222 I. B. M. Taha et al.

l.: Optimal Ratio Limits of Rogers’ Four-Ratios and IEC 60599 Code Methods Using Particle Swarm Optimization

Optimal Ratio Limits of Rogers’

Four-Ratios and IEC 60599 Code Methods Using
Particle Swarm Optimization Fuzzy-Logic Approach
Ibrahim B. M. Taha1, 2 and Ayman Hoballah1, 2
1
Taif University, College of Engineering
Department of Electrical Engineering
City of Taif, Al Hawiyah 21974, Saudi Arabia
2
Tanta University, Faculty of Engineering
Department of Electric Power and Machines
City of Tanta, Seperbay31521, Egypt

Sherif S. M. Ghoneim1, 3
3
Suez University, Faculty of Industrial Education
Electrical Department
City of Suez 43527, Egypt

ABSTRACT
The prediction of fault type in transformers at an early stage is an important aspect for
power system reliability. Checking the transformer status begins with the dissolved gas
analysis (DGA) test. Rogers’ four-ratio and IEC 60599 standard methods are
commonly DGA-based techniques that used to detect the transformer faults. These
methods are very easy to implement, but their diagnostic accuracies are very poor. The
target of this work is to build a new PSO-FS optimization platform identifying the
optimal limits of gas ratios and corresponding rules for each fault type to enhance the
diagnostic accuracies of Rogers’ four-ratio and IEC 60599 code methods. In the
proposed platform, PSO optimization technique specifies the gas ratio limits and
corresponding rules where the fuzzy system identifies the fault types. Based on the
results of the optimization process, the new codes that used to diagnose the fault types
were developed. The performance of the proposed approach investigated using 481
collected datasets. The overall accuracies of the Rogers’ four-ratio and IEC code
models enhanced from 47.19 and 55.09 to 85.65 and 85.03, respectively. However, when
datasets due to mixtures of faults (samples with undetermined or misidentified faults)
are removed from the calculations, the accuracies of the PSO-FS methods are not
significantly better than when using existing DGA methods.
Index Terms —transformer oil, power transformer testing, optimization methods,
fuzzy logic

1 INTRODUCTION ethane (C2H6), ethylene (C2H4), acetylene (C2H2), and carbon

mono-oxide (CO).
EARLY stage detection of transformer faults avoids the
undesired abnormal operating conditions and the outage of the The incombustible gases are carbon di-oxide (CO2), nitrogen
transformer, which affects the system reliability and leads to (N2), and oxygen (O2) [2]. The ratios of concentrated
the reduction of utility’s revenue [1]. The faults in transformer dissolved gases have been used in the most dissolved gas
insulating systems (oil and impregnated paper) are due to analysis (DGA) techniques to diagnose the transformer fault
type. The most commonly DGA interpretation techniques that
excessive stresses such as electrical, thermal and mechanical
built based on the concentrated gas ratios are Dorneneburg,
stresses [1, 2]. These stresses unfortunately degrade the
Roger’s four and three ratios, IEC 60599 standard code
transformer oil to combustible and incombustible gases. The
methods [1], and some other techniques using a graphical
main combustible gases are hydrogen (H2), methane (CH4), approach (e.g. Duval Triangle and Pentagon methods) [3,4].
The ratio and graphical DGA methods have advantages of
Manuscript received on 29 June 2019, in final form 21 September 2019, easy implementations, but they have poor diagnostic
accepted 22 September 2019. Corresponding author: A.Hoballah. accuracies for the transformer fault types [5–7].

DOI: 10.1109/TDEI.2019.008395

Authorized licensed use limited to: University of Canberra. Downloaded on May 01,2020 at 08:16:30 UTC from IEEE Xplore. Restrictions apply.
IEEE Transactions on Dielectrics and Electrical Insulation Vol. 27, No. 1; February 2020 223

