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

net/publication/224182214

Intelligent Agent-Based System Using Dissolved Gas Analysis to Detect

Incipient Faults in Power Transformers

Article  in  IEEE Electrical Insulation Magazine · January 2011

DOI: 10.1109/MEI.2010.5599977 · Source: IEEE Xplore

CITATIONS READS
54 2,984

5 authors, including:

Asghar Akbari Azirani Alireza Setayeshmehr

Khaje Nasir Toosi University of Technology Jundi Shapur University of Technology, Iran, Dezful
92 PUBLICATIONS   709 CITATIONS    16 PUBLICATIONS   488 CITATIONS   

SEE PROFILE SEE PROFILE

H. Borsi Ernst Gockenbach

Leibniz Universität Hannover Leibniz Universität Hannover
206 PUBLICATIONS   2,364 CITATIONS    266 PUBLICATIONS   2,850 CITATIONS   

SEE PROFILE SEE PROFILE

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

Experimental Project (Wind Energy) View project

estimation of aging state of transformers insulation View project

All content following this page was uploaded by Issouf Fofana on 30 March 2014.

The user has requested enhancement of the downloaded file.

 F E A T U R E A R T I C L E

Intelligent Agent-Based System Using

Dissolved Gas Analysis to Detect Incipient
Faults in Power Transformers
Key words: DGA, information fusion, power transformer, diagnosis, evidence theory, multiagent sys-
tem, ontology

Introduction A. Akbari, A. Setayeshmehr, H. Borsi,

Power transformers are considered capital investments in the
infrastructure of every power system in the world. They are the and E. Gockenbach
heart of electric power distribution and transmission systems, Institute of Electric Power Systems, High Voltage
and it is essential that they function properly. Because power Engineering Section (Schering Institute), Leibniz
transformers are critical to the reliable operation of every power Universität Hannover, Hanover, Germany
grid, ways to extend their lives, prevent incipient electrical fail-
ures, and improve preventive maintenance policies have become I. Fofana
increasingly important. As a result, the development of accurate Canada Research Chair on Insulating Liquids and
monitoring and diagnosis systems has been under consideration
for several years [1], [2]. Much work has been done in recent
Mixed Dielectrics for Electrotechnology (ISOLIME),
years to find ways of prolonging transformer life and reducing Université du Québec à Chicoutimi, Québec, Cana-
the cost of failure [3], [4]. Still, new methods for analyzing the da
condition of transformers are needed. In addition to technolo-
gies such as sensing and measuring devices, software architec-
ture is playing an important role in the development of power-
transformer monitoring and diagnostic systems. These systems Condition monitoring and software-
are complex and should fulfill a number of requirements. Some
of these requirements are as follows: based diagnosis tools are central to
• Remote supervision by experts: Because of the com- the implementation of efficient main-
plexity of the diagnosis process, experts in a supervi- tenance management strategies for
sory center need tools to monitor and control the diag-
nosis process of the power-grid transformers. many engineering applications in-
• Combined diagnostic methods: In power-transformer cluding power transformers.
monitoring and diagnostic applications, maintenance
engineers face a complexity and diversity of tasks as-
sociated with this complex domain. Maintenance engi-
neers deal with different types of information from dif- a significant move forward in the next generation of
ferent sources, and these various types of information diagnostic and monitoring systems.
require processing with different diagnostic schemes. • Adaptability to new environments: Intelligent diagnos-
This problem may neither fit the assumptions of a sin- tic systems or expert systems may fail to operate accu-
gle technique nor be effectively solved by the strengths rately in a new working environment. In this situation,
and capabilities of a single technique. Combined diag- the transformer owner should have other applications
nostic techniques can be very useful to experts in de- available to allow the compatibility of the system with
tecting transformer faults and reducing the risk of mis- new environments. Also, with available new monitor-
taken diagnosis. Merging various techniques should be ing and diagnosing hardware and software compo-
nents, it is necessary to develop and update these sys-

