Download as pdf or txt
Download as pdf or txt
You are on page 1of 171

7th International Conference on Engineering Research, Innovation and Education

January 12-14, 2023, Sylhet, Bangladesh

PAPER ID: 077

Online Judge Contest Rating Prediction using Linear


Regression and K-Neighbors Regressor

Nasim Ahmed1,*, Rasel Hasan2 and Enamul Hassan 3


● 1
Sylhet Engineering College, Sylhet-3100, email- n1976348@gmail.com
● 2
Sylhet Engineering College, Sylhet-3100, email- raselhasan.cse11@gmail.com
3
Shahjalal University of Science and Technology, Sylhet-3114, email- enam-cse@sust.edu
*Corresponding author [Nasim Ahmed]

Abstract: In Competitive Programming, online programming contests are a


Keywords: very trending topic nowadays. Every contestant in a rated contest is very
optimistic to have positive rating changes. If a contestant gets a predicted
● Linear Regression; rating change before participating in the main contest, that may help him to
● K-Neighbors improve in their official standing, rating and practice strategy. To tackle this
challenge, we introduce a model that will predict upcoming contest ratings for
Regressor; a user. We have used two different models (Linear Regression and K-
● Online Neighbors Regressor). We have also used users’ previous contest participation
data in our model by using API of Online Judges (here we have demonstrated
Programming our work using Codeforces API), that is easy to handle and updates data from
Contest; each participation of a contestant. However, the model can be used for all the
Online Judges that organize rated programming contests. Our model gave an
● Rating Prediction; accuracy of ~90%. Moreover, the results show that the method can generalize
● Continuous well and can be applied to domain independent datasets.
Dataset;

1. Introduction
A major portion of the “Computer Science” field depends on programming. Any student of
computer science must have to the least knowledge of programming. A student builds up this skill
through practicing more and more. The best way of practicing programming is competitive
programming. It encourages the students to practice effectively and to raise their mental ability.
Several organizations and websites arrange a number of programming contests in various forms.
Most of the programming contests are online based and they have some general rules and
judgment criteria. The web platform that organizes the contests is naturally called Online Judge.
UVa Online Judge [1] is the first online judge system that has a variety of applications. During
the contest time, the online judges provide problems to the contestants and they have the
mechanism of receiving the solutions (written in programming language) from contestants and
to judge them accordingly. A programming problem’s solution can be judged more accurately

Page | 785
and faster [2] than a manual process. In [3], Luca Bez, Neilor A. Tonin showed how the URI (one
of the most renowned online judges) works and assists professors to judge students easily and
which also helps students to improve their programming skill. A large number of OJ’s have been
developed since the start of the millennium. We found a survey on Online Judge Systems [4] that
was showing that the popularity of different Online Judges had been increasing day by day and
their usage was more in vogue than the manual process. Several Online Judges organize the
contests in rated format; that is, a contestant’s profile will be assigned a numerical value based on
the performance in the contest and this will affect his overall rating. Each OJ, that organizes rated
contests may have its own rating calculation criteria but most of the times they follow a common
phenomenon (i.e. Elo Rating) [5] that works on the basis of elementary statistics and splits the
contestants into different rooms according to their rank in the contest.
It will be quite beneficial if the contestant gets a predicted rating change in a specific contest
before participation. We had found several works related to the rating prediction, but they used
to predict rating after the contest had been over and before rating updates from the OJ’s authority.
WasyIF built a chrome browser extension [6] that could predict the rating changes of a contestant
just after the contest had been over but not before the contest. NBHEXT (No ButtHurt EXTension)
[7] is another type of browser extension that allows contestants to see their rating changes at
Codeforces OJ at any time; even when the contest is over and ratings are not updated yet. The
purpose of our work was to contribute to this matter. The model will give a much accurate
prediction of rating changes to a contestant before the contest has been started. We had trained
and tested our model by using Codeforces API and we found the model gave ~90% accurate
results.
2. Theory
Regression is one of the most momentous in the field of AI. Regression aims to define how a
dependent variable changes with the change of independent variable by fitting a line to the
observed data. Our proposed approach needed Linear Regression and K-Neighbors Regression.
Hence, we will try to describe in detail about both of them.
Linear Regression: It is the simplest type of Regression algorithm. It is composed of some core
concepts of statistics and machine learning. Linear regression aims to fit a line on observed data
based on an initial Base value and Slope Coefficient. The simple mathematical representation of
Simple Linear Regression is at formula (1):

𝑌 = 𝑏0 + 𝑏 1 × 𝑋 (1)

Where, y = dependent variable, b0 = Y-intercept, b1 = slope coefficient, x = independent variable.


In the case of Multiple Linear Regression (we have used in our model), it tries to estimate how
the dependent variable changes with the change of independent variables. For Multiple Linear
Regression (we have used in our model) or Polynomial Regression, the structure of the formulas
are a bit complex but they can fit the line on observed data more accurately and also prediction
results give more accuracy on data. In a Multiple Linear Regression model, we try to approximate
how the dependent variable changes with the change in independent variables. Multiple Linear
Regression model allows multiple independent variables.

Page | 786
Linear Regression

Dependent Variable
1500
1480
1460
1440
1420
0 5 10 15 20 25
Independent Variable(s)

Fig.1. Linear Regression

Figure 1 is about Linear Regression, where the Black Arrow is the regression line and the blue
dots represent the data points. The accuracy of a Linear Regression model depends on how the
Line fits the data points.
K-Neighbors Regression: This regression algorithm works on the basis of storing all the available
events for a particular circumstance and estimates the relationship between independent
variable(s) and dependent variable using similarity measures. A simple implementation of this
algorithm is to estimate by calculating the average numerical target of the K nearest neighbors.
The algorithm can fit well with a certain value of K that should be observed and determined
carefully.

3. Proposed Method
In order to predict rating for a contestant in a rated contest, the developed model entails multiple
steps. They are Data collection, Data pre-processing, Implementation of regression model, their
training & testing, Results & Comparison respectively.

Comparing
Remove irrelevant results
OJ Linear
features
Regression

Remove Irrelevant Output from


API KNR highest
Methods
accuracy model
Regression
Data Pre-processing Results & Comparison
Data Collection

Fig. 2. System’s Block Diagram

3.1 Data collection

Page | 787
The first step of our work was gathering necessary data of contestants from their previous contests
history. For simplicity of demonstration, the API of online judges has been used. Codeforces OJ
and some other well-known OJ provide API [8], is an inbuilt API provided by the OJ’s for their
developers. To extract any information from the API of the OJ’s, one needs to go to the relevant
webpage of the OJ’s website and to follow the necessary steps. In our model, we have fetched the
past data of about 35000+ users.
3.2 Pre-processing of data
After collection of data, we processed the data through a pre-processing procedure. This includes
irrelevant features removal, unimportant method elimination, assigning value to variables etc. For
instance, when we had called the “user.rating” method from Codeforces API, some unnecessary
and irrelevant data were being fetched and we had to remove them.

3.3 Regression
After preprocessing the data, we had defined our models and initialized those using python
Machine Learning libraries. We had trained the models using the data collected from API. During
training, we had taken the previous contests’ types and the difficulty as the independent variables
and contests’ ratings as the dependent variable. After training the models, they initialized the
predicted rating for a contestant in a rated contest. We had set the number of neighbors in between
1-3 for the K-Neighbors Regression. By fitting the models with dependent variable and
independent variables, our model had completed its training phase and after it was ready for
testing.
3.4 Prediction
It was the final phase of our proposed system. At this step, the models predicted the rating for a
specific contestant in a rated contestant based on various features. Both the Linear Regression
model and the KNeighborsRegressor model could make the prediction. We then calculated the
accuracy of both models using machine learning algorithms and mathematical approaches and
then we took the output of the model which provided the highest accuracy.

4. Result and Discussions


This section will give the results we found and the performance of our proposed system. As
described earlier, our system was based on two separate regression models.

Fig.3. Collected Data from OJ’s API


Figure 3 shows the snapshot of how data was fetched from the API of one of the most popular
OJ’s (codeforces). Then we had preprocessed the data according to our previously mentioned
approach. After we had implemented the regression models and trained them.

Page | 788
The final results provided an accuracy level of ~90%. For accuracy calculation of both models,
we had used concept of MAPE (Mean Absolute Percentage Error) and Median-APE of statistical
analysis.
The value of Absolute Percentage Error (APE) and MAPE was calculated by the following
formula (3) and formula (4) respectively:
|𝑨𝒄𝒕𝒖𝒂𝒍 𝑹𝒂𝒕𝒊𝒏𝒈−𝑷𝒓𝒆𝒅𝒊𝒄𝒕𝒆𝒅 𝑹𝒂𝒕𝒊𝒏𝒈|
𝑨𝑷𝑬 = × 𝟏𝟎𝟎% (3)
𝑨𝒄𝒕𝒖𝒂𝒍 𝑹𝒂𝒕𝒊𝒏𝒈
𝟏 𝒏
𝑴𝑨𝑷𝑬 = ∑ 𝑨𝑷𝑬 (4)

Page | 789
𝒏 𝒊=𝟏 𝒊

Table 1. Accuracy Measurement Table using MAPE

Actual Predicted Absolute Difference Of Predicted Absolute Percentage


Rating Rating Rating & Actual Rating Error (APE)

1430 1445 15 1.4

1470 1477 7 0.47

1468 1477 9 0.61

1422 1432 10 0.70

MAPE 0.795

Table (1) is showing how we had calculated the Absolute Difference of Predicted Rating and
Actual Rating and how we had calculated the APE and MAPE. Finally we had calculated the
accuracy by using the formula (5):
𝑨𝒄𝒄𝒖𝒓𝒂𝒄𝒚 = 𝟏𝟎𝟎 − 𝑴𝑨𝑷𝑬 (5)
The following formula was used to calculate the overall performance given by a specific
Regression Model.

Accuracy Comparison
96
94
92
Accuracy(%)

90
88
86
84
Linear KNNR(k=1) KNNR(K=2) KNNR(K=3)
Regression
Regression Model

Fig.4. Accuracy Comparison of Various Regression Models

From our testing dataset we found the model had achieved the accuracy level ~90%. Figure (4)
showing the accuracy of used models in our proposed system. It is demonstrating the highest

accuracy was given by the KNNR with Neighbors = 3 on the average test cases whereas the Linear
Regression model was giving 89% accurate results. However, we had taken the average of them
for the overall testing data.
5. Conclusion

Page | 790
Rating in a rated contest or overall profile rating plays an important role to a competitive
programmer’s career and his improvement. It is worthy for every contestant to get positive rating
changes in the contest. If they get a good approximation of their future contest, they will be
benefited by it. Our purpose was to contribute on this subject through the proposed model. The
model can give a well accurate prediction for a future contest.

Fig.5. Developed Web Application


We have built a web application using the model and the snapshot is in Figure 5. The model can
predict with an accuracy level ~90%. We certainly conclude that the accuracy score can be
improved by using more relevant features and larger datasets. We are also working to implement
a well-defined domain independent dynamic dataset and hopefully that will lead to a good
consequence.
References
Miguel A. Revilla, Shahriar Manzoor, Rujia Liu., The UVa Online Judge Experiences,
Olympiads in informatics, 2008, vol.2, 131-148

Andy Kurnia, Andrew Lim, Brenda Cheang. Online Judge, Computers and Education.2001, 299-
315
Jean Luca Bez, Neilor A. Tonin, Paulo R. Rodegheri. (2014). URI Online Judge Academic: A
tool for Algorithms and Programming Classes.

Szymon Wasik, Maciez Antczak, Jan Badura, Artur Laskowski, Tomasz Sternal. A survey on
Online Judge Systems and their Applications.2018, ACM computing surveys, Vol.51, 1-34

Arpad E. Elo. 1978. The rating of chess players, past and present.

WasyIF. 2021. CF-rating prediction, Available in: https://github.com/WasyIF/CF-rating-


prediction1
Roman Rubanenko, Maxim Molchanov, Aleksey Kholovshuk. 2015. No ButtHurt EXTensionfor
Codeforces. Available in: https://github.com/Rubanenko/NBHEXT

Mike Mirzayanov (Codeforces, Worldwide).2010. https://codeforces.com > apiHelp

Page | 791
7th International Conference on Engineering Research, Innovation and Education
January 12-14, 2023, Sylhet, Bangladesh

PAPER ID: 082


A New Approach to Assist Wayfinding and Situation
Awareness for the Visually Impaired
Sheezanul Hassan, Md. Liakot Ali and Muhammad Sheikh Sadi
Department of Computer Science and Engineering, Bangladesh University of Engineering and
Technology, Palashi, Dhaka 1205, Bangladesh
Corresponding author: 1018312010@iict.buet.ac.bd

Abstract: Safe and independent mobility is one of the major daily challenges faced by the visually
impaired. To navigate a new area safely, they need to know the location of obstacles and other
things in their path. They struggle with object detection and obstacle avoidance, making it
challenging for them to perform their day-to-day tasks. The visually impaired should be aware of
obstructions and other interference to navigate properly in new or unfamiliar situations. However,
establishing secure and safe mobility for the visually impaired is a difficult undertaking that must
be done accurately and efficiently. Recognizing currency is another severe problem that the visually
impaired confront because different notes, in our country, have similar surfaces and sizes. The sizes
and colors of the most recent notes are causing major problems for the visually impaired. In this
paper, we propose an object and currency detection system that will benefit the visually impaired
to detect objects and Bangladeshi currency notes and make the appropriate decision regarding their
situation in both indoor and outdoor aspects. The model proposed in the system has been evaluated
with real-world objects for evaluating the performance of the proposed method. The experimental
analysis shows that the system has an average accuracy of 99.44%, a recall of 99.38%, a precision
of 99.39%, and an F1 score of 99.37% after experimenting with 5281 images of real objects. This
paper also explores how it outperforms other existing works in related fields.
Keywords: Visually impaired; Wayfinding; Situation awareness; Safe mobility; Convolutional
neural network;

1. Introduction

Experts use the term "visual impairment" to refer to any form of vision loss, ranging from partial
vision loss to total blindness. Severe or complete loss of vision may result from disease or damage
to one or more processing areas in the brain or eye and cannot be entirely recovered by medical
intervention, surgery, glasses, or contacts (Salvin, 2016). According to a report from the
International Agency for the Prevention of Blindness, around 43 million people are blind in the
world and around 295 million people have moderate to severe vision impairment (Basheshankar
et al. 2022) Vision impairment is more prevalent in least-developed countries like Bangladesh. In
Bangladesh, the number of blind people aged 30 and above is 750,000 (Ahmed et al., 2019).
Navigating through unfamiliar environments is one of the challenges for the visually impaired,
especially if they have completely lost their eyesight (Kumar, 2021). They fail to detect and avoid
obstacles in their path, thus causing them emotional suffering, undercutting their independence,
and exposing them to fracture (Jafri et al., 2018). Identifying currency-note is another major issue
for the visually impaired (Rahman et al., 2019). Hence, detecting obstacles for safe navigation
and being able to detect different banknotes can immensely improve the day-to-day life of a
visually impaired person.

Page | 792
For assisting the visually impaired in their navigation, several systems or devices are currently
available in the literature such as Smart Cane, Ultra Cane, and Ray Mobility Aid (Pionke, 2022).
Smart Cane is a device that uses ultrasonic sensors to detect obstacles for navigation. It also uses
audio-based GPS software to help blind people to reach their destination (Designboom, 2019).
But the system can only detect obstacles above chest height. It is also costly and heavy and GPS
navigation can produce inaccurate results. It is also not suitable for closed places due to constant
vibration (Ronnie, 2021). Ultra-Cane is another device that uses ultrasonic sensors for navigating
narrow paths and obstacles for the visually impaired. However, it is expensive and heavy, and the
vibration can get very strong (Lonergan, 2022). Ray Mobility Aid is another device for use within
9.35 feet and produces an audible signal or vibration when an obstacle is at a closer distance.
Nonetheless, this device does not detect drop-offs or non-obstacle hazards and is very expensive
(Butzke, 2021). The acceptance rate of these available devices and tools is relatively low due to
the high cost, being unable to precisely identify obstacle types, the identification of different
objects in both daylight and dark environments, and the detection of currency banknotes. It is still
now a challenging issue for the visually impaired to attain safe mobility. Hence, there is room to
create a stand-alone system that can handle the aforementioned problems.

In this paper, we have proposed a Convolutional Neural Network (CNN) based detection system
that can detect chairs, tables, beds, toilet signs, and staircases in both indoor and outdoor
environments for the safe navigation of the visually impaired. Our proposed system can also
detect Bangladeshi banknotes with high accuracy. It acts as a wayfinding and situation awareness
system for the visually impaired.

This paper has been organized as follows. Section 2 discusses the related work regarding
navigation and object detection for the visually impaired. In section 3, the methodology of our
proposed object and currency detection system has been explained in detail. The experimental
analysis of the system has been illustrated and the comparison of the performance of our system
with existing systems has been discussed in section 4. The paper is concluded in section 5.

2. Related Work
Several works have been done to solve the challenges of the visually impaired. These are outlined
shortly as follows.
In (Ahmed J.U. et al., 2019), the authors proposed a Recurrent Neural Network (RNN) and CNN-
based system for obstacle avoidance system and used SqueezNet for the training purpose. The
system was used to detect obstacles, stairs, walking, turns, and elevator for the visually impaired.
However, the experiment was done in a limited environment (from the first floor to the third floor)
and the system showed a low accuracy of 95%.
In (Dahiya et al., 2020), faster region-convolutional neural networks (Faster R-CNN) with
Resnet50 were used to identify Restrooms, ATMs, Metro stations, and Pharmacies. However, the
system achieved 92% average accuracy. The system also did not consider the Precision, Recall,
and F1 scores of the tested datasets.
Tasnim et al., (2021) proposed a currency recognition system using CNN architecture with Keras
library and Tensorflow API. The system can detect 8 types of Bangladeshi bank notes. But the
system’s result is only based on average accuracy without considering the precision, recall, and

Page | 793
F1 score of the testing datasets. Furthermore, the average accuracy itself is lower at 92% than our
proposed system and the system can detect 8 classes of Bangladeshi banknotes instead of 10
classes. Also, obstacle detection was not incorporated into the system.
Using the pre-trained CNN model Restnet50, Afif et al. (2020) created an indoor signage and
door detection system for the visually impaired. Their technique first trained the neural network's
feature extraction background on one dataset, then transferred the newly learned features to a
second target network to be trained on a different dataset. The system can identify five states
namely Exit, Wc, disabled exit, Confidence zone, and door with an impressive accuracy of 99.8%.
But the system did not consider other important parameters for evaluation like recall and F1 score.
For indoor object detection, Noman et al. (2020) suggested a system that consists of a camera
with a processing unit using Tensorflow API and Time-of-Flight sensors. However, the system
only achieved an average object detection accuracy of 73.34% and considered only indoor objects.
Thomas and Meehan (2021) used data augmentation techniques to simulate partial currency
images, used the CNN model to detect currency, and achieved 94% accuracy. However,the system
cannot detect obstacles for the visually impaired.
Matsumura and Premachandra (2022) suggested a system that generates three-dimensional point
clouds from depth information using deep learning which can recognize stairs with an accuracy
of 97.3 percent. However, the system cannot detect other obstacles and currency for the visually
impaired. A stick and a YOLO (You Only Look Once) based algorithm were utilized by Kumar
and Jain (2021) to detect items using IOU (Intersection of Union). After the collected pictures of
the path are annotated, the Yolo v3 network is used to train the model with 300 images from each
of roughly 25 classes. The path detection is executed through a Novel tracker system that utilizes
an offline-trained neural network. The accuracy of the proposed model with the wearable mask
was found to be 81% while with the stick it is found to be 96.14 %. No currency detection was
incorporated into the system.

3. The Proposed Methodology


The methodology of the proposed system, to detect chairs, tables, beds, stairs, toilet signs, and
banknotes, is divided into the following parts.
3.1 Data Collection
For the proposed method, data for furniture, stairs, toilet signs, and banknotes have been collected
from Kaggle, Coco datasets, Google datasets, and ImageNet. Banknote data is also collected by
capturing images with a camera in real life environment. The numbers of collected data for
training and testing the proposed system are shown in Table 1. Data on banknotes have been
divided into 10 classes. Here, data are collected for a total of 15 classes of objects.
3.2 Data Preprocessing
All data are resized in 224 X 224 pixels. Data prefetching has been done to improve the processing
performance of data before feeding into the CNN model. Data normalization was performed for
better learning of the system. Data augmentation technique has been used to avoid overfitting.
After these steps, data is ready to transfer into the model.

Page | 794
Table 1: Dataset Summary

Object Class Train Test


Chair 1522 234
Table 1210 248
Bed 960 229
Stairs 818 207
Toilet Signs 800 226
Bank Notes 22692 4137

3.3 Object Detection


The proposed system uses EfficientNet, a CNN architecture, and a scaling method, for object
detection and transfer learning. Compound coefficients are used by EfficientNet to uniformly
scale all depth, breadth, and resolution dimensions. This is called the compound scaling method.
On the CIFAR-100 (91.7%), Flowers (98.8%), and three additional transfer learning datasets,
EfficientNet transfers well and can achieve state-of-the-art accuracy while using orders of
magnitude of fewer parameters. An inverted bottleneck Conv called MBConv, formerly known
as MobileNetV2, is the fundamental component of EfficientNet. It was integrated with an in-
depth separable convolution, which lowered the calculation by almost k² compared to typical
layers by linking a substantially fewer number of channels (relative to expansion layers) between
bottlenecks. The differences between compound scaling and conventional scaling are shown in
Figure 1 (Tan et al., 2019). Based on target FLOPS (Floating Point Operations) EffecientNet
architecture is classified into EfficientNetB0 to EfficientNetB7, a total of 8 classes. The proposed
system for object detection for the visually impaired, the EfficientB0 model has been used for
transfer learning. The target FLOPS for EfficientNetB0 is 400M.

