IET Generation, Transmission & Distribution

Research Article

Using dissolved gas analysis results to detect ISSN 1751-8687

Received on 5th January 2017
Revised 23rd February 2017
and isolate the internal faults of power Accepted on 8th March 2017
E-First on 3rd July 2017
transformers by applying a fuzzy logic doi: 10.1049/iet-gtd.2017.0028
www.ietdl.org

method
Masoud Noori1, Reza Effatnejad1, Payman Hajihosseini1
1Department of Electrical Engineering, Karaj Branch, Islamic Azad University, Alborz, Iran
E-mail: p_hajihosseini@kiau.ac.ir

Abstract: One of the most important tools for condition monitoring is the gas chromatography test of transformer oil, which is
known as dissolved gas analysis (DGA). In this research, the DGA results of >3000 power transformers operating in Iran's
power grid were carefully studied and from among them, the results related to transformers suspicious of being faulty were used
to validate the fault detection accuracy of the presented fuzzy inference system (FIS). In most of the previously published
papers, the detection and isolation of transformer faults has been based on one or two of the following parameters: absolute
concentrations of free and dissolved gases in transformer oil, total dissolved combustible gases, total combustible gases, ratios
of some gases to each other, and the rates of gas increase. However, in this research, most of these parameters have been
used for fault detection and isolation, according to the IEC 60599 standards. Also, no attempt has been previously made to
detect the decomposition of insulation papers of transformers; but the presented FIS is able to detect this fault as well. The
overall performance accuracy of the presented system is F1 = 91.2%, which seems to be a suitable value.

