This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2016.2608842, IEEE
Transactions on Power Electronics

A Data-Driven Fault Diagnosis Methodology in

Three-Phase Inverters for PMSM Drive Systems
Baoping Cai, Member, IEEE, Yubin Zhao, Hanlin Liu, Student member, IEEE, and Min Xie, Fellow, IEEE

Abstract—Permanent magnet synchronous motor and

I. INTRODUCTION
power electronics-based three-phase inverter are the major
components in modern industrial electric drive system, ELECTRICAL actuators are important components of an
such as electrical actuators in an all-electric subsea all-electric subsea Christmas tree which is used to control the
Christmas tree. Inverters are the weakest components in flow of offshore oil and gas, out of the well. This kind of
drive system, and power switches are the most vulnerable actuator consists of two permanent magnet synchronous motors
components in inverters. Fault detection and diagnosis of (PMSMs), i.e. a “drive” motor and a “clutch” motor, as shown
inverters are extremely necessary for improving drive in Fig. 1 [1].The drive motor forces subsea valve open against a
system reliability. Motivated by solving the uncertainty closing spring. Once the valve is fully open, the drive motor
problem in fault diagnosis of inverters, which is caused by stops and the clutch motor ensures that the subsea valve
various reasons, such as bias and noise of sensors, this remains by means of a friction-based mechanism. The speed for
paper proposes a Bayesian network-based data-driven the PMSMs in electric actuator is slow for the sake of safe
fault diagnosis methodology of three-phase inverters. Two operation during oil production. These actuators with motors
output line-to-line voltages for different fault modes are are ones of the most important components for the all-electric
measured, the signal features are extracted using fast subsea Christmas tree system. The failures of actuators can lead
Fourier transform, the dimensions of samples are reduced to devastating consequences, such as oil spill and subsequent
using principal component analysis, and the faults are ocean pollution [33].
detected and diagnosed using Bayesian networks.
Simulated and experimental data are used to train the fault
diagnosis model, as well as validate the proposed fault
diagnosis methodology.

Index Terms—fault diagnosis, open-circuit, three-phase

inverter, Bayesian networks, PMSM

Fig. 1 Subsea electric actuator with PMSMs.

Manuscript received xxxx, 2016; revised xxxx, 2016; accepted These motors are predominantly fed from sinusoidal pulse
September 9, 2016. Date of current version September 10, 2016. This width modulation (SPWM) three-phase inverters for various
work was supported by Hong Kong Scholars Program (No. speed operation. The inverters use power switches, such as
XJ2014004), National Natural Science Foundation of China (No. insulated-gate bipolar transistors (IGBTs) and
51309240), Specialized Research Fund for the Doctoral Program of metal-oxide-semiconductor field-effect transistors (MOSFETs),
Higher Education (No. 20130133120007), China Postdoctoral Science to control the frequency and shape of desired AC voltage.
Foundation (No. 2015M570624), Applied Basic Research Programs These power switches have the advantages of high efficiency,
of Qingdao (NO.14-2-4-68-jch), Science and Technology Project of
fast switching, and easy control of the gate-signal
Huangdao District (No. 2014-1-48) and Fundamental Research Funds
for the Central Universities (No.14CX02197A). Recommended for communications; however, they become faulty because of
publication by Associate Editor xxxxxxx. aging, overloading, or unpredicted operational conditions, and
Baoping Cai is with College of Mechanical and Electronic are the most vulnerable components in inverters [2]. In this
Engineering, China University of Petroleum, Qingdao, Shandong, context, fault detection and diagnosis of inverters become
China, and Department of Systems Engineering and Engineering extremely necessary.
Management, City University of Hong Kong, Kowloon, Hong Kong, Most of faults in inverters are related to power switches, in
(e-mail: [email protected]; [email protected]). the form of short circuit (SC) and open circuit (OC). A SC fault
Yubin Zhao is with College of Mechanical and Electronic of power switch is usually caused by over-voltage,
Engineering, China University of Petroleum, Qingdao, Shandong,
over-temperature, or improper gate signal so that both switches
China, (e-mail: [email protected]).
Hanlin Liu is with Department of Systems Engineering and in one leg are tuned on. The SC fault will eventually cause a
Engineering Management, City University of Hong Kong, Kowloon, short circuit of source voltage and the flow of high currents. In
Hong Kong, (e-mail: [email protected]). most cases, the SC fault is detected by the standard protection
Min Xie is with Department of Systems Engineering and system, such as fuse and breaker, and the initial SC fault
Engineering Management, City University of Hong Kong, Kowloon, become OC fault. The SC fault is destructive and makes the
Hong Kong, (e-mail: [email protected]).
1

0885-8993 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2016.2608842, IEEE
Transactions on Power Electronics

