing the wavelet transform on dyadic for the current time step. The approximation coefficients are
scales. The DWT is constructed from the foundation of calculated as in (2) where i represents the time index, j
multiresolution analysis [34] and piecewise approximations of represents the decomposition stage, and f (i ) is the input

Authorized licensed use limited to: Pontificia Universidad Javeriana. Downloaded on October 23,2023 at 03:41:30 UTC from IEEE Xplore. Restrictions apply.
7080 IEEE JOURNAL OF EMERGING AND SELECTED TOPICS IN POWER ELECTRONICS, VOL. 9, NO. 6, DECEMBER 2021

signal. A running summation is made by adding the new

signal f (i ) to the previously computed approximation and
subtracting the signal outside its filter length. In (3), the detail
coefficients d j [i ] at stage j are calculated by a difference
between approximations from the finer scale j − 1
  
a j [i ] = 2− j/2 · a j [i − 1] + f [i ] − f i − 2 j (2)
  
d j [i ] = 2− j/2 · a j −1 [i ] − a j −1 i − 2 j −1 . (3)

B. Long Short-Term Memory Recurrent Neural Networks

An RNN is formed by allowing a feed-forward neural
network to contain cyclical feedback onto itself. The recurrent
portion of an RNN consists of a single hidden layer with input
from previous time steps incorporated with the current time Fig. 3. LSTM networks cell node [37].
step as in (4). Therein, vector x contains the input sequence,
and F and H denote the number of feature inputs taken each ten-node fully connected layer. h t could also be sized to the
time sample and the number of hidden nodes, respectively. wih desired output vector size if the LSTM cell is the last layer in
and wh  h are the network weights between the input vector and the neural network architecture. For example, a neural network
previous state, respectively, at each hidden layer node h. The with an LSTM in the last layer that is used to estimate the
superscript values correspond to the time sample t or previous speed and position of an object on a 1-D plane would have
time sample t −1. Finally, ah and bh correspond to the input to an h t of size 2 for the LSTM cell.
the nonlinear activation layer and the activation layer output, Fig. 3 can be summarized from the operations in (7) set. f t
respectively. (“forget” gating) is an intermediate variable for the portion of
The activation layer operation is performed by (5) as in cell memory Ct that should be forgotten at time t considering
the artificial neural networks/convolutional neural networks new input from x t and the LSTM cell’s previous output h t−1 .
architectures, where θ is a nonlinear activation function; i t (“input” gating) and gt (“gate” gating) are intermediate vari-
usually the activation layer operation is the sigmoid or softmax ables that govern the effect of x t ’s new information on Ct given
function. Equation (6) translates the hidden state to an output the previous network output h t . Intermediate variable ot is
o of the next layer by multiplication of the hidden states by used to transform the cell state Ct into the output vector h t that
weighting matrix who exposes the cell memory to the rest of the network. Weighting
matrices W f , W i , Wg , and Wo are the recurrent multiplicative
F H
weights applied to the concatenated vector [h t−1 , x t ]. Finally,
aht = wih x it + wh  h bht−1 (4)
b f , bi , bg , and bo are constant bias weight adjustments before
i=1 h  =1
  applying nonlinear activation functions. The recurrent weights
bht = θh aht (5)
and biases are obtained through network training
H
   
yot = who bht . (6) f t = σ W f · h t−1 , x t + b f
   
h=1 i t = σ Wi · h t−1 , x t + bi
 
Traditional RNNs are susceptible to an issue during training ot = σ (Wo · h t−1 , x t + bo )
known as vanishing gradients. The LSTM RNN variant has    
gt = tanh Wg · h t−1 , x t + bg
a unique structure that effectively addresses the vanishing
gradient of the traditional RNN. Fig. 3 shows the LSTM node Ct = f t ∗ Ct−1 + i t ∗ gt
cell implementation. The sample x t is concatenated with the h t = ot ∗ tanh(Ct ). (7)
previous output vector h t−1 of a previous state. The variable
The result of the rescaling activation functions σ and tanh
Ct−1 contains the hidden cell state memory that is passed
is a network that is able to retain information in the long term
between LSTM cells and allows the gradient loss with respect
over many iterations of gradient propagation without decaying.
to the cell state to pass through to the next iteration without
In addition, this network is robust to input noise in the short
successive multiplication operations applied as in the tradi-
term from the input gating operations applied to the current
tional RNN architecture. The σ (sigmoid) and tanh operations
sample. As an example, in the case of pulsed dc signals, the
isolate the internal cell state during backpropagation training
coil gun pulse starts transient generated in past samples should
to determine what information should be forgotten with new
precede the decay transient of the coil gun pulse stop signal
input, what information is relevant to include from the current
within a subsequent sample.
sample block, and what information should be passed from the
network to the output vector h t . The variable x t is the size of
the input block. Ct is the size of the memory cell count that is C. Autoencoder
selected through experimentation. h t will be sized to the next An AE is a type of neural network designed and trained to
layer’s neural network node count. In this article, h t will be reproduce its input. It has internal bottleneck layers describing
a vector of size 10 if the layer following the LSTM cell is a a code to represent the input into a minimum latent space