November/December — Vol. 26, No. 6 0883-7554/07/$25/©2010IEEE 27

 tems for power transformers. Because the lifetimes of methods [8]–[11]. The advantages to the ratio methods are that
electronic components and sensors used for diagnostics they are quantitative and independent of transformer oil volume.
and monitoring are much shorter than those of power However, ratio methods can produce incorrect interpretations or
transformers, the diagnosis system should be modular none at all. Therefore, they should be used in conjunction with
with the possibility of easily adding a new module. other diagnostic methods such as the Key Gas method. In ad-
Taking into account these requirements, the usability of mul- dition, minimum levels of gases must be present before results
tiagent software systems appears to be essential. Agent technol- of the ratio methods become significant. One drawback of the
ogy is a very promising approach to support the new generation gas ratio methods (Dörnenburg, Rogers, IEC) is that some DGA
of monitoring and diagnosis systems. In this contribution, an results may fall outside the ratio codes, and no diagnosis can
intelligent agent-based platform was implemented. It integrates then be given (unresolved diagnoses). This does not occur with
two or more dissolved gas analysis (DGA) techniques, with the the Triangle method because it is a closed system rather than an
aim of combining their strengths and reducing their weaknesses. open one.
Flexible software systems require a distributed architecture in In this contribution, two ratio methods (IEC and Rogers), a
which each element may work independently. This advanced ar- pure ANN (artificial neural network) method, and Duval’s Tri-
chitecture can be implemented using intelligent agent technolo- angle method are considered. Ratio methods have several dis-
gy to ease the development of independent software components advantages.
including a wide variety of artificial intelligence techniques for • They do not always yield an analysis.
different situations. This is in contrast with conventional central- • They are not always correct.
ized stand-alone applications. • They are dependent on the preservation system.
• Dörnenburg’s method has fallen out of favor because it
DGA Diagnostic Methods misses too many incipient faults.
Gases in oil always result from the decomposition of electri- • Solid insulation is handled separately using carbon
cal insulation materials (oil or paper), as a result of faults or monoxide and carbon dioxide ratios.
chemical reactions in the equipment. Such incipient electri- To improve the diagnostics, these two ratio methods were
cal failures are one cause of certain hydrocarbons breakdown implemented using ANN. These two ANN implementations
called gassing of oil. Gassing of oil is defined scientifically as do not use any transformer failure history data. The ANN was
the chemical decomposition of certain hydrocarbons under the used just to map diagnosis codes to corresponding failures. The
effect of electrical and thermal stresses [5]. Because the result- third implemented method was a pure ANN method, trained
ing fault gases dissolve in the oil, the technique of DGA was with some different available known DGA test data. The fourth
developed to detect in the early stage defects on the surface of method used was Duval’s Triangle method, one of the most used
the solid insulation. Dissolved gas analysis is probably the most DGA methods. All of these methods have been widely reported
used tool for detecting faults in electrical equipment in service. in the literature [7]–[13]. However, to facilitate understanding of
The procedures used in the past to analyze gassing of trans- materials presented in this contribution, there is a short review
former oil have generally been inadequate. The fundamental indicating advantages and weaknesses of these techniques.
shortcomings of DGA are underscored by IEEE [6] as follows:
“Many techniques for the detection and the measurement of gas- The IEC Ratio Method
The IEC method, widely used by the utilities, is based on a
es have been established. However, it must be recognized that
analysis of these gases and interpretation of their significance combination of ratios (based on Rogers Ratios), gas concentra-
is at this time not a science, but an art subject to variability.” tions, and rates of gas increase. Three gas ratios are used to de-
Furthermore, “The result of various ASTM testing round robins termine incipient failures. Each combination of diagnosis code
indicates that the analytical procedures for gas analysis are diffi- corresponds to a certain condition of the power transformer
cult, have poor precision, and can be wildly inaccurate, especial- (Table 1).
ly between laboratories.” Finally, “However, operators must be A three-layer back-propagation ANN (Figure 1) was used to
cautioned that, although the physical reasons for gas formation implement this DGA method. Table 2 shows the input and out-
have a firm technical basis, interpretation of that data in terms put patterns for this ANN as introduced in [12]. The ANN has
of the specific cause or causes is not an exact science, but is the three inputs and eight outputs and was trained according to this
result of empirical evidence from which rules for interpretation table. The nine output values can be described as O1 to O9.
have been derived.” Initial Rogers Ratio Method
About 20 DGA interpretation techniques have been devel- Four ratios are considered to detect incipient failures in the
oped so far [7]. Commercial packages implement such methods transformer. These ratios are encoded depending on their value
individually. Several individual DGA users have also devel- (Table 3). The combination of codes provides Rogers fault diag-
oped their own graphical software. Because all these methods nostic (Table 4). A three-layer back-propagation ANN was used
are based on heuristic and not scientific formulation, combined to implement this DGA method (Figure 2).
DGA diagnostics may reduce the risk of mistaken diagnosis and Table 5 shows the input and output patterns for this ANN as
enhance accuracy. Among the available DGA techniques, the introduced in [12]. The ANN has four inputs and 12 outputs and
most used are IEC, Rogers, Duval’s Triangle, and the Key Gas was trained according to this table.

28 IEEE Electrical Insulation Magazine

 Table 1. Fault Types According to the IEC Ratio Method.

Fault codes for ratios

No. Fault type C2H2/C2H4 CH4/H2 C2H4/C2H6

1 No fault 0 0 0

2 <150°C thermal fault 0 0 1

3 150°C –300°C thermal fault 0 2 0

4 300°C –700°C thermal fault 0 2 1

5 >700°C thermal fault 0 2 2

6 Low energy partial discharges 0 1 0

7 High energy partial discharges 1 1 0

8 Discharge of low energy 1-2 0 1-2

9 Discharge of high energy 1 0 2

Pure ANN Method

The third implemented method was a three-layer pure back-
propagation ANN. This ANN method can be used to recognize Figure 1. The back-propagation artificial neural network for the
the hidden relationships between the dissolved gases and the IEC dissolved gas analysis method; O1 = no fault, O2 = thermal
fault types through training process. These hidden relationships fault <150°C, O3 = thermal fault 150–300°C, O4 = thermal
may not be apparent and may not be recognized by normal DGA fault 300–700°C, O5 = thermal fault >700°C, O6 = low energy
methods. This ANN includes input, output, and hidden layer with partial discharges, O7 = high energy partial discharges, O8 =
10 nodes. In this ANN approach, the transformer condition was discharge of low energy, O9 = discharge of high energy.
classified as normal, discharge of low energy fault, discharge of
high energy fault, partial discharge fault, and three types of ther-
the early 1970s by Michel Duval [13]. It is based on the use of
mal faults. Five key gases (H2, CH4, C2H2, C2H4, C2H6) and the
three gases (CH4, C2H4, and C2H2) corresponding to the increas-
total combustible gas (TCG) were selected as inputs (Figure 3).
ing levels of energy necessary to generate gases in transformers
Duval’s Triangle Method in service. This method has proven to be accurate and dependable
Duval’s Triangle diagnostic method for oil-filled power over many years and is now gaining in popularity. One advan-
equipment, mainly transformers, was developed empirically in tage of this method is that it always provides a diagnosis, with a

Table 2. Training Data for the IEC Ratio Method.