Figure 1: Model Scaling of EfficientNet. (a) Is a baseline network example. (b)–(d) are
conventional scaling that only increases one dimension of network width, depth, or resolution.
(e) Compound scaling method that uniformly scales all three dimensions.

Page | 795
EfficientNetB0 has been used as the base model for transfer learning of the system. The weights
of the EfficientB0 model are already learned based on Imagenet datasets. For transfer learning
weights of trainable layers have been frozen. To learn from extracted features by EfficientB0 on
the datasets new layers have been added. For furniture, stairs, and toilet signs detection firstly an
AveragePooling layer of resolution 1x1 with 1280 channel, has been added. A Flatten layer is
added as input before adding an FCL Dense layer of 256 neurons. Relu activation has been used
for this Dense layer. A Dropout layer of 0.6 dropout rate has been added after the Dense layer. A
Dense layer of 5 (due to 5 class) neurons with Softmax activation has been added after that as the
final layer. For banknote detection similar model has been used except the dropout rate was 0.5
instead of 0.6. The model has been summarized in Table 2. Our proposed system architecture is
illustrated in Figure 2.
Table 2: Summary of the proposed system model

Layer Output Size Parameters


EfficientNetB0 (7, 7, 1280) 4,049,571
max_pooling2d_3 (MaxPooling2D) (1, 1, 1280) 0
flatten (Flatten) (1280) 0
dense_6 (Dense) (256) 327,936
dropout_3 (Dropout) (256) 0
dense_7 (Dense) (5) 1,285

Bed

Chair

Stairs

Table
Datasets
Navigation)
Toilet Sign

Taka 1

Taka 2

Taka 5

Taka 10

Taka 20

Taka 50

Taka 100

Taka 200

Taka 500

Taka 1000

Figure 2: Architecture of the proposed model

4. Experimental Analysis
This section describes the experimental setup, results, and discussion.

Page | 796
4.1 Experimental Setup
Python programming language and Tensorflow library with Keras API (version 2.10.0) have been
used to execute the code for training and testing. The code was executed on an Intel Core i7 10700
CPU and an RTX 2060S GPU with a system memory of 16 GB.
4.2 Results and Discussion
Two separate models were used for testing: one for furniture, stairs, and toilet sign detection and
another for banknote detection. ModelCheckpoint function was used to save the model only with
the best (lowest) validation loss during training. Sparse categorical cross-entropy was used as the
loss function and the Adam optimizer was used as the optimization algorithm. Performance has
been evaluated by precision, recall, and F1 score. The number of epochs for training was 250 and
200 with a batch size of 30 for furniture, stairs, toilet sign detection, and banknote detection
respectively. The Experimental Result for furniture, stairs, and toilet sign (Model 1) detection has
been shown in Table 3. Figure 3 represents the comparison of training and validation accuracy
during training. Figure 4 represents the training and validation loss for Model 1. From Table 3, it
can be seen that the precision, recall, and F1 scores for detecting bed are 99.56%, 100%, and
99.78% respectively on 229 samples. For chair, precision, recall, and F1 score are 99.57%,
99.57%, and 99.57% on 229 respectively on 234 samples. For stairs, precision, recall, and F1
score are 100%, 100%, and 100% respectively on 207 samples. Precision, recall, and F1 score are
100%, 99.59%, and 99.79% respectively on 248 samples for stairs. For toilet signs, precision,
recall, and F1 score are 100%, 100%, and 100% respectively on 226 samples. Weighted average
precision, recall, and F1 score for all 5 classes are 99.826%, 99.825%, and 99.825% respectively.
The confusion matrix showing the true positive, true negative, false positive, and false negative
predictions have been shown in Figure 5.

Figure 3: Training and validation accuracy Figure 4: Training and validation loss for
of CNN model 1 CNN model 1

Page | 797
Table 3: Accuracy scores of obstacle detection

Class Precision Recall F1 Score Samples


Bed 99.56% 100% 99.78% 229
Chair 99.57% 99.57% 99.57% 234
Stairs 100% 100% 100% 207
Table 100% 99.59% 99.79% 248
Toilet Sign 100% 100% 100% 226
Score (macro) 99.828% 99.834% 99.831% 1144
Score (weighted) 99.826% 99.825% 99.825% 1144
Overall Accuracy 99.82% 1144

The experimental result for banknote detection (Model 2) has been shown in Table 4. Figure 6
represents the comparison of training and validation accuracy during training. Figure 7 represents
the training and validation loss for Model 2. From Table 4, it can be seen that the precision, recall,
and F1 scores for detecting 1 taka are 99.66%, 100%, and 99.83% respectively on 300 samples.
For 10 Taka precision, recall and F1 scores are 99.38%, 100%, and 99.68% respectively on 482
samples. For, 100 Taka precision, recall, and F1 score are 100%, 97.72%, and 98.85%
respectively

Figure 5: Confusion Matrix of CNN model 1


on 484 samples. For, 1000 Taka precision, recall, and F1 score are 100%, 100%, and 100%
respectively on 479 samples. For 2 Taka precision, recall and F1 scores are 100%, 97.93%, and
98.95% respectively on 388 samples. For 20 Taka precision, recall, and F1 scores are 100%,
98.96%, and 99.48% respectively on 484 samples. For, 200 Taka precision, recall, and F1 score
are 100%, 100%, and 100% respectively on 127 samples. For 5 Taka precision, recall and F1
scores are 99.10%, 99.55%, and 99.33% respectively on 446 samples. For 50 Taka precision,
recall, and F1 scores are 96.59%, 100%, and 98.26% respectively on 481. For, 500 Taka precision,
recall, and F1 scores are 98.72%, 98.93%, and 98.82% respectively on 466 samples. Weighted
average precision, recall, and F1 score for all 10 classes of banknotes are 99.26%, 99.25%, and
99.25% respectively. The confusion matrix showing the true positive, true negative, false positive,
and false negative predictions have been shown in Figure 8. Figure 9 shows the overall
comparison of precision, recall, F1 score, and accuracy of all of the fifteen classes.

Page | 798
Table 4: Accuracy scores of banknote detection

Class Precision Recall F1 Score Samples


1 Taka note 99.67% 100.00% 99.83% 300
10 Taka note 99.38% 100.00% 99.69% 482
100 Taka note 100.00% 97.73% 98.85% 484
1000 Taka note 100.00% 100.00% 100.00% 479
2 Taka note 100.00% 97.94% 98.96% 388
20 Taka note 100.00% 98.97% 99.48% 484
200 Taka note 100.00% 100.00% 100.00% 127
5 Taka note 99.11% 99.55% 99.33% 446
50 Taka note 96.59% 100.00% 98.26% 481
500 Taka note 98.72% 98.93% 98.82% 466
Score (macro) 99.34% 99.31% 99.32% 4137
Score (weighted) 99.26% 99.25% 99.25% 4137
Overall Accuracy 99.25% 4137

Figure 6: Training and validation Figure 7: Training and validation loss


accuracy of CNN model 2 for CNN model 2

Page | 799
Figure 8: Confusion Matrix of CNN model 2

Graphical Representation of Accuracy,


Precision, Recall and F1 scores of 15 classes

F1 Score

Recall

Precision

Accuracy

99.32% 99.34% 99.36% 99.38% 99.40% 99.42% 99.44%

Figure 9: Graphical Representation of Accuracy, Precision, Recall, and F1 Scores for all of the
fifteen classes
A comparison of the performance of the proposed system and some other existing systems is
shown in Table 5. Different parameters namely method, coverage area, number of classes the
system can detect, whether the system can detect obstacles for navigation and currency, and
accuracy are used for the comparison. From the table, it can be seen all systems either proposed
obstacles/navigation or currency but not both. Most systems except Kumar and Jain (2021) can
detect fewer classes. All systems except Afif et al. (2020) have lower accuracy. Our proposed
system achieved 99.44% accuracy, 99.39% precision, 99.38% recall, and 99.37% f1-score for a
total of 15 classes on a test dataset of 5281 images.

Page | 800
Table 5: Comparison of performance between the proposed system and the existing system

Method based on Coverage Class Obstacles Currency Accuracy


Area /Navigation detection
Convolutional Neural Network Indoor 5 Yes No 95%
(CNN)
and Recurrent Neural Network
(RNN) (Ahmed, F. et al. 2019)
Faster region-convolutional Indoor 4 Yes No 92%
neural and
networks (Faster R-CNN) outdoor
with Resnet50 (Dahiya et al.,
2020)
Convolutional Neural Network - 8 No Yes 92%
(CNN) (Tasnim et al., 2021)
Convolutional Neural Network Indoor 5 Yes No 99.80%
(CNN) and
with Resnet50 (Afif et al., outdoor
2020)
Tensorflow API with Time-of- Indoor - Yes No 73.34%
Flight sensor (Noman et al.,
2020)
Convolutional Neural Network - - No Yes 94%
(CNN) (Thomas and Meehan,
2021)
Deep learning (Matsumura and Indoor 1 Yes No 97.30%
Premachandra, 2022) and
outdoor
YOLO (You Only Look Once) Indoor 25 Yes No 96.14%
based and
algorithm and outdoor
IOU (Intersection of Union)
(Kumar and Jain, 2021)
Convolutional Neural Network Indoor 15 Yes Yes 99.44%
(CNN) and
with EfficientNetB0 (our outdoor
proposed method)

5. Conclusions
Visual impairment is a global problem. The main goal of this paper is to contribute to finding an
acceptable, cost-effective, and user-friendly object detection system to mitigate the daily
challenges faced by the visually impaired every day. This paper uses EfficientNetB0 for transfer
learning that achieved an overall high accuracy of 99.44%, Precision of 99.39%, Recall of
99.38%, and F1 score of 99.37% for a total of 15 classes after testing on a total of 5281 images.
It considers both indoor and outdoor scenarios for navigation. However, this paper considers the

Page | 801
indoor environment only for currency recognition. A recommended enhancement of the proposed
method could be to incorporate outdoor environments for currency detection. Moreover, in the
future, the proposed method can be integrated with a physical device so that the visually impaired
can use this in real time.

Acknowledgment

We, the authors, are grateful to the Institute of Information and Communication Technology
(IICT), Bangladesh University of Engineering and Technology (BUET) for providing all possible
support to perform this research.

References

Afif, M., Ayachi, R., Said, Y., Pissaloux, E., Atri, M. et al., "Recognizing signs and doors for
Indoor Wayfinding for Blind and Visually Impaired Persons", 5th International Conference
on Advanced Technologies for Signal and Image Processing (ATSIP), 2020, pp. 1-4.
Ahmed, F., Mahmud, M. S., Yeasin, M. et al., "RNN and CNN for Way-Finding and Obstacle
Avoidance for Visually Impaired", 2nd International Conference on Data Intelligence and
Security (ICDIS), 2019, pp. 225-228.
Ahmed, J. U., Ahmed, S., Ahmed, A., Abdul Majid, M. et al., “Caring Vision: Fighting Against
Blindness in Bangladesh”, FIIB Business Review, 2019, 8(1); 17–24.
Basheshankar, A., Lande, A., Nandeshwar, A., Sarkar, A., Udapure, A., Chandankhede, P. H. et
al., "Assistive Cane for Visually Impaired People", 10th International Conference on
Emerging Trends in Engineering and Technology - Signal and Information Processing
(ICETET-SIP-22), 2022, pp. 1-5.
Butzke, Z., 2021. Assistive Devices for the Visually Impaired for Ground-Level Object and
Surface Condition Detection Literature Review. Available in
https://users.wpi.edu/~zrbutzke/Images/ButzkeLitReview2021Final2.pdf, (accessed 27
October 2022).
Dahiya, D., Gupta, H., Dutta, M. K. et al., "A Deep Learning based Real Time Assistive
Framework for Visually Impaired", International Conference on Contemporary Computing
and Applications (IC3A), 2020, pp. 106-109.
Designboom, K. M. I., 2019. Blind inventor creates 'smart cane' that uses google maps to navigate
visually impaired people, Available in https://www.designboom.com/technology/wewalk-
smart-cane-google-maps-visually-impaired-09-12-2019/, (accessed 27 October 2022).
Jafri, R., Campos, R. L., Ali, S. A., Arabnia, H. R. et al., "Visual and Infrared Sensor Data-Based
Obstacle Detection for the Visually Impaired Using the Google Project Tango Tablet
Development Kit and the Unity Engine", IEEE Access, 2018, vol. 6, pp. 443-454.
Kumar, N., Jain, A., "Smart Navigation Detection using Deep-learning for Visually Impaired
Person", IEEE 2nd International Conference On Electrical Power and Energy Systems
(ICEPES),2021, pp. 1-5.

Page | 802
Kumar,L., WeCapable, 2021. Daily life problems, struggle and challenges faced by blind people,
Available in https://wecapable.com/problems-faced-by-blind-people/, (accessed 27 October
2022).
Lonergan, J., NCBI, 2022. Labs product reviews – UltraCane, an electronic white cane for the
blind, Available at https://www.ncbi.ie/product-reviews-ultracane-an-electronic-white-
cane-for-the-blind/, (accessed 27 October 2022).
Matsumura, H., Premachandra, C., "Deep-Learning-Based Stair Detection Using 3D Point Cloud
Data for Preventing Walking Accidents of the Visually Impaired", IEEE Access, 2022, vol.
10, pp. 56249-56255.
Noman, M., Stankovic, V., Tawfik, A. et al., “Portable offline indoor object recognition system
for the visually impaired”, Cogent Engineering, 2020, Vol. 7, Issue 1.
Pionke, J., 2022. Libguides: Blind/Visual Impairment: Common Assistive Technologies,
Available in https://guides.library.illinois.edu/c.php?g=526852&p=3602299, (accessed
27 October 2022).
Rahman Sarker, M. F., Israfil Mahmud Raju, M., Marouf, A. A., Hafiz, R., Hossain, S. A.,
Hossain Khandker Protik, M. et al., "Real-time Bangladeshi Currency Detection System for
Visually Impaired Person", International Conference on Bangla Speech and Language
Processing(ICBSLP), 2019, pp. 1-4.
Ronnie, W., World of Technology, Video Games & Digital Entertainment, 2021.We walk
Smart Cane: The review of the cane for the visually
impaired, Available inhttps://techgameworld.com/wewalk-smart-cane-the-
review-of-the-cane-for-the-visually- impaired/, (accessed 27 October 2022).
Salvin, J. H., 2016.Visual impairment (for teens) - nemours kidshealth, Available in
https://kidshealth.org/en/teens/visual-impairment.html, (accessed 27 October 2022).
Tan, Mingxing, Quoc Le. et al., "Efficientnet: Rethinking model scaling for convolutional neural
networks", International conference on machine learning, PMLR, 2019, pp. 6105-6114
Tasnim, R., Pritha, S. T., Das, A., Dey, A. et al., "Bangladeshi Banknote Recognition in Real-
time using Convolutional Neural Network for Visually Impaired People", 2nd International
Conference on Robotics, Electrical and Signal Processing Techniques (ICREST), 2021, pp.
388-393.
Thomas, M., Meehan, K., "Banknote Object Detection for the Visually Impaired using a CNN",
32nd Irish Signals and Systems Conference (ISSC), 2021, pp. 1-6.

Page | 803
7th International Conference on Engineering Research, Innovation and
Education
January 12-14, 2023, Sylhet, Bangladesh

PAPER ID: 087

Bangla Hand Sign Recognition using Convolutional Neural


Network and Transfer Learning

Nayan Kumar Nath*, Reza Md Zakiul Karim Dhruba and Kazi Ashique Elahi
Department of Computer Science and Engineering, Sylhet Engineering College, Sylhet
Email: nayancsesec@sec.ac.bd, rmzkdhurba@gmail.com ,kaziashike@gmail.com

Keywords: Abstract: Sign languages are a kind of language that is widely used by people
all over the world who are hearing and voice disordered. To convey the
● Bangla Sign meaning, they use the visual-manual system. There is a common delusion
about sign language that it's the same everywhere. But this is not the actual
Language (BSL);
scenario. Today, people use nearly between 138 and 300 various types of
● Transfer Learning sign language throughout the world. Remarkable works are recorded in
automatic recognition of American, British, and French sign languages with
(TL);
high accuracy. There are some most widely spoken languages in the world.
● Convolution Neural Bangla is one of them. Few remarkable works on Bangla Sign Language
(BSL) are recorded in literature. In our study, we introduced a dataset for BSL
Network (CNN);
excerpted from “Bengali Sign Language Dictionary”. Also, we proposed a
● Data Labeling; model using Transfer Learning (TL) and Convolution Neural Network (CNN)
to train and test the dataset. Our proposed model showed an accuracy of 90.1%
● XML(extensible
after training in 20000 steps.
markup language);
● MobileNet.

1. Introduction
The World Encyclopedia defined sign language as “Non-phonetic means of personal
communication, using hand symbols, movements, and gestures”( Mcarthur,T.,2018). So, sign
language becomes the only feasible mean of their communication. It is used as a primary means
of communication among deaf people or people with impaired hearing. Furthermore, there is no
standard sign language internationally. There are multiple sign languages such as American,
Arabic, French, Spanish, Chinese, Indian, etc. There are approximately 167 million people in
Bangladesh. Among them, almost 13 million people are suffering from some degree of hearing
impairment. In which 3 million people are suffering from severe to intense hearing loss leading
to disability. Very few people understand sign language properly. Bangla Sign Language
Anthology (BSLA) is the standard when it comes to Bangladesh. Bangla Sign Language (BSL)
has a structure that does not match the sign languages of other countries. In BSL, for expressing
alphabets generally, both hands are used. This is why an interpreter is really essential to
understand what they are trying to convey. Currently; there are many eminent fields of research.
Sign Language Recognition (SLR) is one of them. It can help bridge the communication gap
between deaf and normal people with the help of innovative technologies. This problem is
magnified when it comes to Bangla Sign Language. Bangla is the sixth most spoken language in
the world. And also an official language of Bangladesh(Karim,M.,et. al.,2013). It is also the
spoken language of some states of India. About 2.4 million deaf people live in Bangladesh
Page | 804
(Yasir,et. al.,2014).However, the research on BdSL recognition did not progress compared to
other sign languages(Barik,et.al.,2014;Holden,et.al.,1999;Volger, et.al.,1997).Recently,a few
research works on the recognition of BdSL have been reported in the literature
(Ayshee,et.al.,2014;Rahman,et.al.,2014;Uddin, et.al.,2016)
In a study, The authors used a neural network for recognizing handwritten digits and numerals
(Zaki,et. al.,2011).whereas in another study authors developed a model to detect Bangla Sign
Language using the Convolution Neural Network (CNN). An available sign language-related
standard dataset was used. The dataset was split into 80:20 ratios where 80% was used for training
purposes and the rest of them were for test purposes. CNN was able to detect and classify the
signs in the images into Bangla digits and numbers. The accuracy achieved was
98.75%(Hossain,et. al.,2020).In the approach, to extract appearance-based hand features and
track the position in 2D and 3D, Microsoft Kinect was used. The classification results are secured
by comparing sequential pattern boosting (SP-boosting) with a hidden Markov model (HMM)
approach. This resulted in an accuracy of 99.9% on 20 different isolated gestures on their
specifically constructed data set and 85.1% on a more realistic one with 40 gestures. (Cooper,et.
al.,2014).
Although,various researcher work on BdSL recognition and reported moderate accuracy.
However, in comparison with other sign languages that were widely spoken, the research on BdSL
recognition did not progress as much . In our opinion, this is also due to the shortage of publicly
available, reliable, and large datasets. The objective of this study was i) to collect a large set of
the alphabet of BdSL, ii) CNN-based model with transfer learning to train the new model from a
pre-trained model, and iii) to improve the performance of BdSL recognition to human acceptance
level.
2. Materials and Methods
2.1 Sample preparation
In our model, we first take snaps of all the signs that we want to train our engine with. The second
step is to mark the point of interest to increase our model’s accuracy along with that we also label
the data. After that, we will get all the files on a folder named accordingly and with each image
file, there will run an XML file with their point of interest coordinates. Then those data will be
fed into a script provided by Google named “generate_tfrecord.py”. That will make a record file
for each test and train data. Then we will get the models from the web which is relevant for our
specific task from https://github.com/tensorflow/models .Then we will set upour pipeline for the
Transfer Learning process. After setting up all those configurations we will start our training. The
more steps we use the more accurate & slow the result will be. So we have used 20,000 steps to
find the perfect balance between accuracy and speed. After finishing the model training we can
use Open CV to detect and show the result in real-time. In this research, we built CNN based
model using TensorFlow API which uses transfer learning to train the new model from a pre
trained model. We detected the signs using python and Open CVand showed the detected signs
in live time.

Page | 805
Figure 1: Proposed Model for BdSL recognition system
2.2 Experimental procedure
In this process, we use the webcam to collect the images of 53 different signs and for each sign,
we collected 15 images including 9 digits, 9 vowels, 28 consonants, and 7 words. We used Jupyter
notebook and programmed it to take snapshots of all the signs. Python was used to code the
program. This way the dataset generation will be easier containing randomness and variation
which was beneficial for our detector. So this first step will give us the complete dataset for our
project.Data were cleaned, rearranged, and modified into readable for the algorithms. To fit into
the neural network, Pre-processing reshaped and rescaled raw data. Particularly cleaning reduces
noise in the raw data. We started labeling the collected images with software for testing and
training purposes. While labeling we only select the hand part and add a respective tag to it. The
labeling information is stored in an XML file.