Authorized licensed use limited to: Pontificia Universidad Javeriana. Downloaded on October 23,2023 at 03:41:30 UTC from IEEE Xplore. Restrictions apply.
MA et al.: DC FAULT DETECTION AND PULSED LOAD MONITORING 7081

known as the encoder. Then, a decoder maps the code to a

reconstruction of the original input from its minimum latent
space representation.
Since only the essential information is passed through the
bottleneck layer, the network should be robust against noise
perturbations, and the output should be denoised. AE usage
in feature extraction is implemented by stripping the decoding
layers from the architecture and attaching the encoder section
to another network topology. This process is also known as
transfer learning, where portions of one network can be used
in another network. The encoder portion isolates important
signal information to pass to another network structure. The
training algorithm can converge on a solution quickly because Fig. 4. LSTM RNN AE-based architecture.
the encoder portion has its weights already optimally set.
The key usage of the AE architecture in the proposed
monitoring solution is to perform fault detection. The loss neural networks perform a signal reconstruction that can be
used to train the network is the reconstruction error between compared with the input measurement. Finally, the detector
the original signal and the artificial reconstruction. To work as evaluates the reconstruction error and error residuals to flag
a fault detector, an AE can be trained only on valid data sets faults if present.
without faults. Any signal the AE has not been sufficiently An eight-stage input stream of SWT detail coefficients was
trained on should yield large reconstruction errors. A different shown to sufficiently perform classification. It can produce an
input will yield an undeterminable and unpredictable output 8 by N image of detail coefficients on the signal from the
from the trained network. output of the high-pass filters, where N is the length of the
A fault can be flagged based on a threshold determined from sample. Each vertical slice time index of the SWT image is
the reconstruction errors of the network validation data. One fed to the neural network as an 8 by 1 vector, corresponding
way to recover this threshold is an analysis of the receiver to eight stages of detail decomposition.
operating characteristic (ROC) curves. An ROC graphically Both the input and output LSTM networks will consist
represents the diagnostic ability of a binary classifier system of 25 hidden nodes, sized for memory constraints of micro-
(i.e., fault or no-fault determination) as its discrimination controller implementations. The network is trained in two
threshold is varied. To develop an ROC curve, the thresh- stages. In the first stage, the top portion of the LSTM network
old is varied iteratively, and false-positive rates (fprs) and is trained on labeled data sets corresponding to the salient
true-positive rates (tprs) are recovered using the reconstruction features of pulsed signals, such as its initiation, ramp rates,
errors from validation data containing both faults and normal or completion characteristics. This allows for the system to
transients. A detector with a zero fpr and a tpr of one is produce a secondary classification output of the signal, which
considered an ideal binary classifier because it catches all may be beneficial for other load monitoring applications,
faults and flags no normal data sets as abnormal. In real- such as labeled data segmentation or to register events to
world systems, the ideal classifier may not be achievable, and a hypervisor monitoring system. The encoder node size is
the system engineer must develop guidelines for acceptable fixed to the number of waveform segment classification types.
performance that balances fpr and tpr. The second stage of training is unsupervised, requiring the
system to reproduce the original input signal. The decoder
consists of a ten-node fully connected layer and rectifier-linear
D. Proposed Load Monitoring System layer to mimic the reconstruction filters of an inverse SWT
The machine learning-based neural networks are the foun- operation. Overall, the LSTM AE programming for real-time
dation for the proposed load monitoring system. The detail implementation is shown in Fig. 5 that will be applied in
coefficient output from the signal’s SWT is effective for select- experimental validation later.
ing the relevant time and frequency information on transients
as an input into a classification system. The SWT was used for E. Fault Detection Method
the feature extracted input into LSTM RNN-based AE trained The fault detection subsystem in Fig. 4 is made by error
to reproduce the original unfaulted load signals. The wavelet residual calculations shown in Fig. 6 in a similar fashion
transform coefficients will be different in normal operation to calculate the approximation coefficients. It is calculated
than from the faulted condition. by respective small, medium, and large windows performing
The proposed load monitoring system consists of three the same recursive function shown in (8). The accumulated
distinct stages in Fig. 4. The input-processing stage measures window error awe(i ) is computed from the squared dif-
the load current and performs an SWT feature extraction for ference between the LSTM AE signal reconstruction fˆ[i ]
input into the LSTM encoder–decoder neural network in the and the time-delayed original signal to compensate for the
second stage. The encoder performs a dual function to achieve fact the LSTM AE input was delayed during the wavelet
waveform segment classification and to encode the signal into transformation. J is the total number of stages of wavelet
minimum latent space representation. In the last stage, the decomposition. The running summation subtracts off the last

Authorized licensed use limited to: Pontificia Universidad Javeriana. Downloaded on October 23,2023 at 03:41:30 UTC from IEEE Xplore. Restrictions apply.
7082 IEEE JOURNAL OF EMERGING AND SELECTED TOPICS IN POWER ELECTRONICS, VOL. 9, NO. 6, DECEMBER 2021

Fig. 5. LSTM AE programming code architecture.

