IEEE TRANSACTIONS ON POWER ELECTRONICS 1

Reliability Score Benchmarking and Resistive Loss

Profile-based Open-circuit Fault Diagnosis
Approach for Motor Drive System
Ali Arsalan , Behnaz Papari , Laxman Timilsina , Grace Muriithi , Ali Moghassemi , SM Imrat Rahman
, Elutunji Buraimoh , Gokhan Ozkan , Christopher S. Edrington
DISTRIBUTION STATEMENT A. Approved for public release; distribution is unlimited. OPSEC8438

I. A BSTRACT mathematical model of the practical system is obtained by

various observer-based approaches [3]–[5], parity equations
In recent years, data-driven methods have shown promise
[6], and state estimation schemes [7], [8]. Then, the model-
in diagnosing various open circuit fault (OCF) modes in
generated outputs are compared to the actual measured values
inverter drive applications. However, existing studies primarily
of the system for OCF diagnosis [2]. This approach offers
evaluate the reliability of these methods based solely on clas-
a relatively fast diagnosis without any additional hardware.
sification accuracy, neglecting critical real-time factors such
However, model-based schemes are not preferred due to the
as computational delays and data transfer latency associated
high dependence on the accuracy of the system model. This
with the data-driven approach and communication protocols,
leads to problems associated with modeling complexity and
respectively, which can affect real-time operational reliability.
parameter uncertainties in the working environment.
This paper addresses these gaps by proposing a universally ap-
For signal-based schemes that don’t use a system’s math-
plicable reliability score criterion that integrates classification
ematical model, fault indicators are identified by analyzing
accuracy with system timing profiles. Additionally, a model
time or frequency domain signals using various signal pro-
predictive control strategy employing active thermal manage-
cessing techniques such as wavelet analysis and Fast Fourier
ment (ATM) is applied to the drive system, enabling a detailed
Transform (FFT). These techniques, however, can increase the
analysis of the impact of OCF modes on the junction tem-
diagnostic time and computational burden [9] in signal-based
perature of MOSFETs. Moreover, a novel feature extraction
methods. Moreover, these methods also suffer performance
dataset is introduced, leveraging resistive/conduction loss data
degradation due to load fluctuations and unbalanced operating
from the ATM scheme without requiring signal preprocessing.
conditions as well [2], [10]. Signal-based techniques can also
The proposed reliability score quantification and the dataset’s
be further divided into current-based [11], [12] and voltage-
diagnostic potential are validated using various data-driven
based [13], [14] methods based on the signal under test. In this
classifiers. The most reliable classifier, achieving a 99.95%
regard, the current-based schemes have no extra hardware pre-
diagnosis accuracy, is further tested under diverse operating
requisite and operate independently of converter parameters.
conditions using a control hardware-in-the-loop (CHIL) setup
Regarding diagnostic speed, current-based schemes require
on an OPAL-RT testbed and Raspberry Pi.
Index Terms—Reliability score quantification, online fault relatively more time, generally one fundamental period, to
diagnosis, two-level three-phase inverter, ATM-based model pre- ensure higher detection accuracy, leading to slow diagnosis
dictive control (MPC). [15]. However, voltage-based methods exhibit better accuracy
and higher diagnostic speed because the terminal voltage of
power switches expresses the OCF features first compared
II. I NTRODUCTION
to current-based methods [16]. Moreover, voltage-based tech-
HREE-PHASE voltage source inverters (VSI) have exten- niques have complimentary hardware requirements, which, in
T sive applications in the automotive and electrical power
sectors. Within these applications, OCFs in the power switches
return, increases the overall complexity of system topology,
design cost, and sensor-based fault vulnerability points.
of VSI account for up to 38% of total inverter faults [1]. On the other hand, data-driven approaches can be applied to
These faults pose significant challenges to a sustained system time-varying and complex nonlinear systems without relying
operation by causing extreme mechanical vibrations, electro- on the system’s model, pre-tuned thresholds, and load patterns.
magnetic torque distortions, and current fluctuations. This, in The limitation of data scarcity associated with this approach
return, results in excessive thermal stress on healthy switches can be addressed by collecting the system’s data in different
and subsequent failures of other adjacent system components. operating conditions using reliable real-time simulators such as
Therefore, various substantial methods for OCF diagnosis OPAL-RT, Typhoon, and Speedgoat. Moreover, compared to
(detection and localization) are developed to ensure a reliable, the sturdy algorithms of signal-based and model-based meth-
safe, and continuous fault-tolerant operation of VSI. These ods, data-driven schemes exhibit high versatility and superior
methods can be broadly categorized as model-based, signal- generalization performance, making this approach suitable for
based, and data-driven approaches [2]. diagnosing various adjacent fault conditions in VSI.
In model-based fault diagnosis techniques, initially, an explicit Generally, data-driven methods use a data-based model to
IEEE TRANSACTIONS ON POWER ELECTRONICS 2

establish a mapping relationship, similar to system dynamics, the conventional reliability quantification approach can not
between the input features and output labels [17]. Then, based guarantee the real-time diagnosis performance of a data-
on this correlation, the trained data-driven model is employed driven approach. Moreover, for a data-driven algorithm to
to distinguish between different system states. Moreover, a work effectively, the sum of computational and communication
highly consistent, clean, correctly labeled and balanced data delays should always be less than the feature extraction time-
set can enhance the learning curve of a data-driven model. window size. Therefore, a data-driven approach with a simplis-
Therefore, various signal pre-processing techniques, such as tic feature extraction scheme and a data-driven classifier with
FFT, principle component analysis (PCA), wavelet packet less computational burden is always preferred. In this regard,
transform, and FFT+Relief, are used for input data dimen- various previously discussed data-driven schemes that use sig-
sionality reduction and highly correlated feature extraction in nal preprocessing techniques for feature extraction are exposed
[16], [18]–[22]. These methodologies utilize random vector to high computational load. On the other hand, in the proposed
functional link (RVFL), Bayesian networks, multiclass rel- scheme, the inherent segregation property of an ATM-based
evance vector machine, long short-term memory (LSTM), scheme is used as leverage to capture the dynamics of each
and extreme learning machine (ELM)-based RVFL as OCF semiconductor device in a three-phase inverter, which further
diagnosis algorithms for two-level and multilevel inverters. acts as a reliable dataset for OCF diagnosis without the need
Apart from using frequency domain features, various time for any signal preprocessing technique. So, the contributions
domain statistical features as diagnostic signals can also be of this work can be summarized as follows:
used as in [23], [24] with a data-driven classifier for abnor- 1) Existing data-based approaches typically use conven-
malities detection. Moreover, in [25], [26], convolution neural tional classification accuracy as the primary reliability
network (CNN)-based fault diagnosis methods are proposed, metric. However, this reliability score can demean at
where time-domain diagnostic signals are converted into 2-D certain vulnerable instances, such as at high RPMs,
images for OCF feature extraction. So, these feature extraction during the real-time operation of MDS. To address this
schemes can be categorized into frequency, time, and image- limitation, we propose a comprehensive reliability score
based approaches. Furthermore, in [27], various data-driven that incorporates not only classification accuracy but also
schemes have been evaluated for fault diagnosis in power factors like feature extraction time, model computation
electronics applications, showing that deep learning algorithms time, and data transfer delay, ensuring a more robust
(i.e., LSTM, CNN, recurrent neural network, auto-encoder) evaluation in realistic scenarios before online diagnosis.
perform better than machine learning (i.e., support vector 2) The OCF is a single power switch can induce increased
machine, ELM, decision tree) and neural network algorithms thermal stress on the healthy switches, which is rarely
(i.e., multilayer perceptron, backpropagation neural network, discussed in the existing literature. Therefore, this work
radial basis function, and fuzzy neural network). However, this also illustrates the real-time influence of various OCF
proposition can vary based on the application. modes on the junction temperature of power switches.
The diagnostic speed of these data-based schemes relies 3) Combining data processing techniques like FFT with
heavily on feature extraction window size, data-driven model complex data-driven models can improve classification
computational time, and data communication delay. In this accuracy but often results in high computational costs
regard, several past studies used the fixed window size to [9]. To address this, this work introduces a simplified
achieve balanced OCF diagnosis, as in [16], [20], [28], [29], conduction loss profile-based dataset derived from ATM-
which is suitable for a motor drive system (MDS) with a short based MPC as a novel OCF feature extraction method.
speed range. However, for a wide-speed domain, it can lead This approach leverages predefined control layer signals,
to inadequate or excessive diagnosis data, particularly when eliminating additional computational and hardware de-
the fundamental period is significantly smaller or larger than mands. Furthermore, the proposed dataset’s effectiveness
the window size, which may result in misclassifications or in OCF diagnosis is evaluated using a reliability score
slow diagnosis. Therefore, to overcome these shortcomings, a quantification scheme across various data-driven classi-
dynamic feature extraction time window as a function of the fiers.
fundamental period is preferred in this work, where the input In the rest of this paper, section III explains the MDS mod-
diagnostic data transfer rate will be proportional to the varying eling with ATM-based MPC and conduction loss estimation.
fundamental period of MDS. The proposed loss-based feature extraction dataset acquisition
Moreover, in existing data-based approaches, the diagnosis and reliability score formulation are elaborated in section IV.
performance solely considers the predicted output and ground Finally, the thermal impact analysis and OCF diagnosis via
truth for accuracy evaluation during offline validation. Under real-time experimentation are showcased in section V.
this approach, a highly accurate data-driven classifier may
work perfectly during online diagnosis when the sum of the
computational latency of the data-driven model and commu- III. S YSTEM L AYOUT AND FAULT L ABELLING
nication delay is less than the feature extraction time-window The typical structure of the MDS with two-level, three-
size. However, at high RPMs with a fundamental period of a phase VSI is shown in Fig. 1. The control signals are generated
few milliseconds, the delay can exceed the window size, which via an ATM-based MPC, which considers the electrical and
may lead to congestion in the communication framework, thermal attributes of a power converter for cost function
overruns, slow diagnosis, and consistent false alarms. Hence, minimization [30]. A simplified RC network known as a
IEEE TRANSACTIONS ON POWER ELECTRONICS 3