Many artificial intelligence techniques utilized to enhance limits and rules enhancing the diagnostic accuracy of the
the diagnostic accuracy of the DGA methods such as artificial Rogers’ four-ratio and IEC 50699 standard code methods from
neural network (ANN), fuzzy logic system (FS), neuro-fuzzy 47.19 and 55.09 to 85.65 and 85.05%, respectively.
system, support vector machine, and clustering method [8, 9].
In [9], Islam et al presented a new DGA method based on 2 PROBLEM FORMULATION
clustering technique (Duval triangle method) to enhance the The problem is formulated as a constrained optimization
accuracy of the conventional method such as Duval and IEC problem that aims to maximize the fault diagnostic accuracy
60599. The method called Parzan window and used to using PSO-FS approach. The objective function and
estimate the probability of different fault categories. The constraints are presented in Equation (1) and Equation (2).
proposed method built based on the percentage of the five key The input fuzzy sets are specified by membership functions
gas concentrations. The proposed method succeeded in (μ(R)) within the range [0, 1]. Each gas ratio (R) is
diagnosing of difficult cases compared to the other methods. categorized into membership functions (mf1, mf2, mf3 …)
The diagnostic accuracy of Rogers’ four-ratio and IEC 60599 based on the ratio limits (L1, L2, L3 …) according to the
methods improved by categorizing the undetermined cases as DGA method as shown in Figure 1. Therefore, based on the
determining cases by adding new rules and changing the gas value of certain gas ratio, one of the membership functions has
ratio limits to refine Rogers’ four-ratio and IEC 60599 a value in the range [1, 0] and the others are zero values.
methods. The accuracy of these refining methods still low and
more effort required for enhancing the diagnostic accuracy.
μ(R) mf1 mf2 mf3 mf4
The main contribution of this work is to propose particle
1
swarm optimization fuzzy logic system (PSO-FS) platform to
enhance the accuracy of both Rogers’ four-ratio and IEC
60599 code methods. The target is maximizing the fault
0
diagnostic accuracy by introducing new gas ratio limits and L1 L2 L3 R
rules. Where all efficient optimization algorithms must obtain Figure 1. Membership functions of gas ratio.
the same global best solution, PSO is selected as an
optimization technique in this study to catch the optimal limits The proposed PSO-FS approach provides the optimal limits
of Rogers’ four-ratio and IEC 60599 code methods. PSO is of each five combustible gas ratio, which maximize the fault
considered as efficient, fast, and simple optimization diagnostic accuracy. The main objective is defined by:
technique [10]. Thus, a modified PSO is used to identify the
Maximize:
optimal gases ratio limits (membership limits of fuzzy system)
and the corresponding crisp rules of the fuzzy system model to 𝑁𝑡𝑟𝑢𝑒
% 𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 *100 (1)
𝑁 𝑎𝑙𝑙
meet the actual transformer fault types. The membership
functions determine the degree of presence of each gas ratio Subject to:
within the fuzzy sets in order to estimate the correct fault type.
The degrees of presence are converted into rules based on the 𝐿𝑥𝑖 1 𝐿𝑥𝑖 ∀ 𝑖 1 … 𝑛𝑥 , 𝑥 1….𝑚 2
specified limits of each gas ratio. Furthermore, the obtained
where, 𝑁𝑡𝑟𝑢𝑒 and 𝑁𝑎𝑙𝑙 are the number of true estimated fault
rules of the fuzzy system model are converted into distinct
types by the proposed PSO-FS model and the total number of
code, which uses to identify transformer fault type. According 𝑥 𝑥
to the collected data, the fault types are classified into six samples, respectively. 𝐿𝑖 1 and 𝐿𝑖 are the limits of the x
categories that are partial discharge (PD), low energy gas ratio. n and m are the number of limits of x gas ratio and
discharge (D1),high-energy discharge (D2) low temperature the total number of gas ratios, respectively.
fault (T1), medium temperature fault (T2) and high The optimal solution tends to specify the matrix of control
temperature fault (T3). variables (P), which includes the membership function limits
The enhancement of fault diagnostic accuracy was carried (L) of all gas ratios and the fault type (F) corresponding to all
out using 481 faulty dataset samples that collected from the available N rules as expressed by:
electricity utilities and literatures as explained in Section 5. 𝑃 𝐿11 ⋯ 𝐿1𝑛 1 𝐿21 ⋯ 𝐿2𝑛 2 ⋯ 𝐿𝑚 𝑚
1 ⋯ 𝐿𝑛 𝑥 𝐹1 𝐹2 ⋯ 𝐹𝑁 (3)
Each sample contains information of the main five
combustible gases (H2, CH4, C2H6, C2H4 and C2H2) and
corresponding actual fault type. The 481 samples merged, 3 PSO-FS APPROACH
shuffled and divided randomly into two sets during training PSO, as an accurate heuristic optimization technique, has
and testing process. The first set has 385 samples (80% of the ability to discover the search area specified by the limits of
total samples) that are used during the iterative optimization control variables for optimal solution of constrained and
process within the suggested models (training samples) while unconstrained optimization problems [11]. The optimal
the second set has 96 samples (20% of total samples) that are solution represents the position of global best particle in the
used during the testing stage to validate and verify the swarm. The particles move in the search space based on the
proposed models. The results have achieved that the new ratio local experience of each particle (local best position) and the

Authorized licensed use limited to: University of Canberra. Downloaded on May 01,2020 at 08:16:30 UTC from IEEE Xplore. Restrictions apply.
224 I. B. M. Taha et al.: Optimal Ratio Limits of Rogers’ Four-Ratios and IEC 60599 Code Methods Using Particle Swarm Optimization

global experience of the swarm (global best position). The Step 3: Build the FS model using the memberships, rules, and
target is to maximize the objective function (diagnostic proposed fault types. The output corresponding to each dataset
accuracy of fault types) that satisfies all constraints. A is one of the six values 1, 2, 3, 4, 5 and 6, in which it
modified PSO algorithm utilized to obtain the optimal solution represents one of the fault types PD, D1, D2, T1, T2, and T3,
utilizing a selection strategy for determining the local and respectively. The FS built using Sugeno-type fuzzy inference
global best positions based on adaptive penalization for the system for each particle in the swarm.
infeasible solutions [11]. Particles modify their positions Step 4: Apply the FS model to determine the fault types
during the iterative optimization process using the following (output fuzzy set) corresponding to the actual ratios in the
relationships as in: training dataset. The outputs of all particles used to evaluate
𝑃𝑖𝑘 1
𝑃𝑖𝑘 𝜒𝑉𝑖𝑘 1
(4) the objective function, which described by Equation (1).
Figure 3 presents the architecture of fuzzy-logic system
𝑉𝑖𝑘 1
𝑤𝑘 1
𝑉𝑖𝑘 𝐴 𝑉𝑙𝑖𝑘 𝑉𝑖𝑘 𝐵 𝑉𝑔𝑖𝑘 𝑉𝑖𝑘 (5) approach.
where, 𝜒 is a factor uses to improve the swarm conversion. Step 5: Construct a modified fitness function utilizing the
w is an adaptive weight factor to control the motion of swarm objective function and constraint violations for each particle
towards the global best solution. Both A and B are random based on self-adaptive penalty function [11].
𝑘
values between 1 and 2. 𝑉𝑙𝑖 and 𝑉𝑔𝑖 are the local and global
𝑘 Step 6: Select and save the local best position of each particle
𝑘 1 and global best position of the swarm in each iteration.
best positions in k iteration. 𝑉𝑖 and 𝑉𝑖 are the velocities of
𝑘