Fig. 6. Fault detection subsystem for LSTM AE.

error outside window size k in the calculation of the current

awe value just as in the approximation coefficient calculation Fig. 7. Circuit schematic of pulse load hardware prototypes.
of (2). The running summation is almost the averaged window
TABLE I
error; however, the division by the window count is not
H ARDWARE D ESIGN PARAMETERS OF DC P ULSE L OADS
required as the operation is always a constant scaling factor
during the calculation
   2
e[i ] = f i − 2 j −1 − fˆ[i ]
awe[i ] = e[i ] + awe[i − 1] − e[i − k]. (8)

The small window accumulates the error from the previ-

ous 16 samples, the medium window accumulates the error
from the previous 32 samples, and the large window accumu-
lates the error of the previous 128 samples. The small window A. Naval Pulsed DC Loads
can be used to detect short duration glitches or shunt faults,
and the longer windows can be used to detect series arcing The scaled-down MVDC ship power system is shown in
behavior. The threshold settings on the accumulated errors can Fig. 7. Two typical loads, including electromagnetic coil gun
be recovered from the ROC analysis. The presence of a fault (Load 1) and fixed impedance (Load 2), are built up, and
is determined by the logical OR of the three window detectors their potential fault locations and respective monitoring points
and the previous fault detection measurement. The detector are shown up in this circuit schematic. The coil gun load
thresholds are adjusted during testing to tradeoff type 1 (false represents pulsed dc loads, which will test the load monitoring
positive) and type 2 (false negative) classification errors. classification performance. Shunt fault and IGBT fault are
generated in the electromagnetic coil gun, and the arcing
fault is created in the fixed impedance load intentionally.
III. E XPERIMENTAL R ESULTS
They are all used to demonstrate the effectiveness of the dc
A 1.4-GHz 64-b quad-core ARM Cortex-A53 processor fault detection scheme. Design parameters of dc pulse load
is used in load monitoring and fault detection hardware hardware are shown in Table I.
implementation running at 20 kHz (50 μs step). A Texas The electromagnetic coil gun is a typical pulsed dc load
Instruments (TI) DSP TMS320F28335 feeding this processor assembled to emulate an electromagnetic rail gun or solid-state
as analog to digital converters (ADC) input running at 20 kHz laser on naval ship dc power systems. The scaled-down coil
because there are no ADCs available on this ARM processor gun power by 375-V/15-A dc power supply is built up, and
board. This section begins by describing the hardware setup of its potential fault types and locations are shown up in this
naval pulsed dc loads in Section III-A. Then, the experimental circuit schematic Fig. 7. The shunt fault is generated after the
result of coil gun load monitoring classification, shunt fault, full-bridge parallelly connected with the capacitor of the coil
IGBT fault, and series arcing fault detection are described in gun by step change of resistance. The IGBT fault is made by
Sections III-B–III-E. creating a disturbance in the gate driver signal on switch S1,

Authorized licensed use limited to: Pontificia Universidad Javeriana. Downloaded on October 23,2023 at 03:41:30 UTC from IEEE Xplore. Restrictions apply.
MA et al.: DC FAULT DETECTION AND PULSED LOAD MONITORING 7083

Fig. 8. Electromagnetic coil gun load hardware prototype.

