See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/305670375

Improving Fault Isolation in DC/DC Converters Based with Bayesian Belief

Networks

Article  in  IFAC-PapersOnLine · December 2016

DOI: 10.1016/j.ifacol.2016.07.130

CITATION READS

1 551

5 authors, including:

Abbass Zein Eddine Iyad Zaarour

University of Bordeaux Lebanese University
10 PUBLICATIONS   14 CITATIONS    22 PUBLICATIONS   90 CITATIONS   

SEE PROFILE SEE PROFILE

Abbas Hijazi
Lebanese University
44 PUBLICATIONS   257 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

wireless sensor network View project

Fault detection and isolation in DC-DC converters based on Bayesian Belief Network (BBN) View project

All content following this page was uploaded by Abbass Zein Eddine on 02 August 2018.

The user has requested enhancement of the downloaded file.

Preprints of the 4th IFAC International Conference on FrS1T1.1
Intelligent Control and Automation Sciences, Reims,
France, June 1-3, 2016

Improving Fault Isolation in DC/DC Converters Based with Bayesian Belief

Networks
Abbass Zein Eddine**, Iyad Zaarour*, Francois Guerin**, Abbas Hijazi*, Dimitri Lefebvre**

*Lebanese University,Hadat, Beiruth 1003, Lebanon (abhijaz@ul.edu.lb).

** GREAH – Université Le Havre, 25 rue Philippe Lebon, 76058 Le Havre Cedex, France
(e-mail: {francois.guerin;dimitri.lefebvre}@univ-lehavre.fr and abbass.zein-eddine@etu.univ-lehavre.fr )

Abstract: This paper lies in the domain of Fault Detection and Isolation (FDI). A Bayesian Naïve
Classifier (BNC) structure is selected and used as a first attempt to use Bayesian Belief Networks (BBNs)
for DC/DC power converter FDI. In order to highlight the BNC capabilities, it is compared to the well
known and used FDI method based on Proportional Observer (PO). This comparative study is based on
real data collected from a Zero Volt Switch (ZVS) Full Bridge Isolated Buck converter.
Keywords: Fault Detection and Isolation (FDI), DC/DC converter, Bayesian Belief Network, Observer
design, Bayesian Naïve Classifier.

(Guerin et al. 2011) by using an additional measurement

1. INTRODUCTION
depending on the use of a magnetic near-field probe.
Nowadays, electrical machines are used everywhere in our
daily life. Electrical power represents the core of any new
invention. Actually, It is the main power that keeps our life
going on. These facts, increase the demand of continuous
electrical energy, and therefore, the demand of reliable
electrical supplying systems in order to satisfy the
consumers. Continuity requires a fault-free system which can
prevent and treat any fault occurrence. For that, it is
necessary to have a Fault Detection and Isolation (FDI)
system, that can detect the occurrence of the faults as soon as
possible, and then isolate the type of the detected faults.
Power converters play an important role in the electrical
energy lifecycle and are used in almost all the electrical
engineering systems. Thus, in our work we are going to focus
on such equipments, especially the ZVS full bridge isolated
Buck converter. Such a converter is used in some topologies Fig. 1. Topology of the SMSE.
of Systems of Multi Sources of Energy (SMSE) such as the
one shown in Fig.1 (Guerin et al 2012). This DC/DC In this paper, a Bayesian Naïve Classifier (BNC) structure is
converter is used to manage the coupling and decoupling of selected and used as a first attempt to use Bayesian Belief
the energy sources on the DC-bus according to the load Networks (BBNs) for DC/DC power converter FDI. BBNs
demand and available power. have been used for FDI and they proved a high efficiency
(Zhao et al. 2016 ). The capabilities of BBN is highlighted
Many studies have been done to detect and isolate faults in with a comparative study between BNC as a simple form of
power converters. A method for MOSFET faults in a ZVS BBN, and the Proportional Observer (PO) as a well known
full bridge isolated Buck converter using the DC link current and used FDI method (Blanke et al. 2003), and as the
patterns as the signatures of these faults was proposed in simplest form of observers. The obtained results, show
(Kim et al. 2008). Authors in (Chen et al. 2011) describe clearly that the BNC surpasses the PO and prove its
another method based on fast Fourier transform of magnetic capabilities for DC/DC power converter FDI.
near-field of dc-dc converters. In (Meziane et al. 2015) a
model-based approach FDI method was proposed, following 2. STUDIED SYSTEM AND CONSIDERED FAULTS
a sliding mode observer based on a residual generation that
was applied on a three-cell power converter. Moreover, in 2.1. ZVS full bridge isolated Buck converter model
(Guerin et al. 2009), a set of residuals were generated using
parity space algorithm according to a variable structure state The structural diagram of the ZVS full bridge isolated Buck
space model in order to detect sensor faults in ZVS full converter is represented on the Fig.2. These DC/DC
bridge isolated Buck converter. This work was completed in converters are isolated (HF transformer TR1) Buck
converters (D5, D6, D7, D8, L, Ce, Re) with a full bridge