Figure 2: Samples of data collection Figure 3: Data Labeling

2.3 Train Test Split

In this step, we tested the model that we trained. We took a total of 780 images and predicted the
result. For each sign, we took 15 images. We assigned 13 of them for training purposes and the
rest 2 for test purposes. So, a total of 674 images are for training and 106 images are for testing.
The training and test data contained images and targets.

Page | 806
Figure 4: Recognized BdSL letters,digits,words
Figure 4: Recognized BdSL letters,digits,words
3. Calculation
If we have an activation map of size W x W x D, a pooling kernel of spatial size F, and stride S,
then the size of output volume can be determined by the following formula:

This will yield an output volume of size Wout x Wout x D.


4. Result and Discussions
With CNN and TL we were able to get a pretty nice number for accuracy. We used 20,000 steps

pp. 2

Page | 807
Figure 5:(a)Training the model for 100 steps (b)After training the model for 100 steps
for our training steps and for Transfer learning, we used TensorFlow MobileNets with pipeline
After training 100 steps we got a loss of 0.752. so the accuracy of the model after 100 steps is
24.8%. When we trained the model for 20000 steps we got a data loss of 0.099 so the model’s
accuracy at that stage is 90.1%.
Steps Loss Accuracy 0.8
(%) 0.639 0.7
. .5
100 0.752 24.8 80 0.6
2000 0.639 36.1
0.419 0.5
4000 0.590 41 60
6000 0.556 44.4 0.329 0.4
8000 0.419 58.1 40 0.3
10000 0.329 67.1 . .
. . 0.2
12000 0.203 79.7
20
14000 0.187 81.3 0.1
16000 0.151 84.9
0 0
18000 0.131 86.9

100
2000
4000
6000
8000
10000
12000
14000
16000
18000
20000 0.099 90.1

Accuracy Loss
Table 1: Models Accuracy Table
Figure 6: Training the model for 100 steps
5. Conclusion

The research that we have conducted using the TL with CNN model gave us considerably solid
results that can ensure the credibility of our work. The mean validation accuracy that we received
is 90.1% and the mean validation loss is 0.099 which proves that this model is ideal for such
recognition purposes. The inclusion of deaf and mute people in all aspects of our society is very
important for us as a nation to further develop and improve and dispel all boundaries and obstacles
that hinder our advancement. Our research will serve to fulfill that exact purpose to remove all
obstacles, mainly the hindrance of the communication gap between normal people and deaf
people. As a result, these people will no longer be wrongly judged and viewed otherwise
negatively in society. They will no longer be left out and will always be treated as equals.
References
Ayshee ,T. F., Raka, S. A., Hasib ,Q. R., et al.,Fuzzy rule-based hand gesture recognition for
bengali characters,IEEE International Advance Computing Conference (IACC). 2014; 484–
489.
Bangla Sign Language Dictionary, Bangladesh Sign Language Committee, National center for
special education, Ministry of social welfare, Govt. the people’s republic of
Bangladesh.,1997.
Begum,M.,Shahnaj,F.,Aktar,T.,Akter ,Y. A., Nath,N. K.,A Neural Network-based Recognition
of Isolated Handwritten Basic Bangla Alphabets and Numerals Using Spatial
Relationships,Australian Journal of Business, Social Science and Information Technology
(AJBSSIT).2018,4(1).

Page | 808
Birk ,H., Moeslund ,T. B., Madsen ,C. B.,Real-time recognition of hand alphabet gestures using
principal component analysis,Proceedings of the Scandinavian conference on image
analysis, 1997, 1;261–268.
Brooks,R.,2018.A Guide to the Different Types of Sign Language Around the World, Avaiable
in:https://www.k-international.com/blog/different-types-of-sign-language-around-the-
world/ (accessed 12 july 2022)
Cooper,H.,Ong,E.J., Pugeaul, N. t, Bowden,R.,Sign language recognition using sub-units,”
Springer Gesture Recognition.2017;89–118.
Holden,E.-j., Owens, R.,, Roy G. G.,et al., Adaptive fuzzy expert system for sign recognition,in
Proceedings of the International Conference on Signal and Image Processing SIP’2000.
1999: Las Vegas, 1999.
Hossain,S.,Sarma ,D.,Mittra,et al.,Bengali hand sign gestures recognition using convolutional
neural network,Second International Conference on Inventive Research in Computing
Applications(ICIRCA).2020.
Islalm,M.S.,Rahman ,M. M., Rahman ,M. H., et al.,Recognition Bangla Sign Language using
Convolutional Neural Network,International Conference on Innovation and Intelligence for
Informatics,Computing,and Technologies (3ICT).2019.
Karim,M., Kaykobad,M.,Morshed,M.,Technical challenges and design issues in Bangla language
processing. IGI Global, 701 E Chocolate Avenue, Hershey PA, 17033, USA, 2013.
Krizhevsky,A.,Sutskever,I.,Hinton,G.E.,Imagenet classification with deep convolutional neural
networks,Advances in neural information processing systems;2012;1097–1105.
Mcarthur,T.,2018. Sign Language,Avaiable in https://www.encyclopedia.com/literature-and-
arts/language-linguistics-and-literaryterms/language-and-
linguistics/signlanguage#:~:text=sign
%20language%20Non%2Dphonetic%20means,or%20people%20with%20impaired%20hearing.
((accessed 12 july 2022).)
Nagi,J.,Ducatelle, F., Di Caro ,G. A., et. al.,Max-pooling convolutional neural networks for
vision-based hand gesture recognition,International Conference on Signal and Image
ProcessingApplications (ICSIPA), IEEE. 2011; 342–347.
Rahaman,M. A.,Jasim,M.,Ali,M.H.,et al.,Realtime computer vision-based bengali sign language
recognition,17th International Conference on Computer and Information
Technology(ICCIT).2014;192–197.
Uddin,M.A.,Chowdhury,S.A.,Hand sign language recognition for bangla alphabet using support
vector machine,International Conference on Innovations in Science, Engineering and
Technology(ICISET).2016;1–4.
Vogler ,C., Metaxas,D.,Adapting hidden markov models for asl recognition by using three-
dimensional computer vision methods, IEEE International Conference on Systems, Man, and
Cybernetics.Computational Cybernetics and Simulation. 1997,1;156–161.
Yasir,R.,Khan ,R. A.,Two-handed hand gesture recognition for bangla sign language using lda
and ann, 8th International Conference on Software, Knowledge, Information and
Management Applications(SKIMA). 2014;1–5.
Zaki, M. M., Shaheen, S. I. (2011). Sign language recognition using a combination of new
vision based features. Pattern Recognition Letters. 32(4);572–577. https://doi.org/ 10.1016/
j . patrec. 2010.11.013

Page | 809
7th International Conference on Engineering Research, Innovation and Education
January 12-14, 2023, Sylhet, Bangladesh

PAPER ID: 147

Automated Glaucoma Detection from Retinal Fundus Image


Rezaul Islam * Md. Liakot Ali Muhammad Sheikh Sadi
Bangladesh University of Bangladesh University of Bangladesh University of
Engineering and Technology Engineering and Technology Engineering and Technology
Palashi, Dhaka 1205, Bangladesh Palashi, Dhaka 1205, Bangladesh Palashi, Dhaka 1205, Bangladesh
0419312017@ict.buet.ac.bd liakot@iict.buet.ac.bd sadi@iict.buet.ac.bd
*Corresponding author [Rezaul Islam]

Abstract: Glaucoma is an autistic eye disease that causes irreversible blindness


Keywords:
worldwide. One of the dangerous characteristics of glaucoma is that it gradually
increases, with no pain and no noticeable symptoms to the patients in its early stage. It
● Glaucoma;
damages the optic nerve that carries information from the eye to the brain and results in
● Retinal fundus permanent vision loss. Early-stage detection of glaucoma is indispensable to preventing
this disease. Automated glaucoma detection from the retinal fundus image can play a
image;
significant role in this aspect. This paper presents an efficient method to analyze the
● cup-to-disk ratio; retinal fundus images such as adapting thresholding, histogram analysis, wavelet
transformation, localization of optical disk (OD) and optical cup (OC), and then
● histogram analysis;
calculating cup-to-disk ratio (CDR). Finally, it categorizes glaucoma into two classes:
● wavelet early stage and extreme stages through deep learning.
transformation;

1. Introduction

Glaucoma is a disease that destroys the optic nerve which is fundamental to enduring vision (Nawaldgi,
2016). According to the World Health Organization (WHO), it is the second leading reason for the cause
of irreversible blindness among the population of any age globally (Ruby et al., 2014). Every day a huge
number of people in the world lose their vision permanently due to glaucoma. In 2020, about 80 million
people have glaucoma worldwide, and this number is expected to increase to over 111 million by 2040
(Allison et al., 2020). Most glaucoma patients are unaware of the disease till it is reached an extreme stage.
It is usually a slowly progressing condition, has no specific symptoms, and no pain in the early stage.
Therefore, glaucoma is called the silent killer of sight. Though glaucoma cannot be cured fully, it can be
controlled and minimize the risk of vision loss by being detected in an early stage. Benzebouchi et al.,
(2019) show that about 50% of blindness, which is caused due to glaucoma, can be prevented if it is detected
and diagnosed early. Therefore, early detection of glaucoma is a significant challenge and is the primary
treatment for preventing vision loss. Manual analysis of glaucoma detection is time-consuming and costly.
In these cases, the accuracy varies by different factors. Hence, there arises the need for an automated method
using a combination of image processing (retinal fundus image), artificial intelligence, and computer vision.
Several types of research using retinal fundus image analysis have been proposed and are ongoing for
automated glaucoma detection. The main challenge here is how to localize and segment the optical disk
(OD) and optical cup (OC) efficiently from retinal fundus images. Figure 1 shows the OD and OC in a
fundus image. Previously, different machine-learning techniques and superpixel clustering algorithms are
used for OD and OC segmentation. Finally, a neural network is used to classify glaucoma depending on the CDR
calculation. However, the maximum accuracy is achieved at 92.3% here. A little work has been donein
terms of achieving good accuracy and efficient segmentation of OD and OC. Hence, there is still scope

Page | 810
for research in efficient segmentation of OD and OC and classifying the glaucoma screening with better
performance.

Figure 1: Fundus images showing its optical disk and optical cup
This paper presents an efficient image processing algorithm that detects glaucoma automatically with
optimal accuracy. Here, we have used different image processing algorithms such as adapting thresholding,
histogram analysis, wavelet transformation, and deep learning.
This paper is organized as follows. Section II reviews the existing works on automated glaucoma screening
from the retinal fundus image of an eye. Section III describes the proposed methodology for automated
glaucoma detection. Section IV shows the data collection and results respectively. In the end, Section V
concludes the paper.
2. Literature Review
This section describes the current existing state-of-art methods in the proposed area of research.
Khalil et al. (2014) proposed Machine learning techniques with image processing are being proposed for
automated screening of glaucoma. Here, Optical Coherence Tomography (OCT) is used to predict
glaucoma, and they have achieved a limited accuracy of 85% only.
Phasuk et al. (2019) developed an automated method called DENnet for glaucoma screening using a deep
learning algorithm. After localizing the OD and OC, Artificial Neural Network (ANN) is used for the
classification of glaucoma. However, this method achieved a low accuracy for glaucoma detection which
is 94%.
Sunanthini et al. (2022) describe the performance comparison of different CNN algorithms such as
AlexNet, GoogLeNet, and ResNet architectures to detect glaucoma automatically. Here, the authors show
the accuracy varies in different CNN algorithms.
Patil et al. (2020) developed a convolutional neural network (CNN) for the early detection of glaucoma.
Here, the authors used a small number of datasets which is used for training and testing, an d the achieved
accuracy, in this case, is 97%.
Pavithra et al. (2017) used a PIC Microcontroller whereas Aruchamy et al. (2015) used a DSP board to
implement the portable glaucoma detection system. However, both platforms have limitations in terms of
computational speed and power consumption.
T. Afroze et al. (2021) have built a CNN-based model for detecting automated glaucoma with an inception
v3 algorithm. They have used two publicly available datasets which contain a large number of images for
training and tasting. Their proposed model achieved a lower accuracy of 85% and a value of 0.9387 for
AUC.
Navya et al. (2020) focused on the segmentation of OD and OC using morphological function-based image
processing techniques with publicly available datasets and classified the level of glaucoma disease as
normal, moderate, and extreme. However, they obtained an accuracy of 88% only.

Page | 811
Khaing et al. (2021) proposed a method for glaucoma detection based on Deflation-Inflation Gradient
Vector Flow (ADI-GVF) segmentation with Exclusion Method (EM) initialization. The authors conducted
experiments to detect the boundaries of OD and OC from the retinal fundus images and calculate the CDR value.
The proposed method has tested a mobile phone-captured dataset and two standard datasets. The proposed
method shows that the proposed algorithm performs better than the other state-of-art methods. However,
the ADI-GVF method predicts glaucoma incorrectly and does not work when the initial curve islarger.
3. Methodology
In this section, we have described the proposed method for automated glaucoma detection. The overall
process of detecting glaucoma is depicted in Figure 2. As shown in this figure, the proposed method consists
of an input phase, pre-processing phase, a segmentation phase, and finally classification phase. The input
phase consists of a fundus camera or capturing image devices such as mobile phones or digital cameras
which are used to capture glaucoma images. In the pre-processing phase, there are several techniques used
to process images such as resizing, noise removal, histogram equalization, and channel extraction. The
Segmentation phase consists of Region of Interest (ROI) extraction, Optical Disk (OD) and Optical Cup
(OC) separation, and adapting thresholding. After segmenting the OD and OC, Cup to disk ratio is calculated
finally. In the Classification phase, glaucoma is classified through the neural network. The details of these
steps are explained as follows.

Figure 2: The proposed method for glaucoma detection

Page | 812
3.1 Preprocessing Images
Input images are captured with a fundus camera or mobile phone attached to retinal lances. After choosing
the image from the dataset, image preprocessing is performed. Acquired images may be noisy. In the
preprocessing method, image resizing, noise removal, and color channel extraction operation are
performed. We have converted RGB images into grayscale images for further processing. Grayscale images
contain the pixel information of the fundus image. A grayscale image is helpful for further processing of
segmentation.
3.2 Segmentation
The primary goal of image segmentation is to simplify the image for easier analysis. There are several steps
for fundus image segmentation to detect glaucoma such as ROI extraction, OD, and OC separation. Fro m
the histogram analysis of images, maximum index values are selected to localize the optic disk. As the
unwanted noise is filtered from the fundus images in preprocessing step, it contains bright lesions or fringes
which may be considered an optic disc. The image size is reduced and cropped. Then the region with the
largest vessel density is finally selected as the optic disk location.
The segmentation of OD and OC is a challenging task. If an image has different lighting conditions in
different areas, then adapting the thresholding algorithm works better. Here, the algorithm determines the
threshold for a pixel based on a small region around it. Since, fundus images have distributed pixel
intensities, adapting a thresholding algorithm can be used effectively to segment the optical disk. In this
method, a threshold value is selected and if the pixel value is smaller than the threshold value, it is set to 0,
otherwise, it is set to a maximum value.
3.3 Glaucoma Classification
After identifying the disk and cup boundary, the glaucoma classifier was developed with the convolutional
neural network. We have used a 4-layer neural network with an input layer, two hidden layers, and one
output layer. For each input fundus image, 20 features are extracted. To feed these 20 features, the input
layer has 20 neurons. As the output result, we want it to be either a normal or Glaucoma indication, the
output layer consists of only one neuron. We used two hidden layers with 20 neurons in each layer as we
could get better results with two hidden layers than one. The architecture of the neural classifier used is
shown in Figure 8.

Figure 3: Architecture of glaucoma neural classifier

Page | 813
4. Experimental Results Analysis
We have prepared two different models for our work named GLDIS and OPTPIC where these two models
are a combination of the publically available dataset of Drishti-GS1 [14] and the Bangladesh Eye Hospital
(BEH) dataset [15]. These two models are available in the GitHub repository [16] and are shown in Table
1. From Table 1, it can be seen that the GLDIS model contains a total of 40 images where 25 images are
glaucomatous and 15 images are non-glaucomatous and the OPTPIC model contains a total of 60 images
where 25 images are glaucomatous and 35 images are non-glaucomatous. All the images of these two
datasets are of varying sizes and different formats.
Table 1: Dataset description

Model Name Glaucomatous Non-glaucomatous Total image

GLDIS 25 15 40

OPTPIC 25 35 60

We have used 50 retinal fundus images for training with glaucomatous and non-glaucomatous images and
50 fundus images for testing with glaucomatous and non-glaucomatous for our experimental analysis. We
have calculated the optical disk ratio and optical cup ratio. Finally, we have calculated CDR based on the
OD and OC ratio. The Calculation of CDR is shown in detail in Table 2.
Table 2: Calculation of CDR

Image name Image ROI OD OC Diagnosis

Dataset: CDR: .299212


GLDIS Normal
Image No.1

Dataset: CDR: .497539


GLDIS Glaucoma
Image No.4 detected

Dataset: CDR: .506057


GLDIS Glaucoma
Image No. 7 detected

Dataset: CDR: .750809


OPTPIC Extreme stage
Image No.100

Page | 814
The Confusion matrix gives a clear idea of values like True Positives, False Positives, True Negatives, and
False Negatives. The relevant confusion matrix is shown in Figure 3. As shown in Figure 3, the correctly
classified samples can be identified in the diagonal directions of the confusion matrix marked in the green
square.

Figure 3: Confusion Matrix of proposed system


The proposed method shows well performance with 95% accuracy, 97% precision, 92% recall, and .94 %
F1-score which are shown in Figure 4. Precision evaluates the number of positive class expectations that
have a place with the positive class. Recall evaluates the number of positive class expectations made out of
all positive models in the dataset. F1-score provides a way to combine both precisions and recall into a
single measure.

Figure 4: Graphical representation of the performance evaluation of the proposed method


The comparison of the performance among the proposed system and several existing systems is shown in
Table 3. We have chosen different parameters like classification method, the feature used, and the number
of images, and accuracy to complete the comparison table. From Table 3, it can be observed that most
of the existing systems (Kumbhare et al., 2014), (Noronha et al., 2014), and (Nayak et al., 2009) used
a small number of the dataset and achieved high accuracy. (Phasuk et al., 2019), (T. Afroze et al.) and
(Acharya etal., 2017) worked with large datasets but their accuracies are low. Here, our main focal
point is to find a new algorithm that outperforms existing methods concerning accuracy for the dataset
that contains a large number of fundus images. In the proposed method, we have used two datasets
consisting of 100 fundus images and achieved a high accuracy of 95%.

Page | 815
Table 3: Performance analysis of the proposed system in comparison to existing systems

References Classification method Method/Features Used # of images Accuracy

[5] Optical Coherence Different Machine Learning Techniques - 85 %


Tomography (OCT) (ML)

[6] Artificial neural networks The Area under the ROC Curve (AUC) 650 94%

[11] CNN based Inception V3 6072 85%

[17] Random Forest and Wavelet features 63 94.70%


ANN, SVM and K-NN

[18] SVM Raw pixel, FFT, intensities, B-spline 575 80%

[19] Support Vector Machine High Order Spectral (HOS) and wavelet 60 95%
(SVM )
features

[20] Random Forest HOS and texture features 60 91%

Proposed Convolutional neural Adapting Thresholding, wavelet features 100 95%


network(CNN)
Method

5. Conclusion
This paper developed a novel method for classifying glaucoma disease using deep learning where different
image processing algorithms such as adapting thresholding and wavelet transformation are used. A
Convolutional Neural Network (CNN) model is used for the detection and categorization of glaucoma -
affected levels. The experimental analysis shows that the system has an average accuracy of 95%, a
precision of 97.5%, a recall of 92.2%, and an F1 score of 94.3% after experimenting with 100 images from
two datasets. This paper also explores how it outperforms existing works in the related field. In the future,
this research can be extended by experimenting with glaucoma classification with a larger dataset by
augmenting our data that contains thousands of retinal images and by working with other stages of glaucoma
like moderate stages. The accuracy of the system can be improved as well.
Acknowledgement
We, the authors, are grateful to the Institute of Information and Communication (IICT), Bangladesh
University of Engineering and Technology (BUET) for providing all possible support to perform this
research.