1 Introduction available [3]. In some of the research works, the accuracy of

conventional techniques in detecting the internal faults of power
Power transformers are one of the most important and expensive transformers has been lower or much lower than the performance
equipment used in power networks. Thus, it is crucial to maintain of intelligent systems [4–6].
these devices and to take care of them in the best possible way. For Also, in the fuzzy methods presented in similar research works,
several decades, time-based maintenance operations have been almost no attempt has been made to detect the molecular
performed on power transformers by taking them offline. However, decomposition of transformer paper (fault ‘P’) along with the other
since two decades ago, power transformers have been maintained faults. For example, in [7] which has used fuzzy approach for fault
through online condition monitoring which is named condition- diagnosis, there are some basic problems. According to this paper,
based maintenance. For this purpose, a power transformer in detecting decomposition of insulation paper fault (code F1),
undergoes online tests as a way of monitoring its condition [1]. only guesswork and conjecture has been used and the transformer
One of these tests is the chemical gas chromatography test of has not been opened. Also, the carbon dioxide (CO2)-to-carbon
transformer oil. In this test, several oil samples are taken from a
monoxide (CO) ratios for some samples are above 3; which,
transformer at different times, and the concentrations of dissolved
according to IEC standards, do not indicate a cellulose paper
gases in them are measured by a qualified technician. Based on the
decomposition fault in the relevant transformers. In another work
interpretation of these values by the experts, the internal faults of
in [8], the faults of oil heating and deterioration and paper
the transformer are determined. These faults include the electrical
decomposition have been listed under one code (code F1:
faults [partial discharge (PD), electrical discharges with low energy
cellulosic and/or oil decomposition, overheated cellulose and/or
(D1), and electrical discharges with high energy (D2)], thermal
oil) and have not been segregated. This means that it is not clear
faults [hot spots with a temperature of <300°C (T1), hot spots with
for the author whether the mentioned fault is the deterioration of oil
a temperature of >300 and <700°C (T2), and hot spots with a
or the decomposition of paper.
temperature of >700°C (T3)], a combination of electrical and
However, in this research, the fault of molecular decomposition
thermal faults (DT), and the molecular decomposition and gradual
of transformer paper has been diagnosed by the fuzzy inference
deterioration of a transformer's insulation paper (P). The common
system (FIS) along with the other faults.
methods used for interpreting the dissolved gas levels include the
Most of the papers on this subject are weak in terms of the
Dornenburg, Rogers, key gases, Duval triangle, scoring system,
number and variety of empirical data they use [9]. For example, in
IEEE C57.104, and IEC 60599 methods and guidelines. The IEC
[10], the data of only 11 transformers have been considered.
60599 standard has been used in this research. The mentioned
However, to evaluate the accuracy of the performance of the
approaches consider one or two of the following values in their
designed FIS in this work, the results of gas chromatography tests
interpretations: absolute concentrations of free and dissolved gases
conducted during the years of 2001–2016 on >3000 power
in transformer oil, total dissolved combustible gases (TDCG) [2],
transformers installed in Iran's national power grid system have
total combustible gases (TCG) [2], ratios of some gases to each
been studied. From this data set, the dissolved gas analysis (DGA)
other, and the rates of gas increase. One of the advantages of this
results of 185 power transformers have been selected and
research over similar works is that, according to the IEC 60599
examined.
regulations, it has used most of the above-mentioned parameters in
In most similar research works, absolute concentrations of free
diagnosing the internal faults of power transformers.
and dissolved gases, gas ratios, or the TDCG or TCG values have
In evaluating the above-mentioned parameters and detecting the
been used, but the rates of gas increase (maybe the most
transformer faults, intelligent systems are suitable substitutes for
determining parameter) have been ignored. Perhaps, the main
experts. One of the greatest drawbacks of methods that rely on
reason for this is the lack of regular condition monitoring of
experts is the probability of error in the expert's judgment and
suspicious transformers during several consecutive years and,
diagnosis. Also, there are few experts, and they are not always
IET Gener. Transm. Distrib., 2017, Vol. 11 Iss. 10, pp. 2721-2729 2721
© The Institution of Engineering and Technology 2017
consequently, the lack of appropriate data in this regard. For
example, only the dissolved gas concentrations and the ratios of
some gases have been used in [10]. However, in this research, in
addition to the absolute dissolved gas concentrations and the gas
ratios, the annual rates of gas increase have also been used to
design the FIS. This shows the comprehensiveness of the data used
in this research.
Keep in mind that a FIS is sensitive to the quality of the devised
fuzzy rules [11]. In fact, if the fuzzy rules are the more accurate
and comprehensive, the more accurate fault diagnosis by the
designed FIS will be. One of the advantages of this work is the
formulation of 221 fuzzy rules that adequately cover all the
conditions and requirements mentioned in the IEC 60599
guidelines.
Following Section 1, the gas chromatography test has been
explained in Section 2. In Section 3, the tables and criteria
presented in the IEC 60599 standard and the procedure for
diagnosing and isolating the internal faults of power transformers
based on them have been described. The specifications of the
designed FIS have been presented in Section 4. The data and the
empirical results used in this research have been fully introduced in
Section 5. In Section 6, the confusion matrix and the F1 score have
been introduced; and, by applying this criterion, the degree of
accuracy of the designed FIS has been determined. Finally, the
conclusions and some suggestions on the subject have been
presented in Section 6.
2 Gas chromatography (or equally DGA) test
Power transformer oil is usually of mineral type and is considered
as an electrical insulation. This fluid is also used for heat transfer
by convection cooling [12, 13]. When an electrical or thermal fault
occurs in the transformer, the oil molecules adjacent to the fault
location decompose and release some gases [14]. Some of these
gases are dissolved in the oil and some remain free. In the DGA
test, the transformer oil must be sampled first according to the IEC
60567 guidelines and then, complying with the relevant
transportation standards, the samples should be taken to an oil
laboratory. In the laboratory, the gases are separated by a gas
chromatography device, and the concentrations of free and
dissolved gases are measured and recorded [15]. Based on the
conditions of a transformer, the sampling and testing process is
repeated several times during appropriate time periods and the
results are recorded. In the next step, by applying one of the
commonly used methods, the DGA test results are analysed by an
expert technician, and the health of the transformer is evaluated. If
a faulty transformer is detected, the type of its internal fault should
also be diagnosed. Sometimes, the transformer fault is a
combination of several faults; which may be detected by a careful
expert. The gases released due to the internal transformer faults
are: CO2, CO, hydrogen (H2), methane (CH4), acetylene (C2H2),
ethylene (C2H4), and ethane (C2H6). All of these gases are
combustible except CO2. There are different methods for
diagnosing the internal faults of a transformer by examining the
DGA data related to the concentrations of free and dissolved gases,
TDCG, TCG, gas ratios and the rates of gas increase. A general list
of faults that can be detected by these techniques has been given in
Table 1.
Table 1 General list of detected faults by interpretation
techniques of DGA test results
Number Fault abbreviation Description
1 P paper decomposition
2 PD partial discharge
3 D1 discharges of low energy
4 D2 discharges of high energy
5 T1 thermal fault <300°C
6 T2 300°C< thermal fault <700°C
7 T3 thermal fault >700°C
8 DT discharges and thermal
2722
The commonly used methods for interpretation of gas values
include the method of Dornenburg, Rogers, key gases, Duval
triangle, scoring system, and the IEEEC57.104 and IEC 60599
standards. In this investigation, the IEC 60599 guidelines have
been used; which are described in the next section.
3 IEC 60599 guidelines
This method consists of the following steps:
Step 1: If the absolute concentration of at least one of the dissolved
gases is higher than the normal values listed in Table 2, and at the
same time, the rate of increase of the same gas exceeds the normal
rates given in Table 3, it will mean that the transformer is faulty;
otherwise, the transformer will be considered as normal and
healthy. In fact, fault detection is carried out in this step. It should
be mentioned that the concentrations and rates of increases of C2H2
gas are different in transformers with communicating on-load tap-
changer (OLTC) where the tap-changer is immersed in the main oil
tank and communicates with it, than in transformers without the
OLTC mechanism. These differences have been shown in Tables 2
and 3.
Step 2: After detecting the faultiness of a transformer, Table 4
should be consulted to find the fault type. All the three conditions
listed in the columns of this table must be fulfilled for the
mentioned gas ratios in order to identify the corresponding fault
type.
Step 3: If the type of fault is not isolated through Table 4, the
graphical rectangles or cubes in Fig. 1 should be consulted to find
the fault type.
Note that no explanation has been given in the first, second and
third steps regarding the ‘DT’ fault.
Step 4: Finally, if the fault type is not diagnosed by even referring
to the graphical rectangles or cubes in Fig. 1, the Duval triangle in
Fig. 2 should be consulted. It should be mentioned that the type of
transformer fault will definitely be detected by using this triangle.
The mentioned ranges for the faults in the Duval triangle have been
tabulated in Table 5.
It should be mentioned that no explanation has been given in
the four mentioned steps regarding fault ‘P’. The IEC 60599
guidelines have emphasised that, if the ratio of CO2 concentration
to CO concentration is <3 and the ratio of the annual rate of CO2
increase to that of CO is also <3, it will mean that the insulation
paper of transformer is being gradually decomposed and destroyed
(P) [16].
4 Fuzzy inference system
It is beneficial to use intelligent systems in fault detection
operations in which there is a possibility of uncertainty and error.
In this research, fuzzy logic has been employed for deciding on the
health and wellbeing of transformers and classifying their internal
faults. Fig. 3 illustrates the block diagram of a FIS [17, 18].
As is observed in Fig. 3, a FIS includes the following
components:
i. Fuzzifier: The goal of fuzzification is to map a set of real data
to a set of values in the [0, 1] interval by means of membership
functions [18, 19]. In this research, all inputs have been
defined based on Tables 2–4 and Figs. 1 and 2. Therefore, real
normal and abnormal values of the inputs have been extracted
from these tables and then mapped and fuzzified using
membership functions to normal and abnormal ranges in [0, 1]
interval.
ii. Fuzzy inference engine (decision-making unit): By performing
special mathematical operations, this unit uses fuzzy rules to
convert fuzzified inputs to fuzzified outputs [20].
iii. Defuzzifier: The aim of defuzzification is to map a set of data
in the [0, 1] interval to a set of real data by using membership
functions [18, 19].
IET Gener. Transm. Distrib., 2017, Vol. 11 Iss. 10, pp. 2721-2729
© The Institution of Engineering and Technology 2017
Fig. 1 Graphical rectangles or cubes [16]
Note 1: The arrow indicates increasing temperature; Note 2: The axes are limited to values of ten for clarification of presentation, but actually extend to unlimited values; Note 3:
Each of the cases defined in Table 3 is represented by a volume or ‘box’ on the 3D graphic; and Note 4: The coordinates of each box are the same as in Table 3. It is more convenient
to use this representation with the help of a computer software package