inverter shut down immediately. Further operation is forbidden However, this is not a major drawback because large amounts
and a repair is required. An OC fault of power switch is of data under both healthy and faulty operating condition can be
generally cause by loss of gate signal or disconnection of wire obtained using reliable simulation tools. Recently, data-driven
because of thermal effects. The OC fault does not make drive fault diagnosis methods of inverters are investigated, and some
system shutdown, but can lead to the current imbalance in both techniques are used for diagnosis. Khomfoi et al. [16] proposed
the faulty and healthy phase, which degrades the performance fault diagnosis and reconfiguration methods for multilevel
of drive system significantly. Different from SC faults, OC inverter using neural network. Output phase voltage of the
faults do not trigger the standard protection system of inverters, inverter are used as diagnostic signals to detect faults and their
and can lead to secondary faults at the other components if no locations. Zidani et al. [18] proposed a fault detection and
fault detection and diagnosis system is available. diagnosis method in a PWM voltage source inverter induction
Most fault diagnosis methods for OC and SC faults use motor drive using fuzzy logic. The technique requires the
output current or voltage signals for diagnostic purposes, measurement of the output inverter currents to detect
namely, current-based methods or voltage-based methods. intermittent loss of firing pulses in the inverter power switches.
Current-based methods are independent of inverter parameters Masrur et al. [19] proposed a machine learning technique for
and they do not require additional sensors [11]. For example, fault diagnostics in induction motor drives using structured
Park et al. [5] proposed a simple fault diagnosis scheme for neural network. Most common types of faults including single
brushless direct-current motor drives. The algorithm was switch open circuit faults, post short-circuits, short circuits and
achieved based on the measured phase current information and the unknown faults can be detected and isolated using the
operating characteristics of the motor, and it could recover the proposed method. Wang et al. [20] proposed a fault diagnosis
control performance quickly by fast fault detection and strategy for cascaded H-bridge multilevel inverter system based
reconfiguration of system topology. Estima et al. [9-10] on the principle component analysis and the multiclass
researched multiple OC fault diagnosis methods in voltage-fed relevance vector machine. Kamel et al. [21] presented fault
pulse width modulation (PWM) motor drives by using the diagnosis and on-line monitoring schemes of OC faults for
reference current errors. However, the current signals are load grid-connected single-phase inverters using adaptive neuro
dependent, and current-based methods have low diagnostic fuzzy inference system algorithm.
accuracy under no-load and light load operations [17]. Bayesian network (BN) is a probabilistic graphical model
Although voltage-based methods require additional voltage that represents a set of random variables and their conditional
sensors to measure the line-to-line voltages, they have some dependencies via a directed acyclic graph, and is considered to
inherent advantages such as higher immunity to false alarms, be one of the most useful models in the field of probabilistic
loads, and noise. For example, Alavi et al. [12] proposed a fault knowledge representation and reasoning since it was
diagnosis method for isolating SC faults in MOSFETs in introduced by Pearl in early 1980s [22]. Recently, BNs are
three-phase voltage source inverters by analyzing the PWM increasingly used in the field of fault detection and diagnosis
switching signals in voltage spaces. Freire et al. [11] proposed a because it can solve the uncertainty problem. Cho et al. [23]
new voltage-based approach—without additional sensors—for proposed a fault detection and isolation method of induction
OC fault diagnosis in closed-loop controlled PWM AC voltage motors using recurrent neural networks and dynamic BNs. The
source converters, by using the information contained in the neural network was used to train the data from the system under
reference voltages available from the control system. normal operating conditions and known faulty conditions, and
Generally, fault diagnosis methods can be categorized into the BNs were employed to produce random residuals. Morgan
model-based methods, signal-based methods, and data-driven et al. [24] proposed a methodology for detecting and
methods [3, 4]. Over the last several years, model-based and diagnosing faults found in heavy-duty diesel engines based
signal-based fault diagnosis of inverters for drive systems are upon spectrometric analysis of lubrication samples, and the
widely researched. An et al. [6] proposed a fast-diagnostic output from the classification algorithm was diagnosed using a
method for OC faults in inverters without sensors by analysis of BN with a specifically designed structure. Cai et al. [25]
switching function model of the inverter under both healthy and proposed a multi-source information fusion based fault
faulty conditions. Subsequently, they proposed a current diagnosis methodology for ground-source heat pump system by
residual vector-based fault diagnosis method in order to get rid using BN method. Namaki-Shoushtari et al. [26] proposed a
of effects of load [7]. Jung et al. [8] proposed an OC fault BN-based data-driven method with consideration of process
diagnosis method in voltage source inverters for PMSM drive knowledge and training data for control performance diagnosis
system by employing model reference adaptive system when some of the abnormality data are sparse or not available
techniques. Zhang et al. [13, 14] proposed a simple method for in historical database.
single switch and double switches OC fault diagnosis method Uncertainty is a big problem in fault diagnosis of inverters.
based on three-phase current distortions in voltage source The uncertainty can be caused by various reasons, such as bias
inverters for vector controlled induction motor drives. Trabelsi and noise of sensors [27]. Motivated by solving these
et al. [15] proposed a fault detection technique for OC faults of uncertainty problem, a BN-based data-driven fault diagnosis
IGBTs in voltage source inverter-fed induction motor drives by methodology in three-phase inverters for PMSM drive systems
analyzing the PWM switching signals and the line-to-line is proposed. Two output line-to-line voltages for different fault
voltage levels during the switching times. modes are measured, the signal features are extracted using fast
Different from model-based and signal-based methods, Fourier transform (FFT), the dimensions of samples are
data-driven fault diagnosis does not require a known model or reduced using principal component analysis (PCA), and the
signal patterns, but requires a large number of historical data. faults are detected and diagnosed using BNs.
2

0885-8993 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2016.2608842, IEEE
Transactions on Power Electronics

The rest of this paper is organized as follows: Section II voltage Uab = 0. If OC fault occurs in switch T1, when the gate
describes a three-phase voltage source inverter and analyze the signals g1 = 1 and g3 = 1, the output current ia > 0 and ib > 0,
fault modes. Sections III proposes a BN-based fault diagnosis switch T3 and diode D2 conduct, the output line-to-line voltage
methodology. Section IV develops a fault diagnosis system and Uab = -Ud.
validates it using simulated and experimental data. Sections V With different combinations of open-circuit power switches,
summarizes the paper. there are many fault modes. When all the six switches are
healthy, the inverter is in normal state, which is defined as a
II. SYSTEM DESCRIPTION AND FAULT ANALYSIS special fault mode. When a single OC fault for one switch
Fig. 2 illustrates a typical three-phase voltage source inverter occurs, there are 6 faults modes. When OC faults for two
for PMSM drive system. It is composed by the parallel switches occur simultaneously, there are 15 fault modes.
connection of three inverter legs. Each leg consists of two Similarly, when OC faults for three switches occur
power switches Ti (IGBTs or MOSFETs) with corresponding simultaneously, there are 20 fault modes. However, it is rare
antiparallel connected diodes Di. These diodes are used to that three or more switches fail simultaneously. The current
provide a negative current path through the switches. The work focuses on single OC and double OC faults. Therefore,
power switches are controlled by corresponding gate signals gi. there are 22 fault modes altogether under both healthy and
When the gate signal is equal to 1 the switch conducts, and faulty operating condition. The simulated and experimental
when the gate signal is equal to 0 the switch does not conduct. It line-to-line voltages under different fault modes are provided in
is noted that the gate signals in one leg must work in a Fig.3 and Fig. 4, respectively. From these figures, the
complementary way with a dead time to prevent a simultaneous experimental line-to-line voltages Uab are noisy than simulated
conducting of power switches and short circuit of DC bus ones, which are caused by disturbance, sensor noise or sensor
voltage. The inverter switching pattern are determined by bias.
modulation strategy, such as PWM, SPWM and space vector 40
pulse width modulation (SVPWM). In the current work, 30
SPWM is used to control the gate signals. SPWM is generated 20
by comparing a voltage reference signal (with amplitude Ar and 10
Uab (V)