Fig. 1. The Cyber-physical system of MDS based on ATM-based MPC.

. (a) (b)

(a)

. (c) (d)

Fig. 2. Various OCF modes of inverter (a) SS-OCF (b) SDS-OCF (c) CDS-
OCF (d) PDS-OCF

(b)

TABLE I Fig. 3. Conduction losses in phase A during Normal operation

L ABELING VARIOUS O PERATION M ODES OF I NVERTER

Operation Modes Label Operation Modes Label

A. Conduction Loss Estimation
S1, S4 10
Normal Operation 0 S2, S3 11 Each phase of a two-level, three-phase inverter consists of
CDS-OCF
S3, S6 12 two power transistors (Upper and lower) and two antiparallel
S1 1 S4, S5 13 diodes (Upper and lower). The upper switch and lower diode
S2 2 S1, S6 14 can only conduct the positive current into machine winding,
S3 3 S2, S5 15 whereas the negative current can only flow through the upper
SS-OCF
S4 4 S1, S3 16
diode and lower switch, as illustrated in Table II. The power
S5 5 S2, S4 17
S6 6 S3, S5 18
transistor and the diode conduct current consecutively in each
PDS-OCF switching period, as it can be correlated to Fig. 3a, which
S1, S2 7 S4, S6 19
SDS-OCF S3, S4 8 S1, S5 20 shows the conduction losses of the upper power switch and
S5, S6 9 S2, S6 21 lower diode during the positive half-cycle in normal mode.
Moreover, conduction losses are influenced by several factors
besides current values, including the switch brand and its
associated on-state resistance dynamics under varying junction
lumped parameter thermal network (LPTN) is used to model temperatures and drain-to-source current values. This loss
the thermal states of semiconductor devices analytically [30]. pattern almost remains the same for all the power transistors
Moreover, the OCF associated with the inverter’s semiconduc- and the concerted diodes during normal operation. The power
tor devices can mainly occur in single or two-power switches. loss in the diode is higher than in the power transistor due
Therefore, in this study, all possible combinations for single- to the larger weighted voltage drop during the forward bias
switch and double-switch OCF are considered, which can be condition. These weighted values are determined through
summarized as single-switch OCF (SS-OCF), series double- curve fitting applied to the Ids vs. Vds data provided in Fig.
switch OCF (SDS-OCF), cross-double-switch OCF (CDS- 7 of the SiC half-bridge module’s datasheet (Cree/Wolfspeed
OCF), and parallel double-switch OCF (PDS-OCF) as shown CAB530M12BM3, 1200V, 530A) [31]. Similarly, the con-
in Fig. 2 with different assigned labels in Table I. duction losses for the power transistor are estimated using
IEEE TRANSACTIONS ON POWER ELECTRONICS 4

Fig. 4. Normalized On-state Resistance of SiC-MOSFET for different Drain-

to-Source Current and Junction Temperature.

TABLE II
Fig. 5. Conduction loss in MOSFET-1 during normal operation and OCF
C URRENT C ONDUCTION S EQUENCE FOR MOSFET S AND D IODES IN
P HASE -A UNDER VARIOUS OPERATING CONDITIONS
T
j
S1 S2 Current Conduction pattern (Phase-A) Where Ids is drain-to-source current, and Rds(on) is obtained
State State Direction S1 S2 D1 D2 via (1). For (2b), IF is the current flowing through the anti-
ON OFF ia > 0 Y es No No No parallel diode, and VF is the voltage drop across the diode
ON OFF ia < 0 No No Y es No
under forward-bias condition. The total power loss for each
OFF ON ia > 0 No No No Y es
of the semiconductor module per sampling period can be
OFF ON ia < 0 No Y es No No
OFF OFF ia > 0 No No No Y es
calculated by using (2c). Conduction losses vary significantly
OFF OFF ia < 0 No No Y es No in the presence of OCFs in power transistors, resulting in
distinct dynamic behaviors depending on the fault category. In
this context, Fig. 3b illustrates the impact of conduction losses
the relationship between Ids and the normalized on-state on D1 and S2 for a specific OCF case. The results show that
resistance of the MOSFET, as shown in Fig. 2 of the same the OCF influences the switching sequence, with S2 remaining
datasheet. However, since the datasheet provides these values continuously ON for more than half of the negative half-cycle
only for specific junction temperatures, the on-state resistance of phase A. This phenomenon is discussed in detail in the
for intermediate junction temperatures is calculated using the following section.
following approach,
 α Tj −Tj(ref ) B. Conduction Loss Dynamics During OCFs
Tj Tj(ref )
Rds(on) = Rds(on) 1+ (1) Fig. 5 shows the variation in conduction losses of power
100
switches and anti-parallel diodes in phase-A when OCF oc-
T
j(ref )
Here, Tj represents the junction temperature, Rds(on) is curred in S1 . During normal operation, the conduction losses
the on-state resistance at a predefined junction temperature behave consecutively for each power switch and diode, where
(as specified in the datasheet), and α denotes the thermal the diode has a higher loss value than the power switch. The
factor, which is the rate of change of on-state resistance with total conduction losses for the power switch and diode in one
respect to the junction temperature. Using these parameters, cycle remain the same for both MOSFETs of each phase.
Tj
Rds(on) refers to the on-state resistance corresponding to the Whereas, when OCF occurred in S1 , the conduction loss for
required junction temperature. The junction temperature, Tj , is S1 and D2 turns to zero because conduction of the positive
calculated using the LPTN-based Cauer-type thermal network current of phase-A is no longer possible. Moreover, there is
described in [30]. The calculated value of Tj is then used an increase in the conduction losses for D1 and S2 due to the
Tj
to estimate Rds(on) using (1), as illustrated by the dotted associated rise in a negative current. Therefore, after the OCF
lines in Fig. 4. Additionally, the current values for each in S1 , the total losses in one cycle for the upper MOSFET
MOSFET switch and diode under various switching states are higher because the diode conducts with inherently higher
can be determined using the conduction pattern of Phase-A, conduction losses. In addition, as these characteristics are
as shown in Table II. Hence, the conduction losses for the independent of any triggering thresholds and load variations,
MOSFET switch and diode can be formulated as given in they can be used as a reliable feature for fault diagnosis. When
(2a) and (2b), respectively. an SS-OCF occurs in one phase, the impact on frequency,
magnitude, and phase also propagates to the remaining phases
mf t 2 j T
Pcond = Ids ∗ Rds(on) (2a) as well, which can be examined via FFT algorithm. In this
d
Pcond = IF ∗ VF (2b) regard, Fig. 6 shows the propagating impact of OCF in S1 , on
total mf t d
the conduction losses for the remaining MOSFETs, in phases
Pcond = Pcond + Pcond (2c) B and C. In Fig. 6, the highlighted instances where losses
IEEE TRANSACTIONS ON POWER ELECTRONICS 5