particles in current and previous iterations. Step 7: If the specified stop criteria are not satisfied, the
modified position of particles is used to recalculate the
In case of Rogers’ four-ratio method, 72 combinations of
associated fitness functions, where the iterative process is
different rules created based on the four-ratios where two
continued until the optimal solution is obtained as shown in
ratios divided into three limits and the other two ratios divided
Figure 3.
into two and four limits. In IEC 60599 standard code method,
each ratio divided into three limits (memberships). Therefore,
Fuzzification Stage Inference Engine Defuzzification Stage
there are 27 combinations of different rules associated with the (convert crisp inputs (Specify the fuzzy output based (convert the fuzzy sets
variation of the three ratios and three limits of each ratio. (DGA ratios) into
fuzzy sets)
on the degree of match between
fuzzy input and the rules)
into a crisp value
(Fault types))
In the proposed PSO-FS approach, the FS used to obtain the
fault type utilizing the membership functions and rules Knowledge Base
(All possible rules relates the input
corresponding to actual gas ratios. Figure 2 presents the DGA ratios and fault types)
schematic diagram of the proposed approach based on PSO
and Fuzzy logic system (PSO-FS) to maximize the diagnostic PSO Process
(Modify the position of each particle
accuracy. (ratio limits and fault types))

The main steps of the proposed approach can be

Save Global Best Solution
summarized as follows: (Ratio limits and fault type for each rule)

Figure 3. Fuzzy logic system architecture.

Fuzzy Objective
Start Defuzzification function
logic
system of inputs calculation for Step 8: During iterative process, the position of particles
each particle
Initialize (Ratio limits and fault type corresponding to each rule) is
Population and Fuzzy
parameters Fuzzification
inference modified using Equation (4) and Equation (5). Selection
of inputs Stop Yes
engine criteria End strategy described in [11] used to select the local best position
Check limits of
Fuzzy rules
satisfied of the particles and the global best position of the swarm. The
all variables
preparations No final solution tends to improve the fault diagnostic accuracy of
Modify Local and the training dataset.
particle’s global best
positions modifications The stop criteria can be the maximum number of iterations,
fixed fitness function of the global best solution for predefined
Figure 2. Schematic diagram of proposed PSO-FS approach. number of iterations, or acceptable resolution for all control
variables over predefined number of iterations. The iterative
Step 1: Define the PSO-FS parameters and limits of all process was repeated many times with different swarm size to
variables. be sure the optimal solution obtained. Usually, the increasing
Step 2: Initialize all variables for each particle which contains of the number of particles in the swarm tends to enhance the
the ratio limits, rules, and proposed fault types to construct the quality of optimization process. Furthermore, the optimal
membership vector of fuzzy logic model. The control rules solution compared with the results of different DGA methods
depend on the number of limits in each gas ratio (number of to validate the code. All the numerical study cases have been
memberships), which start from one to the number of run on 2.9 GHZ i7 PC with 8 GB of RAM using MATLAB
segments in each gas ratio. software environment.

Authorized licensed use limited to: University of Canberra. Downloaded on May 01,2020 at 08:16:30 UTC from IEEE Xplore. Restrictions apply.
IEEE Transactions on Dielectrics and Electrical Insulation Vol. 27, No. 1; February 2020 225

4 PSO-FA APPROACH Table 1. Ratio limits for current and PSO-FS Rogers’ four-ratio method.
Current Roger
4.1 ROGERS’ FOUR-RATIO METHOD Ratio PSO-FS Limits Code
Limits
The Rogers’ four-ratio method established to diagnose the R1< 1.0 R1 0.22 0
R
transformer fault types based on four-ratios, which constitute R1 1.0 R1 > 0.22 1
from the five combustible gases as follows: 𝑅 1 R 1.0 0
R 1 𝑅 3 1.0 𝑅 3.50 1
C H C H C H CH
R R R R 6 𝑅 3 R 3.50 2
CH C H C H H
𝑅 0.5 R 0.45 0
The implementation of this method was very simple, but the R 0.5 𝑅 1 0.45 𝑅 1.37 1
main problem was its low diagnostic accuracy and its failure 𝑅 1 R 1.37 2
to diagnose some transformer fault cases [5-8].It is known that 𝑅 0.1 R 0.18 5
the ratio methods may lead to out of code result if the amount 0.1 𝑅 1 - 0
of gas in the ratio is insignificant (Undetermined or Zero). In R
1 𝑅 3 - 1
order to stabilize the performance of the computed ratio limits 𝑅 3 R 0.18 2
under the condition that the amount of any gas value was
undetermined or very close to zero, these values assumed Figure 5 presents the schematic diagram of the PSO-FS
equal to 0.0001. In addition, when the gas ratio was found model for Rogers’ four ratios method. The model inputs are
greater than the end of last membership function, the program the Rogers’ four ratios. The first input has two memberships.
put it by default to specify end of last membership function in The range of first membership is [0 0.22] and the range of
Fuzzy system. For example, if the ratio R4 has two second membership is [0.22 1] which converted to code as
memberships where the specified end of second membership presented in Table 1. The other three ratios are built with the
function is 3 and the actual value is 10. This means that the same manner. The output codes from the four ratios applied to
ratio R4 is greater than 3 and then the corresponding code is 2. the main fuzzy system model core (FS ROG) which consists
Therefore, any value greater than 3 was considered as 3 to of 33 rules that describes the different rules in Table 1. The
shrink the membership size.
output block (f(u)) presents one of the six fault types (PD, D1,
4.2 ENHANCED ROGERS’ FOUR-RATIO METHOD D2, T1, T2 and T3).
The PSO-FS approach is used to generate new ratio limits The membership limits of the enhanced PSO-FS Rogers’
and rules of the fuzzy system model. Figure 4 introduces the four-ratio method are presented in Table 1.
variation of total diagnostic accuracy during the iterative
process using PSO-FS that generated the new ratio limits. The
diagnostic accuracy obtained during optimization process is
86.49 % for all training samples.