frequency fr) and a triangular carrier (with amplitude Ac and 0

frequency fc). A modulation index m is defined as the ratio of Ar -10
and Ac [12]. It determined the inverter output voltage. -20
According to different combinations of switching states of six -30
switches, the inverter generate three different line-to-line -40
voltage levels, i.e. +Ud, 0, and -Ud. 0 0.02 0.04 0.06 0.08 0.1
As mentioned above, SC faults can be detected by the Time (s)
standard protection system, whereas OC faults do not trigger (a)
any protection systems of inverters. The voltage and current 40
measurement is mandatory for the feedback control and 30
condition monitoring of subsea electric actuator; therefore, it 20
might be practical to use voltage-based method for fault 10
Uab (V)

diagnosis. Voltage-based fault diagnosis has some inherent 0

advantages such as higher immunity to false alarms, loads, and -10
noise. This current work focuses on the OC faults, and use -20
line-to-line voltage signals for diagnostic purposes. -30
-40
0 0.02 0.04 0.06 0.08 0.1
Time (s)
(b)
40
30
20
10
Uab (V)

0
-10
-20
Fig. 2. Topology of a typical three-phase inverters. -30
-40
Line-to-line voltage of inverters are determined by the gate 0 0.02 0.04 0.06 0.08 0.1
signals gi and phase current ij (j=a, b, c) [6, 8, 14, 15, 28]. The Time (s)
phase current is conducted by both of power switches and (c)
antiparallel connected diodes. Taking line-to-line voltage Uab Fig. 3. Simulated line-to-line voltage Uab (a) Normal state. (b) OC fault
as an example, for a normal operation state, when the gate occurred in T1. (c) OC fault occurred in T1 and T2.
signals g1 = 1 and g3 = 1, the output current ia > 0 and ib > 0, the
power switches T1 and T3 conduct, the output line-to-line
3

0885-8993 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2016.2608842, IEEE
Transactions on Power Electronics

proportion of them are used for model validation. BNs are used
to diagnose the faults of inverters. BNs consist of structure and
parameter models. The BNs structure is determined based on
the relationship between faults and fault symptoms. The BNs
parameter including prior probabilities of root nodes and
conditional probabilities of leaf nodes is learned using the
simulated and experimental data. The BN-based fault diagnosis
model are validated and finally outputted when fault diagnosis
accuracy meets requirement.

(a)

Fig. 5. Flow chart of the proposed fault diagnosis methodology.

(b)
B. Signal feature extraction using FFT
As shown in Fig. 3 and Fig. 4, the simulated and
experimental line-to-line voltage under different fault modes
are difficult to distinguish by a computation unit as a human
does. Therefore, a signal feature extraction technique is
required. Many techniques can be used for signal
transformation, such as FFT, wavelet transform, and Hilbert
transform. In the current work, FFT is adopted since it has a
good identity feature to classify normal and abnormal features.
Assume that the sampled line-to-line voltage U(n) is
extended N-periodic, that is, U(n)=U(n+N) for all n. A discrete
Fourier transform (DFT) is obtained by decomposing the
sequence of voltage values into components of different
(c) frequencies. An FFT computes the DFT and produces exactly
Fig. 4. Experimental line-to-line voltage Uab (a) Normal state. (b) OC fault the same result as evaluating the DFT definition directly. The
occurred in T1. (c) OC fault occurred in T1 and T2. most important difference is that an FFT is much faster than
DFT. DFT is defined by the formula [29]
III. FAULT DIAGNOSIS METHODOLOGY N 1
F ( k )   U ( n )e 2 ikn / N (1)
A. Proposed fault diagnosis methodology n0

The flow chart of the proposed fault diagnosis methodology where F(k) is the frequency-domain output, and k is a positive
for three-phase inverters is illustrated in Fig. 5. Sampled integer in range from 0 to N-1.
line-to-line voltages from experiments and simulations are used It can also be expressed as follows
to train and test the developed fault diagnosis system. It is N 1
 n n 
difficult to obtain a large volume of experimental voltages F ( k )   U ( n )  cos(2 k  )  i sin(2 k  )  (2)
under both healthy and faulty operating condition. Fortunately, n0  N N 
simulated voltages can be easily obtained using reliable The first sample F(0) of the transformed series is the DC
simulation tools, such as Matlab/Simulink/Simpower. The component, more commonly known as the average of the
signal features of these voltages are extracted using FFT, and sampled line-to-line voltage. The second sample F(1) is the
the dimensions of samples are reduced using PCA. A large fundamental component, and other samples F(k) are the
proportion of these data are used for model training, and a small harmonic components.
4

0885-8993 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2016.2608842, IEEE
Transactions on Power Electronics

It is noted that when OC fault occurs in switches T1 or T2, components (PCs), which are ordered so that the first few retain
the magnitude spectrum of line-to-line voltage Uab after FFT is most of the variation present in all of the original variables [30].
same. The two faults cannot be distinguished using the The transformation is, in fact, an orthogonal rotation in p-space.
magnitude spectrum. Fortunately, the phase spectrum of Uab The fundamental PCA method used in the fault diagnosis of
after FFT is different. In particular, the phase of DC component three-phase inverters are illustrated briefly.
have a difference of π. Therefore, the phase information of DC Consider an n×p data matrix, X, where n is the size of
component are added into the magnitude spectrum [20]. Hence, samples, representing a different repetition of an experiment,
the OC faults of switches T1 or T2 can be distinguished. As and p is size of variables, that is, the dimensionality of original
shown in Fig. 6, the DC component involves the phase data. The transformation is defined by a p×p load matrix, P,
information; therefore, the magnitude is negative when OC which maps X to an n×p principal component score matrix, T,
fault occurs in T1. Similarly, the OC failure of switches T3 and given by
T4, and T5 and T6 also lead to the same magnitude spectrums T  X P (3)
of line-to-line voltage Uab and Uac. Since the phase information The load matrix, P, also known as principal component
of DC component are added into the magnitude spectrums, the coefficients, is the eigenvector of covariance matrix R for data
OC faults of T3 and T4, and T5 or T6 can be distinguished. matrix X.
RE  X  E ( X ) X  E ( X )  T
(4)