Fault codes for the ratios

as input pattern Output pattern

No. C2H2/C2H4 CH4/H2 C2H4/C2H6 1 2 3 4 5 6 7 8 9

1 0 0 0 1 0 0 0 0 0 0 0 0

2 0 0 1 0 1 0 0 0 0 0 0 0

3 0 2 0 0 0 1 0 0 0 0 0 0

4 0 2 1 0 0 0 1 0 0 0 0 0

5 0 2 2 0 0 0 0 1 0 0 0 0

6 0 1 0 0 0 0 0 0 1 0 0 0

7 1 1 0 0 0 0 0 0 0 1 0 0

8 1 0 2 0 0 0 0 0 0 0 1 0

9 0.5 0 0.5 0 0 0 0 0 0 0 0 1

November/December — Vol. 26, No. 6 29

 Table 3. Rogers Ratio Codes.

Gas ratio Ranges Code

<0.1 5

0.1–1 0
CH4/H2
1–3 1

>3 2

<1 0
C2H6/CH4
>1 1

<1 0

C2H4/C2H6 1–3 1

>3 2

<0.5 0

C2H2/C2H4 0.5–3 1

>3 2

Figure 2. The back-propagation artificial neural network for

low percentage of wrong diagnoses. The triangle representation Rogers dissolved gas analysis method; O1 = no fault, O2 = ther-
also allows an easy visualization of faults evolution with time mal fault <150°C, O3 = thermal fault 150–200°C, O4 = ther-
(Figure 4). Generally, three types of faults are detectable, i.e., mal fault 200–300°C, O5 = general conductor overheating, O6
partial discharge, high and low energy arcing (electrical fault), = winding circulating currents, O7 = overheated joints, O8 =
and hot spots of various temperature ranges (thermal fault) [13], partial discharge, O9 = partial discharge with tracking, O10 =
[14]. In addition to the six zones of individual faults mentioned flashover without power follow-through, O11 = arc with power
in Table 6 (PD, D1, D2, T1, T2, or T3), an intermediate zone, follow-through, O12 = continuous sparking to floating poten-
DT, has been attributed to combined electrical and thermal faults tial.
in the transformer.

Table 4. Fault Types According to Rogers Ratio Method.

No. CH4/H2 C2H6/CH4 C2H4/C2H6 C2H2/C2H4 Fault type

1 0 0 0 0 No fault

2 1–2 0 0 0 <150°C thermal fault

3 1–2 1 0 0 150–200°C thermal fault

4 0 1 0 0 200–300°C thermal fault

5 0 0 1 0 General conductor overheating

6 1 0 1 0 Winding circulating currents

7 1 0 2 0 Core and tank circulating currents, overheated joints

8 5 0 0 0 Partial discharge

9 5 0 0 1–2 Partial discharge with tracking

10 0 0 0 1 Flashover without power follow-through

11 0 0 1–2 1–2 Arc with power follow-through

12 0 0 2 2 Continuous sparking to floating potential

30 IEEE Electrical Insulation Magazine

 Table 5. Training Data for Rogers Ratio Method.

Input pattern Output pattern

No. CH4/H2 C2H6/CH4 C2H4/C2H6 C2H2/C2H4 1 2 3 4 5 6 7 8 9 10 11 12

1 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0

2 0.5 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0

3 0.5 1 0 0 0 0 1 0 0 0 0 0 0 0 0 0

4 0 1 0 0 0 0 0 1 0 0 0 0 0 0 0 0

5 0 0 1 0 0 0 0 0 1 0 0 0 0 0 0 0

6 1 0 1 0 0 0 0 0 0 1 0 0 0 0 0 0

7 1 0 2 0 0 0 0 0 0 0 1 0 0 0 0 0

8 5 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0

9 5 0 0 0.5 0 0 0 0 0 0 0 0 1 0 0 0

10 0 0 0 1 0 0 0 0 0 0 0 0 0 1 0 0

11 0 0 0.5 0.5 0 0 0 0 0 0 0 0 0 0 1 0

12 0 0 2 2 0 0 0 0 0 0 0 0 0 0 0 1

The three sides of the Triangle are expressed in triangular

coordinates (P1, P2, P3) representing the relative proportions of
CH4, C2H4, and C2H2, from 0 to 100 for each gas. To determine
different zones of Duval’s triangular method, a polygon is de-
fined for each zone. As depicted in Figure 4, seven polygons
are needed to define the different fault zones. For the software
implementation of the Duval’s Triangle DGA method, the Java
programming language was used because of its growing impor-
tance and popularity in modern application development. Java is
platform independent, and a lot of free compilers and tools are
available.

Figure 3. A pure back-propagation artificial neural network

for dissolved gas analysis; O1 = no fault, O2 = thermal fault
<300°C, O3 = thermal fault 300–700°C, O4 = thermal fault
>700°C, O5 = partial discharge, O6 = discharge of low energy,
O7 = discharge of high energy. Figure 4. Duval’s Triangle method.