Copyright © 2016 IFAC 309

(Q1, Q2, Q3, Q4) and ZVS. The full bridge control (Q1, Q2, where d(t) represents the measurement error vector.
Q3, Q4) is realized by a phase shift controller UC3879 𝐴𝑀 , 𝐵𝑀 and 𝐶𝑀 are given by equation (2):
through specialized MOSFET drivers IR2113. The duty cycle
value φ is modified by the phase shift between Va and Vb 𝑅𝑚𝑜𝑠 𝜑 +𝑅𝑚𝑜𝑠 + 𝑟𝑝
0 0
𝐿𝑀
voltages. The phase shift is controlled by an analog DC 𝑛
voltage (between 0V and 5V) which represents the DC/DC 𝐴𝑀 𝜑 = −
𝑛(2𝑅𝑚𝑜𝑠 +𝑟𝑝 )
−
𝑟𝐿 +𝜑 𝑟𝑠 +𝑛 2 𝜑 (2𝑅𝑚𝑜𝑠 + 𝑟𝑝 )
−
1
converter analog voltage control input. 𝐿
1
𝐿 𝐿
1
0 −
𝐶𝑒𝑞 𝑅𝑒𝑞 𝐶𝑒𝑞
0 0
𝑛. 𝜑 2 𝜑 𝑛. 𝜑 0
𝐵𝑀 𝜑 = − , 𝐶𝑀 𝜑 = 0 1 0
𝐿 𝐿 0 0 1
0 0
𝑅𝑒 𝑅𝑐𝑕
𝑅𝑒𝑞 = ; 𝐶𝑒𝑞 = 𝐶𝑒 + 𝐶𝑑𝑐 (2)
𝑅𝑒 +𝑅𝑐𝑕

