IEEE Systems, Man and Cybernetics Magazine - October 2022 - 17

in one setting is exploited to facilitate another one for fast
knowledge acquisition. Transfer learning is generally used
when a new dataset is smaller than the primary dataset
with which a base model is well trained.
Several studies in the literature take advantage of
CNNs for fault diagnosis of TLs. Fahim et al. [18] propose
a self-attention CNN framework and a time-series
image-based feature extraction model for fault diagnosis
of TLs with length of 100 (km) using a discrete
wavelet transform (DWT) for removing the noise of the
faulty voltage and current signals. The work [19] uses a
customized CNN for fault diagnosis of 50 (km) TLs integrated
with distributed generators. In [20], a machine
learning-based CNN for TLs with a length of 280 (km) is
used to execute fault diagnosis using a DWT for feature
extraction.
Shiddieqy et al. [21] propose a methodology that uses
all TL fault features to generate various models for robust
fault detection. They perform various AI approaches,
including CNN, to obtain 100% detection accuracy. In [22],
a scheme to diagnose faults in power
TLs with a length of 200 (km) using
convolutional sparse auto-encoders is
proposed. This approach learns the
extracted features from the voltage
and current signals for fault diagnosis.
Many researchers have applied
ImageNet for hierarchical convolutional feature extraction
of visual object-tracking images.
In this article, we propose utilizing the transfer learning
method to develop a fault-detection system for distinguishing
intact and broken insulator images for the first time.
The contributions of this work are two-fold:
1) implementing an image-augmentation procedure to generate
a large, labeled dataset for insulator image classification
based on the CPLID such that we have a balanced
dataset
2) proposing a transfer learning technique for the balanced
dataset using a VGG-19 CNN and an ImageNet dataset,
which outperform the existing methodologies.
Preliminaries
transfer learning to various datasets in
different domains [23]-[28]. In [29], the
authors show how to transfer deep
CNN knowledge in real-time dataset
classification. The study in [30] proposes
an intelligent approach using a deep
convolutional transfer learning network,
which distinguishes the dynamic
system faults using an unlabeled dataset.
Shao et al. [31] present an intelligent fault diagnosis methodology
for a rotor-bearing system using an updated CNN
with transfer learning. Li et al. [32] propose a deep adversarial
transfer learning network to investigate unknown emerging
faults in rotary machines.
In this study, the VGGNet model is used as a basic
Regional receptive
fields, which are only
a small, focused area
of the input data,
are used to connect
to each node in a
convolutional layer.
CNNs
CNNs are capable of learning hierarchical features independent
from inputs, making them a common solution
for imagery dataset problems. The structure of
CNNs empowers them to have
the fewest requirements for
data preprocessing and handle
da t a s e ts with numerous
features more efficiently in
comparison to most of the artificial
neural networks [36],
[37]. In general, CNNs consist
of three classes of layers: convolution,
pooling, and fully
connected. The convolutional
and pooling layers build convolution
blocks, which are
placed one after another for
the purpose of feature extraction.
Fully connected layers
play the role of classifiers and the role of the output layer
as a fully connected one is to perform the classification
or regression tasks.
CNN architectures are made of regional receptive
model for training on a primary dataset (ImageNet) and
then reused to learn/transfer features for training on an
insulator imagery dataset. Taking advantage of the initial
training, transfer learning allows us to start with the
learned features on the ImageNet dataset and then tune
the weights and structure of the base model to match the
new dataset/task instead of starting the learning process
on the new data from scratch using random weight initialization
[33]. There are also similar studies that perform
transfer learning on various datasets using VGGNet CNNs.
Huang et al. [34] used ImageNet to pretrain a VGG19 network
for DenseBox initialization, which is defined as a
unified end-to-end full CNN framework for object detection.
Li et al. [35] adjust the VGG-19 pretrained on
fields, shared weights, and the pooling operation. Regional
receptive fields, which are only a small, focused area of the
input data, are used to connect to each node in a convolutional
layer. This feature is one of the main advantages of
CNNs that causes a considerable reduction in the number
of parameters in a CNN which, in turn, reduces its training
computational load [38].
Numerous studies have been performed in image classification
problems using CNNs to achieve better performances.
CNNs are also used for protecting TLs in various
aspects, such as their fault identification and distinguishing
intact waveforms from corrupted ones in waveform
images. These studies include two various categories: the
approaches with a focus on imagery datasets recorded
from outdoor TLs [1], [15], [16], [39] and those that contemplate
time-series voltage and current waveforms
October 2022 IEEE SYSTEMS, MAN, & CYBERNETICS MAGAZINE 17
IEEE Systems, Man and Cybernetics Magazine - October 2022

Table of Contents for the Digital Edition of IEEE Systems, Man and Cybernetics Magazine - October 2022