In the sequel, more specifications of the FIS used for fault

detection and isolation in this work have been explained. This FIS
has been designed by using the MATLAB software, and it has the
following characteristics:

Fig. 2 Duval triangle [16]
Where %C2H2 = ((100x)/(x + y + z)) for x = [C2H2] in microlitres per litre (µl/l);
%C2H4 = ((100y)/(x + y + z)) for y = [C2H4] in µl/l; and %CH4 = ((100z)/(x + y + z))
for z = [CH4] in µl/l
iv. Fuzzy rules: Conditional propositions in the form of ‘IF…
THEN’ that use operators such as ‘AND’, ‘OR’ and ‘NOT’ to
formulate the governing rules of the FIS [21]. Fuzzy rules are
generated based on the knowledge of experts, and by using the
linguistic variables [22]. The linguistic variables reduce the
complexity of the fuzzy models [23]. In this research, all of the
rules have been formulated according to Tables 2–4 and Figs. 1
and 2 and based on the normal and abnormal ranges of
absolute concentrations of gases, annual growth of gases and
the gas ratios.
i. The Mamdani model has been selected for the fuzzy inference
engine in this research. The designed FIS consists of 25 inputs,
according to the tables and criteria stated in the IEC 60599
guidelines (Tables 2–4 and Figs. 1 and 2). This system has one
output with nine different states; eight states indicate the faults
(Table 1), and one state points out the health of the transformer.
The different inputs and output states have been listed in
Tables 6 and 7.
ii. In this research, triangular functions have been selected as the
membership functions for the inputs and outputs of the FIS.
From among these different triangular, bell-shaped, trapezoidal
and other membership functions, the best results have been
obtained by means of the triangular functions.
iii. The Zadeh method has been employed for the implication
(AND) and aggregation (OR) operations; which, respectively,
means the ‘Min’ and ‘Max’ operations [24, 25].
iv. The centroid technique has been used for defuzzification.
v. In this system, according to the IEC 60599 guidelines, 221
‘IF…THEN’ fuzzy rules have been set up for the FIS, using
the ‘AND’ operator. Due to the same significance of all these
rules, a weight of 1.0 has been allocated to each rule.
Here, an example is given to explain how the designed fuzzy
system works and the governing rules of this system. Take
transformer T, with a nominal power of 160 MVA, made in 2008
by Iran Transfo Co. and installed in the 230/63 kV power
transmission substation of Boein Zahra, a subsystem of Zanjan
Local Electricity Company. The data obtained from two periods of

Table 2 Normal concentrations of dissolved gases in transformer oil, in µl/l [16]

C2H2 H2 CH4 C2H4 C2H6 CO CO2
all transformers 50–150 30–130 60–280 20–90 400–600 3800–14,000
no OLTC 2–20
communicating OLTC 60–280

Table 3 Normal rates of dissolved gas increase in transformer oil, in µl/l [16]
C2H2 H2 CH4 C2H4 C2H6 CO CO2
all transformers 35–132 10–120 32–146 5–90 260–1060 1700–10,000
no OLTC 0–4
communicating OLTC 21–37

IET Gener. Transm. Distrib., 2017, Vol. 11 Iss. 10, pp. 2721-2729 2723
© The Institution of Engineering and Technology 2017
 the mentioned transformer is the presence of a hot spot or hot spots
with a temperature <300°C (i.e. fault T1). Since the fault has been
diagnosed in the first step, we need not proceed to other steps. In
the fuzzy system designed for fault detection, for conditions such
as those of the mentioned transformer, Rule # 54 in the designed
system's rules set has been defined as follows:

Rule 54: If (C2H6 is abnormal) and (Rate_C2H6 is abnormal) and

(CH4 to H2 is bigger than 1) and (C2H4 to C2H6 is smaller than 1)
then (fault is T1).