Figure 5. Enhanced Fuzzy system of Rogers’ four-ratio method.

Figure 6 presents a case study of using the crisp rules of the

PSO-FS for Rogers’ four-ratio method. The input ratios R1,
R2, R3 and R4 are 2.02, 1.93, 1 and 1.88 respectively which
coverted to code of [1 1 1 2] by new limits in Table 2. The fault
type is 3 (High energy discharge, D2).
Figure 4. Accuracy variation during the iterative process.
Table 2 illustrates the code of fault types corresponding to
The PSO optimization technique tends to modify the limits the modified ratio limits based on new rules of the enhanced
of the current IEC code and Rogers’ four ratios to new limits Rogers’ four-ratio method.
as a result to correct the wrong fault diagnosis of the actual The randomly selected values of the four input ratios are
fault. The old and new ratio limits and codes were presented in 2.024, 1.927, 1 and 1.878, which are corresponding to the
Table 1. In order to illustrate the effectiveness of limits as a code of [11 1 2] as in Table 2. It introduces the same output of
factor to improve the considered DGA methods, 12 randomly the proposed model as fault type 3, which represents high-
selected samples were taken under consideration with the energy discharge fault case (D2).
wrong fault diagnostics and explain how the new limits correct
the diagnostic fault types. Table A1 illustrates selected data 4.3 IEC 60599 CODE METHOD
with gas concentration, ratio values, code for each Rogers’ IEC 60599 code was standardized by international electro-
four ratios and PSO-FS Rogers’ four ratio methods where the technical commission in 1978 [1]. The method uses the five
complete data are summarized in [25]. combustible gases (H2, CH4, C2H2, C2H4 and C2H6) to

Authorized licensed use limited to: University of Canberra. Downloaded on May 01,2020 at 08:16:30 UTC from IEEE Xplore. Restrictions apply.
226 I. B. M. Taha et al.: Optimal Ratio Limits of Rogers’ Four-Ratios and IEC 60599 Code Methods Using Particle Swarm Optimization

generate codes based on their concentration using only three generated to predict the fault type of the remaining 96 samples
ratios, which can be expressed by: (Testing samples).
C2 H2 CH4 C2 H4
R3 R4 R2 = (7)
C2 H4 H2 C2 H6

Figure 7. Accuracy variation during the iterative process of the enhanced IEC
code.

Figure 8 explains the FS model for IEC method with new

limits and codes. The fuzzy system model built using a
Sugeno type with three inputs and one output. Each input has
three memberships while the output (f(u)) has one of six fixed
values 1, 2, 3, 4, 5 and 6, which represented six fault types PD,
D1, D2, T1, T2 and T3 respectively. The first input has three
memberships. The range of first membership is [0 0.3], the
range of second membership is [0.3 1.3] and the range of third
membership is [1.3 4], which converted to code as presented
Figure 6. Rules of PSO-FS model of Rogers’ four-ratio method. in Table 3. The other two ratios were built with the same
manner. The relation between input and output is figured by
26 rules for fault type detection, which extracted from the
Table 2. The fault types and corresponding new codes of the enhanced
Rogers’ four-ratio method.
optimal PSO-FS optimal solution.
Enhanced Roger Code Current Rogers Code
Fault Type
C1 C2 C3 C4 C1 C2 C3 C4
0|1 0 0 5
0|1 1 0 5 0 0 1|2 5
0|1 0 1 5
PD
0 2 0 5
0 0 0 2 0 0 0 5
0 0 2 5
1 2 0 5 Figure 8. Fuzzy system of the enhanced IEC code method.
1 1 1 5
1 0|1 2 5 0 0 1 0
D1 Table 3. Ratios and codes of current and PSO-FS approach for IEC method.
0|1 2 2 5
0 1 2 2 Gas Ratio Current Limits Enhanced Limits Code
1 2 2 2 R 0.1 R 0.3 0
0|1 2 1 5
0 1|2 1|2 0 R 0.1 R 3.0 0.3 𝑅 1.3 1
0|1 1 1 2
0|1 2 1 2 R 3.0 R 1.3 2
D2
0|1 0 2 2
1 1 2 2 R 0.1 R 0.12 1
0 2 2 0
0 2 2 2 R 0.1 R 1.0 0.12 𝑅 0.18 0
1 0 1 2
1 0 0 0|1|2 R 1.0 R 0.18 2
T1 1 0 0 2
0 0 0 1|2 R 1.0 R 1.0 0
T2 0|1 1 0 2 0 1 0 0
R 1.0 R 3.0 1.0 𝑅 3.42 1
T3 0|1 2 0 2 0 1|2 0 1
R 3.0 R 3.42 2

4.4 ENHANCED IEC 60599 CODE METHOD Table 3 presents the gas ratio limits of current and PSO-FS
The procedure, which used to obtain the optimal ratio limits approach for IEC method, while their corresponding codes and
of the IEC 60599 code based on PSO-FS approach, is similar the fault types presented in Table 4. Therefore, the fault type
to the mentioned ones in Rogers’ four-ratio method. Figure 7 can be identified directly using Table 3 and Table 4 using
presents the variation of total diagnostic accuracy during the PSO-FS approach for IEC code. The variation of DGA ratio
iterative process using PSO-FS for the training 385 data limits for both IEC Code and Rogers’ four ratios modify the
samples. At the end of the iterative process, the optimal wrong fault diagnostics to correct fault type based on the
solution represents the ratio limits and rules where new codes actual faults of the collected samples.

