IEEE Systems, Man and Cybernetics Magazine - October 2022 - 16

ImageNet is used to train VGG-19, and then the knowledge
of this deep CNN is transferred. By using a few layers for a
fine-tuning purpose, the newly built deep CNN is capable
of distinguishing the corrupted and intact insulators. This
method is able to diagnose these faults using the aerial
images taken from TLs in different environments. The
original dataset used in this article is the Chinese Power
Line Insulator Dataset (CPLID),
which is an imbalanced dataset
and includes only 3,808 insulator
images. Therefore, a random
image-augmentation procedure is
proposed and applied to generate a
more suitable dataset with 16,720
images. This new dataset allows us
to offer higher detection accuracy
than the original one because it is
a balanced dataset. Training a
deep CNN by using it gives more
power to the system for detecting
the corrupted insulators in different
situations such as rotated,
dark, and blurry images with complex
backgrounds. The comparison results of this study
show the advantages of the proposed method over various
existing ones.
Introduction
Increasing demand for electrical energy results in expansion
of power systems, including TLs. Considering the
vital role of insulators in TLs for mechanical support and
electrical insulation during power grid operations, and
their proneness to faults and breakage, intelligent methodologies
have been proposed to enhance their safety
and reliability [1]-[3]. Due to insulators' exposure to outdoor
environments for a long period of time, they face
" missing faults, " which are defined as broken insulators
(see Figure 1).
Insulator faults are variant and random, and their
occurrences interrupt the safety and stability of the entire
TL operation, thereby imposing tremendous economic
losses. To guarantee the safety and stability of TLs and
intact operation of power grids, insulator fault detection
and intelligent inspection have been considered salient
tasks [1], [4].
Finding insulator faults using a traditional manual patrol
Transfer learning
is generally used
when a new dataset
is smaller than the
primary dataset with
which a base model
is well trained.
is inefficient, time consuming, and wastes human resources.
Hence, it has been gradually replaced by an unmanned
aerial vehicle (UAV) patrol, as discussed in [5] and [6]. In this
new process, workers do not require
investigating TL insulator faults
using telescopes. However, the complexity
and variability of UAV application
scenarios have caused
several new challenges in the autonomous
detection of insulator faults
[1]. Artificial intelligence (AI)-based
methodologies are among the most
important for addressing them.
There are many studies on insulator
fault identification using aerial
images and image processing
methodologies. In their first step,
traditional image processing algorithms
classify insulator images
into classes with specific features such as texture, color,
and shape. Then they take advantage of matching algorithms
to implement fault-detection procedures. Their
drawback is the fact that the features are designed manually
and, therefore, they are inefficient in complicated power
grids with a variety of image features. Moreover, correlation
features of insulators in aerial images are not clear,
and the accuracy of these algorithms is highly dependent
on these features [7]-[10].
Recently, tremendous progress in AI has led to noticeBroken
Insulators
able advancements in image processing methods based
on deep neural networks. CNNs are the most common
deep neural networks that perform pattern recognition
and object-detection tasks, are able to extract image features
automatically, and learn under various environmental
conditions [11]-[13] more than traditional ones. These
CNN-based techniques overcome the limitations of traditional
image processing approaches and conclude higher
performance and better accuracy in object-detection
tasks [14]-[16]. Although CNNs have completely dominated
the image processing area in recent years and their
results are much better than traditional methodologies,
they have some drawbacks, such as requiring big datasets
and considerable time for training. Therefore, an
advanced technique like transfer learning has been proposed,
which makes deep CNNs capable of generating
results faster and transferring knowledge from some
already-trained CNNs to a new CNN that has common
features with the base dataset for the purpose of saving
time and resources [17].
A model gets trained and developed for one dataset
Figure 1. " Missing faults " (broken insulators) of TLs.
16 IEEE SYSTEMS, MAN, & CYBERNETICS MAGAZINE October 2022
first. It is then reused for a second related one, which is
called transfer learning. In other words, a learned matter
IEEE Systems, Man and Cybernetics Magazine - October 2022

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