Page | 816
References
Nawaldgi,S., “Review of Automated Glaucoma Detection Techniques”, 2016 International Conference on
Wireless Communications, Signal Processing and Networking (WiSPNET), 2016, pp. 1435-1438.
Ruby,J., Kumar, P.S., Lepika,J. et al., “Glaucoma Detection and Image Processing Approaches: A Review”,
Journal of current Glaucoma Practice, 2014, 8(1): 36-4.
Allison,K., Patel,D., Alabi,O., “Epidemiology of Glaucoma: The Past, Present, and Predictions for the
Future. Cureus”, 2020, 12(11):e11686.
Benzebouchi, N.E., Azizi,N., Ashour,A.S., “Multi-modal classifier fusion with feature cooperation for
glaucoma diagnosis”, Journal of Experimental & Theoretical Artificial Intelligence, 2019, pp. 841-
874.
Khalil,T., Khalid,S., Syed,A.M.,“Review of Machine Learning Techniques for Glaucoma Detection and
Prediction”, 2014 Science and Information Conference, 2014, pp. 438-442.
Phasuk,S., Poopresert,P., Yaemasuk,A., Suvannachart.P.et al., “Automated Glaucoma Screening from
retina fundus image using deep learning”, 2019 Annual International Conference of the IEEE
Engineeringin Medicine & Biology Society, 2019, pp. 904-907.
Sunanthini,V., Deny, J., Govinda Kumar,E. et al., “Comparison of CNN Algorithms for Feature Extraction
on Fundus Images to Detect Glaucoma”, Journal of Healthcare Engineering, 2022, pp.1-9.
Patil,A., Shetty, R., Jain,S.et al.,“Glaucoma Detection using Convolutional Neural Network”, International
Research Journal of Engineering and Technology (IRJET), 2020, Vol: 7, Issue: 01.
Pavithra,G., Manjunath,T.C., Lamani,D. et al., “Hardware Implementation of Glaucoma using A PIC
Micro-Controller – A Novel Concept for a Normal Case of the Eye Disease”, 2017, International
Conference on Current Trends in Computer, Electrical, Electronics and Communication
(ICCTCEEC), 2017, pp.1104-1109.
Aruchamy,S., Bhattacharjee,P., Sanya,G., “Automated Glaucoma Screening in Retinal Fundus Images”,
International Journal of Multimedia and Ubiquitous Engineering, 2015, pp.129-136.
Afroze,T., Akther,S., Armanuzzaman,M.C. et al., “Glaucoma Detection Using Inception Conv olutional
Neural Network V3”, Applied Intelligence and Informatics, 2021, pp.17-28.
Navya,K., Lekhna,K., Prayla Shyry,S., “Early stage detection of glaucoma and its levels using fundus
images of eye”, Journal of Critical Reviews, 2020, Vol 7, Issue 4.
Khaing,T.T., Ruennark,T., Aimmanee,P. et al, “Glaucoma Detection in Mobile Phone Retinal Images
Based on ADI-GVF Segmentation with EM initialization”, ECTI-CIT Transactions, 2021, vol. 15, pp.
134–149.
Drishti-GS Dataset (IIIT, Hydarabad), Drishti-GS Dataset Portal, available in:
http://cvit.iiit.ac.in/projects/mip/drishti-gs/mip-dataset2/Home.php (accessed 17 November 2022).
BEH (Bangladesh Eye Hospital) Glaucoma Dataset in Google Drive, available in:
https://drive.google.com/file/d/1hi6HcjO5N1MY0GZcFizc1Jle2sUrB8kV/view (accessed 17
November 2022).
Glaucoma dataset in Github repository, Available in: https://github.com/reza38/glaucoma_Dataset
(accessed 17 November 2022).
Kumbhare,P., Turkar,M., Kularkar,R., “Computer Aided Automatic Glaucoma Diagnosis”, International
Journal of Electrical, Electronics and Data Communication, 2014, vol.2, pp. 28-32.

Page | 817
Acharya,U.R., Bhat,S., Koh,J.E.W. et al., "A novel algorithm to detect glaucoma risk using texton and local
configuration pattern features extracted from fundus images", Computers in Biology and Medicine,
2017, vol. 88, pp. 72-8.
Noronha,K.P., Acharya, U.R., Nayak,K.et al., “Automated classification of glaucoma stages using higher
order cumulant features,” Biomedical Signal Processing and Control, 2014, vol. 10, pp. 174–183.
Nayak, J., Acharya, U.R., Bhat, P.S.et al., “Automated Diagnosis of Glaucoma Using Digital Fun dus
Images”, Journal of Medical Systems, 2009, vol.33, pp. 337-346.

Page | 818
7th International Conference on Engineering Research, Innovation and Education
January 12-14, 2023, Sylhet, Bangladesh

PAPER ID: 152

DESIGN AND IMPLEMENTATION OF A LOW-COST IOT BASED


BIOMETRIC ATTENDANCE MANAGEMENT SYSTEM USING
FINGERPRINTAUTHENTICATION AND CLOUD DATABASE.

Md. Hasanur Rahman Sohag 1, Bijan Dev 2 , Md. Sadikul Alam 3 , G.M. SIFAT
IQBAL 4 * Surajit Sinha 5.

1. Alumnus, Metropolitan University, Sylhet. E-mail-hasan.sohag3888@gmail.com.


2. Alumnus, Metropolitan University, Sylhet. E-mail- bijandev13@gmail.com.
3. Alumnus, Metropolitan University, Sylhet. E-mail- samanik97@gmail.com.
4.Student , Metropolitan University, Sylhet. E-mail- gaziiqbal001@gmail.com.
5. Assistant Professor, Metropolitan University, Sylhet. E-mail-
sinha@metrouni.edu.bd.

*Corresponding author [G.M. SIFAT IQBAL ]

Abstract: The low-cost Internet of Things (IoT)-based biometric attendance


Keywords: device is designed for students to scan their fingerprints and authenticate their
attendance and securely record the data in a real-time cloud database.
 Internet of things; Administrators and teachers can maintain student databases and analyze data
 Biometric by generating attendance reports using a web-based user interface for the
parallelly-built system. To verify the accuracy and reliability of the system,
attendance; new techniques are applied in the software and Hardware sections. The
 Real-time database; hardware is designed and constructed with digital electrical components and
the Arduino development platform. Microsoft Visual Studio is mostly used to
 Attendance create the software section's web-based user interface and cloud database. The
management ; system's low development cost makes it a cost-effective electronic attendance
management solution. It makes it easy to evaluate attendance data, prevent
 Cloud database; proxy attendance and save time. When administrators need attendance records,
 Low cost. they may easily access and recover them because of accessibility in the cloud
database. However, the accuracy of the attendance record is ensured by the
usage of fingerprint scanners. The suggested solution is capable of taking the
place of the laborious, ineffective, and outdated pen-and-paper-based
attendance management system.

Page | 819
1. Introduction.
The main objective of this project is to design and implement a low-cost internet of things (IoT)
based biometric attendance management system with fingerprint authentication and a cloud
database, which can take student attendance through the fingerprint authentication process and
store the attendance data in a cloud database in real-time. The traditional way of taking attendance
of students in class is not efficient and effective in modern standards. This process involves the
situation where teachers have to take the attendance of students in each class and calculate the
student attendance percentage throughout the semester and come out with their student attendance
report. But if the teacher lost the attendance sheet, he or she correctly cannot process result and
will be unable to document attendance record.Moreover , the students who are absent in the class
and sometimes, who manage proxy attendance will not be able to learn what the teacher had
taught in the classes, resulting in them knowing nothing about the subject and end up with poor
academic performances. It is essential to look into a student's attendance records when evaluating
their academic performance. Teachers must handle some extra paperwork for attendance
management in the traditional pen-paper-based attendance management process, which wastes a
significant amount of time.
The practice of documenting and maintaining student’s attendance at educational institutions
should more realistic as technology advancement. In this context, some innovative ways for
recording and preserving attendance have been created, the biometric student attendance system
being one of the most prominent. Fingerprint authentication is one of the oldest biometric
techniques but still widely used because of its simplicity and accuracy. Every human being is
born with a unique fingerprint pattern; these characteristics are used to recognize and differentiate
between two different people. The fingerprint recognition and verification methodology has been
used to replace the traditional pen-paper attendance system for more efficiency and efficacy.
After studying the detailed background, we aim to develop the project. Study the current up-to-
date knowledge of IoT, biometric technology etc. through our system blueprint planned. The
design method is divided into two main domains, one is the hardware and another is the software
module. We choose low-cost but durable electronic components to build your hardware modules.
About1,982 BDT is required to construct the total module deducting labor costs. We make sure
that around 10 tests are done to confirm the functionality of the device After several tests the
system is ready for field use.

2. Current attendance management systems.


There are several types of attendance recognition systems that are already in the market one is
RFID(Radio-frequency identification) Based Attendance systems. the Radio Frequency wave has
been widely used in RFID systems to track any implanted tag in a device that uses electromagnetic
and electrostatic coupling for the communication between the tag and reader through modulation
and encoding scheme (Lim, Sim, & Mansor, 2009). RFID is the next wave in the evolution of
computing and it is a technology that connects objects to the Internet which can be used to track
and share data. RFID tags use a silicon microchip to store a unique serial number burned in during
the manufacturing process (M, et al., 2007).The Real-time Face Detection attendance system is a
new approach in automatic attendance management systems, extended with computer vision
algorithms. the system is integrated into an existing Learning Management System (LMS), which
automatically detects and registers students attending a lecture. The system is implemented using

Page | 820
a non-intrusive digital camera installed in a classroom, which scans the room, and detects and
extracts all faces from the acquired images. Similar to the other methods, once faces have been
extracted, they are compared with an existing database of student images. A list will be generated
based on a successful recognition of a student and saved in a database (Shehu & Dika, 210) .
Barcode-based attendance management systems are another popular system in the automatic
attendance field. Barcodes are machine-readable symbols that are used to store bits of data and
all barcode data can be automatically collected and stored in a database. Barcodes represent data
in 1 dimensional barcode or 2-dimensional matrix codes. 1D barcodes refer to the spacing of
parallel lines whereas 2D is in patterns of squares, dots and hexagon patterns. To detect any data,
a barcode requires a white or solid colour background (M.I. & N.M., 2009), the student’s matrix
card will be scanned using the barcode reader as an analogue signal. The scanner will then convert
it to a digital signal to make it available for the computer to receive the data. Then, the data
captured will be stored in a database of the host computer (M.I. & N.M., 2009).
The biometric attendance management system is another reliable and modern attendance
management system. As for this system, physical and behavioral characteristics such as
fingerprints, voice, face, retina, iris, handwriting and hand geometry are used to identify and
verify students' attendance in class. The most popular convenient and secured authentication is
biometric identification using fingerprints. There are two main types of fingerprint attendance
management systems in the market; online fingerprint attendance systems and offline fingerprint
attendance systems. The online system requires a connection with a personal computer (PC) and
all fingerprint templates of users to be verified and stored in a PC database. Hence, this gives a
burden on the PC to store data and eventually increases the response time. The offline fingerprint
attendance system is more convenient as it can complete all processes without support from a PC
(Yongqiang & Ji, 2006).
Our proposed attendance system is built at a very low cost. We use locally available resources
and low-cost but long-lasting electronic components to make it much cheaper than any current
commercial attendance management system. It is 7 to 8 times less than the cost of a high-end
attendance management system currently available in the market. In these market conditions, we
may be able to provide a similar system at a much lower cost.

3. Design Methodology.
The goal of this project is to create an IoT-based electronic attendance device that can take
attendance using a fingerprint identification procedure and save the information in the cloud. The
entire system will be controlled using a web-based user interface that will be created. This will
offer a flexible platform that will take the role of the conventional manual attendance system,
improving the precision and dependability of attendance management. When a student will get
admitted to an educational institute, all the information of students will be stored in the database
and a student profile will be created. For a new class or section, an attendance device will be
allocated and added to the system. After that when students have their first day of classes, or
perhaps a week of orientation, each one is enrolled into the system. The fingerprint of the students
will be scanned and stored in the device that is allocated for the class/section of that particular
student. So, the student will be ready to be assigned to any course as required. The web-based
user interface will provide a platform, from which the admin/teacher can manage the whole
system easily. The admin will be able to create a new teacher profile by

Page | 821
adding an email address, temporary password and other information which will be used by the
teachers to log in and use the system. Teachers will get the password and email address to log in
to the system. Then the teacher will be able to change the password and email address as well as
other information.

Fig. 1 : System classification


Many electronic and hardware components, as well as some software, are used in the design and
development of this project. The system is classified into two section Hardware Section and
Software Section .Hardware Section the hardware section only consists of the IoT based
attendance device. Software Section The software section consists of the web-based user
interface, the database and the user interface for the attendance device.

3.1 . Hardware design.


The electronic attendance device is the single item in the hardware design area. The student
fingerprints could be scanned by the attendance device, and the information is then transmitted
online to a cloud database. The design of the attendance gadget incorporates a number of
electronic modules and parts. This attendance device's construction is broken into four
components based on its functions. Input unit, Control unit, Power unit, and Output unit are the
four of them. the input unit consists of components used to provide data and control signals to the
control unit. The input unit of this attendance device is performing the task of getting user input.
The user input can be the fingerprint data or data from the keypad. The keypad data is used to get
control input, data input and to navigate the user interface of the attendance device. The input unit
components are: DY50 Fingerprint Sensor, Matrix Keypad and Push-Button. This attendance
system's control unit coordinates and manages the system's activities. In order to activate the other
functional elements of the system at the suitable time to carry out their respective functions, it
chooses and retrieves instructions from the main memory in the correct order. It receives user
input, processes the data, and then executes activities as necessary. The operation of every device
is managed by this component. The control unit components are: Node-MCU 32S, IRF9530N
MOSFET, Fixed Resistor and Capacitors.

Page | 822
Fig. 2 : Attendance Device Functional Block Diagram
The power unit consists of the component used to develop this attendance device's power section.
As the attendance device is rechargeable, the power unit has a battery, battery management circuit
and other components that are used to deliver power to other units, namely the Input Unit, Control
Unit and Output Unit of the attendance device. The power unit components are an 18650 Li-ion
Battery, TP4056 Charging Module, and XL6009 DC-DC Boost Module. d. DC-022 Connector.
e. DPST Switch. Any piece of gear that transforms data into readable form is considered an output
device. This attendance device's output unit shows the user the user interface, data, information,
and status that were sent by the control unit. Any specific information, the state of any operation,
or any audio message could be the output. The output unit components are: 2004A Liquid Crystal
Display, PCF8574T I2C Module, LED and Buzzer.

3.2 . Software design method.


The software section of this IoT-based attendance system project, three essential elements are
developed to work in combination with the hardware-based attendance device. They are Web-
based User Interface, Cloud Database and Attendance Device User Interface. Although the
attendance device is technically capable of scanning the fingerprint of the students and sending
the data to a cloud database over the internet, without the device user interface or the database,
the attendance device is useless. Because the device operating system is essential to operate the
attendance device and the database is a must to store the attendance data. Other than that, the web-
based user interface is required for the management of the whole system.
The web-based user interface offers a platform that the administrator or teacher can utilize to
simply administer the entire system. It manages how the attendance system as a whole operates.

Page | 823
This interface is developed using Visual Studio Code IDE, XAMPP software as development
tools and a number of programming languages

Fig. 3 : Software development


In this attendance system project, the cloud database is a relational database that is used to store
all the device information, teacher information, student information and attendance data. This
interface is developed using Visual Studio Code IDE, XAMPP software and Structure Query
Language (SQL). In developing this portion, the code is written using the Visual Studio Code and
to install, run and test the database, XAMPP software is used. Structure Query Language (SQL)
is the main programming language used to develop the database.
The Arduino Development Platform is used to create the user interface for the attendance device.
The Arduino IDE and the Arduino programming language are used to create the program for the
device user interface. An open-source electronics platform called Arduino is built on simple
hardware and software. A motor can be started, an LED can be turned on, and something may be
published online by using an Arduino board to receive inputs like light on a sensor, a finger on a
button, or a tweet. Sending a series of instructions to the microcontroller on the board will allow
you to direct it what to do. To do this, the Arduino Software (IDE) and the Wiring-based Arduino
Programming Language are used.

4. Prototype building.
4.1. Attendance Device building.
This attendance system project's main component is the attendance gadget. It is an electrical
gadget that can connect to the internet, perform fingerprint authentication, and communicate with
a web server and database. The device's circuit diagram displays each component and how they
are connected to one another. In this circuit, a NodeMCU-32s serves as the primary
microcontroller. Below is a schematic of the circuit.

Page | 824
Fig. 4 : Circuit’s diagram.

The controller unit of the circuit consists of NodeMCU-32s where ESP32 is used as a
microcontroller. An IRF9540N MOSFET’s 1st and 2nd pins connected using 1k and 10k resistors
with NodeMCU’s P23(37) and GND(38) respectively. And the IRF9540N MOSFET’s 3rd pin
goes in the positive output terminal of XL6009. A DC-022 Connector is connected to the input
terminal of a charging module and the battery terminal of the charging module (TP 4056) goes in
18560 Li-ion batteries. The input terminal of a DC-DC Boost Module (XL6009) is connected
with the output terminal of the charging module (TP 4056) through a DPST switch. The positive
and the negative output terminal of the DC-DC Boost Module (XL6009) are connected
respectively with Vin(5V) and GND(6).Fingerprint Sensor DY50’s 1st pin Vcc(5v) goes in
MOSFET 2nd pin. The GND(2) pin of the fingerprint sensor goes to the NodeMCU GND(38)
pin, other TXD(3) and RXD(4) are connected respectively with P17(28) and P16(27) of
NodeMCU. 4*4 Matrix Keypad pins are connected respectively with the pins of the
microcontroller. Again, three push buttons connected the keypads R1(1), R2(2), and R3(3) pins
with NodeMCU pin P5(29). A buzzer positive terminal, the 1st pin connected in the schematic
diagram.
Following that, a prototype is put together using common parts and jumper connectors so that it
may test the circuit design as well as the user interface for an attendance device. If the gadget can
perform all necessary tasks and operate the created user device interface, this validates its design.
At this point, adjustments are also made to the circuit diagram and device design to make them
dependable and functional. The illustration below depicts the prototype assembly.

Page | 825
Fig. 5 : Practical setup

4.2. Attendance Device User Interface.


The Arduino IDE and the Arduino programming language are used to create the user interface.
An audiovisual user interface makes up the user interface. All available options and information
are shown, and when necessary, a buzzer alerts the user. The attendance device user interface
block diagram is displayed in the user interface allows for straightforward operation of every
aspect of the attendance device and complete control over it. The user interface features less
complicated functions and is designed to be user-friendly. The gadget keypad inside may be used
to control every function.

Page | 826
Fig. 6 : Front page User Interface .

4.3. Operational Procedures of the device.


The attendance device has four main menus for its operation. They are used to operate and control
the whole system process of the attendance device portion of the attendance management system.
Connect Network: This menu is used to connect the attendance device with a WIFI network. The
device can save the last connected network, so the device can be connected to the saved network
or can be connected to a new network by entering a new SSID and password.
Get Attendance: The get menu is used to record the student attendance. After entering the “Get
Attendance” menu the teacher has to login into the device using the password and then select the
course to take attendance. Then the attendance taking is initiated and students can scan their
fingers to record attendance.
Register Fingerprint: Using this menu, a new student fingerprint can be added to the attendance
device. After entering the “Register Fingerprint” menu the teacher has to login into the device
using the password and then input the student ID to be added. Then the scan fingerprint procedure
will appear to register a new fingerprint bearing that student ID. After scanning the fingerprint
two times the registration process will be completed and the scanned fingerprint template will be
stored in the device.
Delete Fingerprint: From this menu, an existing student fingerprint template can be deleted from
the attendance device. After entering the “Delete Fingerprint” menu the teacher has to login into
the device using the password and then input the fingerprint ID that is to be deleted. This process
will remove the saved fingerprint template bearing that fingerprint ID.

Page | 827
4.4. Web-based User Interface.
In this project on an attendance system, a web-based user interface is created and utilized to
administer and operate the entire system. Some open source and cost-free programming languages
and development tools are used to create the user interface. such as PHP, JavaScript, Cloud
Database, HTML, Cascading Style Sheets (CSS), and Structured Query Language (SQL).The
web-based user interface is a straightforward website that is connected to a cloud database and
published to a web server. An administrator or instructor might access the web-based user
interface by entering their assigned email address and password in a web browser. Later, the
administrator can modify the password and email. The administrator or instructor will see six
parts on the toolbar after entering into the web-based user interface. Admin Panel, Student Profile,
Manage Student, Device Manager, Log Out, and About Us are their titles. The "Admin Profile"
part of the admin panel is where the admin profile is managed. By entering the necessary data in
the "Add New Course" section, a new course may be added. Using the "Erase Existing Course"
feature, the administrator might delete existing courses. The "Manage Course" and "Export

Fig. 7 :Block diagram of Web Based User-Interface.

Attendance" options in the course list are used for administering courses, such as assigning or
removing students from the course and exporting attendance. New students can be added to the
system, student profiles can be updated, and student profiles can be deleted using the "Manage
Student" component. The "Student Profile" area, which contains all necessary information, is
where the added student list may be located. To add a new fingerprint device, go to the "Device
Manager" section. information could be updated as well as a device could be deleted from this
section. Here also the added device list is shown. The “Log Out” option is used for logging out
from the website and it takes the user to the “Login” page. All information about the project
development is given on the “About Us” page.

Page | 828
5. Conclusion.

The project was developed keeping in mind how universities and other education
institutions manage attendance. Having such an automated system instead of manually
recording attendance has several advantages for both institutions and students.
Fingerprint is the key element of authentication system. It is used to take attendance of
students accurately and effectively. Since each person has a unique fingerprint, it is easy
to use and collects data with a high degree of accuracy. Considering that the gadget uses
the internet to connect to web servers and cloud databases. Therefore, it cannot be used
in locations without internet access. Sometimes presence device will not be able to import
or export data from device due to poor internet connection. As a result, attendance process
will be disrupted. The project also created to conserve crucial time for the instructor both
during and outside of class. Because of how simple it is to get attendance performance
reports and how the project does away with the roll calling method in the classroom.
Additionally, the proposed system does away with several drawbacks of the pen-and-
paper-based attendance system,
The system has been in use at Metropolitan University for several months. The users are
very satisfied with the system and give a positive feedback. according to their feedback,
the system can stop the proxy participation, it is quite easy to analyze the attendance
results. Teachers can save a lot of time and work. In a cloud database, it can store and
secure attendance data along with other information. This can reduce the chances of data
loss.
Overall, an inexpensive IoT-based biometric attendance management system that uses
fingerprint authentication and a cloud database can be of great use to educational
institutions in this field as it is highly efficient and secure. The development cost of this
project is considerably lower than traditional biometric attendance systems as the system
was built using mostly free and open-source software and hardware. This opens up the
possibility of practicing an affordable digital presence management system. The
following are some possible future work opportunities to improve the project: Create a
section for students in a mobile application portal, add previous course results to the
database, create a platform for online classes and add student payment history to the
database

Page | 829
References

Lim, T., Sim, S., & Mansor, M. (2009). RFID based attendance system. 2009
IEEESymposium on Industrial Electronics & Applications (pp. 778-
782). Kuala Lumpur,: IEEE.
M, K., Yeop, S., M., Z. A., Abdul, A., M., S. R., & Shah, M. (2007). Smart Attendance
System by Using RFID. Asia-Pacific Confarence On Applied Electtomagnetic
Procedings, (pp. 100-107).
M.I. , M., & N.M., Y. (2009). The Implementation of Wireless Student .
InternationalAssociation ofComputer Science and Information
Technology , (pp. 174-177).
Shehu, V., & Dika, A. (210). Using Real Time Computer Vision Algorithms in
Automatic Attendance . ITI 2010 32nd Int. Conf. on Information
TechnologyInterfaces, (pp. 397-402). Croatia.