According to the normal range of C2H6in Table 2, in the triangular

function considered for this input, interval 0–90 has been defined
as the ‘normal’ range and interval 90–20,090 has been defined as
the ‘abnormal’ range.
According to the normal range of annual growth of C2H6 in
Table 3, in the triangular function considered for this input,
Fig. 3 Block diagram of a FIS [17, 18] intervals −19,910 to 90 and 90–20,090 have been defined as the
‘normal’ and ‘abnormal’ ranges, respectively.
DGA tests performed on the oil of this transformer have been According to the defined ranges for gas ratios in Table 6, in the
tabulated in Table 8. triangular function considered for CH4 to H2 input, intervals 0–0.1,
Based on the first step of IEC 60599 guidelines and by 0.1–0.5, 0.1–1 and 1–2000 have been defined as the ‘smaller than
examining the values in Table 2, it is realised that the 0.1’, ‘between 0.1 and 0.5’, ‘between 0.1 and 1’ and ‘bigger than
concentrations of dissolved ethane (C2H6) in the abovementioned 1’ ranges, respectively. In the triangular function considered for
transformer are higher than normal. According to the values listed C2H4 to C2H6 input, intervals 0–0.2, 0–1, 1–4, 1–2000, 2–2000 and
in Table 2, this gas has had a growth of 157.2 ppm/year (µl/l/year) 4–2000 have been defined as the ‘smaller than 0.2’, ‘smaller than
from 30/05/2010 (previous values) to 02/10/2010 (present values). 1’, ‘between 1 and 4’, ‘bigger than 1’, ‘bigger than 2’ and ‘bigger
So, according to the values in Table 3 of the first step, the growth than 4’ ranges, respectively. Interval 0.55–0.66 has been defined as
of ethane is higher than normal. Next, we refer to Table 4 of the the range of fault T1 of fuzzy system output. Note that all the
first step. Considering the present values, we have: internal faults of the mentioned power transformer have been
(C2H2 /C2H4) = 0, (CH4 /H2) = 2.436 and (C2H4 /C2H6) = 0.033. defined as nine ranges of FIS output, in the total range of 0–1, and
Thus, according to the fourth row of Table 4, the internal fault of according to Table 9.
Now, when the concentrations of dissolved gases in the
mentioned transformer oil are given to this fuzzy system, system
Table 4 Procedure for diagnosing the transformer fault type output becomes 0.6050, which falls in the range of fault T1 from
in step 1 [16] different output ranges of the system. Thus, the internal
Case Characteristic fault C2H2 /C2H4 CH4 /H2 C2H4 /C2H6 transformer fault detected by the designed FIS is the presence of a
PD partial discharges (see NSa <0.1 <0.2 hot spot or hot spots with a temperature of <300°C (i.e. fault T1).
notes 3 and 4) The definition of Rule #54 and its performance and the
D1 discharges of low >1 0.1–0.5 >1 determination of output value 0.6050 based on the abovementioned
energy inputs are shown in Fig. 4.
D2 discharges of high 0.6–2.5 0.1–1 >2
energy 1. If C2H6 = 239 and Rate_C2H6 = 157.2 and CH4 to H2 = 2.436
T1 thermal fault t < 300°C <1 and C2H4 to C2H6 = 0.033, then fault = 0.6050.
NSa >1 but NSa
T2 thermal fault 300°C < t <0.1 >1 1–4
Actually, after draining the oil and checking the interior of
< 700°C
transformer it was observed that, because of improper varnishing
T3 thermal fault t > 700°C <0.2b >1 >4 by the manufacturer, the varnish material has collected at the
Note 1: In some countries, the ratio C2H2/C2H6 is used, rather than the ratio CH4/H2. bottom of transformer and caused the corrosion of collector #6 of
Also in some countries, slightly different ratio limits are used. transformer radiator (see Fig. 5); this has ultimately led to extra
Note 2: CH4/H2 < 0.2 for partial discharges in instrument transformers. CH4/H2 < heating and the creation of a hot spot and the expansion of fault.
0.07 for partial discharges in bushings. Thus, it is confirmed that the designed fuzzy system has
Note 3: Gas decomposition patterns similar to partial discharges have been reported as correctly and truly diagnosed the internal fault of the mentioned
a result of stray gassing of oil. transformer.
Note 4: Gas decomposition patterns similar to partial discharges have been reported as
a result of the decomposition of thin oil film between overheated core laminates at
temperatures of 140°C and above. 5 Experimental results
aNS = Non-significant whatever the value.
One of the advantages of this research is that is has used a
bAn increasing value of the amount of C H may indicate that the hot spot
2 2 sufficiently large set of real data. The results of gas
temperature is >1000°C. chromatography tests performed from 2001 to 2016 on the oil of
>3000 power transformers in Iran's national power network have
been studied in this work. This data bank was established by
Table 5 Limits of faults in the Duval triangle [14] TAVANIR holding company. From this data set, the DGA results of
Limits of zones 185 power transformers were selected and analysed. One hundred
PD 98% CH4 fifty-seven of these transformers included the step-up and step-
D1 23% C2H4 13% C2H2 down power transformers installed in 430, 230, 132 and 63 kV
transmission substations throughout the country; 26 transformers
D2 23% C2H4 13% C2H2 40% C2H4 29% C2H2
were the main step-up and unit step-down transformers installed in
T1 4% C2H2 20% C2H4 some thermal power plants such as Ramine Ahvaz, Ferdosiye
T2 4% C2H2 20% C2H4 50% C2H4 Mashhad, Khalije Fars, Parand, Shazande Arak, Loshan, Zarande
T3 15% C2H2 50% C2H4
Kerman; and 2 transformers were the main step-up power
transformers installed in two hydro power plants (Shahid

