IEEE Electrification - March 2021 - 32

Partially Blocked Rotor Fault
The behavior of a blocked rotor fault is similar to that of a
short circuit fault because the PMM is being used here,
which means that a high current would very easily damage the rotor's magnet. So, only a friction stalling was
applied to the inertial flywheel to produce faulty signatures during the PMM's operation. The flywheel load was
manually disturbed at various times during the motor's
acceleration, deceleration, and steady revolutions, resulting in abnormal current demands.

Verification
The classification accuracy can be improved by the Fourier transform feature extraction, which is demonstrated in
the following current waveform classification test. The
following are current waveform classifications:
1) Pulsed-power load: The trained LSTM networks with
or without feature extraction are both implemented
to compare the performance improvement. One
hundred data sets are captured: 60 training and
40 validation, respectively. Figure 10(a) shows the
four different stages of pulsedpower load operation, including
pulse start; ramp; end; and N/A;
for not applicable. Compared to
the ground truth, an accuracy of
98.75% can be achieved under
feature-extracted classification,
compared with 70.12% without
feature extraction. All of the
waveforms are identified as N/A
because represents the majority
of this load classification.
2) Propulsion-motor drive load: The
motor drive load is sampled at
100 Hz for slow-transient capture. Figure 10(b) shows the four
different stages of motor drive
load operation, including motor
steady, accelerating, decelerating,
and stalled. Compared to the ground truth, an accuracy of 98.53% can be achieved under feature-extracted
classification, compared with 93.75% without feature
extraction. The performance is much better than the
pulsed-power load without feature extraction because
the stage classification of motor drive load is more
evenly distributed, and there is no short period of
stages that exists, such as " start " and " end " in the
pulsed-power load.

undetermined waveform when a fault happens because
an AE never experiences the faulty feature during the
training process:
1) Shunt fault: The validation set of the coil gun includes
30 samples of three-count normal operation pulse trains,
20 samples of a pulse train containing a shunt fault,
and 20 samples of a pulse train containing a gate fault.
The threshold of RMSE is 0.09, and for MLAE it is 0.03 for
zero FPR. Figure 11(a) shows the fault identification of
shunt fault case, including the original and reconstructed
signals, along with the residuals calculated from them. In
the faulty condition in Figure 11(a), the residual has a
positive surge at 3.64 s, which corresponds to the largest
error residuals because the reconstructed signal is totally
different from the original signal. By contrast, in normal
cases, the residual has no large surges because the reconstructed signal is similar to the original signal. This obvious surge is a sign of an anomaly and is used for
differentiating this fault from normal conditions.
2) IGBT gate fault: The IGBT gate fault-identification setpoints for the RMSE and MLAE are the same as its previous shunt fault and verification
data sets. An example IGBT gate
fault behavior and its signal reconstruction are shown in Figure 11(b).
The fault event occurs at approximately 4.8 s, corresponding to the
largest error residuals. Table 2 presents the composite detector performance against the test data set of
both fault categories. In both cases,
an ROC comprehensive review of
varying RMSEs and MLAEs demonstrated that the MLAE was more efficient in the short-duration gate and
in shunt fault identification. One
fault went undiagnosed, but a review
of the undetected gate fault signal
revealed little to no current disturbance in the original sample because
the event occurred during the initial pulse ramping,
when the initial current ramp rate is low.
3) Series arc fault: The validation set of the fixed impedance load includes 10 samples of on-off normal
switching operation and 20 samples containing arcing
faults. Figure 11(b) depicts the fault identification of an
arcing fault case, including the original and reconstructed signals, along with the residuals calculated
from them. Table 2 lists the composite detector performance compared against the test data set. The RMSE
is 0.01, and the MLAE is 0.30 for zero FPR. This detector
scheme flagged all of the series arcing events. An ROC
review disclosed that the RMSE was more efficient in
arcing fault detection.
4) Partially blocked rotor fault: Friction stalling was
applied to the inertial flywheel to produce faulty

In contrast to shunt
faults, series faults
introduce a high
impedance
breakdown of the
conductor material
reducing the overall
load current.

Fault Identification
The STFT feature in faulty condition is different
from in normal condition. An AE can reconstruct only
those signals similar to normal conditions because
an AE is trained only on the features in normal condition. It will reconstruct abnormal signals with an

I E E E E l e c t r i f i cati o n M agaz ine / MARCH 2021

IEEE Electrification - March 2021

Table of Contents for the Digital Edition of IEEE Electrification - March 2021