Page | 830
7th International Conference on Engineering Research, Innovation and Education
January 12-14, 2023, Sylhet, Bangladesh

PAPER ID: 153

Enhancement of the Air Security of a closed System with MQ-3


Sensor
Md. Shamim Hossain , Sumit Roy Pronoy , Md Enamul Hoque
1 1 1*

1
Department of Physics: Shahjalal University of Science and Technology, Sylhet.

Corresponding Author: mjonyh@gmail.com


*

Abstract:
Alcohol is a widely used component in our daily life in different sectors. And in most cases, it is
used within a closed system, like home, wet laboratory, etc. In this work, we developed an MQ-3
sensor-based automatic monitoring system that enables us to detect and track the concentration of
alcohol in the air. We used ESP32 as a microcontroller for this IoT system. Since MQ-3 provides
information on smoke, we enrich the system by introducing its tracking. We computed the stored
data in the server for the necessary alert. We got a satisfactory result in the three of the tested
scenarios.

Keywords: MQ sensor, automatic monitoring system, alcohol concentration, ESP32


microcontroller, three scenarios.

Introduction:
In our home, we are exposed to alcohol by many sources, from food to body spray, everywhere
we have come to contact with alcohol in different kinds of forms. The main ingredient of body
spray and disinfection spray is ethyl alcohol . On average a bottle contains around 40-70% of
[1]

ethanol . In some foods, there are traces of methanol, like lentils (.04ppm), and carbonated
[1]

beverages .056ppm . And whenever a person is smoking inside a room he evaporates around
[2]

0.18ppm of methanol . Again in this pandemic, the widely used hand rub is basically based on
[2]

isopropyl alcohol. Again in public closed places like restaurants, there are many ingredients and
food waste that produces alcohol which is emitted into the air.

On the other hand, if our working place is a wet laboratory where we do chemical-based work
related to alcohol, we will come into direct contact with many kinds of alcohol. In places like
perfume factories, and bioorganic labs the primary alcohols are Ethanol and Methanol . [5]

Whether the place is if it is home or a laboratory, the emission of any kind of alcohol is dangerous
for any type of living body. Because in the case of Methanol the exposure is limited to 200 parts
per million over an 8-hour work shift . And for ethanol and isopropyl alcohol the limit is
[8]

respectively 200 and 1000 ppm . Though their odor limit is around 84-350ppm . High-
[3] [7] [6],[7]

concentration exposure can result in headaches, dizziness, drowsiness, exhaustion, and even loss
of consciousness .[6],[7]

As we know that these alcohols are explosive gas, they have Lower Explosive Limit (LEL) and
Upper Explosive Limit (UEL). Methanol has an LEL of 6% (60000 ppm) and a UEL of 36%
Page | 831
(360000 ppm) by volume in air. Ethanol explosive vapor concentrations in the range of
3.3(33000ppm)–19.0% (190000ppm) by volume, in the air . However, in this case, flammability
[4]

is not a big issue if there is adequate ventilation.

For these reasons, we need to monitor the place’s air quality. Or else we can’t predict when it
becomes harmful. To make our workplace and home become safer we need a device that can
analyze the air, and update us on the percentage and ppm of these types of Alcohol.

There are some air quality monitors in the current market. But in the sense of affordability, most
of them are out of reach. Our goal is to build a cost-efficient molecule analyzer that can analyze
the alcohol concentration in the air.

In this project, we use an IOT-based microprocessor esp32. We choose it for its multifunctionality,
affordability, and portability. It comes with a low-power system on a chip microcontroller with
integrated Wi-Fi and dual-mode Bluetooth. And it has operating temperatures between -40 °C to
125 °C . As we are planning to build our final product as a portable device this is the best choice
[9]

over other microcontrollers. We also use an MQ-3 sensor, it's a tinoxide-based gas detector . And
[9]

with Arduino IDE we build a compact and portable system that can detect Alcohol in any closed
place. It will send the collected data to a server from which one can analyze data over a large period
of time.

This prototype sensor system will collect data; by applying theory and programming through
Arduino ide we analyze it. We visualize and plot this analyzed data with python.

Theory:
From the datasheet of the sensors, we find a log-log graph, which’s x-axis indicates concentration
and the y-axis indicates Rs/Ro. From the graph, we see that the relation between them for different
gas results in a straight line. The relation between them is

log(RsRo) = m log(concentration of c) + b .............. (1)

Where,

Rs = sensor resistance at various concentrations of gas.

Ro = sensor resistance at a given concentration from the datasheet.

m= slope of the line

b= constant

Again Rs can be found through the given equation

Rs = (Vc/V -1) Rl ................ (2)


Rl

Page | 832
Where,

Vc = 3.3 v

Rl = load resistance

V = converted voltage from sensor data.


Rl

From Equation (1), we derive the equation from which we find the concentration in ppm.

C(ppm) = 10((log(Rs/Ro)-b)/m)

Method:
Our device is equipped with an MQ3 sensor that is connected with an esp32 microcontroller.
Through 4.7ohm resistor. To make sure the voltage supply we attach a buck module that constantly
passes 5 volt-12 amp current to the microcontroller. Right now the esp is powered by a direct
source but in the future, we will add a battery source.

When the chemical that is inside of the MQ sensor comes into contact with the alcohol in the air,
it reacts according to the concentration of alcohol. From that reaction, we get an analog value, that
is read from Arduino IDE. We convert that value into voltage. Then we put that value into the
theory equation and from that, we get the concentration of alcohol in ppm.

The data will store on an accessible website. From that, one can monitor the air quality from a
remote place. One can collect data over a fixed period of time. And analyze that for further
improvement of Alcohol emission within the closed system.

1. The circuit diagram of the sensor:

Page | 833
2. Methodology flowchart

Result:
1. Within lab

We take our data from a confined environment with our device and plot it. The resultant graph is
given below.

Page | 834
2. Within household

3. Within restaurant/tong

Page | 835
Analysis:
From the above result graph, which is produced with python, it is clearly shown that the ppm level
of alcohol is below the safety level threshold. As the lab environment is relatively clean and our
device can’t perform well below 0.04ppm, the data plot becomes discontinuous.

This data is about 8 days long and for most of the time, it stays around 4-6 ppm, cause in those
days for covid precaution there is continuous use of the small amount of isopropyl and ethyl
alcohol for cleaning, that leaks the vapors in the air.

Now, if we take a look at our data from a household, the ppm level of alcohol is below the safety
threshold. It ranges from 2 to 6 ppm on average. As it was a closed and clean room, the
environment was preferably better than the lab in regard to alcohol. Again this room was free from
various alcohol sources, that were common in the Lab environment, such as any alcoholic cleaning
product. For this, the ppm level is mostly within 2 to 5 ppm. Which is significantly lower than the
lab.

We are assuming that the reason behind the data was going near and higher than 4ppm is the daily
use of alcohol sources like body spray and hand rub. Also, there was some interference of fume
coming from the kitchen and it was the additional cause to raise the value to the top. It is interesting
to mention that there is a beautiful wave in this graph from the household. This wave indicates the
time cycle of our daily usage of products related to alcohol in our everyday life in this covid time.
It is noted that till 7th April, the first three dates were working days, and it is clearly shown that in
this time the average value of ppm is greater than the rest of the two days, which are holidays.

If we give a close look at our data we will notice that the is lowest from around 11 am to 1 pm. It
indicates that at that time the members of this household are least active or working outside. And
it goes higher up at the end of the day, approximately when the time touches 11 pm to 1 am. It
means that the members are active during this period. Then again the ppm fall to the lowest. During
holidays the time cycle is shifted due to the change of activity of the members.

Now from the third graph, we get the unexpected result. Though our device was surrounded by
crowds, smoke, and fume, it indicates that the quality of air is quite good comparing the situation.
It shows that the value of alcohol in the air is around 20 to 5 ppm and mostly below 3 ppm in the
surrounding area. It is unrealistic considering the environment of the open restaurant. The only
times it raises the values are when someone threw direct smoke into it.

Discussion:
Our sensor system performed well when it was used in the close systems, like a lab and
household. But when we took it outdoor and placed it into an open system, it showcased a poor
performance. It’s not due to the circuit quality or building and analysis method. It happens for
the nature of our sensors. We already know about our sensor, to get a proper result from it, we
have to make sure that it can make a good reaction with the targeted substance from the air. But
when we placed it in an open place, there was always a continuous airflow that make it harder to
make the reaction possible. As a result, it showed poorer performance in open places like
restaurants and tongs.

Page | 836
Though We get good results from close systems that satisfy our goal, It is important to be noted
that our device has some limitations here also. It can’t differentiate one alcohol from another. For
that, we can’t be sure which type of gas reaches a threatening level. But, it can’t make a huge
problem, as the higher safety limit for any alcohol to breathe is below 200ppm. Another limitation
is the accuracy limit of the sensor. Its performance is best within 0.04-4 ppm. If we go higher up
or go down the accuracy falls gradually. Again it has some relation related to the temperature of
the surrounding environment, till now which is not experimented with.

Most of this limitation comes from the sensor, but for a low cost and effective alternative, it is a
good option. With our further work, we will try to upgrade the sensor.

Conclusion:
 Our goal is to monitor the condition of the air, mainly alcohol vapor which has some serious
impacts on our daily life as we use it on a daily basis for pandemic safety; and it can be harmful
ifthe concentration level crosses the limit. We build the prototype as cost-efficient as possible.
 We will examine the relationship between temperature and data. After calibrating the data in
regardto temperature, We are planning to upload the data to a dedicated server. From the server,
one can see the air quality for any given place. We are also planning to retake data from
restaurants, this time we will choose close-door restaurants because they are also identified as
closed systems. Andwe are hoping to get better data this time.
 Further, we will attach a buzzer system that can alarm people when the alcohol concentration
willcross the given threshold limit. Small indication light will also be attached indicates whether
the concentration is within the exposure limit or not. If it is within the exposure limit the green
light will glow. If it is at the exposure limit the yellow light will glow and if it crosses the limit
the redlight will glow and the buzzer will on.
 By multiplying this prototype, we can implant these kinds of devices into different offices and
working places in our city, Sylhet. From a single server, then the authority can monitor the
important places' air quality and can take proper steps according to the data.
 Nowadays road accidents are increasing at an alarming rate. Drunken drivers are one of the
important reasons behind these events. We are planning to redesign the appearance of our device
to make another prototype that will be comfortable and safe for the breath test. our police authority
can use it to check the alcohol percentage in the breath of the bus and truck drivers. We can hope
that it will help to reduce the rate of the accident even if a small bit.
Reference:
Ethanol - an overview | ScienceDirect Topics
Methyl Alcohol - an overview | ScienceDirect Topics
CDC - NIOSH Pocket Guide to Chemical Hazards - Ethyl alcohol
Lower and Upper Explosive Limits for Flammable Gases and Vapors
METHANOL SAFE HANdLING MANUAL - health and safety module
Hazardous Substance Fact Sheet
Isopropyl alcohol, Hazardous Substance Fact Sheet
Hazardous Substance Fact Sheet
MQ-3 Semiconductor Sensor for Alcohol

Page | 837
7th International Conference on Engineering Research, Innovation and Education
January 12-14, 2023, Sylhet, Bangladesh

PAPER ID: 162

Blockchain and Smart Contract Solutions for Fair Payment


and Supply Chain Management in E-Commerce: Systematic
Literature Review

Md. Jabed Dhali*, Johaer Plabon, and Dr. Md Forhad Rabbi


● Department of CSE, Sylhet Engineering College, Sylhet-3100, Bangladesh.
dhalijabed@gmail.com
● Department of CSE, Sylhet Engineering College, Sylhet-3100, Bangladesh.
mdjohaerplabon@gmail.com
● Professor, Department of CSE, Shahjalal University of Science and Technology,
Sylhet-3114, Bangladesh. frabbi-cse@sust.edu
*Corresponding author [Md. Jabed Dhali]

Abstract: Blockchain technology (BCT) is a unique combination of features


Keywords: such as distributed and decentralized structure, consensus mechanism, Smart
Contracting (SC), and asymmetric encryption to ensure network security,
● Systematic transparency, and visibility. Modern industry is fully turning to this technology
Literature Review; for improving trust, auditability, transparency, credibility, etc. FPS and SCM
both are important factors in industry 4.0. There have several studiesexploring
● Blockchain;
the use of BCT in these fields have appeared in recent years. This research
● Smart Contract; presents a review of 20 research articles published between 2018 and2021 that
uses a qualitative method of answering some research questions. This study
● Fair Payment;
intends to review the available literature on FPS systems and supply chain
● Supply Chain; mechanisms in e-commerce, and the methodology taken by researchers,so as
to identify research gaps and recommend for future studies. The main purpose
● E-Commerce.
regarding this research is to analyze the recent works in these fields and come
up with an outcome that will support future research on FPS and supply chain
via BCT.
1. Introduction

In recent year, e-commerce grows rapidly and provides an opportunity to make sales and purchase
products and services for individuals and companies (Fatonah, Yulandari, & Wibowo, 2018).
Recent technologies are provoking significant disruptions and forcing the supply chain
management (SCM) field to develop new business strategy models and fair payment system (FPS)
system in e-commerce to improve trust for all stakeholders. One of the most promising ofthese
technologies is blockchain (BC), a distributive “state-of-the art” technology, which maintains
confidentiality, integrity, and availability (Dutta, Choi, Somani, & Butala, 2020a).

Page | 838
Although BCT applications emerged with bitcoins(Nakamoto, 2008), the current applications
have the potential to disrupt different, traditional industries like banking, real estate, insurance,
healthcare etc. The structure of theBC is organized in a way such that it ensures securityand
transparency of SCM and FPS mechanism. Based onthe decentralization, the SC is an essential
BCTapplication for reconfiguring several business models that works in an automated manner to
transferassets when a determined condition is satisfied (Queiroz et al., 2019). There exist several
works that suggest the application of BCT to e-commerce SCM and FPS or propose models but
there is still a barrier in terms of usability, security, privacy and cost.
The aim of this study is to research the literature of FPS and SCM in e-commerce in order to
understand the trendof using these systems. In this study, we conduct a systematic literature
review (SLR) of business scholarship on BCT from recent publications in these field, in 2018 to
the end of 2021, investigating the first five years of available data. This study isdesigned to
support scholars, decision-makers, and potentiallyleaders to think about the benefits, risks, and
limitations in ourcurrent knowledge of BCT in order to guide future research and application.
The rest of this paper is organized as follows. In Section 2, we justify and present the main aspects
of the SLR methodology based on Kitchenham and Charters. Section 3 provides the answers
addressing the five research questions. It presents the main contributions from the perspectives of
methodology, scope, and insights. Section 4 presents a discussion of the findings and the
conclusions.

2. Methodology
We describe in detail the process followed for the systematicliterature review (SLR) in this section.
The guidelines proposed byBarbara Kitchenham is followed to shape the SLR. The steps of this
approach after summarizing include review protocol, data sourcing, research questions framing,
strategy searching, data extractingand synthesizing.
Defining a set of research questions is the first stage of a SLR. In order to address the issues of
the current research on BCT and SCs for FPS and SCM in e-commerce literature, we devised a
set of the following research questions:
 RQ1: How does the literature define blockchain and SCs?
 RQ2: What benefits of blockchain and SC have the literature show to be applied in FPS and
supplychain?
 RQ3: What research topics have scholars addressin current research in the literature? What is
the impact of BCT on optimizing the FPS and supply chain?
 RQ4: What research actions have the business scholars take for solving the research topic in the
literature?
 RQ5: What are the top challenges and risks associated with blockchain and SCs in the literature?

2.2 Review Protocol


This research developed a review protocol in the preparation phase to set out the techniques to be
used in the review. Most of the broad tasks associated with conducting an SLR like data sourcing,
establishing search techniques,studying selection strategy, processing data extraction, anddata
synthesizing are all included in the review protocol as principal components.
2.3 Data Sources
Data sourcing of this research work consists of the related papers published within the last five
years (2018-2022) in renowned digital libraries which are preferred as the primary data source
for studying and analyzing articles needed forthis research. The electronic digital libraries
selected for this SLR are Google Scholar, ACM DigitalLibrary, IEEE Xplore Digital Library,
SpringerLink, and MDPI respectively.

Page | 839
2.4 Search Process
We searched to the digital libraries to find queries that are focused on FPSs or SCM using a
keyword pattern. The keyword pattern was achieved through identifying related lexical resources
including synonyms and acronyms, identifying relevant terms in the papers and constructing the
search string through Boolean operators as required. We used thefollowing search string for
querying literature and performing paper selection process.
(“blockchain” OR “smart contract”) AND (“securepayment” OR “FPS” OR “secure transaction” OR
“fair transaction”) AND (“supply chain” OR “logistic”)

2.4 Paper Selection


For filter and selection processing, we categorized 232 research findings from thesekeywords into
research publications at first. Then we categorized 62 findings with respect to their title and
abstract as it relates to the research questions. We selected 20 papers whose journal or conference
are included in ScimagoJournal and Country Rank. The research articles discussed Fairness and
securityof FPM and SCM, inclusively or exclusively are finally selected.

3. Result
In this section, we present the answers we identified for each of our researchquestions. In addition,
we outline the limitations and validity of this study, and implications for future research.
3.1 RQ1: How does the literature define blockchain and Smart Contracts?
Blockchain is an open, transparent, distributed databasesystem, is essentially an append-only data
structure maintainedby a set of nodes which do not fully trust each other (Xie et al.,2018), (Zhang,
Hu, Jia, Wang, et al., 2019). (Laxmaiah & Neha, 2018) defined BC as a digitized and
decentralized public ledger of all cryptocurrencytransactions in online cash transactions.In the
article (Hatefi, Bayat, Alaghband, Hamian, & Pournaghi,2022) defined BC as a distributed ledger
which areshared amongst all the involved entities identified by public keys and accessible
through their particular private keys. It is continuously increasing list of records, linked and
secured using cryptography (B. Li & Wang, 2018). Nodes in a BC keep replicas of blocks, eachof
which contains an ordered set of transactions. Due to the consensus mechanism, all nodes will
synchronize the transaction data forthe ledger maintained in each peer to be completelyconsistent.
Once the information is verified and added tothe BC, it will be stored permanently. Modifying
any node is invalid unless more than 51% of thenodes in the system cannot be controlled at the
same time, soBC maintains high stability and reliability (Xie et al.,2018). It is an immutable,
trusted decentralized ledger(Hasan & Salah, 2018), (Zhang et al., 2019).

Page | 840
Smart contracts provide facility of BC to transfer the financial transactions information
automatically to multiple parties according their agreement (Laxmaiah & Neha, 2018). It is a
programmable infrastructure whose instances and statuses are stored in BC (Xiao, Zhou, Guo,
Song, & Chen, 2022). All SC execution records are storedin blocks, to provide stakeholders with
detailed transaction information and a solution in case of dispute (Z. Liu & Li, 2020). Its
mechanism eliminates the needs for trusted third parties (TTP) as each party haveto deposit some
digital assets to the contract and the assets automatically get redistributed among the parties
accordingto the contract (Son et al., 2019). SC ensures the automatic execution oftransactions
based on presupposed conditions, which improvesthe efficiency of the transactions and avoids a
malicious breachof contract (Jiang & Chen, 2021). This behavior leads to opening opportunities
for the use of product-centric enterprise systems.