Each column of the load matrix contains coefficients for one

principal component. The columns are in order of decreasing
component variance. Because the first few components account
for most of the variation in the original data, a new n×k score
matrix, T’, can be expressed by
T ' X P' (5)
where P’ is the first k columns of P, that is a p×k load matrix.
Therefore, PCA reduces the dimensionality of original data
greatly, from p to k.
The criterion for choosing k is to select a cumulative
percentage of total variation which one desires that the selected
PCs contribute, say 90% or 95%. The required number of PCs
is then the smallest value of k for which this chosen percentage
is exceeded. The cumulative percentage of total variation is
Fig. 6 Harmonic magnitude of simulated line-to-line voltage Uab with OC fault defined by [20]
occurred in T1. k

 i

C. Dimensionality reduction using PCA t (k )  i 1

p
 100% (6)
PCA is a technique that reduce the dimensionality of a data  i
i 1
set consisting of a large number of interrelated variables, while
retaining as much as possible of the variation present in the data where λ is the eigenvalue of covariance matrix R for data matrix
set. This is achieved by transforming a set of p correlated X.
variables to a new set of uncorrelated variables, the principal

30
TY1 TY2 TY3 TY4 TY5 TY6 TY7 TY8 TY9 TY10 TY11
20 TY12 TY13 TY14 TY15 TY16 TY17 TY18 TY19 TY20 TY21 TY22
Harmonic magnitude (V)

10

-10

-20

-30
1 2 3 5 4 6 7 8 9 10
Number of principal components
Fig. 7. Low dimensional set of harmonic magnitude for the 22 fault modes

0885-8993 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2016.2608842, IEEE
Transactions on Power Electronics

For the faults of three-phase inverters, line-to-line voltage They can be obtained by experience and judgment of experts or
Uab can indicate the states of switches T1, T2, T3, and T4, but learning using normal and fault data. For the faults of power
not can indicate the states of switches T5 and T6. Another a switches, parameter learning of BNs is adopted with maximum
line-to-line voltage Uac should be used together to distinguish likelihood estimation (MLE) method [31].
the faults of all power switches. Taking the FFT data of Uab in Assume that M = (S, θ) is the proposed BNs with known
Fig. 6 as an example, there are 750 harmonic magnitudes within structure S and unknown parameters θ, D is a complete data set
the frequency 9000 Hz, that is, the harmonic order is 750. For of cases, and each case d ∊ D is complete case, the probability P
distinguishing all faults, the size of variables should be 1500, (d |M) is called the likelihood of M given d. Assume that the
including the integrated FFT data of Uab and Uac. There are 22 cases in D are independent given the model, then the likelihood
fault modes including normal state (TY1), single OC fault (i.e. of M given D is
from TY2 to TY7) and double OC faults (i.e. from TY8 to
L( M D )   P (d M ) (7)
TY22). Hence, the variable sequence is a 22×1500 data matrix. d D
The dimension of variable is too high to construct the structure
It is more convenient to work with the logarithm of the
of BNs. For example, by transforming the data matrix using
likelihood function, called the log-likelihood, defined as
PCA with ten PCs, the dimensionality of original data are
reduced to 22×10, as shown in Fig. 7. The cumulative LL ( M D )   log 2 P ( d M ) (8)
d D
percentage of total variation is 96.27%.
The goal of parameter learning is to estimate each
D. Fault diagnosis using BNs conditional probability density function corresponding to every
1) Structure of BNs: The structure of BNs is developed by node that maximizes the likelihood function. The likelihood
experts manually or learning using normal and fault data. The function is equal or greater than 0, while the log-likelihood
first method is adopted in the current work because there are not function is equal or less than 0. However, maximizing the
so many power switches in three-phase inverters. The proposed likelihood function is equivalent to maximizing the
BNs consist of two layers, i.e. fault layer and fault symptom log-likelihood function. Thus, the optimal BN model
layer, corresponding to two types of nodes, i.e. fault nodes and parameters are estimated from the sequence of training data by
fault symptom nodes. The fault layer consists of six fault nodes maximizing the log-likelihood-based objective function below
representing six power switches, from T1 to T6. Each node has 
two states, i.e. normal (NM) and OC, indicating normal state   arg max L ( M  D )  arg max LL ( M  D ) (9)
 
and open-circuit fault of the corresponding switch given
The PMSM driven by SPWM inverter are controlled by
observed evidence, respectively. The fault symptom layer
frequency fr and modulation index m; therefore, the 22 fault
consists of n fault symptom nodes representing n harmonic
modes and corresponding harmonic magnitudes after PCA with
magnitudes after PCA. The states of each node are the intervals
different frequency and modulation index are used for
of harmonic magnitude, including positive and negative
parameter learning. For example, the training data with
intervals. The causal relationship between fault nodes and fault
frequency from 30 Hz to 80 Hz with the interval of 2 Hz and
symptom nodes is denoted qualitatively using directed arcs. Fig.
modulation index from 0.6 to 0.9 with the interval of 0.001
8 gives a typical two-layers BNs for fault diagnosis. The fault
could be obtained from simulations and experiments. For the
symptom layer consists of ten harmonic magnitudes after PCA.
sake of simplicity, a large proportion of training data are
Each node have 30 states, indicating intervals of harmonic
generated from simulations, and only a small proportion of
magnitude, such as 0–2, 2–4 and so on. The intervals could be
them are obtained from experiments.
different for different inverters, even for different PCs of an
It is noted that the proportion of different fault nodes affects
inverter system. For the sake of simplicity, the fixed interval of
the prior probabilities of root nodes. However, the prior
2 is adopted in the current work.
probabilities themselves cannot affects the diagnostic
performance for the same components. Therefore, a reasonable
proportion of normal state, single OC fault, and double OC
faults, such as 100:10:1, can be set for training data, indicating
that normal state is much more than OC faults, and single OC
fault is more than double OC faults.
3) Fault Diagnosis: After the BNs are constructed and the
network model parameters are estimated from training data, it
can be used for fault diagnosis by probabilistic inference with
new test data. In the current work, Pearl’s belief propagation
algorithm is used for inference [32]. Belief propagation is a
message-passing algorithm originally developed for exact
inference in polytree networks.
Fig. 8. two-layers BNs for fault diangnosis.
For a variable X with parents U and children Y, the message
from node U to child X is denoted by πX (U) and called the
2) Parameter of BNs: After the network structure is causal support from U to X. The message from node Y to parent
determined, the prior probabilities of root nodes and X is denoted by λY (X) and called the diagnostic support from Y
conditional probabilities of leaf nodes are required to specify. to X. For some evidence E, the joint marginal for the family of