𝑟𝑝 , 𝑟𝑠 , 𝐿𝑀 and 𝑅𝑚𝑜𝑠 are respectively the primary resistance,

Fig. 2. Structural diagram of the ZVS full bridge isolated
Buck converter.
Four running phases must be distinguished:
Phase 1: (Q1Q4) closed: from t  0 to t  T0
Phase 2: (Q1d3) closed: from t  T0 to t  T0
Phase 3: (Q2Q3) closed: from t  T0 to t  (1   )T0
Phase 4: (Q2d4) closed: from t  (1   )T0 to t  2T0  T
Let us define iM as the magnetizing current (A) of the HF
transformer, iL as the inductance current (A), ig as the source
current (A), s as the DC/DC converter output voltage (V), Vg
as the source voltage (V) and Vd as the threshold diode
voltage (V).
The authors (Mboup A.B. et al 2008) have developed an
average state space model that depends on the duty cycle
value φ(t), controlled by the analog voltage control input.
Let us define the state, control and output vectors of the
average model as:
 XM = ( IM, IL,S)T
 UM = (Vg, Vd )T
 YM = (Img, ImL, Sm)T
where:
 IM, IL and S stands for the average values of iM, iL,
and s respectively.
 Img is the measured average value of the current
sinked from the source (A).
 ImL is the measured average value of the current
provided to the load (A).
 Sm is the measured average value of the DC/DC
converter output voltage (V).
The average model is represented with the equations (1) and
(2). This model will be used as a reference model in the next
section in order to design the observer used for fault detection
and diagnosis.
𝑋𝑀 = 𝐴𝑀 (φ(t)).𝑋𝑀 (𝑡) + 𝐵𝑀 (φ(t)).𝑈𝑀
𝑌𝑀 (𝑡) = 𝐶𝑀 (φ(t)).𝑋𝑀 + 𝑑 𝑡 (1)
the secondary resistance, the magnetizing inductance of the
HF transformer (TR1) and the MOSFET transistors
(Q1,Q2,Q3,Q4) channel resistance. 𝐿, 𝑟𝐿 , 𝐶𝑒 and 𝑅𝑒 are
respectively the coil inductance, the coil resistance, the
capacity and the resistance of the Buck converter. n is the
ratio of the HF transformer. 𝑅𝑐𝑕 is the load (assumed to be
resistive).
GREAH Research Group of the University of Le Havre
(France), has designed and developed a ZVS full bridge
isolated Buck converter with ability to simulate ten faults
(Fig 3): (1) phase shift controller stuck ON (+15V), (2) phase
shift controller stuck OFF(0V), (3) MOSFET open circuit, (4)
MOSFET driver stuck ON (+15V) (5) MOSFET driver stuck
OFF (0V), (6) MOSFET short circuit, (7) diode short circuit,
(8) coil open circuit, (9) diode open circuit, and (10) clock
frequency deviation including the considered open circuit
faults. This system supports our work with the required real
data.
Fig. 3. ZVS full bridge isolated Buck converter
2.2. Considered faults
ZVS full bridge isolated Buck converter is mainly used in
some topologies of Systems of Multi Sources of Energy
(SMSE) as in (Guerin et al 2013 ). Its main role is to couple
and decouple the sources of energy according to their
availability and the load demand. Such DC/DC converters are
exposed to faults that may be very harmful due to the
environment they work in. Electrical faults can be a short

Copyright © 2016 IFAC 310

circuit, an open circuit, or a leakage. In our work three open
3. OBSERVER DESIGN
circuit faults are considered, MOSFET (Q1) open circuit,
diode (D8) open circuit and coil (L) open circuit. The open 3.1. Luenberger’s Observer
circuit faults are not generally harmful to the controllers and
drivers of the power MOSFETs, and does not cause system Based on the state space model (1) the observer equations
shutdowns, but could lead to the secondary faults at the other are designed. These equations are equivalent to the fault-
components (Diallo et al 2005). free system equations except an additional term comparing
Real data for the three considered faults is collected. Three the actual measured output y to the estimated output 𝑦. The
runs where performed, each one of 8 seconds and correspond equations of the observer are given by (3):
to one of the three faults. Fig 4 shows the three runs. The first
𝑋 = 𝐴𝑜𝑏𝑠 .𝑋𝑀 𝑡 + 𝐵𝑜𝑏𝑠 .[ 𝑈𝑀 𝑌𝑀 ]𝑇
run (a) simulates fault1 which occurs at 𝑡= 4.522 sec, the
second run (b) simulates fault2 which occurs at 𝑡 = 4.358 sec 𝑌𝑀 = 𝐶𝑀 𝑋𝑀 (3)
and the last run (c) simulates fault3 that occurs at 𝑡 = 4.253
sec. These data sets are used for the PO and the BNC. where 𝐴𝑜𝑏𝑠 = 𝐴𝑀 − 𝐿𝑜𝑏𝑠 𝐶𝑀 , 𝐵𝑜𝑏𝑠 = 𝐵𝑀 𝐿𝑜𝑏𝑠 and 𝐿𝑜𝑏𝑠 is
the observer gain.
Source Current (IG)
5

4
An important task in the design of observers is the selection
3
X: 4.522
Y: 2.19
Fault 1 Occurrence
of its gain 𝐿𝑜𝑏𝑠 . Let us define the estimation error e(t)
2