November/December — Vol. 26, No. 6 31

 Table 6. Examples of Faults Detectable by Duval’s Triangle. seven zones contains point R, a Java built-in function called con-
tains() that is applicable for each defined polygon can be used.
Symbol Fault Examples However, if the point is located in the boundary of a polygon, it
Discharges of the cold plasma cannot be recognized using this function.
(corona) type in gas bubbles or To overcome this problem, a small circle with center R and
PD Partial discharges
voids, with the possible formation of radius r can be considered, and the radius r must be selected
X-wax in paper. carefully. In the developed program, the value assigned to r was
Partial discharges of the sparking 5, and there were about 105 points inside it. All points belonging
type, inducing pinholes, carbonized to this circle should be tested to determine which polygon they
punctures in paper. Low energy belong to. Finally, a percentage value can be assigned to each
D1 Discharges of low energy arcing inducing carbonized
polygon according to the number of points that are inside each
perforation or surface tracking of
paper, or the formation of carbon polygon divided by the total number of points of this circle (in
particles in oil. our case 105). This value shows the percentage of the circle in
each of the polygons. If the circle is out of a polygon range, the
Discharges in paper or oil, with
power follow-through, resulting
likelihood will be null. Figure 5 shows the graphical user inter-
in extensive damage to paper or face of the developed software.
D2 Discharges of high energy
large formation of carbon particles
in oil, metal fusion, tripping of the Information Fusion
equipment and gas alarms. Information fusion is the merging of information from dis-
Evidenced by paper turning
parate sources with differing conceptual, contextual, and typo-
T1 Thermal fault, T <300°C brownish (>200°C) or carbonized graphical representations. It is used in data mining and consoli-
(>300°C). dation of data from unstructured or semistructured resources. It
is an area of information technology that aims at making deci-
Thermal fault, 300 < T < Carbonization of paper, formation of
T2
700°C carbon particles in oil. sions based on joint processing of available data and information
obtained from different sources. The information fusion method
Extensive formation of carbon considered in this paper is based on the Dempster–Shafer ev-
T3 Thermal fault, T >700°C particles in oil, metal coloration
(800°C) or metal fusion (>1,000°C).
idence theory. This theory can be considered an extension of
Bayesian inference. This method captures and combines what-
ever certainty exists in the object discrimination capability of
the information sources [15]. It depends on the construction of
To determine the DGA fault according to the seven defined a set, called the frame of discernment (Θ), which contains every
fault zones or polygons, triangular coordinates of R (P1, P2, P3) possible hypothesis. Every hypothesis has a belief denoted by a
are converted to Cartesian coordinates. The location of the point mass probability (m), which assigns beliefs in a hypothesis as
R in a specific zone allows for the identification of the fault cor- shown in equation (1).
responding to the DGA data. To determine which one of the

Figure 5. Example of a diagnosis by dissolved gas analysis based on Duval’s Triangle program module.

32 IEEE Electrical Insulation Magazine

 Table 7. Triangular Coordinates for Duval’s Triangle Zones.
m: ℘ → [0,1] with power set ℘ = 2Θ (1)

Area Points P1 P2 P3 This membership function m has to satisfy the following con-
ditions:
D11 0 0 1

D12 0 0.23 0.77 m(∅) = 0 and m(A) ≥ 0, for all A ⊆ Θ

D1
D13 0.64 0.23 0.13 å m(A) = 1.
D14 0.87 0 0.13 AÍ (2)

D21 0 0.23 0.77 For this application, the frame of discernment (Θ) is Θ = {No
D22 0 0.71 0.29
fault, Thermal fault, Discharge fault}. The second important ad-
vantage of the Dempster–Shafer theory is that it provides a rule
D2 D23 0.31 0.40 0.29 to combine independent evidences m1, m2 into a single more in-
D24 0.47 0.4 0.13
formative hint, m1 ⊕ m2, as follows:

D25 0.64 0.23 0.13 m1 Å m2 (A) = K × å m1 (Ai ) × m2 (B j ),

i, j
DT1 0 0.71 0.29
where Ai Ç B j = A.
(3)
DT2 0 0.85 0.15

DT3 0.35 0.5 0.15 The factor K is a normalization constant that provides the
DT4 0.46 0.5 0.04 å m(A) = 1 condition and can be calculated as follows:
DT AÍ
DT5 0.96 0.0 0.04

DT6 0.87 0.0 0.13 1

K= ,
1 - å m1 (Ai ) × m2 (B j )
DT7 0.47 0.4 0.13 i, j

DT8 0.31 0.4 0.29 where Ai Ç B j = f. (4)

T11 0.76 0.2 0.04
Table 8 shows the different outputs of each DGA method
T12 0.8 0.2 0.0 classified in three categories or states, i.e., no fault, thermal fault,
T1 T13 0.98 0.02 0.0
and discharge fault. Each one of these states for a DGA method
can be calculated through a probability equation. For example,
T14 0.98 0.0 0.02 the implemented IEC method has nine output values that can be
described as O1 to O9, and its three states can be calculated as
T15 0.96 0.0 0.04

T21 0.46 0.5 0.04 No fault = O1

Thermal fault = 1 − ((1 − O2)(1 − O3)(1 − O4)(1 − O5))
T22 0.5 0.5 0.0
T2 Discharge fault = 1 − ((1 − O6)(1 − O7) (1 − O8) (1 − O9)).
T23 0.8 0.2 0.0
The probability of each state is calculated from Table 8 and
T24 0.76 0.2 0.04
used as input for the information fusion process to calculate
T31 0.0 0.85 0.15 three mentioned power transformer states according to the algo-
T32 0.0 1 0.0
T3
T33 0.5 0.5 0.0 Table 8. Classification of Dissolved Gas Analysis Methods Results.