3.2 RQ2: What benefits of Blockchain and Smart Contract have the literature show to be
applied in FPS and SCM?
BCT offers a novel way for bookkeeping to keep track ofall records of the transactions that
occurred in the network(Laxmaiah & Neha, 2018) in a fully distributed manner,which makes
it potential and proved its capability, transparency, decentralized, tamperproof, traceability of
conducting peer-to-peer digital currency tractions to innovateother areas where there is a lack
of trust between involvedparties (B. Li & Wang, 2018), (Z. Liu & Li, 2020), (C. Liu et al.,
2018), (Gou, Shi, Gao, & Wu, 2021). It consists of an encryption algorithm, consensus
mechanism, and SC, featured by authenticity and unforgeability, whose core value lies in
establishing the openand transparent rules of the algorithms, constructing a trustednetwork, and
ensuring security and privacy of transactions (Hasan & Salah, 2018) (Wang et al., 2019)(Xiao
et al., 2022). One of the fundamental objectives of anypayment system is to avoid double-
spending, which is solvableby BCT (Hatefi et al., 2022). It maintains highreliability amongst the
peers involved in communications, andprovides public verifiability, transparency, privacy,
integrity, and redundancy (Hatefi et al., 2022), customers (C. Liu et al., 2018).
In e-commerce, BC provides secure options for making and recording transactions, agreements,
and SCs – anything that needs to be recorded and verified. With its potential, BC enables e-
commerceto be faster, safer, and more reliable to both sellers and customers (Son et al., 2019).
The characteristics of decentralization, persistency, anonymity, and auditability of BCT, the
properties of immutability, transparency, and natural settlement, in fact, completely match the
demand of the cross-border e-business payment field (X. H. Li, 2021). BC has demonstrated
the ability to conduct peer-to-peer digital currency payments ina trustless environment. BCT is
useful in solving many problems such as bottlenecks, delays, and operational risks in the financial
industry. As a safe, reliable, and decentralized architecture, BC is also used to create secure and
unreliablepayment schemes that can help the economy and communitieswithout a TTP. Thus,
these issues are now furtherconsidered according to the importance of anonymity andprivacy in
payment protocols (Hatefi et al., 2022).
3.3 RQ3: What research topics have scholars address in current research in the literature?
In E-commerce, (Xie et al., 2018) stated that security technology has become a major problem
limiting the rapid development and popularization of e-commerce. Existing solutions use BC
protocols to improve the trustworthinessof transactions, but most of them have certain limitations,
suchas lower throughput and higher consensus latency, and these issues make it difficult for BCT
to be widely used. (B. Li & Wang, 2018) stated that most BCT systems lack transactional privacy
Page | 841
protection because all transactions, including trading relationship between pseudonyms and
content transacted, are exposed on the BC. Although many existing BC privacy protection
methods have been proposed, these methods cannot perform well in protecting privacy and
ensuring speed without breaking the authentication protocol.
In e-payment protocol, unconditional security was not guaranteed in previous BCT applications (Zhang et
al., 2019). While the current e-commerce payment systems using conventional online or offline payment
system requirea payment gateway (PG). If the gateway is compromisedor say if the algorithm is decoded
then the entire payment information is at risk. In addition, this incurs PG fees, whichin turn increases the
cost of engaging in e-commerce (Kim & Kim, 2020). (Hatefi et al., 2022) stated that there is an increasing
demandfor the security of electronic payment systems. The importance of features such as avoiding double
spending, tracking malicious users, anonymity, and protecting user privacy are essential to achieve the goals
of e-payment systems. Protectinghonest users’ anonymity and privacy with the capability of tracing and
punishing malicious users without a need for a trusted server is challenging. In subscription-based business
model, (Buckley, Bendiab, Shiaeles, Savage, & Kolokotronis, 2021) argued that the management of
payments for multiple subscriptions has become a very complicated and insecure task for customers;
especially when it comes to renewing payment details when the card is lost, stolen, or expires. (Xiao et al.,
2022) stated that existing centralized ecommerce system requires high storage cost, low processing
efficiency, and the data is vulnerable to attacks and leaks.

In cash-on-delivery system, although it allows consumersto be more proactive in making


payments, it still appears tobe vulnerable for some valid drawbacks, identified by (Son et al.,
2019). First, purchasing process might be influenced by shipping carriers – a TTP. Second, as
payment is made between the customer and shipper which is not the seller, they may take
advantage of the clients’ collected money instead of returning them to the seller.
With the rapid development of information technology, though FPS and asset supply chain
becomethe focus of attention. But (Wang et al., 2019) claimed thatdue to the incomplete third-
party payment mechanism and the intrusion risk of various charging Trojan, it is easy to cause a
trust crisis. Therefore, thechallenging problem is to realize the fairness of payments and the
auditability of assets in a distributed system. In addition, (Shahid et al., 2020) stated that it is

challenging to trackthe provenance of data and maintain its traceability throughout the supply
chain network. In Proof of Delivery system, recent systems lack transparency, accountability
traceability, and credibility, mostly centralized, rely on TTPs to complete the delivery between
sellers and buyers. With the widespread of E-commerce, the need of a trusted system to ensure
the delivery of traded items is essential (Hasan & Salah, 2018).
3.4 RQ4: What research actions have the business scholar take for solving the research
topic in the literature?
(Xie et al., 2018) proposed a BC-based multilevel trusted trading framework in e-commerce
designed to achieve higher trading credibility. It allows all peers in the network to worktogether
to build a public, trusted, decentralized autonomous system that performs better in e-commerce.
(B. Li & Wang, 2018) proposed ring zero knowledge proof based on BC method (RZKPB) – a
Ethereum based public, permissionless, partially justified framework that does not allow financial
information to be stored on the BC in plain text format. (C. Liu et al., 2018) proposed a lightweight
BC platform for managing e-commerce transactions using a consensusmechanism called Practical
Byzantine Fault Tolerance (PBFT).The three-layer shrading is used to improve the scalability,
high computational overhead, and transaction-throughput. (Jiang & Chen, 2021) proposed a BC-
SupportedE-Commerce Platform (BS-EP) for small-and medium-sized enterprises (SME), which
provides trust guarantee for transactions between small-and medium-sized enterprises using
consortium BC and redundant Byzantine Fault Tolerance (RBFT) as a consensus mechanism.
(Hasan & Salah, 2018) proposed BC-enabled Proof of Delivery (PoD) framework to deliver
physical assets between multiple transporters utilizing Ethereum BC. It provides independent of
Page | 842
involvement from any TTPs, dispute settlement and automated remuneration. (Wang et al., 2019)
proposed an auditable protocol for transparent, verifiable and tamper proof transactions
betweentrading entities. The proposal uses three types of SCs designed to obtain reliable
payments between entities and allnetwork members. The protocol requires the use ofa TTP,
called regulator, responsible for authenticatingusers interested in participating in the network, and
registeringan SC with their data in the BC. To eliminate the third-party involvement for payments,
(Son et al., 2019) have proposed a payment system based on SC utilizing Hyperledger fabric. The
SCs in the proposed solutionconsist of a set of multiple rules for users such as delivery cost, time
and place. (Buckley et al., 2021) proposed a novel approach to automate, manage and simplify
the Financial Supply Chain involved in the process of updating and managing paymentsto users’
subscriptions by utilizing the Hyperledger Sawtooth BC framework, which allows the consumer
to enter their payment card details. The card being updated triggers to be updated on subscription
systems automatically.
(Z. Liu & Li, 2020) proposed a product information traceability framework to achieve
effectivetag verification in cross border e-commerce supply chain. It proposes new approach
methods and algorithms which can improve security, and traceability in e-commerce’s supply
chains, and therefore contribute to minimizing counterfeit attacks. (Shahid et al., 2020) presented
a complete solution for Agriculture and Food (Agri-Food) supply chain deploying Ethereum BC
network. The storage system returns a hash of the stored data with efficient, secure and reliable
manner. (X. H. Li, 2021) proposed three levels chain structure and data classification
management to construct an alliance chain with traders, customers, banks and regulators in the
cross-border e-business.
(Laxmaiah & Neha, 2018) presented the workflow of adigital online payment system using the
BC technique, explained the secure sign-in procedure of the BC strategy. The effects of BCT on
the online transaction managementsystem were elaborated in terms of security and usability.
(Ahn, Park, Shin, & Paek, 2019) proposed a payment model to calculate trust and reputation using
the values stored on BC ledger. The model normalizes user evaluations based on each user’s
personal evaluation criteria that changes over time. In addition, the model derives user reputation
and trust among application users by applying psychological factors. -(Zhang et al., 2019)
proposed a novel E-payment protocol, implemented by BC and quantum signature based on six-
qubit entangled states, which is the first time to applyBC to quantum E-payment. Another
quantum system, proposed by (Buckley et al., 2021), is a scheme to upgrade the current online
transaction method to maintain security aspects.By implementation of a quantum gateway and
converting transaction processing system into a distributed system. (Gou et al., 2021) proposed a
novel quantum e-payment protocol for a shopping platform, that used three-qubit entangled states
which can reduce the complexities of quantum resources.
In the article (Hatefi et al., 2022), a e-payment scheme is proposed to protect honest users’
anonymity and privacy with traceability and revocation ofmalicious users without a need for a
trusted server using several algorithms including fair blind digital signature, a secret sharing
scheme, and pseudonyms to maintain anonymity. (Kim & Kim, 2020) proposed a simple e-
commerce payment model using Ethereum BC to eliminate the need for transaction intermediaries
such as public key certificate and payment gateway. The integrity and nonrepudiation of e-
payments was guaranteed using digital signature, and the overall cost of operating e-commerce
services was reduced eliminating the fees for intermediary services. (Xiao et al., 2022) proposed
a decentralized e-commerce transaction system deploying Ethereum and the transactions are
recorded in the Interplanetary File System (IPFS). This literature effectively resists unfair
evaluation and collusion attacks, and proposes an incentive mechanism based on reputation value
to reward and punish nodes, thus promoting the good circulation of the system. (Elghaish et al.,
2022) proposed a Hyperledger fabric and chaincode based novel interconnected financial
management system to solve endemic problems with financial managementpractices adopted
Page | 843
within construction projects. This study provided a financial ‘proof of concept’ system that
addresses the weaknesses of financial management in various phases of construction projects in
decentralized and interconnected way.
3.5 RQ5: What are the top challenges and risks associated with blockchain in the
literature?
In the system proposed by (Son et al., 2019), the privacy protection feature is missing as authors
only focus on SC implementations. In fabric transaction data are exposed to all channel
participants, which the authors were not overcome. However, this leads to dependence on the
belief that the shipper will not take advantage of that key before handingit to the buyer. Therefore,
it cannot completely eliminatethe possibility of commodity fraud. In the proposed solution by
(Hasan & Salah, 2018), allthe entities act honestly by incentivizing the trading

entities. However, the proposed scheme uses the key and the asset whentransporting the asset. The
key and the delivered asset have norelationship with each other, and as a result, carriers can easily
manipulate the delivered asset. (Wang et al., 2019) has also proposed a pre-verification technique
to overcome the limitation in paper (Hasan & Salah, 2018). However, credibility of merchants is
not considered. Additionally, the existing trading networks have information asymmetrybetween
buyers and sellers. The information asymmetry resultsin poor credibility of trading entities and
the end consumers become vulnerable to fraudulent transactions. According to (Shahid et al.,
2020) SCs have still security gaps and they are facing hacking incidents, because they are not
mature enough. Additionally, Location privacy was not considered in this literature. According
to (Zhang et al., 2019), the existing multi-signature schemes in BC cannot both provide anti-
quantum security and keep the computational performance and the scalability of BC.
Additionally, the quantum single signature, quantum Fourier transfer, and quantum blind
signature scheme are also not suitable to be directly applied into multi-signature businesses of
BCs (Cai, Liu, Han, Wang, & Huang, 2021). The system proposed by (Kim & Kim, 2020) requires
a design for multithreading system architecture if the system is to process several payment
transactions.
In the business and information systems engineeringfield, (X. H. Li, 2021) found some BCT
disadvantages. BC scalability related challengesincluding throughput and latency have affected
BCT deeply development and application. The large-scale development andapplication of BC
requires a favorable legal environment, including both international and domestic levels.
Although the combination of BCT with e-commerce platforms can solve the financing and trading
issues of SMEs in (Jiang & Chen, 2021), there are some unresolved problems. It is difficult to
guarantee the authenticity of data before they arerecorded on the BC, which implies that all nodes
are confronted with the threat of source data deception. Tthe manner of auditing the original data
needs to be discussed in future research. Moreover, the application of BC helps solve the problems
faced by SMEs, but it also increases the cost of enterprises simultaneously.

2. Conclusion
E-commerce is getting a rapid popularity due to the advancement of digital transaction, FPS and
secured SCM. BCT promises to provide a secure and permanent solutionin e-commerce industry.
In fact, BC has the potential to offer more opportunities to ensure security. The traditional e-
commerce industryis facing many challenges. This SLR successfully demonstrated a distinct view
of the current scenario regarding FPS and SCM within last five years. In this study, we map out
the stages of BC research, including its potential as a disruptive innovation in society, and its
diffusion within the literature. This paper will help researchers explore about the currentstate-of-
the-art of this robust technology BC applied in FPS system and supply chain in e-commerce. The
studiesin our population demonstrate how BCT is growing up with improving transparency,
efficiency, and security across a widerange of value transactions. However, the researchers
Page | 844
tempered the optimistic tone by discussing the risks and challenges, including technical and
organizational barriers to wider adoptionof the technology. It has also included therole of BC in
managing challenges regarding FPS and SCM.
For articles search for SLR, only Scimago Journal and Country Rank (SJR) had been considered
for this study. In Future studies other electronic databases can be exploredas well. Future research
can also explore the BC use-cases in the other subject areas of e-commerce such as value chain
management, client relationship management, andhuman resource management as well.

Acknowledgment
The authors are thankful to Shahjalal University of Science and Technology as well as its
affiliated college named Sylhet Engineering College. More specifically, a big thanks goes to the
department of Computer Science and Engineering of both of the institutions for supervising and
supporting this research work. Our sincere thanks also go to the anonymous reviewers.In addition,
this research work received no external funding.

References
Ahn, J., Park, M., Shin, H., & Paek, J. (2019). A modelfor deriving trust and reputation on blockchain-
based epayment system. Applied Sciences, 9(24), 5362.

Buckley, D., Bendiab, G., Shiaeles, S., Savage, N., & Kolokotronis, N. (2021). Chainge: A blockchain
solution to automate payment detail updates to subscription services. In 2021 ieee international
conferenceon communications workshops (icc workshops) (pp. 1–6).

Cai, Z., Liu, S., Han, Z., Wang, R., & Huang, Y. (2021). A quantum blind multi-signature method for the
industrial blockchain. Entropy, 23(11), 1520.

Dutta, P., Choi, T.-M., Somani, S., & Butala, R. (2020b). Blockchain technology in supply chain
operations: Applications, challenges and research opportunities. Transportation research part e:
Logistics and transportation review, 142, 102067.

Elghaish, F., Rahimian, F. P., Hosseini, M. R., Edwards, D.,& Shelbourn, M. (2022). Financial
management of construction projects: Hyperledger fabric and chaincode solutions. Automation in
Construction, 137, 104185.

Fatonah, S., Yulandari, A., & Wibowo, F. W. (2018). A review of e-payment system in e-commerce. In
Journal of physics: Conference series (Vol. 1140, p. 012033).

Gajjar, B., & Acharya, J. (2020). Advance online payment security using qubits integrated blockchain.
International Journal of Future Generation Communication andNetworking, 13(1s), 01–07.

Gou, X.-l., Shi, R.-h., Gao, W., & Wu, M. (2021). A novel quantum e-payment protocol based on
blockchain. Quantum Information Processing, 20(5), 1–17.

Hasan, H. R., & Salah, K. (2018). Blockchain-based proofof delivery of physical assets with single and
multiple transporters. Ieee Access, 6, 46781–46793.

Hatefi, Z., Bayat, M., Alaghband, M. R., Hamian, N., & Pournaghi, S. M. (2022). A conditional
privacypreserving fair electronic payment scheme based on blockchain without trusted third party.
Journalof Ambient Intelligence and Humanized Computing, 1–14.
Jiang, J., & Chen, J. (2021). Framework of blockchainsupported e-commerce platform for small and
medium enterprises. Sustainability, 13(15), 8158.

Page | 845
Kim, S.-I., & Kim, S.-H. (2020). E-commerce payment modelusing blockchain. Journal of Ambient
Intelligence and Humanized Computing, 1–13.

Kitchenham, B., Brereton, O. P., Budgen, D., Turner, M.,Bailey, J., & Linkman, S. (2009). Systematic
literature reviews in software engineering–a systematic literature review. Information and software
technology, 51(1), 7– 15.

Laxmaiah, M., & Neha, T. (2018). A novel approach for digitalonline payment system. In International
conference on communications and cyber physical engineering 2018 (pp. 703–712).

Li, B., & Wang, Y. (2018). Rzkpb: A privacy-preserving blockchain-based fair transaction method for
sharingeconomy. In 2018 17th ieee international conference on trust, security and privacy in
computing and communications/12th ieee international conference on big data science and
engineering (trustcom/bigdatase) (pp. 1164–1169).

Li, X. H. (2021). Blockchain-based cross-border e-business payment model. In 2021 2nd international
conference on e-commerce and internet technology (ecit) (pp. 67– 73).

Liu, C., Xiao, Y., Javangula, V., Hu, Q., Wang, S., & Cheng, X. (2018). Normachain: A blockchain-based
normalizedautonomous transaction settlement system for iot-based e-commerce. IEEE Internet of
Things Journal, 6(3), 4680–4693.

Liu, Z., & Li, Z. (2020). A blockchain-based frameworkof cross-border e-commerce supply chain.
InternationalJournal of Information Management, 52, 102059.

Queiroz, M. M., Telles, R., & Bonilla, S. H. (2019). Blockchain and supply chain management integration:
a systematic review of the literature. Supply ChainManagement: An International Journal.

Shahid, A., Almogren, A., Javaid, N., Al-Zahrani, F. A., Zuair,M., & Alam, M. (2020). Blockchain-based
agri-food supply chain: A complete solution. Ieee Access, 8, 69230–69243.

Son, H. X., Nguyen, M. H., Phien, N. N., Le, H. T., Nguyen, Q. N., Tru, P. T., . . . others (2019). Towards
a mechanism for protecting seller’s interest of cash on delivery by using smart contract in hyperledger.
International Journal of Advanced Computer Science and Applications, 10(4).

Wang, S., Tang, X., Zhang, Y., & Chen, J. (2019). Auditable protocols for fair payment and physical asset
delivery based on smart contracts. IEEE Access, 7, 109439– 109453.

Xiao, Y., Zhou, C., Guo, X., Song, Y., & Chen, C. (2022).A novel decentralized e-commerce transaction
system based on blockchain. Applied Sciences, 12(12), 5770.

Xie, W., Zhou, W., Kong, L., Zhang, X., Min, X., Xiao, Z., & Li, Q. (2018). Ettf: A trusted trading
framework using blockchain in e-commerce. In 2018 ieee 22nd international conference on computer
supported cooperativework in design ((cscwd)) (pp. 612–617).

Zhang, J.-L., Hu, M.-S., Jia, Z.-J., Wang, L.-P., et al. (2019). A novel e-payment protocol implented by
blockchain and quantum signature. International Journal of TheoreticalPhysics, 58(4), 1315–1325.

Page | 846
7th International Conference on Engineering Research, Innovation and Education
January 12-14, 2023, Sylhet, Bangladesh

PAPER ID: 173

Dynamics of Soliton in Birefringent optical Fiber

Nabamita Das1*, Md. Jahedul Islam2 and Anindita Guho3


Department of Electrical and Electronic Engineering,
Khulna University of Engineering & Technology Khulna 9203, Bangladesh1,2,
North Western University, Khulna3
Email: nabamitadaspinky@gmail.com1, jahed@eee.kuet.ac.bd2,
guhopayelanindita@gmail.com3
Abstract: Soliton preserves its shape and velocity over long transmission
Keywords: Due to asymmetric core, in single mode optical fiber the refractive index
changes which is known as birefringence. When soliton is passed through
● Soliton; birefringent optical fiber several factors of the fiber results in broadening and
chirping the pulse at the receiver end. So, it requires to have a balance between
● Birefringent; the parameters affecting the transmission to ensure minimum loss in the data
● coupled NLSE; to be sent. In this paper, considering optical fiber as birefringent optical fiber
a balance has been drawn between the parameters affecting the performance
● GVD;
of the soliton like length, effective area, soliton period, and attenuation effects
● SPM; using OptiSystem Software. This work concluded a range of the values of
these parameters for least broadening and chirping. Since soliton follows
coupled non-linear Schrodinger equation the simulated results are compared
with the mathematical and analytical combination of soliton. Several works on
soliton had showed the effects of group velocity dispersion (GVD) and self-
phase modulation (SPM) in the transmission but this paper helps
understanding the transmission setup more clearly having included the
performance affecting parameters besides GVD and SPM.

1. Introduction
Modern world Telecommunication requires transmitting signals in considerably higher frequency
along with greater bit rate and of course having a better data carrying capacity. Fiber optic
communication offers these over copper wire electrical communication, but light pulses used in
fiber optic often change its shape over transmission due to the non-linearity which results in higher
data loss [1]. Then here comes soliton which preserves its shape and velocity over long
transmission because of a delicate balance between nonlinear and linear effects in the medium.
Soliton is very imperative area of research in non-linear optics. In last few decades, several papers
have been published in this area. Most of the works in soliton are theoretical which are related to
the common fact that soliton propagation in optical fiber is assumed to have only a single mode
in single mode [2].
In the single mode optical fiber, refractive index along two foremost axes generally remain same.
While manufacturing, sometimes the core of optical fiber becomes asymmetric. Due to
asymmetric core the refractive index of single mode optical fiber becomes bimodal i.e., the
refractive index along two axes become different. As a result, the group velocities of the pulses
become different [3]. The difference between the group velocities along two axes is known as
birefringence [4]. In general, because of the presence of birefringence, single-mode fibers are
bimodal. But due to the presence of birefringence, the partial pulses will tend to split apart only

Page | 847
when the pulse contains the components with both polarizations. Though soliton maintains its
shape and velocity for long transmission, different fiber parameters affect the transmission
dispersion [5]. So, which parameters are to be considered in modeling soliton in birefringent fibers
are needed to have a detailed knowledge. As it is well known that solitons are nonlinear pulses
which propagate without broadening in the anomalous dispersion region and it can be described
by the coupled nonlinear Schrodinger equation (NLSE) [6].
This work basically emphasizes soliton in birefringent optical fiber. The effect of different fiber
parameters is discussed through simulation approach. In this paper, coupled NLSE is also solved
analytically for soliton in birefringent optical fiber.
2. Mathematical Analysis
Soliton works based on solitary wave principle which can be described by the coupled non-linear
Schrodinger equation. The coupled NLSE for soliton in birefringence fiber without any loss can
be represented by following forms:
2
 u u  1  U 2 2
i    ( u  3 v )u  0 (1)
    2  2
2
 u u  1 U 2 2
i    ( v  3 u )v  0 (1)
    2  2

Where  is normalized group velocity, is the normalized time.