1 =𝑋𝑀 𝑡 − 𝑋𝑀 𝑡 , 𝐿𝑜𝑏𝑠 should be selected such that the

0
0 1 2 3 4 5 6 7 eigenvalues of 𝐴𝑜𝑏𝑠 are larger than those of 𝐴𝑀 in
3
Inductance Current (IL)
magnitude, and lies in the left-hand side of the complex
2
plane, thus the dynamics of the estimation error (𝑒(𝑡) =
1 𝐴𝑀 − 𝐿𝑜𝑏𝑠 𝐶𝑀 𝑒(t)) converges to zero and the estimated
0
0 1 2 3 4 5 6 7 state 𝑋𝑀 𝑡 will rapidly and asymptotically approach
60
Output Voltage (S)
𝑋𝑀 𝑡 . This designed observer will be used to generate
40 residuals. Our method depends on residual analysis, these
20
residuals represent the difference between the system
0
0 1 2 3 4 5 6 7
measured variables (𝐼𝑚𝑔 , 𝐼𝑚𝐿 , 𝑆𝑚 ) with the estimated ones
time (s)

(a) (4):
Source Current (IG)
5

4
𝑟 𝑡 = 𝑦 𝑡 − 𝑦(𝑡) (4)
3 X: 4.358
Y: 2.035 Fault 2 Occurrence
2

1 At each sampling time the measured values are compared to

0
0 1 2 3 4 5 6 7 the estimated one to get a vector r(t)=(𝑟𝐼𝐺 , 𝑟𝐼𝐿 , 𝑟𝑆 ) of
3
Inductance Current (IL)
residuals. This vector will be used to detect the occurrence
2
of the considered faults.
1

0
3.2. Detection and experimental results
0 1 2 3 4 5 6 7

60
Output Voltage (S) Fault detection is based on studying the resulted residuals.
40
Residuals in fault-free case are in average equal to zero, and
different from zero when faults occur. Consequently, a fault
20

is detected when the value of the residuals become different

from zero. Thus, a convenient threshold should be specified
0
0 1 2 3 4 5 6 7
time (s)

(b) in order to detect the faults and avoid false alarms. The
Source Current (IG)
5

4
selected threshold is based on a significant statistical value,
3

2
X: 4.253
Y: 1.947 Fault 3 Occurrence the standard deviation of the residuals in the fault-free case,
1 which represent the deviation of the residuals value from
0
0 1 2 3 4 5 6 7 their mean in the normal case. This value is multiplied by a
3
Inductance Current (IL)
gain k, in order to ensure that the detection is not a false
2 alarm (equation 5).
1
𝑁 𝑥 −𝜇
𝑖=1 𝑖
0
0 1 2 3 4 5 6 7 𝑇𝑕𝑟 = 𝑘. 𝜎 = 𝑘. (5)
Output Voltage (S)
𝑁
60

where µ is the mean of the residual component X, 𝑥𝑖 is its

40

ith value, and N is the number of the measured points. This

step will results in a threshold vector of three components,
20

one for each residual component (𝑇𝑕𝑟𝑟 𝐼𝐺 , 𝑇𝑕𝑟𝑟 𝐼𝐿 , 𝑇𝑕𝑟𝑟 𝑆 ).

0
0 1 2 3 4 5 6 7
time (s)

(c) whenever one of the components is above its threshold a

Fig. 4. Collected real data (Source current Img, Inductance fault is detected.
current ImL, Output voltage Sm).
This observer is tested on the collected real data. The

Copyright © 2016 IFAC 311

resulted residuals were able to detect the occurring faults. Bayesian Belief Network (BBN) is a powerful tool for
Fig 5 presents the generated residuals (blue line) and the knowledge representation and inference under conditions of
detection alarm (red line). The time of occurrence and uncertainty. These networks are defined by specifying their
detection of each fault is presented in Table 1. The average qualitative and quantitative components. The BBN structure
delay for detection is 3 ms. (qualitative) is made of nodes that represent the random
Residual
variables and arcs for the dependency between the nodes, in
RIG
50
X: 4.526
Detection alert
the form of Direct Acyclic Graph (DAG). It can be specified
0 Y: 40
Fault 1 Detection by experts (Wiegerinck 2005), by structure learning
algorithms that study the correlation between the nodes (Pearl
-50

