Abstract
M. Ing.
Fault detection tools have gained popularity in recent years due to the increasing need
for reliable and predictable equipments. Transformer bushings account for the majority
of transformer faults. Hence, to uphold the integrity of the power transmission and dis-
tribution system, a tool to detect and identify faults in their developing stage is necessary
in transformer bushings. Among the numerous tools for bushings monitoring, dissolved
gas analysis (DGA) is the most commonly used. The advances in DGA and data storage
capabilities have resulted in large amount of data and ultimately, the data analysis crisis.
Consequent to that, computational intelligence methods have advanced to deal with this
data analysis problem and help in the decision-making process.
Numerous computational intelligence approaches have been proposed for bushing fault
detection. Most of these approaches focus on the accuracy of prediction and not much
research has been allocated to investigate the interpretability of the decisions derived
from these systems. This work proposes a rough set theory (RST) model for bushing
fault detection based on DGA data analyzed using the IEEEc57.104 and the IEC 60599
standards. RST is a rule-based technique suitable for analyzing vague, uncertain and
imprecise data. RST extracts rules from the data to model the system. These rules are
used for prediction and interpreting the decision process. The lesser the number of rules,
the easier it is to interpret the model. The performance of the RST is dependent on
the discretization technique employed. An equal frequency bin (EFB), Boolean reasoning
(BR) and entropy partition (EP) are used to develop an RST model. The model trained
using EFB data performs better than the models trained using BR and EP. The accuracy
achieved is 96.4%, 96.0% and 91.3% for EFB, BR and EP respectively. This work also pro
poses an ant colony optimization (ACO) for discretization. A model created using ACO
discretized achieved an accuracy of 96.1%, which is compatible with the three methods
above. When considering the overall performance, the ACO is a better discretization tool
since it produces an accurate model with the least number of rules. The rough set tool
proposed in this work is benchmarked against a multi-layer perceptron (MLP) and radial
basis function (RBF) neural networks. Results prove that RST modeling for bushing is
equally as capable as the MLP and better than RBF. The RST, MLP and RBF are used
in an ensemble of classifiers. The ensemble performs better than the standalone models.