0885-8993 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2016.2608842, IEEE
Transactions on Power Electronics

variable X with parents Ui and children Yi is given by with six power switches is modeled using MATALB
P ( XU , E )   E ( X )  X U   X (U i ) Y ( X ) SimPowerSystems, as shown in Fig. 9. The OC faults of
i j
j (10) switches are simulated by disabling the corresponding gate
Here λe (X) is an evidence indicator where λe (x) = 1 if x is signal. The line-to-line voltages Uab and Uac for the 22 fault
consistent with evidence e and zero otherwise. The causal and modes are measured and used for training BNs, when the
diagnostic messages can be defined as frequency varies from 30 Hz to 80 Hz with the interval of 2 Hz
and the modulation index varies from 0.6 to 0.9 with the
 X (U i )  
XU \{U i }
 E ( X ) X U   X (U k ) Y ( X )
k i j
j (11) interval of 0.001. The switching frequency of SPWM is 1 kHz,
and the sampling frequency for experiment and simulation are
 Y ( X )    E ( X ) X U   X (U i ) Y ( X )
j k (12) both 20 kHz, which can reflect the actual waveform well.
U i k j

A node can send a message to a neighbor only after it has

received messages from all other neighbors. When a node has a
single neighbor, it can immediately send a message to that
neighbor. This includes a leaf node X with a single parent U, for
which
 X (U )   E ( X ) X U (13)
X

It also includes a root node X with a single child Y, for which

 Y ( X )   E ( X ) X (14) Fig. 9. Simulink model of the PMSM drive system.
During fault diagnosis, new data after FFT and PCA are
inputted into fault symptom nodes of BNs, and posterior The experimental setup of the PMSM drive system is given
probability of fault nodes are updated using the belief in Fig. 10. A three-phase voltage source inverter with six power
propagation algorithm above. It is noted that the new data switches MOSFET IRF1010 is built. The DRV8301 is a gate
should map to adjacent state space by multiplying by load driver integrated circuit for three-phase motor drive
matrix with adjacent frequency and modulation index. applications, and used to drive the six MOSFETs. Digital signal
Generally, the larger the posterior probability of fault is, the processor TMS320F28035 is used to generate the SPWM
higher the possibility of occurrence of the corresponding fault signals. The PMSM used is ASMA01L03BAK. Two voltage
is. The fault diagnosis methodology can give a probability of a sensors are installed in the DRV8301 drive card to measure the
fault, but not draw a definite diagnostic result. According to line-to-line voltages Uab and Uac. The OC faults are inserted by
engineering experience, two judgment rules are defined to sticking the gate signal of the faulty power switches to the low
determine the fault diagnosis results as follows. state. It is difficult to obtain a large volume of experimental
Rule 1: A single OC fault of power switch with largest voltages under both healthy and faulty operating condition.
posterior probability will be reported if the largest posterior Therefore, only when the frequency is 60 Hz and the
probability is larger than 0.7, or it is 0.5 larger than the second modulation index is 0.8, the line-to-line voltages Uab and Uac
largest one. for the 22 fault modes are measured and used for training BNs.
Rule 2: A double OC fault of power switches with largest and In actual fault diagnosis, the measured voltages are input to
second largest posterior probabilities will be reported if the two software system of fault diagnosis unit, including FFT-based
posterior probabilities are both larger than 0.7, or they are both signal feature extraction unit, PCA-based dimensionality
0.5 larger than the third largest one. reduction unit, and BN-based fault diagnosis unit. They are all
located in the computer.
IV. DEVELOPMENTS AND VALIDATIONS
A. Simulation and experimental setup
To evaluate the performance of the BN-based data-driven
fault diagnosis method, simulation and experiment of a PMSM
drive system with a three-phase voltage source inverter are
conducted. The FFT-based signal feature extraction,
PCA-based dimensionality reduction, and BN-based fault
diagnosis are all performed using Matlab software, and the
functions tic and toc are used to calculate the time. A DELL
Precision T5600 workstation is used for calculation, because
the calculation time for a same program code is a little different
and around a fixed average value, the average values for them Fig. 10. Experimentcal setup of the PMSM drive system.
are obtained as 0.015, 0.028 and 0.172, respectively. It is noted
that these values do not vary greatly when the operating speed The parameters of the PMSM and the inverter used in
is varies. This is good enough for fault diagnosis of three-phase simulation and experiment are given in Table I. It is noted that
inverters for PMSM drive systems used for subsea electric the frequency of the rated speed for the given PMSM is 400 Hz,
actuator. The model of the three-phase voltage source inverter however, for the purposes of subsea electric actuator control as