T34 0.35 0.5 0.15 Method No fault Thermal fault Discharge fault

PD1 0.98 0.02 0.0 IEC O1 O2 , O3 , O4 , O5 O6 , O7 , O8, O9

PD PD2 1 0.0 0.0 Rogers O1 O2 , O3 , O4 , O5 , O6 , O7 O8 , O9 , O10 , O11, O12

PD3 0.98 0.0 0.02 Pure ANN1 O1 O2 , O3 , O4 O5 , O6 , O7

Duval’s Triangle — T1, T2, T3, DT D1, D2, PD, DT

ANN = artificial neural network.

1

November/December — Vol. 26, No. 6 33

 rithm given in Figure 6. After identification of the three states,
the thermal fault and discharge fault can be further subdivided
to the Duval individual faults reported in Table 7 (PD, D1, D2,
T1, T2, or T3) using output of each DGA method (Table 9). This
provides more detail information about the transformer state.
Each detailed fault for a DGA method can be calculated again
through a probability equation. For example, the implemented
IEC method has nine output values that can be described as O1
to O9, and its six fault probability values can be calculated as
follows:

PD = 1 − ((1 − O6)(1 − O7))

D1 = O8
D2 = O9
T1 = 1 − ((1 − O2)(1 − O3))
T2 = O4
T3 = O5.

After computing the likelihood of each fault, the information

fusion algorithm is used again for fusing these values and calcu-
Figure 6. The fusion process. ANN = artificial neural network. lating the detailed fault information. The frame of discernment
(Θ) in this case is Θ = {PD, D1, D2, T1, T2, T3}.

Multiagent Technology
The term agent or software agent is usually defined as a soft-
ware artifact that acts independently to perform tasks without
the intervention of an operator. Agents are independent entities
enabling in-between interactions. Multiagent systems can be
used to model complex systems and can include agents having
common or conflicting tasks. To ensure interoperability of inde-
pendent agents, a standard specification is needed. The Founda-
tion for Intelligent Physical Agents (FIPA) is an international
organization devoted to providing open specifications ensuring
interoperability among agents and agent-based applications
[16]. The FIPA specifications establish a logical reference model
named the FIPA agent management reference model, depicted
in Figure 7, to provide a framework in which agents can be de-
ployed and operate.
The logical reference model is established for agent creation,
registration, communication, location, migration, and suppres-
sion. The reference model includes a set of logic-based entities,
such as
• A message transport service (MTS) to establish com-
munication between agents registered on different plat-
forms,
• A directory facilitator (DF), which provides a yellow-
pages-like service for the available agents. Agents may
register their services with the DF or query the DF to
gain information on other agents,
• An agent runtime environment for defining the notion
of agenthood used in FIPA and an agent lifecycle,
• An agent platform (AP) for deploying agents in a phys-
ical infrastructure, and
• An agent management system (AMS) acting as a white-
pages service for supervisory control over access to the
Figure 7. Foundation for Intelligent Physical Agents (FIPA) agent platform. It keeps an index of all agents currently
agent management reference model sequence. registered with the agent platform.

34 IEEE Electrical Insulation Magazine

Figure 8. A part of the developed Web Ontology Language (OWL) ontology model.

An agent’s skill to communicate provides the fundamental agents communicate by exchanging messages in compliance
means to present social behavior and promotes the design of with the FIPA-ACL language [9].
multiagent systems. A multiagent system comprises a group of
agents that interact with each other by exchanging messages, Representing and Manipulating Knowledge
which are defined in a specific agent communication language The control, monitoring, and diagnosis of power devices in-
(ACL). A standard for an ACL that has been adopted by FIPA is volves a huge amount of data recorded by different sensors and
called FIPA-ACL, which is composed of a message body and a measuring devices. Protocols such as IEC 61850, CIM [17], and
message header. IEC 61400-25 attempt to define some standards for data ma-
The ACL message header has abilities such as request or in- nipulation. These protocols have enhanced the interoperability
form and is used to express the sender’s intention in sending the of different devices and communities. However, an effective
message to the receiver. The header also contains the address of way to manipulate and interpret the collected data for use by
the sender and receiver and some other meta information about maintenance planners or engineers is still an important and ac-
the message’s content. The body of the ACL message contains tive research topic. The adoption of Semantic Web technologies
the message’s content. The content is coded in terms of an ontol- appears to be a sound and promising solution for representing,
ogy that is defined for a specific domain. reasoning, and querying the power system data.
Ontology, which provides a semantic framework for knowl-
edge management, is being considered as a very significant as-
Agent-Based Architecture pect in many applications. Ontology is defined as “an explicit
The proposed DGA diagnostic system was developed using
specification of a shared conceptualization of a domain.” The
the Java Agent Development (JADE) software framework. JADE
first reason why ontology should be used is knowledge repre-
is a software framework that highly simplifies the implementa-
sentation. In other words, ontology contains a set of definitions
tion of multiagent systems through middleware that is in line
and terms (classes, relations, functions, and constants) allowing
with the FIPA specifications. This is a library of classes that soft-
the representation of the hierarchical structure of knowledge in a
ware developers can use or extend while creating agents and a
specific domain and the interplay between different elements.
set of functionalities independent of the specific application and
The possibility of using a knowledge domain is also a major
which simplify the implementation of distributed applications
asset of ontology. Ontology developed for a given application
exploiting the software agent abstraction. JADE implements this
can be used in another one. Moreover, several existing ontolo-
abstraction over a well-known object-oriented language, Java,
gies can be integrated in a larger ontology. In addition, a general
providing a simple and friendly graphical user interface. JADE
ontology can be used and extended to describe a specific domain
of interest [18].