Authorized licensed use limited to: University of Canberra. Downloaded on May 01,2020 at 08:16:30 UTC from IEEE Xplore. Restrictions apply.
IEEE Transactions on Dielectrics and Electrical Insulation Vol. 27, No. 1; February 2020 227
Table 4. Codes of enhanced and current IEC 60599 models. Table 5. The references of the collected datasets.
Enhanced IEC Codes Current IEC 60599 Code Reference PD D1 D2 T1 T2 T3 Sum
Fault Type
C1 C2 C3 C1 C2 C3 [3] 4 3 2 1 10
0 1 0|1 [4] 10 26 49 1 4 20 110
0 1 2 0|2 1 0 [9] 2 1 1 2 1 7
0 0 0 [13] 27 43 57 72 20 28 247
PD
1 0|1 0 [14] 1 2 3
1 0 1 1 1 0 [15] 3 4 4 3 5 19
2 0 0 [16] 1 1 2 1 1 6
1 1 1|2 2 0 1
[17] 1 3 2 1 7 14
D1 2 1 0|1|2 2 0 2
[18] 1 1
2 0 1|2 1 0 1
[19] 3 2 5
1 0 2
[20] 1 4 15 15 35
D2 1 2 0|1|2 1 0 2
[21] 4 1 2 7
2 2 0|1|2
[22] 1 2 1 4
0 2 0
T1 0 2 0 [23] 1 2 3
0 0 1
[24] 1 4 4 1 10
T2 0 2 1 0 2 1
Sum 48 79 126 95 49 84 481
T3 0 0|2 2 0 2 2

The effectiveness of PSO-FS limits for Rogers’ four ratios 5.1 ANALYSIS OF ROGERS’ FOUR-RATIO MODELS
to correct the fault diagnostic of 12 selected samples with gas The results of the comparison between the enhanced model,
concentration, ratio values, current and new IEC code were the refined model [8] and the current Rogers’ four-ratio model
illustrated in Appendix A at Table A2 where the complete data [1] are introduced in Table 6. The results illustrated that the
are summarized in [25]. enhanced model has highest accuracy to diagnose the six-fault
Figure 9 presents a case study of the crisp rules of the PSO- type than that of the refined model and the current Rogers’
FS for IEC method. The values of the three input ratios R3, R4 four-ratio model in most cases. The overall accuracies during
and R2 are 0.1748, 1.262 and 2.194 respectively which the training stage were 86.49, 63.12, and 46.75% for the
coverted to a code of [0 2 1] by new limts in Table 3. The enhanced, refined, and current Rogers’ four-ratio models,
corresponding fault type is 5 (Medium thermal fault, T2). respectively, while the overall accuracies were 82.29, 63.54,
and 48.96%, respectively during the testing stage. The overall
accuracies for all 481 samples (Training and testing samples)
were 85.65, 63.2 and 47.19 for the enhanced, refined, and
current Rogers’ four-ratio methods, respectively. Appendix A
presents randomly selected 12 samples as in Table A3 to
verify the proposed codes, which contains two samples for
each fault type. The data includes the actual collected five
combustible gases, the ratios, the corresponding fault type and
the diagnosis results for PSO-FS Rogers’ and IEC Codes. The
analysis of all data samples summarized in [25].

Table 6. Accuracies of training and test sets for Rogers’ four-ratio.

Training Set Test Set
Fault Type
Enhanced Refined Current Enhanced Refined Current
PD 94.87 38.46 33.33 88.89 22.22 22.22
D1 74.6 9.52 1.59 56.25 18.75 0
Figure 9. Rules of FS model of enhanced IEC method. D2 89.11 70.30 56.44 84 72 64
T1 86.84 85.53 84.21 84.21 94.74 94.74
T2 89.47 60.53 36.84 90 30 20
5 DISCUSSION AND ANALYSIS
T3 86.76 92.65 45.59 94.12 100 52.94
The proposed PSO-FS models built using 481 dataset All 86.49 63.12 46.75 82.29 63.54 48.96
samples that divided into 385 which were used during
optimization process and 96 as unforeseen dataset samples for
testing. The samples were collected from Egyptian Electric 5.2 ANALYSIS OF IEC 60599 CODE MODELS
Utility and literatures. The distribution of the faults according The obtained results using PSO-FS model for IEC 60599-
to each reference was explained as in Table 5. For example, code method are compared with that of the refined IEC code
the samples from reference [3] included 4 samples as PD fault, method in [10] and the current IEC 60599-code method in [1]
3 samples as T1, 2 samples as T2, and only 1 sample as T3. as in Table 7. The results clarified that the overall accuracy