-100
0 1 2 3 4 5 6 7
2000), or it can be identified by both experts and structure
50
RIL
learning as in (Flores et al. 2011). The BBN parameters
0
(quantitative), i.e. the Conditional Probability Tables (CPT's),
are filled with conditional probabilities for each node giving
its parents. In general, those parameters are estimated (fully
-50
0 1 2 3 4 5 6 7

20
RS
or partially) via learning processes from historical data of the
10

0
system, or by expert's knowledge.
These networks rely on inference algorithms to compute
-10

-20
0 1 2 3 4 5 6 7
time (s)
beliefs in the context of observed evidence. Inference is the
(a)
process of updating probabilities of outcomes (CPTs) based
50
RIG
Residual
Detection alert upon the relationships in the model and the evidence known
0
X: 4.361
Y: 40
about the situation at hand. Inference algorithms can be exact
-50
Fault 2 Detection
or approximated (Guo and Hsu 2002).
-100
0 1 2 3 4 5 6 7 The aim of this study is to investigate the capabilities of BBN
at the domain of Fault Detection and Isolation (FDI) in
RIL
50

0
power electronics and to assess its performance compared to
the previously designed observer. Hence, and in order to
-50
0 1 2 3 4 5 6 7 highlight BBN capabilities, we assume that the BBN
structure will depict a simple correlation between the direct
RS
20

10

0
measurement of the system output variables (𝐼𝑚𝑔 , 𝐼𝑚𝐿 , 𝑆𝑚 )
-10 and a decision node (fault type). Specifically, we tend to use
-20
0 1 2 3 4
time (s)
5 6 7
the Bayesian Naive Classifier (BNC) as a first attempt to use
(b) BNNs for DC/DC power converter Fault Detection and
Isolation (FDI). The used BNC is shown in Fig.6 where the
Residual
RIG Detection alert
50

0
X: 4.255
Y: 40
Fault 3 Detection
class node represents the faulty case and can take four values:
-50 ff for fault –free, f1 for fault1, f2 for fault2 and f3 for fault3.
-100

0 1 2 3 4 5 6 7
After specifying the structure, next comes the learning part.
50
RIL The BNC is learned from the direct measured outputs (Fig.4).
0

-50
CLASS
0 1 2 3 4 5 6 7

RS
20

10

-10

-20
0 1 2 3 4
time (s)
5 6 7
𝑰𝒎𝒈 𝑺𝒎
(c) 𝑰𝒎𝑳
Fig. 5. Generated residuals and detection alarm for (a)
fault1;(b) fault2 ;(c) fault3.
Fig. 6. The BNC structure used for detection and isolation.
Table 1. Occurrence and detection time
Page Fault 1 Fault 2 Fault 3 4.1. Detection and isolation algorithm.
Occurrence 4.522 sec 4.358 sec 4.253 sec
Let’s define 𝑆𝐷1 , 𝑆𝐷2 and 𝑆𝐷3 as the sets of the direct
Detection 4.526 sec 4.361 sec 4.255 sec
measurements presented in Fig. 4 (a), (b), and (c)
respectively (i.e. the dimension of each 𝑆𝐷 set is
Its clear that the residuals are generated for the three faults 8000×3). Let 𝐹𝐷1 , 𝐹𝐷2 and 𝐹𝐷3 be the subsets of 𝑆𝐷1 , 𝑆𝐷2 ,
are similar in shape and sign, which make it difficult to and 𝑆𝐷3 respectively, such that each 𝐹𝐷i set include 4000
isolate faults. Thus the PO is able to detect the three faults but points of 𝑆𝐷i related to the fault 𝑓𝑖 , starting from the
not to isolate them. time of 𝑓𝑖 occurrence, in addition to a part of the free-
4. BAYESIAN BELIEF NETWORK fault points that precedes the fault occurrence.