Fig. 6. Conduction loss in MOSFET-3, MOSFET-4, MOSFET-5, and

MOSFET-6 during normal operation and under OCF in MOSFET-1.

become zero are because of the MOSFET irregular behavior Fig. 7. Proposed conduction losses based OCF diagnosis algorithm flowchart
of continuous conduction controlled by switching states, also
shown in detail in Fig. 3b. Finally, these distortions in losses
for all six semiconductor modules with distinct dynamics can they require no additional computation beyond the MDS
be fed to train a data-driven algorithm for OCF diagnosis. control layer.
This simplicity becomes particularly important when de-
IV. P ROPOSED M ETHOD FOR OCF D ETECTION AND ploying ML classifiers in power electronics applications,
L OCALIZATION where high computational loads can lead to overruns. An over-
run occurs when the incoming data frequency from the MDS
The proposed data-driven machine learning (ML) approach exceeds the processing speed of the ML network—resulting
for an OCF diagnosis in a three-phase inverter, leveraging in slow diagnosis and potential false alarms. Therefore, it is
conduction losses as a key feature, is illustrated in Fig. 7. This crucial to consider the diagnostic data rate and the execution
approach involves two main stages: offline training of the ML time of ML classifiers during reliability testing and before
network and its online deployment in a target environment for real-time deployment of the model. To address these chal-
real-time OCF diagnosis. lenges, this work introduces a conduction loss-profile-based
Conduction losses serve as a highly effective feature for dataset as a new feature extraction approach and proposes a
OCF detection and localization due to the inherent ability of novel reliability quantification scheme for evaluating various
the ATM-based MPC approach to segregate electrical losses, ML classifiers in OCF diagnosis. These contributions ensure
enabling optimal thermal stress control for each semiconductor robust, real-time performance while minimizing computational
device. The advantages of conduction losses over traditional overhead during deployment.
three-phase current measurements for OCF diagnosis have
been demonstrated in [32]. Unlike methods that rely on
three-phase current combined with computationally intensive A. Feature Extraction and Preprocessing
processes such as FFT for fault frequency extraction and the The database, providing prior knowledge to an ML model
ReliefF algorithm for feature selection and dimensionality about the operational characteristics of each MOSFET in an
reduction [16], conduction losses offer a more streamlined inverter, acts as a foundation to ensure the accuracy and
and efficient alternative. Since conduction loss dynamics are reliability of the associated ML classifier. In this regard, 22
inherently generated within the ATM-based MPC framework, different operation modes of the inverter are considered, as
IEEE TRANSACTIONS ON POWER ELECTRONICS 6

given in Table I. The operational data of MDS should be

collected for each operation mode while considering various
operating conditions associated with the reference speed and
load torque of MDS. In this regard, for offline training,
initially, the conduction losses of each power switch and diode,
estimated via the 3-phase output current of the inverter under
various operating conditions, are obtained via (2a) and (2b).
h i
Loss Loss Loss
Xswitch = Ps(i) , Ps(i+1) , ....., Ps(i+5) , i = 1 (3) Fig. 8. The structure of feedforward MLP neural network used for multi-class
h i classification
Loss Loss Loss
Xdiode = Pd(i) , Pd(i+1) , ....., Pd(i+5) , i=1 (4)
Loss
where Ps(i) Loss
and Pd(i) represent the conduction losses in as documented in the literature [33]. The architectures of the
each power switch and antiparallel diode. These losses remain data-driven classifiers, including feedforward neural network
zero when there is no conduction, whereas, during the con- (FNN) and LSTM, tested for the proposed OCF feature
duction cycle, loss values continuously switch between zero extraction method, are explained as follows.
and non-zero values in each sampling instant, as shown in
Fig. 3. The conduction cycle duration and the amplitude of B. Classification using FNN
conduction losses are influenced by the reference speed and This study uses the most commonly used feedforward
load torque, respectively. Consequently, MDS operating under Multi-Layer Perceptron (MLP) neural network with four layers
real-world conditions can exhibit rapidly changing and highly (one input, one output, and two hidden layers), as shown in
dynamic conduction loss behaviors, leading to unpredictable Fig. 8. The input of an ith neuron in the hidden layer without
scenarios. These intricate inverter dynamics, involving all bias at nth learning instant can be defined as,
twelve semiconductor devices, demand more advanced ML
J
models and larger training datasets. However, this approach X
significantly increases computational costs and adds complex- neti (n) = mij xj (n), i = 1, 2, ...I − 1 (7)
j=1
ity to the control architecture and data preparation process.
Therefore, for OCF diagnosis, adopting a simplified dataset where J is the total number of neurons for the input layer, m is
with a straightforward ML model offers several advantages. It the weighting factor and xj is the input data from jth neuron.
enhances generalization, reduces computational and communi- Moreover, a rectified linear unit (ReLU) activation function is
cation burdens, and improves the model’s adaptability to han- applied to this sum of the first fully connected (FC) hidden
dle unconventional or underrepresented scenarios efficiently. layer as follows,
To address this, instead of considering the conduction loss
yi (n) = neti (n)+ = max(0, neti (n))
value of each semiconductor device at every sampling period,
(
the average conduction loss over a complete fundamental neti (n), if neti (n) > 0
period is estimated. This preprocessing strategy simplifies the yi (n) = i = 1, 2, ...I − 1 (8)
0 otherwise.
dataset while maintaining diagnostic effectiveness, given the
data extraction window size of one fundamental period. The Finally, the sum of all hidden layer neuron outputs is fed
conduction loss transients in (3) and (4) are reformulated using into the output layer with the softMax activation function
this averaging approach, as outlined below. to predict the class labels. The output layer with softmax
activation function transforms the output values from the FC
T (tsc , tdc ) = 2π/ωe (5)
layer into a probability distribution for all output classes, with
 Ptsc Loss the total sum equal to 1.
P Loss
′
 t=0(i)s(i) ;
 ∀ Ps(i) >0
ns
X(i) = P td (6) eyi (n)
c Loss
 t=0 Pd(i) Loss Softmax(yi (n)) = ŷi = Pn (9)
; ∀ Pd(i) >0 (n)