2724 IET Gener. Transm. Distrib., 2017, Vol. 11 Iss. 10, pp. 2721-2729
© The Institution of Engineering and Technology 2017
Abbaspour and Masjed Soleiman). One hundred sixty-two of these corresponding to that range. After applying all the inputs, the
185 transformers had been condition monitored and found to have diagnosis results of the FIS were compared to the real faults. For
one or more electrical, thermal or chemical faults or a combination example, as shown in Table 10, the designed FIS diagnosed a D2
of these faults. Once the internal faults of these transformers were type fault for power transformer T1 used in the Khamseh
detected, they were taken offline and repaired. The remaining 23 substation of Zanjan Regional Power Company; which agreed with
transformers had no faults and were totally healthy. The interesting the real fault. However, the FIS diagnosed a T1 fault for the power
point about these data is their extensive quantity and variety. IRAN transformer in the Palor-63 substation of Hormozgan Regional
is famous for being a four-season country; and the selected datasets Power Company; which did not match the real fault; this
were related to power transformers that have operated in different transformer had actually two real faults of PD and T2. Finally, the
climatic conditions, including the warm and dry, warm and humid, fault diagnosis accuracy of the FIS has been determined.
cold and dry, cold and humid, moderate and dry, and moderate and The point to consider in analysing the fault of insulation paper
humid climates. A sample of these data has been presented in decomposition is that it is true that one cannot rely solely on the
Table 10. ratio of CO2 to CO gas concentrations to decide the condition of
Using the MATLAB codes, the data in these tables have been the insulation paper and that the transformer paper might be
applied to the FIS as inputs. Through a fuzzy computational healthy even if this ratio is <3; however, IEC 60599 guidelines
process, the FIS generated defuzzified numbers from these data. state that to find out whether the changes of gas ratios relative to
Any defuzzified number belonging to one of the ranges defined for the measurements of the previous period are due to a new fault or
the nine output states specified the type of internal fault not, the new values should be subtracted from the previous values
and the ratio of gases (ratio of growth of gases relative to each
other) should be calculated again. This is particularly true for the
Table 6 Different inputs of the FIS designed in this ratio of CO2 to CO. Therefore, if the ratio of CO2 to CO is <3 and,
research at the same time, the ratio of the annual growth of CO2 to CO is
Number Inputs also <3, it means that the insulation paper of transformer is
1 CO2 gradually being decomposed and deteriorated [16]. Of course, there
2 CO are two tests for determining the condition of transformer paper: an
3 H2 online test which measures the amounts of furanic compounds (2-
furfuraldehyde) and calculates its annual growth and compares
4 CH4
these values against a standard table; and an offline test (performed
5 C2H2 when the transformer is disassembled) which removes a small
6 C2H4 piece of transformer paper and measures its degree of
7 C2H6 polymerisation (DP) [26–28]. In this research, the second test has
been used to evaluate the correct performance of the designed
8 Rate_CO2 fuzzy system; which in most cases, the detection of cellulose paper
9 Rate_CO fault by the presented fuzzy system has matched the results of
10 Rate_H2 measuring the insulation paper's DP after opening the transformer.
11 Rate_CH4 In very few occasions (just two cases, according to Table 12 of our
paper), in which the fuzzy system has alerted the ongoing
12 Rate_C2H2 deterioration of the cellulose paper, the increased concentration of
13 Rate_C2H4 CO gas might have been due to the leakage and entering of air into
14 Rate_C2H6 the tank, oil oxidation, or other causes, ant not because of the
deterioration of insulation paper. It should be mentioned that the
15 CO2 to CO
number of rules in the presented fuzzy system for detecting the
16 Rate_CO to gradual deterioration of the cellulose paper is 3.
17 C2H2 to C2H4
18 CH4 to H2 5.1 Evaluation of fault diagnosis accuracy
19 C2H4 to C2H6 5.1.1 F1 score: The fault diagnosis accuracy can be quantified by
20 Graphical C2H2 to C2H4 means of the F1 score. This criterion has been widely used in
21 Graphical CH4 to H2 machine learning systems. The reason for using the F1 score is that
it simplifies the process of determining the accuracy of system
22 Graphical C2H4 to C2H6 performance by providing a single index that combines the
23 CH4 Percent parameters of precision (positive predictive value) and recall
24 C2H2 Percent (sensitivity or TPR value). After testing each set of trained faults
by using new data, the obtained fault diagnosis results can be
25 C2H4 Percent
arranged in a confusion matrix (Table 11).

Table 7 Different states of output of the FIS designed in this research

Number 1 2 3 4 5 6 7 8 9
states of output no fault P PD D1 D2 T1 T2 T3 DT

Table 8 Results of gas chromatography test performed on the oil of transformer T

TDCG CO2, ppm CO, ppm H2, ppm CH4, ppm C2H2, ppm C2H4, ppm C2H6. ppm Sampling date
637 489 257 81 70 2.369 40 186 30/05/2010
722 1451 341 39 95 0 8 239 02/10/2010

Table 9 Outputs of the applied FIS and the internal faults corresponding to each output
Fault No fault P PD D1 D2 T1 T2 T3 DT
range 0–0.11 0.11–0.22 0.22–0.33 0.33–0.44 0.44–0.55 0.55–0.66 0.66–0.77 0.77–0.88 0.88–1

IET Gener. Transm. Distrib., 2017, Vol. 11 Iss. 10, pp. 2721-2729 2725
© The Institution of Engineering and Technology 2017
Fig. 4 Definition of Rule #54 in the applied FIS and its result

Fig. 5 Varnish material has collected at the bottom of transformer and caused the corrosion of collector #6 of transformer radiator
(a) Healthy and intact power transformer collector, (b) Damaged and corroded collector #6 of transformer radiator