0885-8993 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2016.2608842, IEEE
Transactions on Power Electronics

well as fault diagnosis, the frequency for training BNs varies in Fig. 12(a), with the decreasing of noise, the diagnostic
from 30 Hz to 80 Hz. accuracies for all PCA dimensions increase. The diagnostic
TABLE I accuracies vary greatly from 1 dB to 12 dB, and reach stable
MAIN PARAMETERS OF THE PMSM DRIVE SYSTEM
levels when the noise is smaller than 12 dB. In addition, the
Quantity Value diagnostic accuracies are almost the same when the PCA
DC voltage Udc=36 V dimensions vary from 4 to 7, but they decrease continuously
Rated power of PMSM 0.1 kW when the dimensions vary from 7 to 10. As shown in Fig. 12(b),
Rated torque of PMSM 0.318 N·m the relationship between diagnostic accuracy and sensor bias
Rated current of PMSM 4A are parabolic type of curve. When the PCA dimension increases
Rated speed of PMSM 3000 rpm
Maximum speed of PMSM 4000 rpm
from 4 to 10, the diagnostic accuracies decrease slightly.
Torque constant of PMSM 0.0866 N·m/Arms Therefore, the PCA dimension of 4 is the best choice for the OC
Rotor inertia of PMSM 0.1034 kg·m2×10-4 fault diagnosis of inverters. The results show that the PCA can
Stator phase resistance of PMSM 0.76 ohms extract the principle components which represent main fault
Armature inductance of PMSM 1.82 mH principle features even the fault signals are noisy. It can also be
Number of pole pairs of PMSM 8
Resistance of MOSFETs 0.001 ohms seen that the BN-based fault diagnosis methodology has a
Resistance of diodes 0.001 ohms strong toleration for sensor noise and bias.
In Ref. [20], white Gaussian noise with signal-noise ratio 10
B. Results dB was added to the simulation samples of the cascaded
The validation of the fault diagnosis method is conducted H-bridge multilevel inverter system fault diagnosis using a
using the test data from simulation and experiment. For all the PCA and multiclass relevance vector machine approach, the
22 fault modes, the simulation test data are obtained when the average diagnosis accuracy for testing samples 5 groups, 10
frequency is 60 Hz and the modulation index varies from groups, and 20 groups are 97.30%, 98.11%, and 97.97%. As
0.6005 to 0.9005 with the interval of 0.01. Obviously, the test shown in Fig. 12 (a), when the signal-noise ratio is 10 dB, the
data are totally different from the training data. The cumulative diagnosis accuracy is 98.48%, which is better than previous
percentage of total variation accounted for by first n one.
components and diagnostic accuracy for different PCA 1.1
dimension are shown in Fig. 11. It can be seen that with the
increasing of PCA dimension, the total variation increases 1.0
continuously, from 22.86% to 96.27%. However, the
0.9
Diagnostic accuracy

diagnostic accuracy increases almost linearly when the

dimension varies from 1 to 3, and then reaches a stable level of 0.8
almost 100%. That means three PCs are enough for the purpose
of fault diagnosis. It is noted that when the PCA dimensions are 0.7 Dim=10
3 and 4 the diagnostic accuracies are 99.55% and 100%, Dim=9
respectively; hence, the PCA dimensions of 4 and more are 0.6 Dim=8
adopted to develop the fault diagnosis program and investigate Dim=7
the effects of sensor noise and bias on the diagnostic accuracy. 0.5
Dim=4
1.1
0.4
1.0 0 2 4 6 8 10 12 14 16 18 20
0.9 Signal-noise ratio (dB)
(a)
Variance and accuracy

0.8
1.1
0.7 Dim=10
0.6 1.0 Dim=7
Dim=4
0.5 0.9
Diagnostic accuracy

0.4
0.8
0.3 Variance accounted
Diagnositc accuracy
0.2 0.7
0.1
0 2 6 4 8 10 0.6
Dimension
Fig. 11. Variance accounted for by first n components and diagnostic accuracy 0.5
for different PCA dimension.
A faulty sensor may cause diagnostic performance 0.4
degradation. Sensor noise and bias are the main types of faults. -1.6 -1.2 -0.8 -0.4 0 0.4 0.8 1.2 1.6
The white Gaussian noise from 0 to 20 dB and the bias from Sensor bias (V)
-1.6 V to 1.6 V are added to the measured line-to-line voltage. (b)
Fig. 12. Effects of sensor noise and bias on the diagnostic accuracy for different
The effects of sensor noise and bias on the diagnostic accuracy PCA dimensions. (a) Noise and (b) bias.
with different PCA dimension are ploted in Fig. 12. As shown

0885-8993 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2016.2608842, IEEE
Transactions on Power Electronics

For all the 22 fault modes, the experimental test data are
obtained when the frequencies are 55, 65, and 75 Hz and the
modulation indexes are 0.65, 0.75, and 0.85. Therefore, there
are 9 combinations with different frequencies and modulation
indexes, and totally 198 fault modes (this takes lots of time).
These data are completely different from the experimental
training data. The fault diagnosis is conducted when the PCA
dimension is 4 without extra artificial sensor noise and bias.
The results show that 196 faults are diagnosed correctly and 2
faults with double OC are diagnosed incorrectly. Therefore, the
experimental diagnostic accuracy for single OC fault is 100%,
and the total accuracy is 98.99%.
Taking OC fault of switch T1 as an example, we perform the Fig. 14. FFT-based signal feature extraction results for the experimental voltage
fault diagnosis using BN-based data-driven method. The Uab with OC fault occurred in T1.
experimental voltage and current waveforms with both healthy
and faulty conditions for long-term and short-term are given in
Fig. 13. It can be seen that the positive half cycle phase current
4
disappear for the OC fault occurred in switch T1.