Authorized licensed use limited to: University of Canberra. Downloaded on May 01,2020 at 08:16:30 UTC from IEEE Xplore. Restrictions apply.
228 I. B. M. Taha et al.: Optimal Ratio Limits of Rogers’ Four-Ratios and IEC 60599 Code Methods Using Particle Swarm Optimization

was enhanced from 55.32% with current IEC 60599 code to Table 9. Accuracies of different DGA techniques for 359 samples.
68.57% with the refined IEC. On the other hand, the accuracy DGA technique 481 samples 359 samples
of the enhanced model using PSO-FS increased to 85.71% for 1 IEC Code 55.09 73.82
training datasets. The overall accuracies with testing datasets 2 Rogers’ four ratios 47.19 59.89
were 54.17, 69.79 and 82.29% for current IEC 60599, refined 3 Refining IEC Code 68.82 74.93
IEC, and enhanced model, respectively. 4 Refining Rogers’ four ratios 63.2 69.08
5 Duval 64.61 68.08
Table 7. IEC method comparisons for both training and test sets.
6 Clustering 80.67 77.16
Fault Training Set Test Set 7 Probability 84.2 83.84
Type Enhanced Refined Current Enhanced Refined Current 8 CSUS 77.55 77.16
PD 92.31 66.67 48.729 77.78 33.33 22.22 9 PSO-FS IEC model 85.03 86.35
D1 65.08 31.75 38.10 50 43.75 37.50 10 PSO-FS Rogers’ four ratios 88.65 87.19
D2 90.1 79.21 44.55 88 72 48
T1 86.84 63.16 57.89 89.47 78.95 68.42
T2 89.47 76.32 68.42 90 80 70 6 CONCLUSIONS
T3 91.18 89.71 80.88 94.12 94.12 70.59 Particle swarm optimization (PSO) algorithm and fuzzy
All 85.71 68.57 55.32 82.29 69.79 54.17 logic system (FS) approach was utilized to enhance the
accuracies of the transformer fault diagnosis using Rogers’
four-ratio and IEC 60599 code methods. The proposed models
5.3 COMPARISON BETWEEN DIFFERENT METHODS
were built using 481 collected datasets from Egyptian Electric
In this section, a comparison with most of the efficient fault Utility and literature. The rules of fuzzy logic system in the
diagnostic methods used to verify the proposed models. The optimal solution were converted to codes corresponding to
proposed PSO-FS models of both IEC and Rogers’ four-ratio each fault type based on the gas ratio limits. The overall
methods were compared with Duval triangle three ratios accuracies of the proposed models for IEC and Rogers’ four-
method (graphical method), California State University ratio were 85.03 and 85.65% compared to 55.09 and 47.19%
Sacramento (CSUS) [5], clustering method [6] and probability
of the current IEC 60599 and Rogers’ four-ratio methods,
method [8]. The accuracies of the proposed models of IEC and
respectively. The proposed models were also compared with
Rogers’ four-ratio methods were 85.05% and 85.65%,
other DGA methods such as Duval triangle, clustering, CSUS,
respectively, in which they were higher than that of a Duval
and probability methods, which have overall accuracies of
triangle (64.61%), clustering method (80.67%), CSUS
64.61, 80.67, 77.55, and 84.20% respectively. The accuracies
(77.55%), and the probability method (84.20%) as shown in
of the previous methods were less than the proposed PSO-FS
Table 8. The results clarify that the proposed models for both
models for IEC 60599 and Rogers’ four-ratio methods.
IEC and Rogers’ four-ratio methods have high accuracies
Therefore, the proposed models have high abilities to diagnose
compared to the other methods.
the transformer fault types with simple codes and high
Table 8. Comparison between different methods with 481 dataset. accuracy. However, when datasets due to mixtures of faults
PSO-FS PSO-FS (samples with undetermined or misidentified faults) are
FT Duval Clustering Probability CSUS removed from the calculations, the accuracies of the PSO-FS
IEC Rogers
PD 47.92 93.75 89.58 87.50 89.58 93.75 methods are not significantly better than when using existing
D1 73.42 74.68 65.82 59.49 62.03 70.89 DGA methods.
D2 63.49 85.71 94.44 87.30 89.68 88.10
T1 48.42 89.47 80 89.47 87.37 86.32 REFERENCES
T2 50 39.58 87.50 50 89.58 89.58 [1] Mineral oil-filled electrical equipment in service — Guidance on the
interpretation of dissolved and free gases analysis, IEC 60599Edition
T3 94.12 84.71 85.88 76.47 91.76 88.24
2.1, 2007-05
All 64.61 80.67 84.20 77.55 85.03 85.65 [2] IEEE Guide for the Interpretation of Gases Generated in Oil-Immersed
Transformers, IEEE Standard C57.104-2008, Feb 2009
[3] M. Duval, “A review of faults detectable by gas-in-oil analysis in
The samples data (122 samples), which gave undetermined transformers,” IEEE Elec. Ins. Mag., 18, (3), pp. 8–17, 2002.
fault types with the IEC standard code and considered as [4] M. Duval, and A. DePablo, “Interpretation of gas-in-oil analysis using
mixture faults, were reviewed and removed from 481 samples. new IEC publication 60599 and IEC TC 10 databases,” IEEE Elec. Ins.
Mag., vol. 17, no. 2, pp. 31–41, 2001.
By evaluating the accuracy of the proposed PSO-FS models [5] S. S. M.Ghoneim, I. B.M. Taha and N. I. Elkalashy, “Integrated ANN-
with the new limits and comparing them with all DGA based proactive fault diagnostic scheme for power transformers using
techniques, the results are shown in Table 9. However, when dissolved gas analysis,” IEEE Trans. Dielectr. Electr. Insul, vol. 23, no.
3, pp. 1838–1845, 2016.
datasets due to mixtures of faults (samples with undetermined [6] S. S. M. Ghoneim and I. B. M. Taha,”A new approach of DGA
or misidentified faults) are removed from the calculations, the interpretation technique for transformer fault diagnosis,” Int. Jour. of
accuracies of the PSO-FS methods are not significantly better Elec. Power and Energy Sys., 81, pp. 265–274, 2016.
than when using existing DGA methods. [7] I. B. M. Taha, S. S. M. Ghoneimand and A. S. A. Duaywah, “Refining
DGA methods of IEC code and Rogers four-ratios for transformer fault
diagnosis,” IEEE PES General Meeting, 2016, pp. 17–21.