Table 10a Sample of DGA data used in this research and the fault diagnosis results of some power transformers used in Iran's
electricity power network
Number Comp. Trans. code Present Previous Present gas concentration values, Previous gas concentration
name sampling sampling ppm values. ppm
data data
CO2 CO H2 CH4 C2H2 C2H4 C2H6 CO2 CO H2 CH4 C2H2 C2H4 C2H6
1 Bakhtar 9-1402 2008/24/08 2007/30/08 952 416 387 40 0 2 7 706 266 215 25 0 0 0
T6100-104
2 Azarbayejan SARDROOD2-T5 2014/26/07 2014/14/06 1132 204 1261 77 0 9.303 74 790 166 1014 65 0 11.35 56
3 Azarbayejan SARDROOD2-T6 2014/26/10 2014/01/09 3402 206 379 22 75 20 0 3940 226 343 16 58 17 6.3
4 Zanjan Khamseh – T1 2010/16/10 2009/06/06 3693 529 387 137 31 52 924 2347 360 348 56 17 28 519
5 Gilan Siadati-T2 2003/06/02 2003/09/01 3482 603 25.62 134 0 58 155 3157 431 54.4 104 0 58.85 102
6 Hormozgan Palor-63 2013/02/09 2012/25/02 4545 870 454 10 0.24 58 23 1839 893 67 6.5 0.24 57 22
7 Khuzestan Susangerde asli – 2006/09/08 2006/05/07 474 286 428 1984 9 2626 621 529 298 371 1735 0.82 2214 515
T3
8 Zanjan Khamseh – T2 2013/11/11 2013/01/01 2380 353 138 33 7 20 228 1186 126 0 0 17 20 9
9 Bakhtar 8-1401T1_22-26_ 2008/21/04 2007/19/09 1762 199 22 2 2 2 3 2084 259 35 3 0 2 3

In Table 11, af is the number of true positive diagnoses, bf is the negative diagnoses, and df is the number of true negative
number of false positive diagnoses, cf is the number of false diagnoses. In other words, af are the diagnoses correctly assumed

2726 IET Gener. Transm. Distrib., 2017, Vol. 11 Iss. 10, pp. 2721-2729
© The Institution of Engineering and Technology 2017
Table 10b
Gas increase rate, ppm/year Gas ratio Duval tri. gas percent Real Diagnosed
CO2 CO H2 CH4 C2H2 C2H4 C2H6 CO2 /CO (CO2 C2H2 /C2H4 CH4 /H2 C2H4 /C2H6 CH4% C2H2% C2H4% fault fault
growth/
CO
growth)
249 152 174 15 0 2 7 2.228 1.64 0 0.103 0.286 0.952 0 0.05 P+ P
T1
2908 316 2908 107 0 17 154 5.558 9.194 0 0.061 0.126 0.892 0 0.107 PD PD
−3636 −135 245 43 114 22 −43 16.487 26.901 3.662 0.059 2000 0.191 0.636 0.174 D1 D1
767 124 28.6 59 10 17.59 297 6.981 7.946 0.596 0.354 0.056 0.623 0.141 0.236 D2 D2
4090 2163 −362 376 0 −11.831 665 5.771 10.873 0 5.227 0.375 0.698 0 0.302 T1 T1
1780 −15 254 2.368 0 0.395 0.685 5.224 −118 0.004 0.022 2.522 0.148 0.004 0.849 PD + T1
T2
−558 −122 579 2528 83 4177 1080 1.657 4.619 0.003 4.636 4.227 0.43 0.002 0.568 T3 T3
1392 265 161 38 −12 0 255 6.742 5.26 0.35 0.239 0.088 0.55 0.117 0.333 DT DT
−547 −102 −22 −1.7 3.4 0 0 8.854 5.367 1 0.091 0.667 0.333 0.333 0.333 no no fault
fault

to belong to fault f; bf are the diagnoses incorrectly considered to averaged to get a single performance value in multiclass problems
belong to f; cf are the diagnoses incorrectly supposed not to belong [30]. In macro-averaging, which is calculated by averaging the
to f and df are the diagnoses correctly assumed not to belong to f. results of different categories in multiclass problems, ‘C’ is
Thus, the TPR or the recall value in (1) measures the ratio of the considered as the total number of categories in the multiclass
number of positive cases correctly detected as faults (af) to the total problem, and the macro-averaging procedure is performed as
follows [29, 30]:
number of actual positive cases, by taking parameter cf into
consideration [29, 30]
∑Ci = 1 F1 f
F1 = i
, (0 ≤ F1 ≤ 1) (4)
af C
Recf = (1)
af + cf
The greater the F1 score, the more accurate the fault diagnosis
results. Hence, the objective in fault diagnosis problem is to
Also, the true predictive value or the precision in (2) measures the
maximise F1 score.
ratio of af over the total number of positively predicted outcomes,
by taking into account the false alarms (bf)
5.1.2 Measuring the fault diagnosis accuracy of the designed
FIS by using the F1 score: In this step, the considered FIS has
af
Precf = (2) been implemented and the results of monitoring 185 examined
af + bf transformers have been compiled. The set of actual results and the
prediction results obtained from the FIS have been presented in a
Finally, the F1 score can be obtained from (3) [29] confusion matrix, according to Table 12.
The F1 score is a very suitable index for measuring the degree
2 × Precf × Recf of fault diagnosis accuracy which measures based on confusion
F1f = (3) matrix. For example, for determining the degree of accuracy of
Precf + Recf
isolating fault T1 from the other faults by the designed FIS, first,
Since the precision and recall parameters are only used in binary the F1 score is computed eight times, as follows, and then, by
(two-class) classification tasks, the obtained results need to be averaging these eight scores, the fault diagnosis accuracy for fault
T1 is obtained (Table 13).
For example, the value of F1(T1,NoFault) = 92.7% expresses the
Table 11 Confusion matrix [29, 30] percentage of transformers with the T1 fault, which have been
Actual class correctly differentiated from the healthy transformers. Here, the
f −f procedure for calculating the F1(T1,NoFault) is explained. First, the
predicted class f af bf recall and precision parameters related to F1(T1,NoFault) are
−f cf df calculated based on Table 12 as follows:

Table 12 Confusion matrix for internal power transformer faults

Predicted faults Actual faults
No fault P PD D1 D2 T1 T2 T3 DT
no fault 21 0 0 1 0 3 1 1 0
P 0 37 0 0 0 2 0 0 0
PD 0 0 2 0 0 2 1 4 0
D1 0 0 1 4 0 1 0 0 1
D2 2 0 12 0 4 4 7 1 0
T1 0 0 0 0 0 19 6 3 0
T2 0 1 1 0 0 4 18 0 0
T3 0 0 0 0 0 0 0 15 0
DT 0 0 0 0 1 0 1 0 4

