Vol.:(0123456789) 1 3
Journal of Control, Automation and Electrical Systems
https://doi.org/10.1007/s40313-020-00625-5
Detection and Classifcation of Incipient Faults in Three‑Phase Power
Transformer Using DGA Information and Rule‑based Machine Learning
Method
Mohsen Savari Katooli
1
· Amangaldi Koochaki
1
Received: 28 December 2019 / Revised: 9 May 2020 / Accepted: 27 June 2020
© Brazilian Society for Automatics--SBA 2020
Abstract
Three-phase transformers (TPT) play a signifcant and crucial function in the power networks in order to connect the sub-
systems and deliver the electrical energy to fnal customers. The TPT are one of the most high-priced equipment in modern
power networks, and therefore their working condition should be constantly monitored to prevent their breakdown, power
outages and huge fnancial damage. Accordingly, this paper presents a hybrid method for detection and classifcation of
incipient faults in TPT using dissolved gas analysis techniques (DGAT) information and rule-based machine learning method.
In the developed method, the most informative and important items of DGAT data out of 14 items selected by association
rules mining technique (ARMT) are employed as the input of adaptive neuro-fuzzy inference system (ANFIS). The ARMT is
implemented to select the items, which have maximum information and can train the ANFIS more accurately. Furthermore,
in order to enhance the accuracy of ANFIS and improve its robustness in diferent implementations, black widow optimiza-
tion algorithm is applied for ANFIS training. In order to evaluate the performance of developed method on real issues, two
industrial data collections obtained from Iran-Transfo Company chemical laboratory and Damavand power substations are
used. The obtained results through MATLAB simulations proved that the developed method has high fault detection and
classifcation accuracy, robust function in diferent implementations, short run time and simple structure.
Keywords BWO · Fuzzy rules · Item selection · Learning algorithm · Transformer
List of symbols
C
2
H
6
Ethane
C
2
H
4
Ethylene
CH
4
Methane
C
2
H
2
Acetylene
H
2
Hydrogen
σ Sigma
ARMT Association Rules Mining Technique
ANFIS Adaptive Neuro-Fuzzy Inference System
BRO Bayesian Regularization Optimization
BWO Black Widow Optimization
BGAFG Binary GA With Feature Granulation
BGAFG Binary GA With Feature Granulation
BMOPSO Binary Multi-Objective Particle Swarm
Optimization
CFR Composition of Feature Relevancy
ConvNet Convolutional Neural Networks
CFNN Cascade-Forward Neural Network
CHC-GA GA with cataclysmic mutation, heterogene-
ous recombination and cross-generational
elitist selection
BPA Back Propagation Algorithm
DT Decision Tree
DGAT Dissolved Gas Analysis Technique
DEU Discriminative Embedded Unsupervised
DRA Doernenburg’s Ratio Approach
ED Energy Discharges
FKN Fuzzy K-Nearest Neighbors
GA Genetic Algorithm
GWN GA-Based Wavelet Networks
KRR Kernel Ridge Regression
LM Levenberg–Marquardt
MLPNN Multilayer Perceptron Neural Network
MF Membership Function
MSE Mean Square Error
NB Naive Bayes
* Amangaldi Koochaki
koochaki@aliabadiau.ac.ir
1
Department of Electrical Engineering, Aliabad Katoul
Branch, Islamic Azad University, Aliabad Katoul, Iran