Authorized licensed use limited to: University of Canberra. Downloaded on May 01,2020 at 08:16:30 UTC from IEEE Xplore. Restrictions apply.
IEEE Transactions on Dielectrics and Electrical Insulation Vol. 27, No. 1; February 2020 229

APPENDIX
Table A1. Application of current and PSO-FS limits of Rogers’ Four ratio methods.
PSO‐FS‐ROG Current Rogers' four
H2 CH4 C2H6 C2H4 C2H2 ACT R1 R2 R3 R4
C1 C2 C3 C4 Diagnose C1 C2 C3 C4 Diagnose
117 17 1 3 1 PD 0.058824 3 0.333333 0.145299 0 1 0 5 PD 0 1 0 0 T2
116 14 134 5.3 0.001 PD 9.571429 0.039552 0.000189 0.12069 1 0 0 5 PD 1 0 0 0 T1
266 30.2 4.9 26.2 60.2 D1 0.162252 5.346939 2.29771 0.113534 0 2 2 5 D1 0 2 2 0 D2
1330 10 20 66 182 D1 2 3.3 2.757576 0.007519 1 1 2 5 D1 1 2 2 5 UD
114 41 27 15 42 D2 0.658537 0.555556 2.8 0.359649 1 0 2 2 D2 0 0 2 0 D1
110 62 90 140 250 D2 1.451613 1.555556 1.785714 0.563636 1 1 2 2 D2 1 1 2 0 UD
625 130 47 2 0.0001 T1 0.361538 0.042553 0.00005 0.208 1 0 0 2 T1 0 0 0 0 Nf
12 18 4 4 0.001 T1 0.222222 1 0.00025 1.5 1 0 0 2 T1 0 0 0 1 T2
30.4 117 44.2 138 0.1 T2 0.377778 3.122172 0.000725 3.848684 1 1 0 2 T2 0 2 0 2 UD
15 159 29 87 0.001 T2 0.18239 3 1.15E‐05 10.6 0 1 0 2 T2 0 1 0 2 UD
35 29 7 78 0.001 T3 0.241379 11.14286 1.28E‐05 0.828571 1 2 0 2 T3 0 2 0 0 UD
24 20 4 87 0.001 T3 0.2 21.75 1.15E‐05 0.833333 0 2 0 2 T3 0 2 0 0 UD

Table A2. Application of current and PSO-FS limits of IEC 60599 methods.
PSO‐FS IEC Current IEC 60599
H2 CH4 C2H6 C2H4 C2H2 ACT R1 R2 R3 R4
C1 C2 C3 Diagnose C1 C2 C3 Diagnose
117 17 1 3 1 PD 0.058824 3 0.333333 0.145299 1 0 1 PD 1 0 2 D2
116 14 134 5.3 0.001 PD 9.571429 0.039552 0.000189 0.12069 0 0 0 PD 0 0 0 Nf
266 30.2 4.9 26.2 60.2 D1 0.162252 5.346939 2.29771 0.113534 2 1 2 D1 1 0 2 D2
1330 10 20 66 182 D1 2 3.3 2.757576 0.007519 2 1 1 D1 1 1 2 UD
114 41 27 15 42 D2 0.658537 0.555556 2.8 0.359649 2 2 0 D2 1 0 0 UD
110 62 90 140 250 D2 1.451613 1.555556 1.785714 0.563636 2 2 1 D2 1 0 1 D1
625 130 47 2 0.0001 T1 0.361538 0.042553 0.00005 0.208 0 2 0 T1 0 0 0 Nf
12 18 4 4 0.001 T1 0.222222 1 0.00025 1.5 0 2 0 T1 0 2 1 T2
30.4 117 44.2 138 0.1 T2 0.377778 3.122172 0.000725 3.848684 0 2 1 T2 0 2 2 T3
15 159 29 87 0.001 T2 0.18239 3 1.15E‐05 10.6 0 2 1 T2 0 2 2 T3
35 29 7 78 0.001 T3 0.241379 11.14286 1.28E‐05 0.828571 0 2 2 T3 0 0 2 UD
24 20 4 87 0.001 T3 0.2 21.75 1.15E‐05 0.833333 0 2 2 T3 0 0 2 UD

Table A3. Actual data of 12 randomly selected samples.