IET Gener. Transm. Distrib., 2017, Vol. 11 Iss. 10, pp. 2721-2729 2727
© The Institution of Engineering and Technology 2017
Table 13 Degree of accuracy (%) of diagnosing fault T1 from the other faults by the designed FIS
Measure F1 (T1, no fault) F1 (T1, P) F1 (T1, PD) F1 (T1, D1) F1 (T1, D2) F1 (T1, T2) F1 (T1, T3) F1 (T1, DT) Total F1 (T1)
value, % 92.7 95 95 97.4 90.5 79.2 92.7 100 92.8

Table 14 F1 score for each internal power transformer fault

Measure F1 (no fault) F1 (P) F1 (PD) F1 (D1) F1 (D2) F1 (T1) F1 (T2) F1 (T3) F1 (DT)
value, % 97.7 99.5 73.6 94.5 77.2 92.8 93.6 96.6 95.8

af: Number of transformers with the T1 fault; which were correctly Due to the high importance and cost of power transformers used in
diagnosed by the FIS = 19. power networks, one of the key objectives of the managers in the
bf: Number of healthy transformers which were misdiagnosed by sensitive electricity industry is to properly operate and maintain
the FIS as transformers with the T1 fault = 0. this important equipment. To this end, the strategy adopted by the
managers is ‘condition monitoring’, and the tool for implementing
cf: Number of transformers with the T1 fault; which were
this strategy is the different electrical and chemical tests which are
misdiagnosed by the FIS as healthy transformers =3. performed on various parts of power transformers. One of these
chemical tests, which is performed on the oil of a power
Using the above values, we obtain transformer, is the gas chromatography test. In reality, the number
of defective transformers relative to the whole transformers in the
af 19 power grid is low; hence, not much experience exists in this regard,
Prec T 1, No Fault = = = 1,
af + bf 19 + 0 and there are very few expert technicians with sufficient experience
af in dealing with faulty transformers. Therefore, intelligent systems
19
Rec T 1, No Fault = = = 0.864 that are able to detect and isolate the internal faults of power
af + cf 19 + 3 transformers can be very useful. So in this research, based on the
data collected from the whole country in the last 15 years regarding
so the technical specifications and the results of tests performed on
162 faulty transformers and 23 healthy transformers, a FIS was
2 × Prec × Rec 2 × 1 × 0.864 designed which uses the gas chromatography test results as well as
F1(T 1, NoFault) = = = 0.927
Prec + Rec 1 + 0.864 the tables and procedures given in the IEC 60599 guidelines to
detect and isolate the faults of power transformers. The efficiency
In the same way, F1(T1,P) can be computed as below: of this system, i.e. the fault diagnosis accuracy of the designed FIS
and the matching of its predictions with the reality, was estimated
af 19 to be 91.2%.
Prec T 1, P = = = 1, However, the important point that should be emphasised is that
af + bf 19 + 0
the decisions made regarding the manner of operating and
af 19 maintaining power transformers should not be solely based on the
Rec(T 1, P) = = = 0.904
af + cf 19 + 2 results of this test, but the findings of other complementary
transformer tests should also be taken into consideration in order to
so adopt appropriate decisions in this regard.
One of the limitations of the designed FIS is that it is only able
2 × Prec × Rec 2 × 1 × 0.904 to diagnose one fault in a transformer that has multiple faults; and
F1(T 1, P) = = = 0.95 sometimes, even this one fault is misdiagnosed. Therefore, a
Prec + Rec 1 + 0.904
suggestion for future work on this subject is to optimise the FIS so
With regards to the above values, the value of parameter F1(T1), i.e. that it can adequately diagnose multiple faults in a power
the accuracy of correctly diagnosing fault T1 from the other faults, transformer.
is obtained as
7 References
∑ F1(T 1, others) [1] Suna, H.C., Huanga, Y.C., Huang, C.M.: ‘A review of dissolved gas analysis
F1(T 1) = = 92.8% (5)
8 in power transformers’. Second Int. Conf. on Advances in Energy
Engineering (ICAEE), 2011
[2] IEEE – Institute of Electrical and Electronics Engineer – IEEE guide for the
This means that the designed FIS can isolate the T1 fault with an interpretation of gases generated in oil immersed transformer, IEEE Std
accuracy of 92.8%. Finally, the F1 score for each fault is tabulated C57.104–2008
in Table 14. [3] Lee, J.P., Lee, D.J., Ji, P.S., et al.: ‘Dissolved gas analysis of power
The relatively low values of F1(PD) and F1(D2) compared to transformer using fuzzy clustering and radial basis function neural network’.
Int. Joint Conf. on Neural Networks, Vancouver, Canada, July 2006
other F1 scores are due to the fact that most of the transformers [4] Castro, A.R.G., Miranda, V., Lima, S.: ‘Transformer fault diagnosis based on
that had one of these two faults also had a combination of other autoassociative neural networks’. 16th Int. Conf. on Intelligent System
faults as well. For example, of 18 misdiagnoses of the ‘PD’ fault, Application to Power Systems (ISAP), Hersonissos, Greek, 25–28 September
2011
16 cases were related to transformers which had one or two other [5] Guardado, J.L., Naredo, J.L., Moreno, P., et al.: ‘A Comparative study of
faults in addition to ‘PD’. neural networks efficiency in power transformers diagnosis using dissolved
Finally, it can be concluded that the effectiveness of the deigned gas analysis’, IEEE Trans. Power Deliv., 2001, 16, pp. 643–647
FIS (i.e. degree of fault diagnosis accuracy of this system for the [6] Gaikwad, S.G., Swami, P.S., Thosar, A.G.: ‘Comparison of conventional and
fuzzy logic approach for DGA of EHV transformer’, Int. J. Recent Technol.
internal faults of power transformers) is the average F1 score for Eng. (IJRTE), 2015, 4, pp. 30–37
all the possible faults. So, based on (6), we have [7] Abu-Siada, A., Hmood, S., Islam, S.: ‘A new fuzzy logic approach for
consistent interpretation of dissolved gas-in-oil analysis’, IEEE Trans.
∑ F1(a fault, other faults) Dielectr. Electr. Insul., 2013, 20, pp. 2343–2349
F1 = = 91.2% (6) [8] Abu-Siada, A., Hmood, S.: ‘Fuzzy logic approach for power transformer asset
9 management based on dissolved gas-in-oil analysis’, Chem. Eng. Trans.
(CEt), 2013, 33, pp. 997–1002
[9] Cai, G., Liu, N., Yang, D.: ‘The transformer fault diagnosis based on quantum
6 Conclusion neural network’. Int. Conf. on Computer, Mechatronics, Control and
Electronic Engineering (CMCE), Changchun, China, 24–26 August 2010