c xi
i=1 e
(i)
nd

i = {1, 2, 3, 4, 5, 6} Where yi (n) represents the output from the FC layer’s ith
neuron at n learning sample, and nc is the total number of
Where T is the estimated fundamental time period, and ωe target classes. Finally, the output from softMax is fed into the
′
is the electrical rotor speed (rad/s). Moreover, X(i) represents categorical cross-entropy cost function to minimize the error
the average values of conduction losses for the conduction between predicted labels and ground truth while obtaining the
period of each power switch and diode illustrated as tsc optimal weights for hidden and output layers.
(i) (i)
and tdc respectively. In addition, ns and nd illustrate the X
total number of non-zero values in the conduction period min J = H(y, ŷ) = − yi · log(ŷi ) (10)
of each power switch and diode, respectively. Furthermore, i

the effectiveness of various ML classifiers for the suggested Where yi and ŷi represent the true (one-hot encoded) and
OCF feature extraction method can be assessed using diverse the predicted probability of class i, respectively. The negative
performance metrics such as accuracy, precision, and recall sign turns the cross-entropy minimization into a maximization
IEEE TRANSACTIONS ON POWER ELECTRONICS 7

to normalize the output predictions and a classification layer

with a categorical cross-entropy cost function, as illustrated
in (9) and (10). Moreover, Tensorflow is used to build this
network with hyper-parameters as follows: batch size = 24,
learning rate = 0.1ms; 1ms; 5ms, hidden size = 16, optimizer
= Adam.

D. Reliability Score Quantification

In most past studies, the applicability of ML models in
Fig. 9. The structure of a basic LSTM cell power electronics applications is primarily centered on the ML
network’s accuracy, precision, and recall. However, incorpo-
rating ML-based networks with power electronics systems in
real-world scenarios may pose constraints on system reliability
arising from the computational burden associated with the
ML model’s complexity or their compatibility with the high
diagnostic data rate from power converters. More specifically,
the MDS, characterized by continuously changing fundamental
Fig. 10. LSTM-based neural network architecture periods, may result in overruns during communication with
a complex ML model, specifically at high RPMs, thereby
causing slow and unreliable diagnosis. Therefore, it is crucial
problem. Backpropagation and gradient descent methods are
to consider the timing profiles of these consecutive systems
used to adjust the weights and biases in FNN based on the
in the context of reliability evaluation before their real-world
cross-entropy error. Finally, Tensorflow is used to build this
implementation. In this regard, system status trigger time (tss
tgr )
network with hyper-parameters as follows: batch size= 24,
and tolerance time (Ttol ) is formulated as follows,
learning rate= 0.1ms; 1ms; 5ms, hidden size = 16, optimizer
d
= Adam. tss ml
tgr = β · (Tf p(r) ) + tcom + tex (12a)
Ttol = δ · (Tf p(r) ) (12b)
C. Classification using LSTM Where tss is the time taken to produce the associated opera-
tgr
An LSTM is the extended version of a recurrent neural tional label of the inverter, mentioned in Table I. Furthermore,
network (RNN), preferred due to its ability to capture the long- Tf p(r) denotes the fundamental time period at the rated speed
term dynamic features and temporal patterns of sequential of MDS. tdcom represents the communication delay during two-
data. Therefore, similar to system dynamics, LSTM establishes way data transmission between MDS and ML-model, and
a definitive correlation among different time series input tml
ex signifies the execution time of ML-model to process a
features and better distinguishes between different system singular set of data for predicting an output class. Moreover,
operational states. Hence, a correlation among time series data the first term of (12a) represents the proportion of Tf p (r)
sets, similar to system dynamics, is established. Moreover, used for data accumulation before it is transmitted to the
LSTM is computationally expensive and has a longer training ML model. The value of the weighting factor β depends on
time due to its complex architecture, as shown in Fig. 9, the design parameters of the proposed approach, where the
where the forget gate, input, and output gate are used to time-window size for feature extraction is the function of
solve the vanishing gradient problem in conventional RNN the fundamental period. Generally, it is always preferred to
and FNN. In addition, LSTM first calculates a candidate have an β as small as possible to have a fast OCF detection
′
cell state c(t) for the given input x(t) and output from the without compromising optimal OCF diagnosis performance.
previous timestamp h(t−1) , at each sampling period by using Nevertheless, some past studies have preferred fixed window
′
c(t) = tanh(mch .h(t−1) +mcx .x(t) ). The forget gate, input gate size to achieve balanced diagnosis time, as in [16]. This
and output gate are calculated as follows [34], approach, however, can lead to inadequate OCF feature data or
a slow diagnosis process at some instances of speed. Anyhow,
f(t) = σ(mfh .h(t−1) + mfx .x(t) ) (11a) for the reliability quantification of the approach with fixed
i(t) = σ(mih .h(t−1) + mix .x(t) ) (11b) window size, the first term of (12a) will be replaced with
o(t) = σ(moh .h(t−1) + mox .x(t) ) (11c) the associated time window value. In addition, Ttol signifies
the critical remaining lifetime before an OCF in a MOSFET
where mfh , mfx , mih , mix , moh , and mox are the weights for inflicts permanent damage to MDS’s adjacent components due
each gate. Then, the new cell state c(t) and hidden state h(t) to accelerated aging and increased thermal stress [35], [36].
′
are computed by using c(t) = f(t) ⊗ c(t−1) + i(t) ⊗ c(t) and This critical time can also calculated by using the degradation
h(t) = o(t) ⊗ tanh(c(t) ), respectively. forecasting approach, which can be a promising area of
For the problem of diagnosing OCFs in an inverter, the research for our future work. However, in this proposed study,
architecture of the LSTM-based multi-class classifier used in Ttol is defined as a function of Tf p(r) and the scaling factor δ,
this study is shown in Fig. 10. The network has a softmax layer which is assumed to be equal to 10. For an MDS operating at a
IEEE TRANSACTIONS ON POWER ELECTRONICS 8

maximum speed of 733.04 rad/s (US06 drive cycle maximum

speed), the fundamental period is 2.14 ms. At this speed, fault
conditions would persist for roughly 10 fundamental periods
before existing diagnostic schemes [9], [37]–[40], with an
average feature extraction window of approximately 25 ms,
detect the OCF. For lower speeds, the fundamental period
will be longer, and hence the fault conditions will persist
for fewer fundamental periods. Therefore, we considered the
maximum speed as it provides the maximum threshold of the
number of cycles we can tolerate under faulty conditions. Fig. 11. Motor drive system with ATM-based MPC control implementation
on OPAL-RT testbed
Hence, we chose δ = 10 as a practical and conservative
assumption, reflecting a reasonable weighting factor for the
tolerance time estimation, based on prior research insights.
Moreover, in terms of reliability quantification, Ttol provides
an upper bound for the detection time, i.e., the lesser the value
of tss
tgr compared to Ttol , the better the reliability. Finally, the
accumulated classification efficiency of the ML model, using
conventional gauging parameters, is given as,
avg
A + Pm + Pwavg + Rm avg
+ Rw avg
+ F 1avg
m + F 1w
avg
ηml =
7
(13)
avg avg
Where A denotes accuracy, while Pm , Rm , F 1avg
m , Pw ,
avg
avg avg . (a)
Rw , F 1w indicate the macro-average values and weighted-
average values of precision, recall and F1 score, respectively.
Moreover, during the training of the ML model, it is ensured
that any individual output class value for precision, recall and
F1 score is not completely zero. Finally, the reliability score of
an ML model based on the timing profile and the conventional
efficiency scoring mechanism is developed as follows,
tss td +tml
    