resulting in a negative surge in current. The coil gun consists

of a full-bridge charger used for charging the capacitor on the
coil gun. A DSP chip TMS320F28335 is used to control the
full-bridge according to the feedback voltage Vout and inductor
current i L . The capacitor is charged by constant current 10 A
until Vout is equal to 300 V, and then the full-bridge charger
is disconnected. After 300 ms, a trigger signal is provided
to discharge the capacitor through a coil, which creates an
electromagnetic force to launch a small metallic projectile.
The total length of a single-pulse event is 100 ms. The whole
process is repeated several times to create a periodic pulsed
power load current profile. Fig. 8 shows a photograph of the Fig. 9. Normal cycle of current and recursive Haar SWT detail coefficients
coil gun load hardware prototype in the laboratory, and its profile of electromagnetic coil gun.
design parameters are shown in Table I. A normal cycle of
current, and recursive Haar SWT detail coefficients profile of
coil gun are shown in Fig. 9. Shunt fault and IGBT fault were
introduced into the coil gun system to test the fault detection
later.
A fixed impedance load is assembled to emulate ship service
loads on a naval shipboard system, such as heating or lighting
equipment. A typical operating cycle is a step load is turned
ON and OFF periodically. A contactor is controlled by an
Arduino control board to switch the resistor to load ON or
OFF the current. It is connected directly to the dc grid through
a coupled-inductor dc circuit breaker, which allows automatic
and immediate isolation of the load in case of a shunt fault.
A series arc generator for creating arcing faults by movable
contact is used to test the proposed load monitoring scheme
that is able to detect sustained arc faults. Fig. 10 shows the
laboratory setup for fixed load with a series arcing generator. Fig. 10. Fixed impedance load hardware prototype.
Its design parameters are shown in Table I, and the circuit
schematic is as load 2 of Fig. 7. The encoder portion of the fault detector performs a dual
role as a signal-waveform classifier. This information could be
useful in load monitoring for specific events or to separate data
B. Electromagnetic Coilgun Load Monitoring Classification
into collections for offline prognostic and health monitoring
The first stage performed training of the encoder portion at a later time. The classification node size is set to 4 nodes
of the fault detector using a labeled data set, and the second corresponding to the stages including “start,” “ramp,” “end,”
stage performed unsupervised training of the decoder section and “n/a.” The “start” is the coil gun activated to charge from
to reconstruct the original signal sets. zero current until 2 ms, and then the “ramp” starts between

Authorized licensed use limited to: Pontificia Universidad Javeriana. Downloaded on October 23,2023 at 03:41:30 UTC from IEEE Xplore. Restrictions apply.
7084 IEEE JOURNAL OF EMERGING AND SELECTED TOPICS IN POWER ELECTRONICS, VOL. 9, NO. 6, DECEMBER 2021

“start” and “end,” when the capacitor voltage reaches 300 V,

the full-bridge charger is disconnected, and the current drop to
zero in 1 ms during the “end” period. The other interval period
between each pulse is “n/a.” The training data set consisted
of 20 coil gun pulses sampled at 20 kHz.
The testing accuracy performance of the classifier section
was 98.66%. The labeling performance is shown for one
of the test samples in Fig. 11. The top plot represents the
actual per time index waveform label information for the given
current signal. The range of colors represents the different
classification labels. The middle classification plot represents
the predicted classification label of the time index for the
same given current signal. There is no perceptible difference
between the top and middle plots.
An ideal feature vector extraction for time-series classifica-
tion should have its information conveyed to the classifier at
the event time. The requirement for a classifier to correlate
and retain information between analysis points is achieved
by localizing the feature vector in both time and information.
A single-dimension time-domain signal is insufficient to deter-
mine the classification as the desired number of discriminated
classification groups increase. Therefore, the feature vector
size is supposed to be sufficiently large that allows for discrim-
ination between the classification groups in neural networks. Fig. 11. Coil gun LSTM RNN classification performance.
The LSTM RNN was, therefore, trained with and without
the eight-stage SWT operation to demonstrate the importance achieve a zero false positive rate. The training data’s maximum
of the vector size in the feature extraction step prior to a accumulated window error multiplied by a small gain was used
signal input to the LSTM neural networks. Obviously, the start to set the original threshold set points on the validation set.
and end cannot be labeled accurately without SWT feature A typical shunt fault is shown in the middle pulse in Fig. 12.
extraction. The total accuracy of the LSTM RNN with feature Normal pulse behavior can also be seen in first or third pulses.
extraction was 98.66% and 75.23% without. The top plot shows the original pulse signal, the middle
plot shows the reconstructed signal from the coil gun load
monitor, and the bottom plot shows the instantaneous error
C. Shunt Fault Detection
residual between the original signal and its reconstruction.
There are several sources of electrical faults in a complex In normal cases, the LSTM AE can accurately reproduce
power electronic system though; only three are being recreated the original signal with small accumulated window errors and
here. The first common source of the fault in any electrical small instantaneous error residuals. In a fault case, the original
system is a shunt fault creating a large pulse of current. fault signal contains a sharp faulty pulse noise than during
A source of this fault type is insulation breakdown and failure. normal operation, which the neural network monitor cannot
Most protection relays are easily able to detect a large transient reproduce. The error residuals are an order of magnitude
that exceeds a threshold or if the rate of change of current higher than those encountered during normal operation. The
is too high. However, this kind of detection becomes more shunt fault event can then be localized around the largest error
challenging in a system where pulsed events are part of the residuals in this window.
normal operation. Further complications can happen if the
fault impedance is significant, and the transient current is not
too much larger than the pulsed current for normal operation. D. IGBT Fault Detection
In this setup, the shunt fault modeled is created in a more The IGBT fault is made by the noise on communication
controlled manner. An IGBT is used as a switch to add a fault channels between the controller and the gate driver of the
impedance of 15  into the system, as shown in Fig. 7. Instead power electronics devices. Sources of this type of fault can be
of giving a constant signal to the switch, a 5-ms intermittent an unexpected bug in the programming device, a loose physical
signal is used to produce large current demands exhibited in connection between electrical interfaces, or significant noise
shunt fault behavior. from other controlled loads on the line. Most commercial
The fault detector successively identified 100% of all shunt gate drivers have built-in overcurrent, overvoltage, and under-
faults and flagged no false-positive pulses. Fault detection set voltage protection. These features are meant to protect the
points for accumulated window error thresholds are 0.13, 0.24, device from extreme operating conditions and would not
and 0.98, respectively, from small to large windows. They are react if, for example, some disturbance in the communication
obtained by varying the detector thresholds on a validation channel forces the signals to fall to zero at unexpected times.
data set consisting of 20 normal and 20 faulted cases to In this setup, the DSP adds some random noise to the gate