2728 IET Gener. Transm. Distrib., 2017, Vol. 11 Iss. 10, pp. 2721-2729
© The Institution of Engineering and Technology 2017
[10] Sharma, E.N.: ‘Power transformer diagnosis using fuzzy logic’, Int. J. Latest
Res. Sci. Technol., 2012, 1, pp. 149–151
[11] Németh, B., Laboncz, S., Kiss, I.: ‘Condition monitoring of power
transformers using DGA and fuzzy logic’. IEEE Electrical Insulation Conf.,
Montreal, Canada, 31 May–3 June 2009
[12] Cosemans, H.: ‘Properties behind effective transformer oil cooling’.
Transformer Life Management Conf., 2013
[13] Soni, R., Chaudhari, K.: ‘An approach to diagnose incipient faults of power
transformer using dissolved gas analysis of mineral oil by ratio methods using
fuzzy logic’. Int. Conf. on Signal Processing, Communication, Power and
Embedded System (SCOPES), 2016
[14] Seifeddine, S., Khmais, B., Abdelkader, C.: ‘Artificial intelligence tools
aided-decision for power transformer fault diagnosis’, Int. J. Comput. Appl.,
2012, 38, pp. 1–8
[15] IEC Publication 60567: ‘Oil-filled electrical equipment – sampling of gases
and oil for analysis of free and dissolved gases – Guidance’, 2005–06
[16] IEC Publication 60599: ‘Interpretation of the analysis of gases in transformers
and other oil-filled electrical equipment in service’, 2014
[17] Hmood, S., Abu-Siada, A., Masoum, M.A.S., et al.: ‘Standardization of DGA
interpretation techniques using fuzzy logic approach’. IEEE Int. Conf. on
Condition Monitoring and Diagnosis, Bali, Indonesia, 23–27 September 2012
[18] Chow, M.-Y., Irwin, J.D.: ‘The industrial electronics handbook’ (CRC Press,
1997)
[19] Wang, L.-X.: ‘A course in fuzzy systems and control’ (Prentice-Hall
International, Inc., 2014)
[20] Fuzzy Logic Toolbox™: ‘User's Guide’ (The MathWorks, Inc., 2016)
IET Gener. Transm. Distrib., 2017, Vol. 11 Iss. 10, pp. 2721-2729
© The Institution of Engineering and Technology 2017
[21] Hoang, P., Tomsovic, K.: ‘Design and analysis of an adaptive fuzzy power
system stabilizer’, IEEE Trans. Power Deliv., 1996, 11, pp. 455–461
[22] Ramezani, S., Memariani, A.: ‘A fuzzy rule based system for fault diagnosis,
using oil analysis results’, Int. J. Ind. Eng. Prod. Res., 2011, 22, (2), pp. 91–
98
[23] Espinosa, J., Vandewalle, J., Wertz, V.: ‘Fuzzy logic, identification and
predictive control’ (Springer, London, 2005)
[24] Zadeh, L.A.: ‘Information and control’ (Academic Press, New York, 1965)
[25] Ekel, P., Kokshenev, I., Parreiras, R., et al.: ‘Fuzzy set based models and
methods of decision making and power engineering problems’, Engineering,
2013, 5, (5A), pp. 41–51, doi: 10.4236/eng.2013.55A007
[26] Yamagata, N., Miyagi, K., Oe, E.: ‘Aging effects on mixture of thermally
upgraded paper and Kraft paper in mineral oil’. Int. Symp. on High Voltage
Engineering, Innes House, Johannesburg, 2009
[27] Ariffin, M.F., Ghosh, P.S.: ‘Estimating the age of paper insulation in 33/11 kV
distribution power transformers using mathematical modelling’. 19th Int.
Conf. on Electricity Distribution, CIRED, Vienna, 21–24 May 2007
[28] Miyagi, K., Oe, E., Yamagata, N.: ‘Evaluation of aging for thermally
upgraded paper in mineral oil’, J. Int. Council Electr. Eng., 2011, 1, (2), pp.
181–187
[29] Hajihosseini, P., Salahshoor, K., Moshiri, B.: ‘Decentralized fault isolation by
a combination of transfer entropy and classification methods’, Chem. Eng.
Commun., 2015, 202, pp. 1131–1144
[30] Hajihosseini, P., Salahshoor, K., Moshiri, B.: ‘Process fault isolation based on
transfer entropy algorithm’, ISA Trans., 2014, 53, pp. 230–240
2729