1 − Ttgrtol + ψ(Tfdmin ) 1 − com min
Tfd
ex
+ ηml
Rs = ψ(Ttol )  
3
(14) . (b)
Where, ψ(Ttol) ,Tfdmin ) ∈ {0, 1} represents a bipolar parameter.
ψ(Ttol ) assumes the value of zero if tss Fig. 12. Current waveform and gating signals variation during different modes
tgr > Ttol and ψ(Tfdmin ) will of operation (a) Normal operation (b) Single switch fault
be zero if (tdcom +tml ex ) > T min
fd ; otherwise, they take the value of
one. Moreover, this proposed OCF diagnosis approach uses a
dynamic window size associated with a fundamental period for explains the implementation steps and experimental results for
OCF feature extraction. Therefore, Tfdmin as the most vulnerable this real-time experiment.
instance is used in (14), which indicates the minimum funda-
mental period (which will be at highest possible RPM) for a A. System Modelling
3-phase inverter output current. This period is considered the The MDS with ATM-based MPC is implemented in MAT-
most vulnerable because it has the highest possible diagnostic LAB/Simulink environment integrated with the real-time digi-
data transfer rate from MDS to the ML model and is highly tal simulator, i.e., Opal-RT OP4510. The model is executed
susceptible to overruns. Hence, evaluating the system under its in RT-LAB with real-time hardware synchronization mode.
most vulnerable conditions is crucial to achieving a practically Moreover, the detailed implementation of MDS in OPAL-RT
feasible reliability assessment. Moreover, for the reliability is shown in Fig. 11, where the power converter and PMSM
quantification for the fixed window size scheme, the Tfdmin value are simulated in the FPGA solver with a sampling time of
will be replaced with the associated feature extraction time up to 210ns, and the controller is implemented in the CPU of
window value. OPAL-RT with a sampling rate of 25µs. The OCFs in different
power switches of the inverter are instigated by disabling the
V. R EAL -T IME E XPERIMENTAL VALIDATION gate drive signals from the console subsystem in the real-time
The performance of the proposed OCF diagnosis approach MDS model, as shown in Fig. 12.
is demonstrated through a real-time CHIL platform setup car- Python-based scripts, such as ML multi-class classifiers, are
ried out on an MDS, shown in Fig. 1. This CHIL experiment implemented on Raspberry Pi 4 model B, which aims to detect
aims to illustrate the viability of the presented framework and localize OCFs in power converter, as described in Section
for real-time applications. Moreover, the subsequent section IV. Moreover, for real-time OCF location visualization, a GUI
IEEE TRANSACTIONS ON POWER ELECTRONICS 9

Fig. 13. CHIL experimental setup for Real-time OCF diagnosis

Fig. 14. Inverter output current dynamics during different OCFs

is also modeled in Raspberry Pi using the Tkinter interface,

illustrating each MOSFET’s normal or OCF status in the . (a) (b)
power converter. The OPAL-RT and Raspberry Pi are linked
via User Datagram Protocol (UDP). The real-time values of Fig. 16. Effect of OCFs on conduction losses for phases A, B, and C (a)
SS-OCF (b) CDS-OCF
conduction losses from the control layer of the MDS model
(i.e., OPAL-RT) are transmitted to Raspberry Pi, where an ML-
based OCF diagnosis algorithm runs to predict output class B. Conduction Loss-based Real-time Data Acquisition
labels based on Table I. Finally, depending upon the predicted
label, the operational status of each MOSFET is represented To validate the performance and reliability of the proposed
in the GUI, and an OCF diagnosis signal is also sent back ML approach for OCF diagnosis under real-time system
to the OPAL-RT for data monitoring on the Oscilloscope. dynamics, training data is gathered for various operating
The detailed layout of the proposed strategy via the CHIL conditions from the OPAL-RT-based MDS model illustrated
experiment setup is shown in Fig. 13. in Fig. 11. In this regard, various categories of single and
Moreover, to efficiently utilize the available computation re- double-switch OCFs are instigated, as illustrated in Fig. 14.
sources, the OCF diagnosis algorithm uses multithreading by Moreover, the reference speed is varied between 20rad/s to
concurrently running three threads (utilizing three out of four 200rad/s with an interval of 20rad/s. This period is considered
cores in Raspberry Pi). The first thread realizes the com- due to substantial change in the fundamental period, as shown
munication protocol to exchange the OCF trigger and input in Fig. 15. Moreover, the data associated with conduction
diagnosis data between connected nodes, while the second losses of each power switch and diode for different operating
thread runs an ML-based OCF diagnosis neural network, modes (mentioned in Table I) is obtained at each incremented
generating the OCF DaL signals to be sent back to OPAL-RT interval of reference speed. In this regard, Fig. 16 and 17
by leveraging the same communication protocol implemented show the real-time plots of conduction losses of phases A,
in the first thread. Moreover, the GUI is deployed in the third B, and C for single, cross, series and parallel-switch OCFs.
threat, which receives each MOSFET’s health status signal During normal operation, the losses of all the diodes are
from the second thread. always higher than those of their associated power switches.
However, this proposition can vary after an OCF is instigated
at the time (toc ). The variations in these loss values are
associated with the impacted switching sequence of power
switches after toc . Each type of OCF has a different impact on
these losses, which can be quantified by the ATM approach
discussed in [30]. Moreover, OCF in a power switch turns
conduction losses to zero for that switch and its associated
diode in the same phase, which is true for all OCF modes.
Hence, this real-time data associated with each semiconductor
device during various operating states of MDS is collected
Fig. 15. Foundational time period variation for a range of reference speed and further used for OCF feature extraction, as explained in
section IV. Furthermore, the MDS can have a wide speed
IEEE TRANSACTIONS ON POWER ELECTRONICS 10

(a)

. (a) (b)

Fig. 17. Effect of OCFs on conduction losses for phases A, B, and C (a) (b)
SDS-OCF (b) PDS-OCF
Fig. 18. Thermal imbalance among MOSFETs and increased thermal stress
due to various OCF scenarios (a) SS-OCF (b) SDS-OCF

profile of up to approx. 733.04rad/s (observed in US06 drive

cycle), with a fundamental period varying between 150ms to
MOSFET’s junction temperature (Tj ) due to single-switch
2.14ms, as shown in Fig. 15, particularly in EV applications.
OCFs in all six power switches is shown in Fig. 18a. During
So, using a fixed time window for feature extraction in a
normal operation, the Tj for each MOSFET remains the same.
frequency-varying system can lead to slow diagnosis and
However, when an SS-OCF is instigated, the Tj does not show
misclassifications due to abnormalities in extracted feature
any substantial variation for MOSFET under fault, whereas the
data. Moreover, computing an optimal value of time-window
other MOSFET in the same phase experiences a decline in Tj .
size for a balanced diagnosis in a wide-speed MDS profile can
This decline is because this Tj is associated with the lower
also be very challenging. Conversely, the fundamental period-
losses of the power transistor. Meanwhile, the higher Tj of
based time window approach is much more simplistic and can
MOSFET under fault is because this Tj is associated with the
achieve fast diagnosis at high RPMs when the system is highly
higher losses of the antiparallel diode. Moreover, It can also
susceptible to physical faults. However, for this approach to
be concluded by Fig. 18a that SS-OCF in the upper switch of
work perfectly, it is crucial to consider the timing profiles of
a phase has a higher impact on the Tj of the lower MOSFETs
the data-based approach for reliability score quantification.
of the other two phases compared to the upper switches and
vice versa. In addition, for MOSFETs under SDS-OCF, the Tj
C. Impact of different OCFs on MOSFET junction tempera- drops to ambient temperature for both MOSFETs in all three
ture cases, as shown in Fig. 18b. Meanwhile, the MOSFETs in the
The current distortion with magnitude and phase variations remaining two phases without fault experience a similar rise in
caused by various OCFs can accelerate the aging and the Tj . Moreover, compared with SS-OCF, SDS-OCF puts higher
thermal stress for some semiconductor devices in an inverter. thermal stress on the Tj of MOSFETs without fault.
This rise in temperature is considered a salient feature in In addition, Fig. 19 explains the impact of various CDS-
measuring the health condition of a MOSFET because most OCFs and PDS-OCFs on Tj of MOSFETs. In the case of
electrical faults ultimately reflect damage in the form of CDS-OCFs, the MOSFET with a power switch under fault
increased thermal stress. Therefore, it is very meaningful to has a higher Tj than the other MOSFET in the same phase,
study the thermal impact of various OCFs on MOSFETs to similar to SS-OCF. These higher Tj are normally associated
model better thermal control schemes. So, different single and with the conduction losses of diodes having a higher area
double-switch OCFs mentioned in Table I are instigated in the under the curve. Moreover, in the case of all double-switch
real-time model of the MDS shown in Fig. 11. The real-time CDS faults, the latter switch, such as S4 in the case of
effect of each category of OCFs on the 3-phase output current OCF-S14, always has a higher impact on Tj , as shown in
can also be seen in Fig. 14. Moreover, the impact on each Fig. 19a. This shows the uneven current distribution among
IEEE TRANSACTIONS ON POWER ELECTRONICS 11