November/December — Vol. 26, No. 6 35

 power plant personnel, and field experts are involved in using
and manipulating this information.
Web Ontology Language (OWL), developed under the man-
agement of the World Wide Web Consortium, is currently the
most popular ontology representation language. It is a part of
the growing stack of World Wide Web Consortium recommen-
dations related to the semantic web. An important feature of the
OWL vocabulary is its extreme richness for describing relation-
ships between classes, properties, and individuals [20].
We tried to use a part of the CIM model for power trans-
formers and to extend it to cover our requirements for develop-
ing an OWL ontology model for a DGA diagnosis agent-based
system. Figure 8 shows some part of this OWL ontology that
will support knowledge representation, exchange, and sharing.
Agents need to be able to communicate with users, with system
resources, and with each other if they need to cooperate, collab-
orate, and negotiate. The FIPA-ACL is currently the most used
and studied agent communication language. A content language,
such as FIPA-SL or FIPA-KIF, should be used to express the
content of the message. The FIPA-SL and FIPA-KIF languages
are powerful but different from ontology description in the se-
mantic web area such as OWL. The authors tried to use OWL as
a content language within the FIPA-ACL messages. To facilitate
Figure 9. The framework of the transformer dissolved gas analy- communication and data sharing among multiple agents and to
sis (DGA) multiagent system. have a seamless knowledge representation, the whole system,
the messages in OWL format can be encapsulated into FIPA-
ACL messages.
Another goal in using ontology is to share common under-
standing of the information structure among people, software Architecture of the Software System
agents, and heterogeneous and widely spread application sys- The agent-based architecture is shown in Figure 9. Four lay-
tems [19]. This criterion is very useful, especially for the power ers were proposed for the agent-based system. There are some
network industry, in which different parties such as substations, agents in each layer that should have interaction with the agents

Figure 10. Screenshot of a part of the software graphical user interface.

36 IEEE Electrical Insulation Magazine

Figure 11. Agent communication.

in a neighboring layer to complete their tasks. The AdminAgent information fusion algorithm. The DiagnosticAgent will extract
has interaction with other agents in other layers in the capacity the content of this ACL message that is in OWL and add it into
of controller. The software system provides a friendly graphi- its knowledge model for further processing. Representing system
cal user interface (Figure 10). The system OWL ontology will knowledge using OWL ontology provides us with the advantage
change according to the user’s settings. of querying using some query language such as RDQL (RDF
Data Query Language) [21]. A part of agent communication in
Agent Communication and Agent Knowledge Base this agent-based system is shown in Figure 11. It shows the FI-
To ease communication and data sharing between multiple PA-ACL messages required to request a DGA measurement and
agents, FIPA-ACL messages were used. For example, for comb- perform the necessary operations such as sending measured data
ing diagnostic results from different DGA agents, Diagnosti- and acknowledging a received message.
cAgent should receive the result of each DGA agent and apply
Results and Discussion
To verify the information fusion process, a database of about
210 known DGA test data from IEC 60599 [22] and [12], [23]–
Table 9. Further Dissolved Gas Analysis Methods Result Classification.
[25] was used and tested by this system. Only 20 data of this
Method PD D1 D2 T1 T2 T3 database are reported in Table 10. Table 11 compares the diag-
nostic reliability of the combined method with four single DGA
IEC O6 , O7 O8 O9 O2 ,O3 O4 O5
methods. It can be seen that the combined method depicts higher
Rogers O8 , O9 O10 O11 O2, O3 ,O4 — — accuracy than the single DGA methods.
As mentioned before, developing a power transformer moni-
Pure ANN1 O2 O3 O4 O5 O6 O7
toring and diagnosing system is a complex task. Some of the
Duval’s Triangle PD D1 D2 T1 T2 T3 requirements that should be fulfilled in this system were also
mentioned. The proposed software architecture takes into ac-
1
ANN = artificial neural network.
count the requirements emphasized in the introduction:

November/December — Vol. 26, No. 6 37

 Table 10. Dissolved Gas Analysis Test Data.

No. H2 CH4 C2H2 C2H4 C2H6 TCG Known fault

1 980 73 0 12 58 1123 Discharge

2 650 53 0 20 34 757 Discharge

3 17000 110000 16000 89000 84000 316000 Thermal

4 800 1393 3000 2817 304 8314 Discharge

5 4906 8784 9671 9924 1404 34689 Discharge

6 57 77 0 21 19 174 Thermal

7 22 40 1 6 36 105 Thermal

8 64 23 1 83 19 190 Thermal

9 274 27 97 33 5 436 Discharge

10 1 27 1 4 49 82 Thermal

11 1450 940 61 322 211 2984 Thermal

12 37800 1740 8 8 249 39805 Discharge

13 9340 995 7 6 60 10408 Discharge

14 36036 4704 10 5 554 41309 Discharge

15 40280 1069 1 1 1060 42411 Discharge

16 1270 3450 8 1390 520 6638 Thermal

17 3420 7870 33 6990 1500 19813 Thermal

18 360 610 9 260 259 1498 Thermal

19 960 4000 6 1560 1290 7816 Thermal

20 24700 61000 1560 42100 26300 155660 Thermal

• The combination of different DGA diagnostic methods • Remote supervision by experts, which can be imple-
integrable in separate software agent. Information fu- mented effectively using the proposed agent-based sys-
sion helps in processing the optimal diagnostic; tem.
• Adaptability to new environments. Agents are modular In the proposed system, KnowledgeSharingMobileAgent can
and proactive objects, which makes them suitable for move between different diagnosing systems and distribute a new
changing environments; and DGA measurement with its known fault information. A diag-
nostic result or other information necessary to the remote expert
can be transferred by a mobile agent in an OWL format that is
defined by common power transformer ontology as depicted in
Figure 8.
Table 11. Comparison of Diagnosis Results of Different Methods.