To solve (1) and (2) following hypotheses are considered for an exact solution

i
qx, t   P1(x, t)e 1 (2) (3)

i
rx, t   P2 (x, t)e 2 (3) (4)

Where P1( x, t) represents wave forms and l represents the phase components of pulses. Thus

l  kl x  wlt  l (4)

For l  1, 2 here in (5), kl represents the frequency of the two solitons, wl are the wave numbers

and l are the phase constants.

Substituting (3) and (4) reduces (1) and (2) respectively to


2
Pl 1  Pl Pl 2 3 2 2 Pl

i  wl pl  ( 2  2ikl  kl Pl )  Pl  Pl Pl  il  lkl Pl  0 (5)


t 2 x x 3 x

Page | 848
For l  1, 2 and l  3  l . Now, the real and imaginary parts of the equation (6), respectively
are
 2
 1 2  3 2 2 1  Pl

 
  wl  lkl  kl Pl  Pl  P P 
(6)
0
2 3 l l 2 x2
Pl Pl  Pl
k
l  l 0 (7) (8)
t x x

For bright solitons,

p i 
P(x, t)  Al sec h le l (8) (9)

For l  1, 2 where Al represents the amplitude of the soliton. Also,

  Bx  vt  (9) (10)

Where B is the two pulse’s width and v indicate the velocity with which the two polarized pulses
travel.
Now, the real part of the equation (6) which is given by (7) reduces to
  1 2 2 2 Pl 1 2 Pl 2
 w  k  k A P A B sec h   P (2P  1) A B sec h 

  l l l 2 l  l l l 
    2 l l l
(10)
2 2 P 2Pl 3 3P
A A sec h l  A sec h l  0
3
l l l

While the imaginary part of the equation (6) given by (8) reduces

 P

 vP   P  k P sec h l tanh A B  0 (11) (12)

l l l l l l
This is the desired solution of coupled NLSE for soliton in birefringent optical fiber.

3. Simulation and Results


The layout for fundamental soliton N=1.The simulation is done in OptiSystem software. For given
20 2
n2  3e m /W, A  93 um2,   1550 nm and   20 ps /km. The fiber length is set to
2
eff 2

Page | 849
631.72 km. Initial pulse has a sec h shape. Pulse power for fundamental soliton is 19.0155 mW
Figure 1 shows the transmitted pulses after 631.72 km propagation in a loss-free fiber. Both partial
pulses have been broadened by GVD and shifted in time with respect to each other which
corresponds to a difference in the arrival time for 1 km of fiber length. This shift can be attributed
to the birefringence.

Figure 1: Transmitted Signal

Figure 2 shows the received signal. Soliton is supposed to have zero dispersion but in practical
design soliton faces some dispersion. But with different parameter change, how soliton behaves
can be found out with following graphs where along the x axis different parameters are plotted
against the ratio of the corresponding parameter to the broadening of the output pulse along the y
axis.

Figure 2: Received Signal

Figure 3 shows that with the increase of fiber length the distortion increases very rapidly in 0 to
300 km but after 300 this effect is slowed down and over 1000 km the distortion decreases which
indicates that for a transmission line over 1000 km the abnormal behavior is exhibited.

Page | 850
Distortion means the difference between broaden of input signal and broaden of received signal
with respect to input broaden.

Figure 3: Distortion vs. fiber length

Figure 4 represents the relation of the effective area and the distortion for fiber length of 631.72 km.
Here we can see that with the changing effective area distortion increases.

Figure 4: Distortion vs. effective area

This paper emphasized getting a verified result of how different parameters of soliton in a
birefringent fiber should be changed to maintain limited distortion. Change in the length of the
transmission line shows significant change in the distortion. Since longer length is the foremost
requirement in fiber communication, signal distortion with length changing demands special
observation. Figure 3 indicates the distortion in the received signal when the length of the fiber is
being changed. Though the signal is being broadened up with length, the rate is quite low. But
increase of bit rate would cause significant losses in signal power here. This paper leads to a
conclusion about the distortion rate with length changing which guides the decision of how much
distortion may cause if a certain amount of fiber length is increased.

Page | 851
As figure 3 shows, the distortion is almost proportional to the fiber length while fiber length is in km
and distortion is innegative exponential unit. Soliton will hold its shape but while designing a
birefringent fiber transmission line, this research will help to take precaution about the distortion
loss. Figure 4 alsoshows the distortion effect while effective area of the fiber is changed keeping the
length of the fiber fixed. While designing a transmission line, fiber design has the foremost
importance since the major cost of the transmission is in the fiber. Effective are changing effect will
thus be handyin designing the fiber. Figure 4 exhibits the effect in changing the effective area of the
fiber in micrometer unit here. As the graph says, 20-40 and 60-80 micrometer effective area of the
fiber shows less distortion rate compared to others. So, design cost can be calculated based on data
raterequired with these results. This paper concludes that the relation between the length (along x
axis) and the distortion (along y axis) in the signal is expressed with the following equation:
2
y  8 exp(16x )  4 exp(12x)  4 exp(11)

And the relation between effective area (along x axis) and the distortion (along y axis) is expressed
with the following equation:
2
y  9 exp(14x )  9 exp(12x)  exp(11)

4. Conclusion
The entire model was designed in the software named OptiSystem to observe and have results.
Soliton maintains its shape, but other parameters affect the signal transmission. This paper talks
about those effects but circuit complexity and distortion cannot be minimized at a certain level. 100
percent efficiency from a transmission line also cannot be expected. But the distortion whichleads to
cross talks and inter channel interference, changes with length and effective area of the fiber. Thus,
this paper emphasized designing fiber while keeping the distortion possibility in knowledge.
References
Bhuvaneshwari, M., Dr. Hasan, S., Razak, A., Study on soliton pulse and its characteristics for fiber
optic communication, International Journal of Engineering & Technology. 2018, 7(4); 2845-
2847.
Xie, C., Karlsson, M., Andrekson, P.A. et al., Influences of Polarization-Mode Dispersionon Soliton
Transmission Systems, IEEE Journal of selected topics in quantum electronics.2002, 8(3); 575-
590.
Matsumoto, M., Akagi, Y., Hasegawa, A., Propagation of solitons in fibers with randomly varying
birefringence: Effects of soliton transmission control, Journal of Lightwave Technology. 1997,
15(4); 584-587.
Menyuk, C. R., Stability of solitons in birefringent optical fibers. II.Arbitrary amplitudes,Journal of
Optical Society America B. 1988, 5(2); 392-402.
Zhang, X., Wang, X., Soliton propagation in birefringent optical fibers near the zero- dispersion
wavelength, International Journal for Light and Electron Optics. 2004, 115(1);36–42.
Zhang, P., Hu, C., Zhou, Q. et al., Nonlinear control for soliton interactions in optical fiber systems,
Nonlinear Dynamics. 2020, 101; 1215-1220.

Page | 852
7th International Conference on Engineering Research, Innovation and Education
January 12-14, 2023, Sylhet, Bangladesh

PAPER ID: 201

Broadband mid-IR supercontinuum generation in As2Se3


based chalcogenide suspended waveguide: A new design
and analysis

Morol Musfiqur Rahman1, Md. Jahedul Islam2 and Nayem Al Kayed2


Department of Electrical and Electronic Engineering,
Khulna University of Engineering & Technology, Bangladesh1,2
Chittagong University of Engineering & Technology, Bangladesh3
Email: rahman1903555@stud.kuet.ac.bd1, jahed@eee.kuet.ac.bd2,
alkayed16@gmail.com3

Abstract: We numerically modeled a 1 cm long suspended rib waveguide


Keywords: that is compatible with photonic integrated chip (IC) production for broadband
mid-infrared (MIR) supercontinuum generation (SCG). The As2Se3 ChG glass
● Suspended rib system serves as the waveguide's core, and LiNbO3 glass and air serve as its
waveguide; bottom and top claddings, respectively. Only the anomalous absorption control
is pumped in the given waveguide structure, which is optimized for this.
● Supercontinuum
According to numerical analysis, the proposed geometry could provide a
generation; wideband SCG into the MIR, scaling the wavelength from 0.95 μm to 3.5 μm,
while utilizing an authentically low pulse peak power of 150 W and a
● Chalcogenide glass
commanding width of 50 fs, assuming pump at 1.55 μm wavelength. For the
materials; suggested waveguide structure, we developed the customized GVD curve. The
pump wavelength of 1.55 μm, spectral power, and temporal density, which
● Mid-infrared;
confirms the formation of the supercontinuum.

1. Introduction
Supercontinuum (SC) production is the conversion of laser light into super-wide continuous
optical spectrum with an extremely large spectral bandwidth (i.e. poor temporal coherence). SC
production in bulk BK7 glass was first found in 1970, and it has since been the subject of many
studies in a wide range of nonlinear optic waveguides [1]. Because of the design flexibility of
micro structured fibers, SC production has been detected over a far larger range of source
parameters than was previously feasible with planar waveguides or conventional fibers. SC
sources have established several operations in spectroscopy, pulse compacting, optic coherent
tomography, bio-imaging, frequency metrology, and the construction of adjustable ultrafast
femtosecond laser sources [2]. The spectral slicing of broadband SC spectra has also been
proposed in the telecommunications environment as a simple approach to construct multi-
wavelength optic sources for dense wavelength division multiplexing (DWDM) operations. Over
the past several decades, micro structured-based fiber design has been extensively employed to
produce SCG spectra into the MIR by manipulating the light mode confinementand dispersion
changes by changing the design's different structural characteristics [3].

Page | 853
However, a significant issue with this method is that when building long lengths of
microstructured fibers, there is a lack of shot to shot coherence in the output SCG spectra. Short-
length, scalable, and inexpensive CMOS compatible integrated photonic device manufacturing
is needed to solve this issue [4]. In place of micro structured fibers, the planar waveguide is a
possible solution to overcome these problems in the design of SCG sources [5]. In recent years,
several research groups have examined theoretical and experimental Supercontinuum generation
development in the mid-infrared (MIR) area utilizing a planar waveguide and micro structured
optical fiber. SCG coverage from 2 to 10 m using a Chalcogenide glass (ChG) rib waveguide
composed of GeAsSe/GeAsS glasses pumped at 4.184 µm with 330-fs pulses and aninput peak
power of 4.5 kW projected MIR SCG expansion between 1.2 and 7.2 µm in 2.5 mm long As2Se3
waveguide by pumping at 2800 nm wavelength with 200-fs pulses and input peak power of 2500
W [6]. Again a 30 mm long ChG step-index fiber constructed of As2Se3 with spectrum coverage
up to 16 µm and a peak power of 2890 kW using pump pulses of 170-fs at 9.8 m wavelength
[7]. A SCG ranging from 2 to 5 µm in a 12 mm long silicon suspended rib waveguide by pumping
with a 300-fs pulse width at 4 m wavelength and an average power of10 mW [8]. By pumping
100-fs pulses at 3.3 µm with an input peak power of 900 W [9] demonstrate computationally the
SCG spectrum expansion from 1.76 to 14.42 µm in a 0.87 mm long suspended As2Se3 rib
waveguide. A 20 mm long As2Se3 channel waveguide that creates a SCG spectrum ranging from
3.5 to 15 µm using a 170-fs pulse width pumped at 6 µm with a 10 kW input peak power [10].
Following recent technical advancements, researchers have become interested in on-chip
complementary metal oxide semiconductor (CMOS) compatible integrated photonic devices for
the production of effective broadband SCG into the MIR [11]. Designing CMOS compatible on-
chip broadband photonic devices using ChG materials with the help of their strong Kerr nonlinear
characteristics can result in ultra-broadband spectrum coverage that extends well into the MIR
[12].
2. MODELING AND METHOD

(a) (b)
Figure 1: (a) 2D and (b) 3D view of proposed suspended rib waveguide.

A model of a 1 cm long suspended rib waveguide that is compatible with photonic IC production
for broadband MIR supercontinuum generation, the As2Se3 ChG glass system serves as the
waveguide's core, and the bottom and top claddings are LiNbO3 glass and air, respectively. The
2D and 3D view of the proposed suspended rib waveguide are shown in Fig.1.
The necessary linear refractive indices of As2Se3 ChG glass and LiNbO3 glasses can be calculated
using the Sellmeier equation.

n(𝜆) = √1 + ∑𝑁 𝛼𝑘∗𝜆2 (1)


𝑘=1 𝜆2−𝛽𝑘2

Page | 854
where λ = the measurement for is in micrometers (m). The finite-element analysis (FEA) is a
widely used and well-known computational technique for characterizing an optical waveguide
with a specific dispersion of refractive index.
The Sellmeier fitting coefficients can be measure from the table:
Table 1: Sellmeier fitting coefficients of As2Se3 and LiNbO3
Material As2Se3 LiNbO3
N 𝛼𝑘 𝛽𝑘 𝛼𝑘 𝛽𝑘
K=1 4.994872 0.24164 2.21715 0.01764
K=2 0.120715 19 1.2290 0.054914
K=3 1.712369 0.48328 12.614 474.600

The refractive index of As2Se3 ChG glass varies as a function of wavelength according to the
Sellmeier equation [7].
4.994872𝜆2 0.120715𝜆2 1.712369𝜆2 (2)
𝑛(𝜆) = √1 + + +
𝜆2−0.241642 𝜆2−192 𝜆2−0.483282

The refractive index of LiNbO3 glasses varies as a function of wavelength according to the
Sellmeier equation [8].
2.21715𝜆2 1.2290𝜆2 12.614𝜆2
𝑛(𝜆) = √1 + + + (3)
𝜆2−0.017642 𝜆2−0.0549142 𝜆2−474.6002

where ‘λ’ is described in micrometers (µm). Effective area neff values are used to calculate group
velocity dispersion (GVD), which is an important metric to consider when characterizing a
waveguide for SCG production can be determine by equation (3)
𝐷( 𝜆) = − 𝜆 𝑑2𝑅𝑒(𝑛𝑒𝑓𝑓) (4)
𝑑𝜆2
𝑐
where ‘c’ is the speed of light. In numerical simulation, a generalized nonlinear Schrodinger
equation (GNLSE) will be used to examine supercontinuum production for the pulse's slowly
shifting envelope are shown in equation (4)
∂A α 𝑖𝑘+1 ∂A𝑘 𝑖 ∂ ∞ ′ 2 ′
∑8
= − 𝐴+ 𝑘≥2 𝑘! 𝛽𝑘 ∂Tk+ iγ(1 + ω 0 ∂T ) X(A∫−∞ 𝑅(𝑇)|A(z, T −𝑇 )| 𝑑𝑇 ) (5)
∂z 2

The pulse envelops evolve across the whole wavelength range of the waveguide in a retard
period with reference T = t − 𝛽 Z travelling at the group velocity 𝑣 = 1 . β (k≥2) and
1 𝑔 β1 𝑘

associated attenuation of the Suspended rib waveguides waveguide is. The nonlinear parameter
is defined as following equation (5)
𝑛 ω
γ = 𝑐𝐴2 0 (6)
𝑒𝑓𝑓

(∬ |E2|dXdy)2
where, n2 =Kerr-nonlinearity at pump wavelength and A eff = is mode effective
(∬ |E|4dXdy)1

area for fundamental mode. In our proposed model Mode effective area and nonlinearity curves
are shown in Fig. 4.

Page | 855
Figure 2: Field profile at 1.55 μm of proposed waveguide.

Figure 3: Mode effective area and nonlinearity curves of proposed SRN waveguide structure.

In Raman scattering As2Se3 materials shows sensibility, the complete nonlinear response
function R(t) is depicted as
𝑅(𝑡) = (1 − 𝑓𝑅)δ(t) + 𝑓𝑅ℎ𝑅) (7)
2 2 −1
τ +τ
ℎ𝑅(𝑡) = 1 2 𝑡 )𝑒 τ2 (8)
τ1 τ2 τ1
(𝑠𝑖𝑛
2

Page | 856
where the values of parameters for As2Se3 glass are described as f𝑅 = 0.148, τ1 = 23-fs, τ2 = 164.5-
fs.

3. Results and Discussion


The results of calculating the impact of changing H from 600 nm to 750 nm with a 50 nm step
while maintaining W = 3500 nm and S = 150 nm of the waveguide for pumping at 2.8 m
wavelength will be obtain. The H = 600 nm and S = 150 nm will be obtained by maintaining
constant W variation from 3000 nm to 3500 nm with a step of 250 nm. The field profile of quasi-
TE mode at 1.55 μm of proposed suspended rib waveguide is shown in Fig. 3. By pumping
different wavelength in proposed structure group velocity dispersion, spectral density, temporal
density curve will be determine [7]. The numerical study is carried out using the simulation
program COMSOL Multiphysics®, which is based on FEA. To begin, all of themode
propagation constants β(ω) for the fundamental quasi-TE mode must be computed using
COMSOL, and then these values must be used to determine any refractive index distribution over
a specific range of the required wavelength. By utilizing MATLAB simulation software,the
optimized waveguide structures the SCG in the MIR region will be investigated using the

Figure 4: SC spectrum at the waveguide output is shown by pumping at 1.55 um in anomalous


dispersion region with a low input peak power of 5W.

(a) (b)
Figure 5: (a) The tailored GVD curve for the proposed SRN waveguide structure. Vertical
dotted line represents the pump wavelength of 1.55 um. (b) The temporal evolution for
respective proposed waveguides.

symmetrized split-step Fourier method (SSFM). SC spectrum at the waveguide output is shown
in the Fig. 4.
Through the use of COMSOL, the GVD was estimated using Eq. 5 of an effective refractive index
for the wavelength range of 0.5um to 5um. The Sellmeier Equations for SRN and LiNbO3 given
in Eqs. 3 and 4 were used to generate the neff value.
Page | 857
The waveguide structure used to generate the GVD curve in Fig. 5 has dimensions of H = 0.5um
and W = 3um, and the predicted dispersion at pump wavelength is 89.4 ps/nm/km. It can be seen
in the image that two ZDW are formed at 1.2um and 2.5um, and they are crucial for the growth
of SCG in the longer wavelength area.

4. Conclusion
At the output of the proposed waveguides, the SCG spectrum is generated by pump pulses with
low to moderate peak power at wavelengths of 1.55 um and 2.8 um. The SCG expansion utilizes
a very low peak power of 5W at a wavelength of 1.55 um to covers the wavelength range of 1.2-
2.6 um. The acquired SCG expansion covers the wavelength range of 1.5-5.1 um and requires a
pump source of 2.8 um with a moderate peak power of 1kW to expand beyond 2.6 um.
However, by considering edibility into account, the architectural variation and impacton SCG
creation are also shown.
References

B. Wu, Z. Zhao, X. Wang, Y. Tian, N. Mi, P. Chen, Z. Xue, Z. Liu, P. Zhang, X. Shen,
Q. Nie, S. Dai, and R. Wang, “Mid-infrared supercontinuum generation in a suspended-core
tellurium-based chalcogenide fiber,” Optical Materials Express, vol. 8, no. 5, pp. 1341-
1348, 2018.
T. S. Saini, N. P. T. Hoa, K. Nagasaka, X. Luo, T. H. Tuan, T. Suzuki, and Y. Ohishi, “Coherent
midinfrared supercontinuum generation using a rib waveguide pumped with 200-fs laser
pulses at 2.8 μm,” Applied Optics, vol. 57, no. 7, pp. 1689-1693, 2018.
H. Ou, S. Dai, P. Zhang, Z. Liu, X. Wang, F. Chen, H. Xu, Baohua Luo, Yicong Huang, and R.
Wang, “Ultrabroad supercontinuum generated from a highly nonlinear GE–Sb– Se Fiber,”
Optics Letters, vol. 41, no. 14, pp. 3201-3204, 2016.
R. Kou, T. Hatakeyama, J. Horng, J. Kang, Y. Wang, X. Zhang, and F. Wang, “Mid-IR
broadband supercontinuum generation from a suspended silicon waveguide,” Optics
Letters, vol. 43, no. 6, pp.1387-1390, 2018.
U. Møller, Y. Yu, I. Kubat, C. R. Petersen, X. Gai, L. Brilland, D. Méchin, C. Caillaud,
J. Troles, B. L.Davies, and O. Bang, “Multi-milliwatt mid-infrared supercontinuum generation
in a suspended core chalcogenide fiber,” Optics express, vol. 23, no. 3, pp. 3282-3291,
2015.
D. D. Hudson,M. Baudisch, D. Werdehausen, B. J.Eggleton, and Biegert,“1.9 octave
supercontinuum generation in a As2Se3 step-index fiber driven by mid-IR OPCPA,” Optics
Letters, vol. 39 no. 19, pp. 5752-5755, 2014.
C. Chaudhari, T. Suzuki, and Y. Ohishi, “Design of Zero Chromatic Dispersion Chalcogenide
As2Se3 Glass Nanofibers,” Journal of Lightwave Technology, vol. 27, no. 12, pp. 2095–
2099, Jun. 2009.
M. R. Karim, B. M. A. Rahman, and G. P. Agrawal, “Dispersion engineered Ge11.5As24Se64.5
nanowire for supercontinuum generation: A parametric study,” Optics express, vol. 22, no.
25, pp. 31029-31040, Dec. 2014.