(a)

Fig. 20. Increased thermal stress characterized by heightened thermal cycling

under OCF scenario

of various state-of-the-art ML networks. These networks’

frameworks are then optimized to achieve fast convergence
using hit and trail on various hyperparameters, including
batch normalization, drop-out layers, number of hidden layers
(b) and neurons, batch size, and learning rate. Moreover, the
ML network structures are kept as simplistic as possible to
Fig. 19. Thermal imbalance among MOSFETs and increased thermal stress reduce the computational burden while avoiding overfitting or
due to various OCF scenarios (a) CDS-OCF (b) PDS-OCF
underfitting issues. The architecture of these neural networks
is already explained in Section IV. Moreover, the validity of
MOSFETs due to OCFs, whereas, in normal operation, the the proposed OCF feature extraction approach for various ML
current is always balanced. Moreover, in the case of PDS- classifiers, including FNN, LSTM, random forest (RF), and
OCF, it can be seen in Fig. 19b that the lower MOSFET of decision tree (DT), is illustrated in Table III. In this regard,
the phase without fault has the highest Tj when the upper Fig. 21 signifies the convergence rate of accuracy for FNN and
two switches of the remaining phases are under OCF and vice LSTM neural networks, considering different learning rates
versa. This is because, in the case of PDS-OCF, the controller for one hundred epochs, with FNN showing better accuracy
always tries to avoid the scenario where all top or bottom and rapid convergence. A higher learning rate is preferred
MOSFETs are OFF. Due to PDS-OCF, either the top two or as it can lead to fast training; However, it can also cause
bottom two MOSFETs of two-phase are under OCF; if the convergence issues. Furthermore, comparing the accuracy of
third parallel MOSFET also goes into OFF condition, then all ML classifiers in Table III, FNN has the highest value for
the output of the inverter will be zero, hence maximizing training and testing data sets. Whereas, in the case of training
the MPC cost function. Finally, comparing all four categories time, DT has the least duration with a value of 0.067s.
of OCFs shows that the PDS-OCF propagates the highest
thermal impact on the MOSFET without fault. In addition to E. Real-Time Reliability Analysis
the uneven thermal stress distribution among switches caused
Moreover, various reliability score parameters derived from
by OCFs, Fig. 20 illustrates the elevated thermal stress in
real-time testing are also mentioned in Table III. In this
the form of increased thermal cycling, as highlighted using
regard, tml
ex is obtained by running each trained ML classifier in
a box plot for the MOSFETs in phase-A under a specific
Raspberry Pi for a single input data frame. As a result, FNN
OCF scenario. The figure demonstrates not only the rise in
signifies the least computational burden, whereas tml ex of RF
junction temperature due to the OCF but also the increased
shows the highest execution time of 7.0ms. Furthermore, when
variation/spread in junction temperature after reaching its peak
tml min
ex is compared to Tfd , i.e., 2.14ms, FNN and DT are prone
values, indicating intensified thermal cycling effects. Thus,
to illustrate superior computational performance in terms of
this thermal impact quantification can further be used for the
overruns compared to LSTM and RF. Moreover, during the
degradation forecasting of healthy MOSFETs due to OCFs and
actual implementation of such diagnosis approaches, the delay
lead to critical time estimation mentioned in (12b).
associated with the communication framework must also be
considered. In this regard, Fig. 22 shows the real-time detec-
D. Offline Validation tion of an SS-OCF in S1 based on the proposed OCF feature
The obtained MDS data is split into training and testing extraction approach with communication and computational
data sets with a 70:30 ratio for the training and evaluation latency during data transmission between OPAL-RT and Rasp-
IEEE TRANSACTIONS ON POWER ELECTRONICS 12

TABLE III
P ERFORMANCE C OMPARISON A MONG VARIOUS ML C LASSIFIERS

ML Learning Rate (∆t) Reliability Score Quantification

Characteristics
Classifier 0.1ms 1ms 5ms tml
ex (ms) Rs (%)
Accuracy [99.88, 99.88] [100, 100] [100, 100]
FNN Training Time (s) 39.14 14.6 5.18 0.6 77.31
Epoch 100 40 14
Accuracy [94.01,93.94] [98.59,98.55] [98.59, 98.45]
LSTM Training Time (s) 268.05 268.22 268.82 1.9 60.59
Epoch 100 100 100
Accuracy [100, 100], (No. of Trees = 3)
7.0 56.39
Training Time (s) 0.214, (No. of Trees = 3)
RF
Accuracy [100, 99.92], (No. of Trees = 2)
5.9 57.21
Training Time (s) 0.167, (No. of Trees = 2)
Accuracy [100, 99.83]
DT 0.9 71.98
Training Time (s) 0.067

(a)

(a)

(b)

(b)

Fig. 21. Accuracy curves for two neural networks. (a) FNN (b) LSTM

berry Pi. The communication delay is not always fixed due to (c)
the dynamic nature of the communication network, as shown Fig. 22. Real-time OCF diagnosis for different ML classifiers. (a) 3-phase
in Fig.22b, where the diagnosis delay for four different ML current under OCF-S1 (b) OCF diagnosis time (c) Accuracy dip.
classifiers is compared. The detection sequence based on tml ex
remains the same, with FNN showing the least OCF diagnosis
time of approx. 1.0ms.
These different diagnosis periods are primarily associated
with tmlex , and the bidirectional data transfer delay for the less overall accuracy, slow diagnosis, probable congestion in
UDP framework used in the experimental testbed, which the communication network at high RPMs, and false alarms.
is approx. 0.5ms. This value is obtained via various data- Therefore, it is important to consider the timing profiles during
transfer loopback tests between OPAL-RT and Raspberry reliability quantification as in (14) to scale the practical effec-
Pi. The delay can vary for other communication protocols tiveness of an ML-based approach across MDSs spanning a
such as Transmission Control Protocol (TCP), control area wide speed range. So, considering the temporal analysis of ML
network (CAN) buses, FlexRay communication, and local networks and the communication framework, the reliability
interconnect network (LIN). Moreover, Fig. 22c shows the score for each ML classifier is given in Table III. Thus,
real-time accuracy dip for different classifiers caused by the considering the offline validation results and reliability scores,
misclassifications due to OCF detection latency. The classifier FNN is opted as a reliable ML classifier to further validate the
with the highest latency has the highest dip, thus resulting in performance of the proposed OCF diagnosis scheme.
IEEE TRANSACTIONS ON POWER ELECTRONICS 13

. (a) (b)
Fig. 23. CHIL OCF diagnosis results for an SS-OCF
Fig. 25. CHIL OCF diagnosis results (a) CDS-OCF (b) PDS-OCF