Copyright © 2016 IFAC 312

Moreover we suppose that the sum of the number of
free-fault points in the three 𝐹𝐷 sets is equal to 4000.
Those sets are labeled offline by adding a column to each set,
representing the class of each point (f1 for fault1, f2 for
fault2, f3 for fault3, and ff for fault-free). Subsequently, part
of these data should be used for the BNC learning. Let 𝐿1𝐷 ,
𝐿2𝐷 , and 𝐿3𝐷 be the data part of 𝐹𝐷1 , 𝐹𝐷2 , and 𝐹𝐷3 respectively
used for learning, such that each 𝐿𝑖𝐷 set represents 50% of the
𝐹𝐷𝑖 set.
After parameter learning, the BNC is ready for inference. The
remaining 50% of 𝐹𝐷1 , 𝐹𝐷2 , and 𝐹𝐷3 are denoted by 𝐴𝑓1 , 𝐴𝑓2 ,
and 𝐴𝑓3 respectively. These three sets will be used for
validation. Each observation 𝑑 𝑡 = (𝐼𝑚𝑔 , 𝐼𝑚𝐿 , 𝑆𝑚 ) in 𝐴𝑓1 ,
𝐴𝑓2 , and 𝐴𝑓3 is passed through the BNC, which in its turn will
compute the probability p(t) of the responsibility of each fault
to the given observation d(t):
p(t)={ 𝑃𝑓 𝑓𝑟𝑒𝑒 𝑡 = 𝑃 𝑓𝑓𝑟𝑒𝑒 𝑟 𝑡 , 𝑃𝑓1 𝑡 = 𝑃 𝑓1 𝑟 𝑡 ,
𝑃𝑓2 𝑡 = 𝑃 𝑓2 𝑟 𝑡 , 𝑃𝑓3 𝑡 = 𝑃(𝑓3 /𝑟(𝑡)) }.
The case with maximum probability will be considered as the
BNC classification. The detection is done according to this
immediate classification, that is whenever a point is classified
as faulty (i.e. f1, f2 or f3), a fault is detected. The set of
recorded points after fault detection will be used for isolation.
Subsequently, let 𝐵𝑓1 , 𝐵𝑓2 , and 𝐵𝑓3 be the sets of p(t)'s for the
d(t) points that belong to in 𝐴𝑓1 , 𝐴𝑓2 , and 𝐴𝑓3 respectively.
From each of these sets, another equivalent set (𝐶𝑓 𝑖 ) is
generated such that:
c(t) = {𝑓𝑖 𝑃𝑓 𝑖 𝑡 = max 𝑝 𝑡 𝑤𝑕𝑒𝑟𝑒 𝑖 = 1,2,3 𝑜𝑟 𝑓𝑟𝑒𝑒}.
This process will lead to three new sets 𝐶𝑓1 , 𝐶𝑓2 , and 𝐶𝑓3 that
will be used for isolation. In each set, the frequency of each
𝑓𝑖 will be computed, and the fault with the maximum
frequency will be selected as the isolated fault.
4.2. Experimental results
The discussed algorithm is tested on the remaining 50% of
the considered part (𝑆𝐷 sets), that is the 𝐴𝑓 i sets. The results
are shown in Fig 7. It’s clear that the proposed BNC
algorithm is able to detect and isolate the three faults. In (a),
(b), and (c) the free-fault points (the beginning of each graph)
are all correctly classified. As well as, in the remaining part
of each graph, a few number of points are misclassified. The
graphs are more explained in the confusion matrix (table 2).
For an instance, let’s consider Fig 7. (a). First, for detection,
the first point classified as faulty is at t= 698 ms after 3 ms
from fault1 occurrence (at t=695 ms), and this point is
classified as f2 which is actually not the real case (i.e. f1).
But as we said in the algorithm, the detection alarm is on at
the time of the first point classified as faulty even if it is
wrongly classified. Second, for isolation, the set of points
recorded after this detection, are classified (𝐶𝑓 i sets), that is
the green line in (a) from t=699 ms (first point after point of
detection) till t= 2695 ms (last point in the set), which is
coincident with the dotted red line (actual occurring fault f1),