Harmonic magnitude (V)

-1

-6

-11

-16
2 1 3 4
Number of principal components
Fig. 15. PCA-based dimensionality reduction result for the experimental
voltage Uab with OC fault occurred in T1.
(a)

100

80
Posterior probability (%)

60

40

20

0
(b) T1 T3 T2 T4 T5 T6
Fig. 13. Experimental line-to-line voltage Uab and phase current Ia with OC Fault nodes
fault occurred in T1. (a) Long-term waveforms (b) Short-term waveforms. Fig. 16. Posterior probabilities of the fault nodes for the experimental OC fault
occurred in T1.
The FFT-based signal feature extraction results of the
line-to-line voltage Uab are shown in oscilloscope directly for V. CONCLUSION
better comparison as given in Fig. 14. The new data are mapped
This paper proposes a BN-based data-driven fault diagnosis
to adjacent state space by multiplying by load matrix with
methodology in three-phase inverters for PMSM drive systems.
adjacent frequency and modulation index. When the PCA
Two output line-to-line voltages for different fault modes are
dimensions is 4, the dimensionality reduction result is shown in
measured, the signal features are extracted using FFT, the
Fig. 15. After inputting these data input BN models, as shown
dimensions of samples are reduced using PCA, and the faults
in Fig. 16, we can see that the posterior probability of fault node
are detected and diagnosed using BNs. The proposed fault
T1 is 100% and posterior probabilities of other fault nodes are 0,
diagnosis methodology has been validated in both simulation
indicating that OC fault occurs in the switch T1.
and experiment, which show that the methodology performs
quite well. The results also indicate that the diagnostic accuracy

0885-8993 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2016.2608842, IEEE
Transactions on Power Electronics

increases almost linearly when the dimension varies from 1 to 3, [16] S. Khomfoi, and L.M. Tolbert, “Fault diagnostic system for a multilevel
inverter using a neural network,” IEEE Trans. Power Electron., vol. 22,
and then reaches a stable level. Therefore, the PCA dimensions no. 3, pp. 1062–1069, May 2007.
of 4 are adopted to develop the fault diagnosis program. The [17] S. Khomfoi, and L.M. Tolbert, “Fault diagnosis and reconfiguration for
effects of sensor noise and bias on the diagnostic accuracy with multilevel inverter drive using AI-based techniques,” IEEE Trans. Power
different PCA dimension are also researched. The results show Electron., vol. 54, no. 6, pp. 2954–2968, Dec. 2007.
[18] F. Zidani, D. Diallo, M.E.H. Benbouzid, and N.S. Rachid , “A fuzzy-based
that the PCA dimension of 4 is the best choice for the OC fault approach for the diagnosis of fault modes in a voltage-fed PWM inverter
diagnosis of inverters. It can also be seen that the BN-based induction motor drive,” IEEE Trans. Ind. Electron.,vol. 55, no. 2, pp. 586–
data-driven fault diagnosis methodology has a better toleration 593, Feb. 2008.
for sensor noise and bias. [19] M.A. Masrur, Z. Chen, and Y. Murphey, “Intelligent diagnosis of open and
short circuit faults in electric drive inverters for real-time applications,”
The proposed fault diagnosis methodology are also suitable IET Power Electron.,vol. 3, no. 2, pp. 279–291, Mar. 2010.
for various power switch-based electric devices, such as [20] T.Z. Wang, H. Xu, and J.G. Han, E. Elbouchikhi, and M.E.H. Benbouzid,
multilevel inverter and converter. Future works can be directed “Cascaded H-bridge multilevel inverter system fault diagnosis using a
at current signal-based fault diagnosis methodology, and the PCA and multiclass relevance vector machine approach,” IEEE Trans.
Power Electron.,vol. 30, no. 12, pp. 7006–7018, Dec. 2015.
toleration of loads can be investigated using BN method. [21] T. Kamel, Y. Biletskiy, and L.C. Chang, “Fault diagnosis and on-line
monitoring for grid-connected single-phase inverters,” Electr. Power Syst.
REFERENCES Res., vol. 126 pp. 68–77, Sep. 2015.
[22] J. Pearl, “Bayesian networks: a model of self-activated memory for
[1] D. Abicht and J.V.D. Akker, “The 2nd generation DC all-electronic subsea
evidential reasoning,” In: Proceedings of the 7th Conference of the
production control system,” In: Proceedings of Offshore Technology
Cognitive Science Society, University of California, Irvine, 1985.
Conference held in Houston, Texas, USA, 2-5 May 2011.
[23] H.C. Cho, J. Knowles, M.S. Fadali, and K.S. Lee, “Fault detection and
[2] Y.T. Song, and B.S. Wang, “Survey on reliability of power electronic
isolation of induction motors using recurrent neural networks and dynamic
systems,” IEEE Trans. Power Electron., vol. 28, no. 1, pp. 591–604, Jan.
Bayesian modeling,” IEEE Trans. Control Syst. Technol., vol. 18, no. 2,
2013.
pp. 430–437, Mar. 2010.
[3] Z.W Gao, C. Cecati and S.X. Ding, “A survey of fault diagnosis and
[24] I. Morgan, H.H. Liu, B. Tormos, and A. Sala, “Detection and diagnosis of
fault-tolerant techniques—Part I: fault diagnosis with model-based and
incipient faults in heavy-duty diesel engines,” IEEE Trans. Ind. Electron.,
signal-based approaches,” IEEE Trans. Ind. Electron., vol. 62, no. 6, pp.
vol. 57, no. 10, pp. 3522–3532, Oct. 2010.
3757–3767, Jun. 2015.
[25] B. Cai, Y. Liu, Q. Fan, Y. Zhang, Z Liu, S. Yu, and R. Ji, “Multi-source
[4] Z.W Gao, C. Cecati and S.X. Ding, “A survey of fault diagnosis and
information fusion based fault diagnosis of ground-source heat pump
fault-tolerant techniques—Part II: fault diagnosis with knowledge-based
using Bayesian network,” Appl. Energy, vol. 114, pp. 1–9, Feb. 2014..
and hybrid/active approaches,” IEEE Trans. Ind. Electron., vol. 62, no. 6,
[26] N.S. Omid and B. Huang, “Bayesian control loop diagnosis by combining
pp. 3768–3774, Jun. 2015.
historical data and process knowledge of fault signatures,” IEEE Trans.
[5] B.G. Park, K.J. Lee, R.Y. Kim, T.S. Kim, J.S. Ryu and D.S. Hyun,
Ind. Electron., vol. 62, no. 6, pp. 3696–3704, Jun. 2015.
“Simple fault diagnosis based on operating characteristic of brushless
[27] N. Mehranbod, M. Soroush, and C. Panjapornpon, “A method of sensor
direct-current motor drives,” IEEE Trans. Ind. Electron., vol. 58, no. 5, pp.
fault detection and identification,” J. Process Contr., vol. 15, no. 3, pp.
1586–1593, May. 2011.
321–339, Apr. 2005.
[6] Q.T. An, L.Z. Sun, K. Zhao, and L. Sun, “Switching function model-based
[28] U.M. Choi, F. Blaabjerg, and K.B. Lee, “Study and handling methods of
fast-diagnostic method of open-switch faults in inverters without sensors,”
power IGBT module failures in power electronic converter systems,”
IEEE Trans. Power Electron., vol. 26, no. 1, pp. 119–126, Jan. 2011.
IEEE Trans. Power Electron., vol. 30, no. 5, pp. 2517–2533, May 2015.
[7] Q.T. An, L. Sun, and L.Z. Sun, “Current residual vector-based
[29] J.G. Nash, “Computationally efficient systolic architecture for computing
open-wwitch fault diagnosis of inverters in PMSM drive systems,” IEEE
the discrete fourier transform,” IEEE Trans. Signal Process., vol. 53, no.
Trans. Power Electron., vol. 30, no. 5, pp. 2814–2827, May. 2015.
12, pp. 4640–4651, Dec. 2005.
[8] S.M. Jung, J.S. Park, H.W. Kim, K.Y. Cho, and M.J. Youn, “An
[30] I.T. Jolliffe, “Principal component analysis,” Springer, second edition,
MRAS-based diagnosis of open-circuit fault in PWM voltage-source
2002.
inverters for PM synchronous motor drive systems,” IEEE Trans. Power
[31] F.V. Jensen and T.D. Nielsen, “Bayesian networks and decision graphs,”
Electron., vol. 28, no. 5, pp. 2514–2526, May. 2013.
Springer, second edition, 2007.
[9] J.O. Estima and A.J.M. Cardoso, “A new algorithm for real-time multiple
[32] A. Darwiche, “Modeling and reasoning with Bayesian networks,”
open-circuit fault diagnosis in voltage-fed PWM motor drives by the
Cambridge University Press, 2009.
reference current errors,” IEEE Trans. Ind. Electron., vol. 60, no. 8, pp.
[33] A.K. Aadland, and K. Petersen, “Subsea all electric,” In: Proceedings of
3496–3505, Aug. 2013.
Offshore Technology Conference held in Houston, Texas, USA, 3-6 May
[10] N.M.A. Freire, J.O. Estima, and A.J.M. Cardoso, “Open-circuit fault
2010.
diagnosis in PMSG drives for wind turbine applications,” IEEE Trans.
Ind. Electron., vol. 60, no. 9, pp. 3957–3967, Sept. 2013.
[11] N.M.A. Freire, J.O. Estima, and A.J.M. Cardoso, “A voltage-based Baoping Cai (M’15) received B.S.
approach without extra hardware for open-circuit fault diagnosis in degree in Mechanical Design,
closed-loop PWM AC regenerative drives,” IEEE Trans. Ind. Electron.,
vol. 61, no. 9, pp. 4960–4970, Sept. 2014.
Manufacturing and Automation from
[12] M. Alavi, D.W. Wang, and M. Luo, “Short-circuit fault diagnosis for China University of Petroleum in
three-phase inverters based on voltage-space patterns,” IEEE Trans. Ind. 2006. He received his Ph.D. degree
Electron., vol. 61, no. 10, pp. 5558–5569, Oct. 2014. in Mechanical and Electronic
[13] J.H. Zhang, J. Zhao, D.H. Zhou, and C.G. Huang, “High-performance
fault diagnosis in PWM voltage-source inverters for vector-controlled
Engineering from China University
induction motor drives,” IEEE Trans. Power Electron., vol. 29, no. 11, pp. of Petroleum in 2012. He is a lecturer
6087–6099, Nov. 2014. in the same university from 2013 to
[14] F. Wu and J. Zhao, “A real-time multiple open-circuit fault diagnosis 2014. Currently, he is an associate
method in voltage-source-inverter fed vector controlled drives,” IEEE
Trans. Power Electron., DOI10.1109/TPEL.2015.2422131.
professor in China University of
[15] M. Trabelsi, M. Boussak, and M. Gossa, “PWM-switching pattern-based Petroleum, as well as a “Hong Kong Scholars” postdoctoral
diagnosis scheme for single and multiple open-switch damages in VSI-fed fellow in City University of Hong Kong.
induction motor drives,” ISA Trans., vol. 51, no. 2, pp. 333–344, Mar.
2012.