. (a) (b)
(a)
Fig. 24. CHIL OCF diagnosis results (a) SS-OCF (b) SDS-OCF

F. Real-time FNN-based CHIL OCF Diagnosis

In order to further substantiate the practical applicability
of the proposed approach, various case study scenarios from
Table I are implemented via the CHIL experimental setup. The
real-time data acquisition results related to the 3-phase output (b)
current and OCF diagnosis signal are displayed by using an
Fig. 26. CHIL OCF diagnosis for all operational states in Table I. (a) 3-phase
oscilloscope (Tektronix TDS2024C) connected to the analog current (b) OCF diagnosis signal
output port of OPAL-RT. The CH1, CH2, and CH3 correspond
to phase-A, B, and C currents, while CH4 is linked to the fault
diagnosis signal obtained via Raspberry Pi. In this regard, III, indicating the classifier’s sensitivity concerning different
Fig. 23 and Fig. 24a, illustrate two scenarios of SS-OCFs, operating conditions. Hence, the false alarms caused by the
where there is no phase-A positive current component for real-world operating conditions associated with MDS can be
OCF in S1 and no negative current component for OCF in subdued by using a majority vote mechanism (MVM). In
S4. Moreover, t0 represents the time when OCF has been this scheme, the classifier’s output is considered for multiple
instigated, while t1 is the time of OCF detection. The fault instants after the first trigger, rather than the disturbance trigger
diagnosis signal shows the associated label value according for a single instant, reducing the chances of false alarms but
to Table I. Furthermore, Fig. 24b shows the OCF diagnosis also increasing the diagnosis time. So, by adopting the MVM
for SDS-OCF in Phase-A and B. Since both the upper and approach, the accuracy of the proposed OCF diagnosis can be
lower switches of phases A and B are under OCF therefore, further improved.
the current for both phases has become zero. Moreover, the
associated fault label, i.e., 7, is also successfully detected. For G. Comparative Analysis with Prior Studies
the scenarios of CDS-OCF and PDS-OCF in Fig. 25a and
Fig. 25b, respectively, the OCF in the upper switch reduces VI. C ONCLUSION
the positive current of the associated phase to zero, while the This study provides a practically more viable approach for
OCF in the lower switch eliminates the negative current of reliability score benchmarking of data-driven schemes used for
the associated phase. Moreover, the assigned labels to each OCF diagnosis in MDS. Moreover, the conduction loss pro-
category of faults are shown as a Fault diagnosis signal. files associated with each semiconductor device in the power
Finally, to demonstrate the comprehensive effectiveness of the converter are proposed as a new system’s operational feature
proposed OCF diagnosis approach, all OCF categories outlined extraction approach for OCF diagnosis, reflecting the enhanced
in Table I are randomly executed in real-time, as shown applicability of ATM-based schemes. Moreover, these loss
in Fig. 26. The classifier shows misclassifications in certain profiles further estimated the thermal impact of OCFs on
instances, particularly during the transition states. However, each MOSFET. The proposed approach is validated via the
during the steady state, the OCFs are perfectly diagnosed. CHIL experiment as well, showing an accuracy of 99.95% for
The overall accuracy of the suggested FNN-based classifier, 21 different modes of OCFs. Moreover, the reliability score
as illustrated in Fig. 26b, is less than that shown in Table quantification approach considers real-time timing profiles as
IEEE TRANSACTIONS ON POWER ELECTRONICS 14