Method Diagnosis accuracy (%) Conclusions

The development of condition-monitoring and software-
Pure ANN1 80
based diagnosis tools is central to the implementation of effi-
Rogers 85 cient maintenance management strategies for many engineering
applications including power transformers. Due to the increasing
IEC Ratio 75
number of methods and techniques and acquired data involved
Duval’s Triangle 85 in the process, the need for more flexible systems is expected.
In this contribution, an intelligent agent-based expert system
Combined method 100 was implemented. By combining DGA diagnostic methods, the
ANN = artificial neural network.
1 risk of a mistaken diagnosis was reduced, and the likelihood of

38 IEEE Electrical Insulation Magazine

identifying faults before they lead to transformer failures was to Creating Your First Ontology,” Tech. Rep., Stanford, CA: Stanford Uni-
maximized. versity, Mar. 2001.
[19] M. A. Musen, “Dimensions of knowledge sharing and reuse,” Comput.
Multiagent systems appear to be a promising approach to Biomed. Res., vol. 25, pp. 435–467, 1992.
support a new generation of monitoring and diagnosis systems [20] D. Gasevic, Model Driven Architecture and Ontology Development.
and to overcome the complexities. Agent-based systems are able Berlin: Springer, 2006.
to generate complex system-level performance from relatively [21] S. Powers, Practical RDF. Sebastopol, CA: O’Reilly & Associates Inc.,
2003.
simple individual agents. [22] M. Duval and A. Pablo, “Interpretation of gas-in-oil analysis using new
Because ontology engineering in power systems is still a IEC publication 60599 and IEC TC10 databases,” IEEE Elect. Insul.
young domain, there is still a real need for advanced research. Mag., vol. 17, no. 2, pp. 31–41, Mar./Apr. 2001.
Integrating OWL ontology into multiagent systems can generate [23] N. Gao, G. J. Zhang, Z. Qian, Z. Yan, and D. H. Zhu, “Diagnosis of DGA
based on fuzzy and ANN methods,” in Proceeding of 1998 International
more powerful monitoring and diagnosis systems for the power Symposium on Electrical Insulating Materials, 1998, pp.767–770.
system domain. The proposed software system can be extended [24] O. Vanegas and Y. Mizuno, “Diagnosis of oil-insulated power apparatus
for other diagnostic approaches on power transformers, such as by using neural network simulation,” IEEE Trans. Dielectr. Electr. Insul.,
thermal model and partial discharge, to present an agent-based vol. 3, pp. 290–299, 1996.
[25] Y. Zhang, X. Ding, Y. Liu, and P. J. Griffin, “An artificial neural network
solution for complete diagnosis and monitoring of power trans- approach to transformer fault diagnosis,” IEEE Trans. Power Del., vol.
formers. 11, no. 4, pp. 1836–1841, Oct. 1996.