this coincidence means that the point are correctly classified
Copyright © 2016 IFAC
as their actual class. Finally, the frequency of each case is
computed and the most frequent classified case will be
selected as the occurring fault (isolated). Table 2 ( third
column) indicates that the frequency of fault1 in Fig 7 (a)
after detection is 1987 which is the largest between the
frequency of fault2 equal to 11 and frequency of fault3 equal
to 0. Therefore f1 is the isolated fault which is the real case.
5. PO Vs BNC
The aim of this study is to introduce the BBN as an efficient
method for Fault Detection and Isolation (FDI) in DC/DC
power converters via a comparison with Proportional
Observer (PO). As an FDI method, three aspects should be
taken into consideration. The first aspect is the latency of
detection, the second is the accuracy of isolation, and the last
is the simplicity of the method. For the first aspect, PO
detects the three faults in an average of 3 ms while for the
BBN the experiment shows that fault1 is detected after 3 ms,
fault2 is detected after 3 ms and fault3 is detected after 4 ms
which gives an average of 3.3 ms. As for the second aspect
which is isolation, for the PO, it is impossible to isolate the
three faults using simple residual analysis, while BBN shows
a very high aptitude in isolation. Finally, the simplicity of
the method is the last aspect. BNC is the simple form of
BBNs and its complexity is studied in (Zheng et al. 2005).
Actually, in our case, the time complexity of the learning
phase will not be considered because it is done offline. What
we are interested in is the time complexity of classifying a
single point (observation) which equals to O(kl) where k is
the number of the class values and l is the number of
attributes. In our case, k=4 and l=3, then O(kl)=O(12). On the
other, hand the Proportional Observer (PO) complexity is
equal to the complexity of the matrices multiplication
O(mnp) where the first matrix of dimension is m×n, the
second matrix is n×p, and the matrices addition is O(mn).
Thus:
O(observer)= O(calculating 𝑋) + O(calculating 𝑌) + O(calculating
r)= [ O(3 × 3 × 1) + O(3 × 2 × 1) + O(3 × 1)] + [ O(3 × 3 × 1)] +
[O(3 × 1)]=O(30).
Consequently, the complexity of the BNC is much better than
the PO which is a pro. This comparison highlights the high
capabilities of the BNC and subsequently the eligibility of
using Bayesian Belief Networks in the domain of Fault
Detection and Isolation (FDI) within DC/DC power
converters.
f3
BNC classifications
Actual fault type
X: 698
Y: 3
f2 Detection Time
Fault type
f1
ff
0 500 695 1,000 1,500
Time (ms)
2,000 2,500 2695
(a)
313
 f3
BNC classifications Diallo, D., Benbouzid, M.E.H., Hamad, D. and Pierre, X.,
Actual fault type
2005. Fault detection and diagnosis in an induction
machine drive: A pattern recognition approach based on
f2
concordia stator mean current vector. Energy
X: 682
Y: 3 Conversion, IEEE Transactions on, 20(3), pp.512-519.
Flores, M.J., Nicholson, A.E., Brunskill, A., Korb, K.B. and
Fault type

