IEEE Systems, Man and Cybernetics Magazine - October 2022 - 22

imbalanced large dataset to prove the effect of making
this dataset balanced.
Simulation Results and Comparisons
This section describes results of the proposed transfer
learning-based method for classification of insulator images
into two categories of broken and intact insulators. Simulations
are run on a PC with Microsoft Windows 10. This
PC uses an Intel Corei7-4710 MQ at a 2.50-GHz processor
with 8 GB of random-access memory. The programming
language used in this study is Python version 3.8. Keras 2.4
library [56] with TensorFlow 2.3 backend [57] is used to
design VGG-19, and scikit-learn library [58] is used for classification
modules.
In this study, we compared our proposed methodology
using transfer learning and image-augmentation
techniques with five benchmarks from the literature
and that were implemented for this study, which are
offered in [1].
In [1], a modified model based on You Only Look Once
(YOLO) is presented for insulator fault detection in aerial
images with compound backgrounds. Initially, aerial
images with few faults are collected in various scenes,
and then a new dataset is established using Photoshop.
To enhance feature reuse and propagation in low-resolution
feature layers, a Cross Stage Partial Dense YOLO
(CSPD-YOLO) model is presented based on YOLO-v3 and
the Cross Stage Partial Network. Finally, the feature pyramid
network and improved loss function are applied to
the CSPD-YOLO model to improve the accuracy of insulator
fault detection.
In this article, the CSPD-YOLO model and compared
models in [1] are trained and tested on the generated
dataset for comparison. We used these three networks
to perform the first step of comparison in this study.
Table 1 lists these comparison results. The purpose of
this comparison was to show that the proposed method
can perform well in half the time, which is needed for
other methodologies.
As indicated in Table 1, not only have the accuracy, precision,
and F1 score been improved using our proposed
transfer learning method, but this methodology also
reduced the fault-detection time of insulator datasets. In
this table, the running time shows the required time for
one image classification between two classes of intact or
broken insulator. The definitions of accuracy, precision,
recall, and F1 score are as follows:
Accuracy =
TP TN FP FN
TP TN
++ +
+
Precision =
Recall =
F12
Score
TP FP
TP
+
TP FN
TP
+
RecallPrecision
= # RecallPrecision#
+
.
(1)
(2)
(3)
(4)
In (1)-(3), TP is the number of correct predictions of
minority samples, TN the number of correct predictions
Table 1. A comparison among existing CNNs and the proposed
transfer learning and VGG-19 method by using the original CPLID.
Neural
Network
YOLO-v3 [59]
YOLO-v4 [59]
CSPD-YOLO [1]
Transferred VGG-19
Number of
Images
3,808
3,808
3,808
3,808
Accuracy
(%)
93.31
96.38
98.18
99.28
Precision
(%)
94
98
99
99.62
Recall
(%)
94
95
98
99
F1
Score (%)
94
97
99
99.3
Test Time
(s)
0.01
0.01
0.011
0.006
Table 2. The proposed transfer learning and VGG-19 method:
a comparison between the original and (balanced) augmented CPLID for training
while using 20% of the augmented CPLID for testing.
Neural Network
Transferred VGG-19
Transferred VGG-19
Train Dataset
3,808
(Original)
13,376
(Balanced
augmented)
22 IEEE SYSTEMS, MAN, & CYBERNETICS MAGAZINE October 2022
Test Dataset
3,344
(20% of
Augmented)
-
99.93
99.2
99.41
99.2
0.007
Accuracy
(%)
96.04
Precision
(%)
95.1
Recall
(%)
95.19
F1 Score
(%)
95.14
Test Time
(s)
0.007
IEEE Systems, Man and Cybernetics Magazine - October 2022

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