References
[1] Y. Liang, Z. Wang, and Y. Liu, “Power transformer DGA data process-
ing and alarming tool for on-line monitors,” in IEEE/PES Power Systems Alireza Akbari received his BSc and MSc
Conference and Exposition, 2009, pp. 1–8. degree in 1997 and 1999 in computer en-
[2] D. Arvind, S. Khushdeep, and K. Deepak, “Condition monitoring of gineering from Azad University of Najafa-
power transformer: A review,” in Proceedings of the IEEE/PES Transmis- bad in Iran and his second MSc degree in
sion and Distribution Conference and Exposition, 2008, pp. 621–624.
[3] S. Kovacevic, N. Dominelli, and B. Ward, “Transformer life extension 2004 in distributed computing system from
by continuous online oil conditioning,” Proceedings of the Electrical Brunel University in England. Currently he
Insulation Conference and Electrical Manufacturing and Coil Winding is a PhD student at the Schering Institute
Technology Conference, 2003, pp. 271–275. at the University of Hannover, Germany.
[4] T. K. Saha, “Review of modern diagnostic techniques for assessing
insulation condition in aged transformers,” IEEE Trans. Dielectr. Electr. His main research interests are monitoring
Insul., vol. 10, no. 5, pp. 903–917, 2003. and diagnostic software systems for high-
[5] I. Fofana, A. Bouaicha, M. Farzaneh, J. Sabau, and D. Bussières, “Decay voltage apparatus, intelligent and agent-based systems, and
products in the liquid insulation of power transformers,” IET Electr. computer applications in power systems.
Power Applic., vol. 17, no. 2, pp. 458–464, April 2010.
[6] IEEE Guide for the Interpretation of Gases Generated in Oil-Immersed
Transformers—Section 1.2 Limitations, IEEE C57.104, 1991.
[7] L. Lewand, “Techniques for interpretation of data for DGA from trans-
formers,” presented at the Doble Engineering, IEEE PES Conference, Alireza Setayeshmehr was born in 1969
http://www.transformerscommittee.org/info/F06/F06-DGA.pdf, 2006.
[8] R. Roger, “IEEE and IEC codes to interpret incipient faults in transform- in Iran. He received his BSc degree in
ers using gas in oil analysis,” IEEE Trans. Electr. Insul., vol. 13, no. 5, pp. 1993 from Ferdosi University and his
348–354, 1978. MSc degree in electrical engineering in
[9] “IEEE Guide for the Interpretation of Gases Generated in Oil-Immersed 1996 from the Tarbiat Modarres Univer-
Transformers,” IEEE Power and Energy Society, 1992.
[10] Interpretation of the Analysis of Gases in Transformers and Other Oil- sity in Iran. Since 1996 he has worked as
Filled Electrical Equipment in Service, IEC 60599 Ed. 1.0, Mar. 1999. a member of the academic staff of Cham-
[11] M. Duval and J. Dukarm, “Improving the reliability of transformer gas- ran University, Ahvaz, Iran. Since October
in-oil diagnosis,” IEEE Electr. Insul. Mag., vol. 21, no. 4, pp. 21–27, July/ 2003, he has worked as a PhD student at
Aug. 2005.
[12] D. Sarma and G. N. S. Kalyani, “ANN approach for condition monitor- the Schering Institute of High Voltage Techniques and Engineer-
ing of power transformers using DGA,” in TENCON 2004. 2004 IEEE ing at the University of Hannover, Germany. His main research
Region 10 Conference, 2004, pp. 444–447. interests are monitoring and diagnostics of high-voltage power
[13] M. Duval, “Fault gases formed in oil-filled breathing EHV power trans- transformers.
formers: The interpretation of gas analysis data,” presented at the IEEE
PAS Conference, Paper no. C 74 476-8, 1974.
[14] A. Akbari, A. Setayeshmehr, H. Borsi, and E. Gockenbach, A Software
Implementation of the Duval’s Triangle Method. Vancouver, Canada:
ISEI, 2008. Issouf Fofana (M ’05, SM ’09) received
[15] L. A. Klein., Sensor and Data Fusion: A Tool for Information Assessment
and Decision Making. Bellingham, WA: SPIE Press, 2004. his electro-mechanical engineering degree
[16] F. Bellifemine, G. Caire, and D. Greenwood, Developing Multi-Agent in 1991 from the University of Abidjan
Systems With JADE, Wiley, 2007. (Côte d’Ivoire) and his MS and PhD de-
[17] Energy Management System Application Program Interface (EMS API) grees from the Ecole Centrale de Lyon,
- Part 301: Common Information Model Base, IEC 61970-301 Ed. 1.0,
Nov. 2003. France, in 1993 and 1996, respectively. He
[18] N. F. Noy and D. L. McGuinness, “Ontology Development 101: A Guide was a postdoctoral researcher in Lyon in

November/December — Vol. 26, No. 6 39

 1997 and at the Schering Institute of High Voltage Engineer- of the Ministry of Energy in Iran. Currently he is professor of
ing Techniques, University of Hanover, Germany, from 1998 to high-voltage engineering and academic director at the Schering
2000. He was a Fellow of the Alexander von Humboldt Stiftung Institute of the University of Hanover. He is a member of VDE,
from November 1997 to August 1999. He joined the Univer- different CIGRE Task Forces, and national working groups for
sité du Québec à Chicoutimi (UQAC) as an associate researcher standardization.
in 2000. He is currently associate professor at UQAC, Quebec,
Canada. Since September 2005, Fofana has held the Canada Re-
search Chair, tier 2, of insulating liquids and mixed dielectrics
for electrotechnology (ISOLIME). He is registered as a profes- Ernst Gockenbach (M ’83, SM ’88, FM
sional engineer in the province of Quebec and is currently a ’01) received his Diplom degree in 1974
member of the Technical Committee of the IEEE CEIDP and and PhD in 1979 from the Technical Uni-
the International Advisory Committee of the IEEE ICDL. He is versity of Darmstadt. From 1979 to 1982,
a member of the IEEE Task Force on atmospheric icing perfor- he worked at the High Voltage Test Labo-
mance of line insulators and a member of the ASTM D27 Task ratory of the Switchgear Factory, Siemens
Group to develop a test method for partial discharge inception AG, Berlin, Germany, and was responsible
voltage. He has authored or coauthored over 130 scientific pub- for the High Voltage Outdoor Test Field.
lications and holds three patents. From 1982 to 1990, he worked with E.
Haefely AG in Basel, Switzerland, as chief
engineer for high-voltage test equipment. Currently he is profes-
sor of high-voltage engineering and director of the Schering In-
Hossein Borsi received his Dipl.-Ing. de- stitute of High Voltage Technique at the University of Hannover.
gree in electrical engineering in 1972, He is member of VDE and CIGRE, chairman of CIGRE Study
Dr.-Ing. degree in 1976, and habilitation Committee D1 Materials and Emerging Test Techniques, and
with “‘Venia legendi’ for Hochspannungs- a member of national and international Working Groups (IEC,
meßtechnik” in 1979 to 1985. He was a IEEE) for Standardization of High Voltage Test and Measuring
professor of power engineering at the Uni- Procedures.
versity of Mashhad, Iran, from 1980 to
1982, dean of the Faculty of Engineering
from 1981 to 1985, and scientific adviser

40 IEEE Electrical Insulation Magazine

View publication stats