Authorized licensed use limited to: Pontificia Universidad Javeriana. Downloaded on October 23,2023 at 03:41:30 UTC from IEEE Xplore. Restrictions apply.
MA et al.: DC FAULT DETECTION AND PULSED LOAD MONITORING 7085

Fig. 12. Original and reconstructed coil gun signal of shunt fault. Fig. 13. Original and reconstructed coil gun signal of IGBT fault.

signal labeled as S1 and S1, where the actually commanded

signals get intermittently overwritten by a 0 signal. The IGBT
fault can be inserted by fluctuating the top leg of the IGBT
H-bridge, S1, during load capacitor charging for 5 ms. S1 is
never closed when S1 is closed.
The fault detector accuracy achieved 95% of all IGBT
faults. IGBT fault detection set points for accumulated window
error thresholds are the same as shunt fault, and the original
threshold setpoints are based on the same validation set.
A typical IGBT fault is shown in the middle pulse in Fig. 13.
The top plot shows the original pulse signal, the middle plot
shows the reconstructed signal from the coil gun load monitor,
and the bottom plot shows the instantaneous error residual
between the original signal and its reconstruction. In this case,
the coil gun load monitor can accurately reproduce the original
signal with accumulated window errors and instantaneous error
residuals. The IGBT fault event can be localized around the
largest error residuals in the window.

E. Series Arcing Fault Detection

Series arcing fault as a common source of faults in dc
loads can be generated due to loose electrical connections or a Fig. 14. Original and reconstructed fixed load signal of series arcing fault.
mechanical fault creating an air gap in the path of conduction.
Such a fault will not generate a large transient in current; in
fact, the dc current may remain nearly the same or decrease One of the contacts is fixed while the other is connected to the
slightly. Therefore, a coupled-inductor dc breaker or any dc barrel of a micrometer screw gauge and can be moved laterally
breaker looking only at the transients or time value of current away from or toward the fixed contact. The micrometer barrel
will not be able to automatically react to this kind of fault. allows precise movement of the contact, and therefore, the
An arcing device, shown in Fig. 10, is designed for arcing air gap can be controlled and set to the desired value. The
generation. The contacts are made from copper rods machined apparatus is mounted on a high-temperature resistant plastic
into a conical shape. Arcing creates very high temperatures, base. The location of this arcing device within the system is
and therefore, the rods are embedded in a ceramic housing. marked as movable contact in Fig. 7.

Authorized licensed use limited to: Pontificia Universidad Javeriana. Downloaded on October 23,2023 at 03:41:30 UTC from IEEE Xplore. Restrictions apply.
7086 IEEE JOURNAL OF EMERGING AND SELECTED TOPICS IN POWER ELECTRONICS, VOL. 9, NO. 6, DECEMBER 2021