Detection time

Mascaro, S., 2011. Incorporating expert knowledge when

f1 learning Bayesian network structure: a medical case
study. Artificial intelligence in medicine, 53(3), pp.181-
204.
Guerin, F., Labarre, C. and Lefebvre, D., 2011, September.
ff
0 500 679 1,000 1,500 2,000 2,500 2,679 Magnetic near-field measurement for FDI of ZVS full
bridge isolated buck converter. InDiagnostics for
Time (ms)

(b)
f3
BNC Classification
Electric Machines, Power Electronics & Drives
Actual fault type
(SDEMPED), 2011 IEEE International Symposium
on (pp. 344-349). IEEE.
Guerin, F. and Lefebvre, D., 2009. Residual analysis for the
f2
X: 630
Y: 3
diagnosis of hybrid electrical energy systems. IFAC
Safeprocess 2009, Invited Session on “Diagnosis and
Fault type

Detection time

Environmental Issues.
f1
Guerin, F., Lefebvre, D. and Loisel, V., 2012. Supervisory
control design for systems of multiple sources of
energy. Control Engineering Practice, 20(12), pp.1310-
1324.
ff
0 500 626 1,000 1,500 2,000 2,5002,626
Guérin, F. and Lefebvre, D., 2013, December. Adaptive
Time (ms)
generalized PID controllers and fuzzy logic coordinator
(c) for load sharing in SMSE. In Decision and Control
Fig. 7. Experimental results. (CDC), 2013 IEEE 52nd Annual Conference on (pp.
Table 2. Confusion matrix 5588-5593). IEEE.
Guo, H. and Hsu, W., 2002, July. A survey of algorithms for
Actual Class
Fault free Fault1 Fault2 Fault3 real-time Bayesian network inference.
Classification In AAAI/KDD/UAI02 Joint Workshop on Real-Time
Fault-free 2000 2 3 3 Decision Support and Diagnosis Systems. Edmonton,
Fault1 0 1987 0 9 Canada.
Kim, S.Y., Nam, K., Song, H.S. and Kim, H.G., 2008. Fault
Fault2 0 11 1997 2
diagnosis of a ZVS DC–DC converter based on DC-link
Fault3 0 0 0 1986
current pulse shapes. Industrial Electronics, IEEE
Transactions on, 55(3), pp.1491-1494.
6. CONCLUSION AND PERSPECTIVES Mboup, A.B., Guerin, F., Ndiaye, P.A. and Lefebvre, D.,
This paper discusses the ability of using Bayesian belief 2008, June. Multimodel for the coupling of several dc/dc
networks in the domain of Fault Detection and Isolation power converters on a dc bus. In Industrial Electronics,
(FDI) within DC/DC power converters, via a comparative 2008. ISIE 2008. IEEE International Symposium on(pp.
study between a simple well-known and used FDI method 1507-1512). IEEE.
represented by Proportional Observer (PO) versus the Meziane, H., Labarre, C., Lefteriu, S., Defoort, M. and
simplest form of the BBN represented by the Bayesian Naïve Djemai, M., 2015. Fault detection and isolation for a
Classifier. Both methods were implemented and tested on multi-cellular converter based on sliding mode
real data. The BNC shows acceptable latency, in addition, a observer. IFAC-PapersOnLine, 48(21), pp.164-170.
high advance on the PO in both simplicity and isolation Pearl, J. Causality: Models, Reasoning, and Inference.
accuracy. In the future this work will be extended in order to Cambridge. 2000.
have a complete solution by adding a reconfiguration part Wiegerinck, W., 2005. Modeling Bayesian Networks by
that ensures continuity. Learning from Experts. In BNAIC (pp. 305-312).
Zhao, Y., Wen, J., Xiao, F., Yang, X. and Wang, S., 2016.
REFERENCES Diagnostic bayesian networks for diagnosing air
handling units faults-Part I: faults in dampers, fans,
Blanke M., Kinnaert M., Lunze J., Staroswiecki M., 2003.
filters and sensors. Applied Thermal Engineering.
Diagnosis and fault tolerant control. Springer Verlag,
Zheng, F. and Webb, G.I., 2005. A comparative study of
New York.
semi-naive bayes methods in classification learning.
Chen, Y., Pei, X., Nie, S. and Kang, Y., 2011. Monitoring
In Proceedings of the fourth Australasian data mining
and diagnosis for the DC-DC converter using the
conference (AusDM05) (pp. 141-156).
magnetic near field waveform. IEEE Trans. Ind.
Electron, 58(5), pp. 1634–1647, (Apr).

Copyright © 2016 IFAC 314

View publication stats