Sample No. H2 CH4 C2H6 C2H4 C2H2 R1 R2 R3 R4 ACT PSO-FS Rogers’ PSO-FS IEC 60599
1[20] 870 77 73 54 14 0.948 0.74 0.259 0.089 PD PD PD
2 [16] 1651 90 33 45 2 0.367 1.364 0.044 0.055 PD PD PD
3 [4] 1330 10 20 66 182 2 3.3 2.758 0.008 D1 D1 D1
4 [4] 4 1 2 7 52 2 3.5 7.4 0.25 D1 D2 D2
5 [10] 60 5 2 21 21 0.4 10.5 1 0.083 D2 D2 D1
6 [15] 200 48 14 117 131 0.29 8.35 1.12 0.24 D2 D2 D2
7 [13] 6 197 476.9 16.9 0 2.42 0.035 0.00005 32.83 T1 T1 T1
8 [13] 24.7 20.4 117 71 6.8 5.73 0.61 0.096 0.83 T1 T1 T1
9 [10] 1270 3450 520 1390 8 0.151 2.673 0.006 2.717 T2 T2 T2
10 [10] 360 610 259 260 9 0.425 1.004 0.035 1.694 T2 T2 T2
11 [13] 0 3.4 0 54 0 0.0003 54000 0.00001 3400 T3 T3 T3
12 [13] 11 13 2 127 0 0.154 63.5 0.000007 1.18 T3 T3 T3

[8] S. I. Ibrahim, S. S. M. Ghoneim and I. B. M. Taha, “DGALab: an [16] Zhu Y. Li, F. Wang and L.-q.Geng, “Transformer fault diagnosis based
extensible software implementation for DGA,” IET Gen., Trans.& Dist., on naive Bayesian classifier and SVR,” IEEE region 10 Conf.
Vol. 12 Iss. 18, pp. 4117–4124, July 2018. (TENCON), Nov 2006. p. 1–4.
[9] K. Bacha, S. Souahlia and M. Gossa, “Power transformer fault diagnosis [17] D.V.S.S.S. Sarma G.N.S. Kalyani, “ANN approach for condition
based on dissolved gas analysis by support vector machine,” Electric monitoring of power transformers using DGA,”IEEE Reg. 10 Conf.
Power System Research, 83, (1), pp. 73–79, 2012. (TENCON), 2004, pp.444–447.
[10] Md M. Islam, G. Lee and S. N.Hettiwatte, “Application of Parzen [18] G. Zhang, K. Yasuoka ,S. Ishii, “Application of Fuzzy Equivalent
Window estimation for incipient fault diagnosis in power Matrix for Fault Diagnosis of Oil-immersed Insulation,”IEEE 13th Int.
Transformers,”IET High Voltage, vol. 3, no. 4, pp. 303–309, 2018. Conf. Dielectr. Liquids (ICDL), 1999, pp. 400–403.
[11] A. Hoballah, I. Erlich,“PSO-ANN approach for transient stability [19] J. T. Hu, L. X. Zhou and M.L. Song, “Transformer Fault Diagnosis
constrained economic power generation,” PowerTech, 2009, pp. 1–6. Method of Gas Hromatographic Analysis Using Computer image
[12] A. Hoballah and I. Erlich, “Dynamic stability and network constrained analysis,”Int. Conf. Intelligent Sys. Design and Eng. App., Jan. 2012, pp.
optimal spinning reserve allocation,” 2011 IEEE Power and Energy 1169–1172.
Society General Meeting, 2011, pp. 1–5. [20] O.E Gouda, Saber Salem and Salah Hamdy El-Hoshy, “Power
[13] Egyptian Electricity Holding Company (EEHC) Reports, 1991-2016. transformer incipient faults diagnosis based on dissolved gas analysis,”
[14] A. Sanjay and A. K. Chandel,“Transformer incipient fault diagnosis TELKOMNIKA Indonesian J. Electr. Eng., vol. 17, no. 1, pp. 10–16,
based on probabilistic neural network,”2012 Students Conf. Eng. Jan 2016.
Sys.SCES), Mar. 2012. p. 15. [21] S. Seifeddine, B. Khmais and C. Abdelkader, “Power transformer fault
[15] M.-H. Wang, “A novel extension method for transformer fault diagnosis based on dissolved gas analysis by artificial neural
diagnosis,” IEEE Trans. on Power Del., 18(1), pp. 164–169, 2003. network,”1st Int. Conf. Renewable Energies and Vehicular Technology
(REVET), 2012, pp. 230–236.

Authorized licensed use limited to: University of Canberra. Downloaded on May 01,2020 at 08:16:30 UTC from IEEE Xplore. Restrictions apply.
230 I. B. M. Taha et al.: Optimal Ratio Limits of Rogers’ Four-Ratios and IEC 60599 Code Methods Using Particle Swarm Optimization
[22] Zheng Qiaogen, Ke Wang and Yiyi Zhang, “Optimal dissolved gas
ratios selected by genetic algorithm for power transformer fault
diagnosis based on Support vector machine,” IEEE Trans. Dielectr.
Electr. Insul, vol. 23, no. pp. 1198–1206, Apr 2016.
[23] M. Rajabimendi and E.P. Dadios, “A hybrid algorithm based on neural
fuzzy system for interpretation of dissolved gas analysis in power
transformers,”IEEE Reg. 10 Conf. (TENCON), 2012, Nov. 2012.
[24] R. Soni and K. R. Chaudhari, “A Novel Proposed Model to Diagnose
Incipient Faults of Power Transformer Using Dissolved Gas Analysis by
Ratio methods,”Int. Conf. on Comp. of Power, Energy, Information and
Communication, April 2015.
[25] https://drive.google.com/open?id=1t01Fl6KFKTKLz2l_CW6UDjTJZcw
pmnCW.

Authorized licensed use limited to: University of Canberra. Downloaded on May 01,2020 at 08:16:30 UTC from IEEE Xplore. Restrictions apply.