The fault detector accuracy achieved 95% of all arcing [2] L. J. Petersen, “Next generation integrated power system: The backbone
faults. Fault detection set points for accumulated window error of the electric warship(hybrid electric drive: A near term opportunity),”
in Proc. IEEE Int. Electr. Mach. Drives Conf., May 2009, pp. 14–15.
thresholds were 0.02, 0.34, and 0.70, respectively, from small [3] T. Ericsen, “The second electronic revolution (it’s all about Control),”
to large windows. The testing data set for the fixed load IEEE Trans. Ind. Appl., vol. 46, no. 5, pp. 1778–1786, Sep. 2010.
included 10 samples of normal switching between ON and [4] L. J. Petersen, D. J. Hoffman, J. P. Borraccini, and S. B. Swindler,
OFF and 20 samples of the fixed load switching under series “Next-generation power and energy: Maybe not so next generation,”
Nav. Engineers J., vol. 122, no. 4, pp. 59–74, Dec. 2010.
arcing. Fig. 14 shows an exemplar plot of the LSTM AE [5] L. J. Petersen, M. Ziv, D. P. Burns, T. Q. Dinh, and P. E. Malek,
current estimation for series arcing. There are no sharp current “U.S. Navy efforts towards development of future naval weapons and
swings that occur during series arcing; however, the LSTM AE integration into an all electric warship (AEW),” in Proc. IMarEST
Engine Weapon Int. Symp., London, U.K., Sep. 2011, pp. 1–16.
has difficulty reproducing the noisy arcing behavior resulting
[6] J. Petersen and T. S. Ericsen, “A history of silicon carbide (SiC)
in large-reconstruction errors over the sample window. The advancement: Basic research to product applications,” IEEE Transp.
large residual spikes that occur during fixed load switching Electrific. Community, Jun. 2017. [Online]. Available: https://tec.ieee.
are a good sign that the network is not overfitted and has org/newsletter/june-2017/a-history-of-silicon-carbide-sicadvancement-
basic-research-to-product-applications
not predicted when the load should be ON or OFF without [7] M. M. Steurer et al., “Multifunctional megawatt-scale medium voltage
observing the event first. DC test bed based on modular multilevel converter technology,” IEEE
Trans. Transport. Electrific., vol. 2, no. 4, pp. 597–606, Dec. 2016.
IV. C ONCLUSION [8] G. Sulligoi, A. Vicenzutti, and R. Menis, “All-electric ship design:
This article proposed a novel machine learning-based From electrical propulsion to integrated electrical and electronic power
systems,” IEEE Trans. Transport. Electrific., vol. 2, no. 4, pp. 507–521,
algorithm—An SWT-fed LSTM RNN-based AE network to Dec. 2016.
detect dc faults and monitor load conditions applied to naval [9] E. Skjong, R. Volden, E. Rødskar, M. Molinas, T. A. Johansen, and
pulse loads. The proposed solutions identify patterns and J. Cunningham, “Past, present, and future challenges of the marine ves-
extract wavelet features that can be toward analyzing data, sel’s electrical power system,” IEEE Trans. Transport. Electrific., vol. 2,
no. 4, pp. 522–537, Dec. 2016.
rather than specific problems, such as traditional nonintrusive [10] A. Boveri, F. D’Agostino, A. Fidigatti, E. Ragaini, and F. Silvestro,
load monitoring and short time FT-based fault detection [38], “Dynamic modeling of a supply vessel power system for DP3 protection
[39], leading to adaptable and more generic methods and system,” IEEE Trans. Transport. Electrific., vol. 2, no. 4, pp. 570–579,
Dec. 2016.
require less expert knowledge. A drawback of the more
[11] N. Haje Obeid, A. Battiston, T. Boileau, and B. Nahid-Mobarakeh,
traditional fault detection method, such as short-time FT-based “Early intermittent interturn fault detection and localization for a
implementation, is that the analysis window size must be permanent magnet synchronous motor of electrical vehicles using
chosen using expert knowledge of the pulse’s transient ramp wavelet transform,” IEEE Trans. Transport. Electrific., vol. 3, no. 3,
pp. 694–702, Sep. 2017.