a crucial parameter of the system for practical implementation, [10] Z. Liu, L. Fang, D. Jiang, and R. Qu, “A machine-learning-based fault
hence making it a more reliable method for data-driven-based diagnosis method with adaptive secondary sampling for multiphase drive
systems,” IEEE Transactions on Power Electronics, vol. 37, no. 8, pp.
OCF diagnosis system design for power converters. 8767–8772, 2022.
The proposed OCF diagnosis approach is independent of any [11] H. Zhao and L. Cheng, “Open-switch fault-diagnostic method for back-
system parameters, and it can be extended for a multi-level to-back converters of a doubly fed wind power generation system,” IEEE
Transactions on Power Electronics, vol. 33, no. 4, pp. 3452–3461, 2017.
inverter due to the inherent segregation property of the ATM [12] B. Gou, X.-L. Ge, Y.-C. Liu, and X.-Y. Feng, “Load-current-based
scheme. In addition, the thermal impact analysis can also be current sensor fault diagnosis and tolerant control scheme for traction
further used for degradation forecasting of healthy switches inverters,” Electronics Letters, vol. 52, no. 20, pp. 1717–1719, 2016.
[13] D. Zhou, H. Qiu, S. Yang, and Y. Tang, “Submodule voltage similarity-
due to OCFs for a more sophisticated critical time estimation based open-circuit fault diagnosis for modular multilevel converters,”
in reliability score quantification, which will be a focus of IEEE Transactions on Power Electronics, vol. 34, no. 8, pp. 8008–8016,
our future work. Moreover, OCF localization will also be 2018.
[14] D. Zhou, S. Yang, and Y. Tang, “A voltage-based open-circuit fault
associated with an adaptive fault-tolerant operation to develop detection and isolation approach for modular multilevel converters with
a resilient control approach for our upcoming research work. model-predictive control,” IEEE Transactions on Power Electronics,
vol. 33, no. 11, pp. 9866–9874, 2018.
[15] S. Khojet El Khil, I. Jlassi, A. J. Marques Cardoso, J. O. Estima, and
ACKNOWLEDGMENTS N. Mrabet-Bellaaj, “Diagnosis of open-switch and current sensor faults
in pmsm drives through stator current analysis,” IEEE Transactions on
This work was supported by Clemson University’s Virtual Industry Applications, vol. 55, no. 6, pp. 5925–5937, 2019.
Prototyping of Autonomy Enabled Ground Systems (VIPR- [16] B. Gou, Y. Xu, Y. Xia, Q. Deng, and X. Ge, “An online data-driven
method for simultaneous diagnosis of igbt and current sensor fault of
GS), a US Army Center of Excellence for modeling and three-phase pwm inverter in induction motor drives,” IEEE Transactions
simulation of ground vehicles, under Cooperative Agreement on Power Electronics, vol. 35, no. 12, pp. 13 281–13 294, 2020.
W56HZV-21-2-0001 with the US Army DEVCOM Ground [17] G. Zhiwei, C. Cecati, S. X. Ding et al., “A survey of fault diagnosis
and fault-tolerant techniques—part ii: Fault diagnosis with knowledge-
Vehicle Systems Center (GVSC). DISTRIBUTION STATE- based and hybrid/active approaches,” IEEE transactions on industrial
MENT A. Approved for public release; distribution is un- electronics, 2015.
limited. OPSEC8438. The authors would like to express their [18] B. Cai, Y. Zhao, H. Liu, and M. Xie, “A data-driven fault diagnosis
methodology in three-phase inverters for pmsm drive systems,” IEEE
gratitude for the valuable resources and facilities provided by Transactions on Power Electronics, vol. 32, no. 7, pp. 5590–5600, 2016.
the Secure Energy and Automation Laboratory (SEAL) and [19] T. Wang, H. Xu, J. Han, E. Elbouchikhi, and M. E. H. Benbouzid,
Real Time Controls and Optimization Laboratory (RTCOOL). “Cascaded h-bridge multilevel inverter system fault diagnosis using a pca
and multiclass relevance vector machine approach,” IEEE Transactions
The authors would like to acknowledge the Warren H. Owen on Power Electronics, vol. 30, no. 12, pp. 7006–7018, 2015.
Distinguished Professorship Endowment for its support of the [20] S. Ye, J. Jiang, J. Li, Y. Liu, Z. Zhou, and C. Liu, “Fault diagnosis and
research effort. tolerance control of five-level nested npp converter using wavelet packet
and lstm,” IEEE Transactions on Power Electronics, vol. 35, no. 2, pp.
1907–1921, 2019.
R EFERENCES [21] Z. Xue, K. Xiahou, M. Li, T. Ji, and Q. Wu, “Diagnosis of multiple
open-circuit switch faults based on long short-term memory network
[1] X. Zhou, J. Sun, H. Li, M. Lu, and F. Zeng, “Pmsm open-phase fault- for dfig-based wind turbine systems,” IEEE Journal of Emerging and
tolerant control strategy based on four-leg inverter,” IEEE Transactions Selected Topics in Power Electronics, vol. 8, no. 3, pp. 2600–2610, 2019.
on Power Electronics, vol. 35, no. 3, pp. 2799–2808, 2019. [22] Y. Xia, Y. Xu, and B. Gou, “A data-driven method for igbt open-circuit
[2] Z. Gao, C. Cecati, and S. X. Ding, “A survey of fault diagnosis and fault diagnosis based on hybrid ensemble learning and sliding-window
fault-tolerant techniques—part i: Fault diagnosis with model-based and classification,” IEEE Transactions on Industrial Informatics, vol. 16,
signal-based approaches,” IEEE Transactions on Industrial Electronics, no. 8, pp. 5223–5233, 2019.
vol. 62, no. 6, pp. 3757–3767, 2015. [23] S. U. Jan, Y.-D. Lee, J. Shin, and I. Koo, “Sensor fault classification
[3] S. Zhuo, A. Gaillard, L. Xu, C. Liu, D. Paire, and F. Gao, “An observer- based on support vector machine and statistical time-domain features,”
based switch open-circuit fault diagnosis of dc–dc converter for fuel cell IEEE Access, vol. 5, pp. 8682–8690, 2017.
application,” IEEE Transactions on Industry Applications, vol. 56, no. 3, [24] A. Arsalan, L. Timilsina, B. Papari, G. Muriithi, G. Ozkan, P. Kumar,
pp. 3159–3167, 2020. and C. S. Edrington, “Cyber attack detection and classification for inte-
[4] N. Wassinger, E. Penovi, R. G. Retegui, and S. Maestri, “Open-circuit grated on-board electric vehicle chargers subject to stochastic charging
fault identification method for interleaved converters based on time- coordination,” Transportation Research Procedia, vol. 70, pp. 44–51,
domain analysis of the state observer residual,” IEEE Transactions on 2023.
Power Electronics, vol. 34, no. 4, pp. 3740–3749, 2018. [25] L. Wen, X. Li, L. Gao, and Y. Zhang, “A new convolutional neural
[5] Y. Yu, Y. Zhao, B. Wang, X. Huang, and D. Xu, “Current sensor fault network-based data-driven fault diagnosis method,” IEEE Transactions
diagnosis and tolerant control for vsi-based induction motor drives,” on Industrial Electronics, vol. 65, no. 7, pp. 5990–5998, 2017.
IEEE Transactions on Power Electronics, vol. 33, no. 5, pp. 4238–4248, [26] S.-H. Kim, D.-Y. Yoo, S.-W. An, Y.-S. Park, J.-W. Lee, and K.-B.
2017. Lee, “Fault detection method using a convolution neural network for
[6] W. Hwang and K. Huh, “Fault detection and estimation for elec- hybrid active neutral-point clamped inverters,” IEEE Access, vol. 8, pp.
tromechanical brake systems using parity space approach,” Journal of 140 632–140 642, 2020.
Dynamic Systems, Measurement, and Control, vol. 137, no. 1, p. 014504, [27] A. Moradzadeh, B. Mohammadi-Ivatloo, K. Pourhossein, and A. Anvari-
2015. Moghaddam, “Data mining applications to fault diagnosis in power
[7] J. Poon, P. Jain, I. C. Konstantakopoulos, C. Spanos, S. K. Panda, electronic systems: A systematic review,” IEEE Transactions on Power
and S. R. Sanders, “Model-based fault detection and identification for Electronics, vol. 37, no. 5, pp. 6026–6050, 2021.
switching power converters,” IEEE Transactions on Power Electronics, [28] Y. Xia and Y. Xu, “A transferrable data-driven method for igbt open-
vol. 32, no. 2, pp. 1419–1430, 2016. circuit fault diagnosis in three-phase inverters,” IEEE Transactions on
[8] D. Xie and X. Ge, “A state estimator-based approach for open-circuit Power Electronics, vol. 36, no. 12, pp. 13 478–13 488, 2021.
fault diagnosis in single-phase cascaded h-bridge rectifiers,” IEEE Trans- [29] T. Shi, Y. He, T. Wang, and B. Li, “Open switch fault diagnosis method
actions on Industry Applications, vol. 55, no. 2, pp. 1608–1618, 2018. for pwm voltage source rectifier based on deep learning approach,” IEEE
[9] Y. Xia and Y. Xu, “A transferrable data-driven method for igbt open- Access, vol. 7, pp. 66 595–66 608, 2019.
circuit fault diagnosis in three-phase inverters,” IEEE Transactions on [30] G. Ozkan, P. H. Hoang, P. R. Badr, C. S. Edrington, and B. Papari,
Power Electronics, vol. 36, no. 12, pp. 13 478–13 488, 2021. “Real-time thermal management for two-level active rectifier with finite
IEEE TRANSACTIONS ON POWER ELECTRONICS 15

control set model predictive control,” International Journal of Electrical

Power & Energy Systems, vol. 131, p. 107057, 2021.
[31] w. CREE, “Cab530m12bm3,” Datasheet, 2021. [Online]. Available:
https://assets.wolfspeed.com/uploads/2021/05/CAB530M12BM3.pdf
[32] A. Arsalan, B. Papari, S. I. Rahman, L. Timilsina, A. Moghassemi,
G. Muriithi, G. Ozkan, C. S. Edrington, and E. Buraimoh, “Machine
learning approach for open circuit fault detection and localization in ev
motor drive systems,” WCX SAE World Congress Experience, Tech.
Rep., 2024.
[33] M. Sokolova and G. Lapalme, “A systematic analysis of performance
measures for classification tasks,” Information processing & manage-
ment, vol. 45, no. 4, pp. 427–437, 2009.
[34] S. Hochreiter and J. Schmidhuber, “Long short-term memory,” Neural
computation, vol. 9, no. 8, pp. 1735–1780, 1997.
[35] S. Dusmez, H. Duran, and B. Akin, “Remaining useful lifetime estima-
tion for thermally stressed power mosfets based on on-state resistance
variation,” IEEE Transactions on Industry Applications, vol. 52, no. 3,
pp. 2554–2563, 2016.
[36] A. Nordelöf, E. Grunditz, S. Lundmark, A.-M. Tillman, M. Alatalo,
and T. Thiringer, “Life cycle assessment of permanent magnet electric
traction motors,” Transportation Research Part D: Transport and Envi-
ronment, vol. 67, pp. 263–274, 2019.
[37] Y. Xia, Y. Xu, and B. Gou, “A data-driven method for igbt open-circuit
fault diagnosis based on hybrid ensemble learning and sliding-window
classification,” IEEE Transactions on Industrial Informatics, vol. 16,
no. 8, pp. 5223–5233, 2020.
[38] B. Gou, Y. Xu, Y. Xia, Q. Deng, and X. Ge, “An online data-driven
method for simultaneous diagnosis of igbt and current sensor fault of
three-phase pwm inverter in induction motor drives,” IEEE Transactions
on Power Electronics, vol. 35, no. 12, pp. 13 281–13 294, 2020.
[39] T. Wang, H. Xu, J. Han, E. Elbouchikhi, and M. E. H. Benbouzid,
“Cascaded h-bridge multilevel inverter system fault diagnosis using a pca
and multiclass relevance vector machine approach,” IEEE Transactions
on Power Electronics, vol. 30, no. 12, pp. 7006–7018, 2015.
[40] B. Gou, Y. Xu, Y. Xia, G. Wilson, and S. Liu, “An intelligent time-
adaptive data-driven method for sensor fault diagnosis in induction motor
drive system,” IEEE Transactions on Industrial Electronics, vol. 66,
no. 12, pp. 9817–9827, 2019.