10

0885-8993 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2016.2608842, IEEE
Transactions on Power Electronics

His research interests include reliability engineering, fault

diagnosis, risk analysis, and Bayesian Networks methodology
and application.

Yubin Zhao received B.S. degree in

Mechanical Design, Manufacturing
and Automation from China
University of Petroleum in 2014. He is
current pursuing the M.S. degree with
the College of Mechanical and
Electronic Engineering, China
University of Petroleum.
His current research interests
include automation control and fault
diagnosis of offshore oil subsea
production systems.

Hanlin Liu (S’14) received the B.S.

degree in applied mathematics from
the University of Science and
Technology of China, Hefei, China, in
2014. He is current pursuing the Ph.D.
degree with the Department of
Systems Engineering and Engineering
Management, City University of Hong
Kong.
His current research interests
include system reliability modeling
and related optimization problems.

Min Xie (M’91–SM’94–F’06)

received the Ph.D. degree in quality
technology from Linkoping
University, Linkoping, Sweden, in
1987. Currently, he is a Chair
Professor at City University of Hong
Kong. He has authored or coauthored
numerous refereed journal papers,
and some books on quality and
reliability engineering, including
Software Reliability Modeling
(World Scientific), Weibull Models (Wiley), Computing
Systems Reliability (Kluwer), and Advanced QFD
Applications (ASQ Quality Press).
Prof. Xie was awarded the prestigious LKY Research
Fellowship in 1991. He is an Editor of the International Journal
of Reliability, Quality and Safety Engineering, Department
Editor of IIE Transactions, Associate Editor of the IEEE
Transactions on Reliability, and Reliability Engineering &
System Safety, and is on the editorial board of a number other
international journals.

11

0885-8993 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.