rates. The analysis window is fixed at run time and must [12] J. Liang, S. K. K. Ng, G. Kendall, and J. W. M. Cheng, “Load signature
be selected for each new transient pulse to be analyzed. The study—Part I: Basic concept, structure, and methodology,” IEEE Trans.
wavelet transforms innately provides a zooming feature with Power Del., vol. 25, no. 2, pp. 551–560, Apr. 2010.
a dynamically sized window for fast and slow transients with- [13] A. G. Ruzzelli, C. Nicolas, A. Schoofs, and G. M. P. O’Hare, “Real-
time recognition and profiling of appliances through a single electricity
out operator selection. More importantly, the fault detection sensor,” in Proc. 7th Annu. IEEE Commun. Soc. Conf. Sensor, Mesh Ad
performance improves by using an LSTM AE classifier over Hoc Commun. Netw. (SECON), Jun. 2010, pp. 1–9.
traditional methods without specific operator interaction and [14] D. Das, N. K. Singh, and A. K. Sinha, “A comparison of Fourier
transform and wavelet transform methods for detection and classification
expertise to properly tune the algorithm’s specific parameters. of faults on transmission lines,” in Proc. IEEE Power India Conf.,
The novel load monitoring solution presented herein can be Apr. 2006, p. 7.
applied to any load profile that exhibits repetitive transients [15] S. Lin, L. Zhao, F. Li, Q. Liu, D. Li, and Y. Fu, “A nonintrusive load
identification method for residential applications based on quadratic pro-
during normal operation. The frequency-domain features of gramming,” Electr. Power Syst. Res., vol. 133, pp. 241–248, Apr. 2016.
the load current are extracted under wavelet transform for the [16] K. Satpathi, A. Ukil, and J. Pou, “Short-circuit fault management in
network training to set the network weights and biases. Once DC electric ship propulsion system: Protection requirements, review of
the network training is completed, the LSTM RNN-based existing technologies and future research trends,” IEEE Trans. Transport.
Electrific., vol. 4, no. 1, pp. 272–291, Mar. 2018.
AE will produce both signal classification and signal recon- [17] D. K. Alves, R. L. A. Ribeiro, F. B. Costa, and T. O. A. Rocha, “Real-
struction of the pulse load based on wavelet features of time wavelet-based grid impedance estimation method,” IEEE Trans.
input current. Any faults should yield large reconstruction Ind. Electron., vol. 66, no. 10, pp. 8263–8265, Oct. 2019.
errors for protective action. The experimental results of the [18] D. K. Alves, F. B. Costa, R. Lucio de Araujo Ribeiro, C. Martins de
Sousa Neto, and T. de Oliveira Alves Rocha, “Real-time power mea-
load monitoring and fault detection for naval pulse loads are surement using the maximal overlap discrete wavelet-packet transform,”
discussed that demonstrate the effectiveness of this method. IEEE Trans. Ind. Electron., vol. 64, no. 4, pp. 3177–3187, Apr. 2017.
[19] Y. Q. Chen, O. Fink, and G. Sansavini, “Combined fault location
ACKNOWLEDGMENT and classification for power transmission lines fault diagnosis with
The authors wish to thank Lynn Petersen at the Office integrated feature extraction,” IEEE Trans. Ind. Electron., vol. 65, no. 1,
pp. 561–569, Jan. 2018.
of Naval Research for support of this work through ONR [20] T. Ince, S. Kiranyaz, L. Eren, M. Askar, and M. Gabbouj, “Real-time
Grant N00014-18-1-2315. Distribution A. Approved for public motor fault detection by 1-D convolutional neural networks,” IEEE
release, distribution is unlimited. DCN # 43-6790-20. Trans. Ind. Electron., vol. 63, no. 11, pp. 7067–7075, Nov. 2016.
[21] M. Valtierra-Rodriguez, R. de Jesus Romero-Troncoso,
R EFERENCES R. A. Osornio-Rios, and A. Garcia-Perez, “Detection and classification
of single and combined power quality disturbances using neural
[1] T. Ericsen, “Future navy application of wide bandgap power semicon- networks,” IEEE Trans. Ind. Electron., vol. 61, no. 5, pp. 2473–2482,
ductor devices,” Proc. IEEE, vol. 90, no. 6, pp. 1077–1082, Jun. 2002. May 2014.

Authorized licensed use limited to: Pontificia Universidad Javeriana. Downloaded on October 23,2023 at 03:41:30 UTC from IEEE Xplore. Restrictions apply.
MA et al.: DC FAULT DETECTION AND PULSED LOAD MONITORING 7087

[22] Z. Chen, R. Zhao, Q. Zhu, M. K. Masood, Y. C. Soh, and Damian Oslebo (Member, IEEE) received the
K. Mao, “Building occupancy estimation with environmental sensors via degree in naval engineering and the M.S. degree in
CDBLSTM,” IEEE Trans. Ind. Electron., vol. 64, no. 12, pp. 9549–9559, electrical engineering from the Massachusetts Insti-
Dec. 2017. tute of Technology, Cambridge, MA, USA, in 2014,
[23] J. Mazumdar and R. G. Harley, “Recurrent neural networks trained and the Ph.D. degree from the Naval Postgraduate
with backpropagation through time algorithm to estimate nonlinear School, Monterey, CA, USA, in 2020.
load harmonic currents,” IEEE Trans. Ind. Electron., vol. 55, no. 9, He is currently an Engineering Duty Officer
pp. 3484–3491, Sep. 2008. with the United States Navy, Washington, DC,
[24] J. Zhang and C. Zong, “Deep neural networks in machine translation: USA. He is an Assistant Program Officer with the
An overview. (Natural language processing) (report),” IEEE Intell. Syst., Naval Sea Systems Command, Washington Naval
vol. 30, no. 5, pp. 16–25, Sep. 2015. Yard, Washington, DC, USA. His research interests
[25] S. Saadatnejad, M. Oveisi, and M. Hashemi, “LSTM-based ECG classi- include machine learning, computer engineering, and digital signal processing
fication for continuous monitoring on personal wearable devices,” IEEE applications.
J. Biomed. Health Informat., vol. 24, no. 2, pp. 515–523, Feb. 2020.
[26] H. D. Huynh, L. M. Dang, and D. Duong, “A new model for stock
price movements prediction using deep neural network,” in Proc. 8th
Int. Symp. Inf. Commun. Technol. SoICT, 2017, pp. 57–62.
[27] A. Nanduri and L. Sherry, “Anomaly detection in aircraft data using
recurrent neural networks (RNN),” in Proc. Integr. Commun. Navigat.
Surveill. (ICNS), Apr. 2016, pp. 5C2-1–5C2-8.
[28] T. de Bruin, K. Verbert, and R. Babuska, “Railway track circuit fault
diagnosis using recurrent neural networks,” IEEE Trans. Neural Netw.
Learn. Syst., vol. 28, no. 3, pp. 523–533, Mar. 2017.
[29] M. Cheng, Q. Li, J. Lv, W. Liu, and J. Wang, “Multi-scale LSTM model
for BGP anomaly classification,” IEEE Trans. Services Comput., early
access, Apr. 10, 2018, doi: 10.1109/TSC.2018.2824809.
[30] S. Ma, M. Chen, J. Wu, Y. Wang, B. Jia, and Y. Jiang, “High-voltage
circuit breaker fault diagnosis using a hybrid feature transformation Atif Maqsood (Member, IEEE) received the B.S.
approach based on random forest and stacked autoencoder,” IEEE Trans. degree in electrical engineering from the Lahore
Ind. Electron., vol. 66, no. 12, pp. 9777–9788, Dec. 2019. University of Management and Sciences, Lahore,
[31] E. Habler and A. Shabtai, “Using LSTM encoder-decoder algorithm Pakistan, in 2013, the master’s degree from Clemson
for detecting anomalous ADS-B messages,” Comput. Secur., vol. 78, University, Clemson, SC, USA, in 2017, and the
pp. 155–173, Sep. 2018. Ph.D. degree with the University of California at
[32] P. Malhotra et al., “LSTM-based encoder-decoder for multi-sensor Santa Cruz, Santa Cruz, CA, USA, in 2020, with a
anomaly detection,” in Proc. ICML Anomaly Detection Workshop, focus on power electronics.
New York, NY, USA, 2016. [Online]. Available: https://arxiv.org/ He is an Electrical and Control Engineer with
abs/1607.00148 Dynapower Company, LLC, South Burlington, VT,
[33] A. Petrosian and F. Meyer, Wavelets in Signal and Image Analysis, From USA. His research interests include Z-source break-
Theory to Practice. Amsterdam, The Netherlands Kluwer, 2001. ers, motor drives, energy storage solutions, and digital signal processing.
[34] S. Qian, Introduction to Time-Frequency and Wavelet Transforms.
Upper Saddle River, NJ, USA: Prentice-Hall, 2002.
[35] J. Teslade. (2005). DWT Filter Bank. [Online]. Available:
https://commons.wikimedia.org/wiki/File:Wavelets_-_Filter_Bank.png
[36] J. Teslade. (2005). SWT Implementation. [Online]. Available:
https://en.wikipedia.org/wiki/Stationary_wavelet_transform#/media/
FileWavelets_-_SWT_Filter_Bank.png
[37] G. Chevalier. The LSTM Cell. Accessed: May 16, 2018. [Online]. Avail-
able: https://commons.wikimedia.org/wiki/File:The_LSTM_cell.png
[38] S. R. Shaw, S. B. Leeb, L. K. Norford, and R. W. Cox, “Nonintru-
sive load monitoring and diagnostics in power systems,” IEEE Trans.
Instrum. Meas., vol. 57, no. 7, pp. 1445–1454, Jul. 2008.
[39] Q. Xiong et al., “Arc fault detection and localization in photovoltaic
Keith Corzine (Fellow, IEEE) received the B.S.E.E.,
systems using feature distribution maps of parallel capacitor currents,”
M.S.E.E., and Ph.D. degrees from the Missouri
IEEE J. Photovolt., vol. 8, no. 4, pp. 1090–1097, Jul. 2018.
University of Science and Technology, Rolla, MO,
USA, in 1992, 1994, and 1997, respectively.
He taught at the University of Wisconsin–
Yue Ma (Student Member, IEEE) received the Milwaukee, Milwaukee, WI, USA, the Missouri
M.S.E.E. degree from the University of Nottingham, University of Science and Technology, and Clemson
Nottingham, U.K., in 2013. He is currently pursuing University, Clemson, SC, USA. He is currently a
the Ph.D. degree with the University of California at Professor with the University of California at Santa
Santa Cruz, Santa Cruz, CA, USA. Cruz, Santa Cruz, CA, USA. His research interests
His research interests include power electronics, include power electronics, motor drives, naval ship
artificial intelligence applications, and transportation propulsion systems, and electric machinery. He holds four U.S. patents related
electrification. to power electronics. He has authored or coauthored over 60 refereed journal
articles and over 120 refereed international conference papers.
Dr. Corzine was the Chair of the IEEE Saint Louis Section in 2010. He has
been involved in a number of the IEEE conferences.

Authorized licensed use limited to: Pontificia Universidad Javeriana. Downloaded on October 23,2023 at 03:41:30 UTC from IEEE Xplore. Restrictions apply.