Page | 858
X. Gai, T. Han, A. Prasad, S. Madden, D. Choi, R. Wang, D. Bulla, and B. L. Davies , “Progress
in optical waveguides fabricated from chalcogenide glasses,” Optics express, vol. 18, no.
25, pp. 26635-26646, Dec. 2010.
P. Ma, D. Y. Choi, Y. Yu, X. Gai, Z. Yang, S. Debbarma, S. Madden, and B. L. Davies, “Low-
loss chalcogenide waveguides for chemical sensing in the mid-infrared,” Optics express,
vol. 21, no. 24, pp. 29927-29937, Dec. 2013.
J. Hu, C. R. Menyuk, L. B. Shaw, J. S. Sanghera, and I. D. Aggarwal, “Maximizing the
bandwidth of supercontinuum generation in As2Se3 chalcogenide fibers,” Optics express,
vol. 18, no. 7, pp. 6722-6739, Mar. 2010.
I. Kubat, C. S. Agger, U. Møller, A. B. Seddon, Z. Tang, S. Sujecki, T. M. Benson, D. Furniss,
S. Lamrini, K. Scholle, P. Fuhrberg, B. Napier, M. Farries, J. Ward, P. M. Moselund, and
O. Bang, “Mid-infrared supercontinuum generation to 125μm in large NA chalcogenide
step index fibres pumped at 45μm,” Optics express, vol. 22, no. 16, pp. 19169-19182, Aug.
2014.

Page | 859
7th International Conference on Engineering Research, Innovation and Education
January 12-14, 2023, Sylhet, Bangladesh

PAPER ID: 207

High-sensitivity Acetone Sensor Using Pt Nanoparticles


Functionalized ZnO

Md Tawabur Rahman 1*, S. M. Sayedur Rahman 2, Rajesh Pathak 3, and Quinn


Qiao 4

1,2Department of Electrical and Electronic Engineering, Khulna University of


Engineering & Technology, Khulna, 9203, Bangladesh
tawabur@eee.kuet.ac.bd, smsayedur26@gmail.com
3 Department of Electrical Engineering, South Dakota State University, Brookings, SD,

USA
rajesh.pathak@jacks.sdstate.edu
4Department of Mechanical and Aerospace Engineering, Syracuse University, NY,
USA
quqiao@syr.edu

* Corresponding author [Md Tawabur Rahman]

Abstract: The development of an acetone sensor is demonstrated using ZnO


Keywords: functionalized with platinum nanoparticles (Pt NPs). The loading of Pt NPs on
ZnO helped to realize high sensitivity even at low concentrations of acetone. A
● Acetone; facile solvent exfoliation method was utilized for preparing ZnO and then
● Gas sensor; functionalized with Pt NPs which is observed by their morphologies on SEM
images. The ZnO-Pt NPs sensors showed an excellent response for acetone in
● ZnO;
the range of 5 - 100 ppm. This can be ascribed to the structure of ZnO-Pt NPs
● Pt NPs; which provides more active sites and diffusion pathways for absorbing acetone
molecules. Further, the n-doping effect of Pt NPs causes to increase in the
● Morphology;
thickness of the electron depletion layer, which enhanced the sensitivity of the
● Sensitivity; gas sensor. Additionally, the resulting sensor showed a fast response and
recovery. Thus, this sensor could be a potential candidate for high-performance
gas sensing applications.

1. Introduction
The release of various hazardous gases including nitrogen dioxide, carbon dioxide, ammonia,
acetone, benzene, toluene, etc., in the air due to the rapid growth of industrialization causes
environmental pollution and poses detrimental effects on human health [1]. In particular, the
concentration of acetone in a human exhaled breath of more than 1.8 ppm causes diabetes, while

Page | 860
it is 0.3–0.9 ppm in healthy individuals [2]. In addition, a concentration of acetone beyond 173
ppm can irritate the eyes, throat, nose skin, and central nerval systems [2]. Moreover, the intake
of acetone in excessive amounts may cause intoxication and death [3]. Thus, acetone in exhaled
breath can be utilized as a biomarker for noninvasive diagnosis of diabetes, overweight, and
systolic dysfunction that are related to each other [4]. Due to the presence of low concentrations
of water, carbon dioxide (CO2), and different VOCs in exhaled breath, it is very difficult to obtain
selective detection of gaseous acetone [5]. Therefore, it is important to develop a facile and highly
sensitive sensor for the detection of acetone in exhaled breath.

Various methods, including gas chromatography, mass spectrum analysis, etc., have been used to
detect acetone in the air. However, these instruments are bulky, expensive, and need skilled
personnel. Metal Oxide Semiconductors (MOS) like SnO2, WO3, and TiO2, etc., have been
recently widely used for developing various gas sensors [5]. Among them, ZnO has been
considered a promising candidate for gas sensing due to its easy preparation, low cost, high
chemical and thermal stability, and favorable sensing properties [6]. Various nanostructures of
ZnO including nanowires, nanorods, nanotubes, nanosheets, and nanoparticles have been
investigated to enhance the sensing capabilities of ZnO. The sensing capabilities of these
structurally modified ZnO materials are unsatisfactory due to their slow reaction times, lack of
selectivity and poor response rates [7]. However, the sensing properties of ZnO were improved
by incorporating pure metals (Au and Pt) as well as transition metals such as Cd, Cu, Cr, Mn, Fe,
Co, and Ni into it [8]. The incorporation of these pure metals reduces the activation energy of
ZnO so that it may react with acetone more effectively and thus, enhance sensing performance.
Furthermore, various metal nanoparticles (NPs) such as Au, Al, Cu, etc., have been also
incorporated into ZnO to further improve acetone sensing capabilities [8]. For example, Koo et
al. used Pt NPs decorated Al-doped ZnO to achieve good sensitivity and quick response toward
acetone sensing [9]. But high thermal energy (450°C) is needed for operating the developed
sensor, which costs large electrical power.

In this work, ZnO decorated with platinum nanoparticles (Pt NPs) has been utilized for developing
an acetone sensor. A facile solvent-mixture technique was utilized for preparing ZnO and then
functionalized with Pt NPs. The Pt NPs incorporated ZnO showed an excellent response under
the exposure of acetone in the range of 5 - 100 ppm at 250 °C. This can be ascribed to the structure
of ZnO-Pt NPs which provides more active sites and diffusion pathways for absorbing acetone
molecules. Further, the n-doping effect of Pt NPs caused an increase in the thickness of the
electron depletion layer, which enhanced the sensitivity of the gas sensor. Additionally, the
resulting sensor showed a fast response and recovery time of about 29 s and 26 s, respectively.
Thus, this sensor could be a potential candidate for high-performance gas sensing applications.
2. Experimental
2.1 Materials
ZnO nanopowder (<100 nm particle size), platinum nanoparticles dispersion, and Poly
(vinylidene fluoride) (PVDF)were purchased from Sigma Aldrich, USA. N-methyl-2-pyrrolidone
(NMP) solvent supplied by Thermo Fisher Scientific, USA. Deionized (DI) water was thoroughly
used to dissolve ZnO nanopowder.

Page | 861
The loading of Pt NPs on the ZnO sample was conducted by following the work reported by
Wang et al. [10]. Briefly, 40 mg of ZnO nanopowder was dissolved in 5 ml of DI water. Then,
10µl of Pt NPs dispersion was added in the ZnO dispersion and the resulting solutions were
ultrasonically agitated at room temperature for 10 min. Finally, the precipitation was collected by
washing with DI water and dried under a vacuum oven for 6 h to obtain Pt NPs loaded ZnO
sample.

2.2 Characterizations
The morphology and microstructure of ZnO and Pt NPs were analyzed by Field emission scanning
electron microscopy (SEM, Hitachi S-4700).

2.3 Sensor Fabrication


A 60 nm of Au was deposited on an alumina substrate using a CRC-150 Sputtering System. A
stainless steel shadow mask was utilized to develop a pattern of interdigitated electrodes (IDEs)
on an alumina substrate with a width of 100 µm. An 80 wt% of ZnO nanopowder, 15 wt% of Pt
NPs, and 5 wt% of PVDF with 200 µl NMP were placed into an agate mortar and ground to form
a homogenous paste. The resulting paste was coated on Au IDEs coated on an alumina substrate
and dried in a vacuum oven at 100 °C overnight. Finally, the as-prepared sensors were stabilized
in the air for 1 h before measurements.

2.4 Gas-Sensing Experiment


The performance of sensor devices was tested in a customized stainless-steel chamber built in the
laboratory. A gas inlet, an outlet, and an electrical feed-through were installed in the chamber.
The sensor was placed in small aluminum metal blocks housed with a heater and an RTD
integrated with an external temperature controller. Nitrogen was used as the purging/balance gas
and the operating temperature was controlled with a temperature controller. The desired
concentrations of target gas were obtained by controlling the flow rates of target gas and balance
gas using mass-flow controllers by keeping a total flow rate of 500 sccm. Before testing, the
chamber was purged with nitrogen gas to remove residual gases in the chamber. A +1 V of input
voltage was applied between two electrodes and the real-time sensor response was acquired using
the source meter (Keithley 2602A). The sensor response is defined as below:
𝐼# − 𝐼$ ∆𝐼 (1)
𝑟𝑒𝑠𝑝𝑜𝑛𝑠𝑒(%) = × 100% = × 100%
𝐼$ 𝐼$

Here, 𝐼$ 𝑎𝑛𝑑 𝑖𝑠𝐼# are the current of the gas sensor before and after exposure to the target gas,
respectively.

3. Result and Discussions


3.1 Morphological and Structural Characterization

Page | 862
Figure 1a shows the image of the obtained ZnO-Pt NPs solution ultra-sonicated for 10 minutes.
It shows a homogeneous solution of ZnO-Pt NPs. Figure 1 b shows the morphology of the Pt NPs
decorated ZnO. A mixture of rod and flake shape morphology of ZnO can be seen in Figure 1 b.
Due to the very small size of Pt NPs, it is difficult to observe the presence of Pt NPs on the surface
of ZnO.

(a) (b)

Figure 1. (a) Image of ZnO-Pt NPs solution sonicated for 10 minutes (b) SEM image of Pt NPs modified
ZnO.
3.2 Sensor Performance
Figure 2 shows the linear I-V curve of the ZnO-Pt NPs gas sensor under air and at 250°C. The
source voltage was varied from -1 V to +1 V. The linear variation between applied voltage and
current is considered evidence for the existence of ohmic contact between ZnO and Au. The
sensor current increased gradually with voltage at 250°C, whereas a rapid increase of current was
observed in the range of - 49.73 µA to 49.98 µA under air. The resulting sensor showed a
resistance of 3.5 Ω under air, while it showed 64 kΩ at 250°C.

Fig. 3a shows the gas sensing performance of the ZnO-Pt NPs sensor demonstrated under various
concentrations of acetone gas in the range of 5 ppm-100 ppm. The ZnO- Pt NPs sensor showed a
substantially increasing current with increasing concentrations of acetone. In particular, the sensor
current of 114.24µA was obtained for 5 ppm acetone, which increased to 152.68µA for 40 ppm.
This can be attributed to the adsorption of acetone molecules on the sensor surface. The current
saturated to 167.81µA and 176.39µA for 50 ppm and 100 ppm of acetone, respectively. Under
high concentrations of acetone (50-100 ppm), the sensor surface become saturated and did not
allow adsorption anymore, leading to an almost constant response. Figure 3b shows the response
percentage variation with different acetone gas concentrations. It is seen that initially, the
response increased from 89 % to 153% for 5 ppm to 40 ppm of acetone. Finally, the maximum
response of 192% was obtained at 100 ppm of acetone.

Page | 863
Figure 2. I-V characteristics of the ZnO-Pt NPs gas sensor under air and at 250°C.

(a) (b)
200

180

160
Response (%)

140

120

100

80

5 10 20 40 50 100
Concentrations of acetone (ppm)

Figure 3. (a) The detection of acetone at various concentrations of acetone in the range of 5 ppm to 100
ppm at 250°C (b) Response percentage versus different concentrations of acetone.

Figure 4 shows the response and recovery curve of the ZnO-Pt NPs gas sensor to 5 ppm of
acetone. The sensor showed a response and recovery time of 29 s and 26 s, respectively, which
demonstrates a faster response and recovery.

Page | 864
1

Figure 4. Response and recovery curve of ZnO-Pt NPs gas sensor to 5 ppm of acetone.

4. Conclusion
In summary, the surface of the ZnO nanostructure was decorated with Pt NPs and utilized for
developing a miniature sensor for detecting acetone. The developed sensor showed high
sensitivity toward acetone in the range of 5 – 100 ppm. Furthermore, high sensitivity of 89% was
obtained even at a very low concentration (5 ppm) of acetone, which is promising. The highly
specified surface area of the ZnO- Pt NPs provides more active sites for acetone adsorption and
improves the kinetics of the surface reactions, which can be ascribed to the catalytic effect of Pt
NPs. Additionally, it showed a faster response and recovery of 29 s and 26 s, respectively, for 5
ppm of acetone. Therefore, the ZnO - Pt NPs is the potential for a practical acetone sensor for
protecting human health and environmental protection.

Acknowledgment
The research described in this work was supported by Sanford/SDSU Collaborative Research
Seed Grant and NSF Innovation Corps (1906755).

References
S. Chowhan, “Potential Hazards of Air Pollution on Environment, Health and Agriculture,” Int.
J. Environ. Sci. Nat. Resour., vol. 28, no. 2, pp. 1–10, May 2021, doi:
10.19080/ijesnr.2021.28.556231.
L. Chen et al., “Enhanced acetone-sensing properties of pt-decorated al-doped ZnO
nanoparticles,” Sensors Actuators, B Chem., vol. 280, no. 1, pp. 109–119, Feb. 2019, doi:
10.1016/j.snb.2018.10.049.
Y. N. Lee et al., Atmospheric chemistry and distribution of formaldehyde and several
multioxygenated carbonyl compounds during the 1995 Nashville/Middle Tennessee Ozone
study, vol. null. 1998.
A. Koo, R. Yoo, S. P. Woo, H. S. Lee, and W. Lee, “Enhanced acetone-sensing propertiesof pt-
decorated al-doped ZnO nanoparticles,” Sensors Actuators, B Chem., vol. 280, no. October
2018, pp. 109–119, 2019, doi: 10.1016/j.snb.2018.10.049.

Page | 865
V. Saasa, T. Malwela, M. Beukes, M. Mokgotho, C.-P. Liu, and B. Mwakikunga, “Sensing
Technologies for Detection of Acetone in Human Breath for Diabetes Diagnosis and
Monitoring,” Diagnostics, vol. 8, no. 1, p. 12, Jan. 2018, doi: 10.3390/diagnostics8010012.
H. Xu, X. Liu, D. Cui, M. Li, and M. Jiang, “A novel method for improving the performance of
ZnO gas sensors,” Sensors Actuators, B Chem., vol. 114, no. 1, pp. 301– 307, Mar. 2006,
doi: 10.1016/j.snb.2005.05.020.
Y. Chen et al., “Highly sensitive and selective acetone gas sensors based on modified ZnO
nanomaterials,” Mater. Sci. Semicond. Process., vol. 148, p. 106807, Sep. 2022, doi:
10.1016/j.mssp.2022.106807.
M. B. Rahmani et al., “Gas sensing properties of interconnected ZnO nanowires,” in
Sensor Letters, Apr. 2011, vol. 9, no. 2, pp. 929–935. doi: 10.1166/sl.2011.1646.
N. Zhang, H. Li, Z. Xu, R. Yuan, Y. Xu, and Y. Cui, “Enhanced acetone sensing propertyof a
sacrificial template based on cubic-like mof-5 doped by ni nanoparticles,” Nanomaterials,
vol. 10, no. 2, Feb. 2020, doi: 10.3390/nano10020386.
P. Wang, T. Dong, C. Jia, and P. Yang, “Ultraselective acetone-gas sensor based ZnO flowers
functionalized by Au nanoparticle loading on the certain facet,” Sensors Actuators, B
Chem., vol. 288, pp. 1–11, Jun. 2019, doi: 10.1016/j.snb.2019.02.095.

Page | 866
7th International Conference on Engineering Research, Innovation and Education
January 12-14, 2023, Sylhet, Bangladesh

PAPER ID: 214

EFFICIENCY MEASUREMENT OF MACHINE LEARNING CLASSIFIERS


TO ANALYZE BUSINESS DATA FOR FUTURE PREDICTION WITH
RESPECT TO GLOBAL DATASET
Partha Pratim Debnath1*, Fahmida Tahmin2, Md. Golam Rashed3 and Dipankar Das4
1,2
Assistant Professor, Bangladesh Army University of Engineering and Technology, Natore,
Bangladesh
3
Associate Professor, University of Rajshahi, Rajshahi-6205, Bangladesh
4
Professor, University of Rajshahi, Rajshahi-6205, Bangladesh
partha@bauet.ac.bd, fahmidatahmin14@gmail.com, golamrashed@ru.ac.bd,
dipankar@ru.ac.bd

Abstract: Accurate prediction of stock market returns is a very challenging


Keywords: task due to volatile and non-linear nature of the financial stock markets. For
prediction of the stock market prices people search for accurate methods and
● Machine Learning tools which will increase their profits, while minimize their risks. With the
● Artificial Neural introduction of artificial intelligence and increased computational capabilities,
programmed methods of prediction have proved to be more efficient in
Network predicting stock prices. In our proposed work, we have implemented various AI
● Linear Regression algorithms for future prediction of stock market and tested the efficiency. For
this, Linear Regression, Artificial Neural Network and Naïve Bayes have been
● Naïve Bayes utilized for predicting the next day closing price for several companies. By using
Classifier linear regression, we got almost 93% of accuracy of prediction. In naïve bayes
method we obtained 29% accuracy. Whereas neural network showed major
error. We used same dataset for every classifier to evaluate the accuracy after
performing data engineering and relevant analysis. We can conclude that linear
regression is best among tested methods during this research for stock market
prediction.

1. Introduction
Predicting how the stock market will perform in future is one of the most difficult things to do.
There are so many factors involved in the prediction – physical factors vs. psychological, rational
and irrational behavior, etc. All these aspects combine to make share prices volatile and very
difficult to predict with a high degree of accuracy.
We propose to predict the future of the business by using machine learning classifiers. We have
intended to use different types of machine learning classifiers to predict the future of the business
by analyzing the dataset and measure the accuracy of the prediction.
Because of global digitization, SMP has entered a technological era. Machine learning in stock
price prediction is used to discover patterns in data [1]. Usually, a tremendous amount of
structured and unstructured heterogeneous data is generated from stock markets. Using machine

Page | 867
learning algorithms, it is possible to quickly analyze more complex heterogeneous data and
generate more accurate results. Various machine learning methods have been used for SMP [2].
In business there are many uncertainties and risk. The stock market appears in the news every
day. Every day it reaches a new closing price, high price, low price, opening price also to quick
changes because of the underlying nature of the financial domain and in part because of the mix
of known parameters (Previous Day’s Closing Price, P/E Ratio etc.) and unknown factors (like
Election Results, Rumors etc.). The rates of investment and business opportunities in stock market
can increase if an efficient algorithm could be devised to predict the short term price of an
individual stock [3]. But a proper choice of algorithm is very important to ensure the optimal
accuracy of prediction. This is due to fact that not all the algorithms shows good performance for
analyzing time series data. Moreover, in a global scenario, where the price data changes randomly,
it is very tough to predict without good analytics. Keeping those facts in mind, our goal is to
implement various algorithms on collected data sets and track the best one for proper prediction.
The goal of this research is to adequately describe a prediction model that is able to predict stock
market trends with sufficient accuracy and profitability. While there have been many attempts at
doing this in the past, we wanted to find out if we could do this using the resources available to
us, free of cost. This thesis reflects our research on the stock market, various ML algorithms used
to predict stock market trends in the past, and the specific features, classifiers, and datasets needed
to do so accurately.
2. Literature Review
2.1 Review of Existing Work
Recently, a lot of interesting work has been done in the area of applying Machine Learning
Algorithms for analyzing price patterns and predicting stock prices and index changes. Most stock
traders nowadays depend on Intelligent Trading Systems which help them in predicting prices
based on various situations and conditions, thereby helping them in making instantaneous
investment decisions [3].
There are many previous work with the classifiers and the level of accuracy is much better with
logistic regression and neural network. If we change the target the level of accuracy is still good.
But there is a drawback they didn’t work with Naive Bayes and the accuracy level isn’t defined.
The ability of Gaussian Naïve Bayes ML algorithm to predict stock price movement has not been
addressed properly in the existing literature, hence this work attempt to fill that gap by evaluating
the performance of GNB algorithm when combined with different feature scaling and feature
extraction techniques in stock price movement prediction.
2.2 Limitation of Existing Work and Justification
Table 1: The classifiers and the accuracy level.

Classifiers Accuracy Level

Linear regression 96%

Neural Network 86%

Naive Bayes Not defined

Page | 868
Here we can see the limitation of previous work. In the previous work there is no clear definition
of accuracy for Naive Bayes classifier. The accuracy level of the other classifiers are not that
much satisfied.
3. Methodology
3.1 Linear Regression
Regression predicts a numerical value [4]. Regression performs operations on a dataset where the
target values have been defined already. And the result can be extended by adding new
information [5]. The relations which regression establishes between predictor and target values
can make a pattern. This pattern can be used on other datasets which their target values are not
known. Therefore the data needed for regression are 2 part, first section for defining model and
the other for testing model.
Mathematically we can represent linear regression as:
y=a0+a1x+ε ........................................................................................................................... (1)
Here,
Y= Dependent Variable (Target Variable).
X= Independent variable (Predictor Variable).
a0 = Intercept of the line (Gives an additional degree of freedom).
a1 = Linear regression coefficient (scale factor to each input value).
ε = Random error
3.2 Naive Bayes
The NB is a classification algorithm and a probabilistic machine learning algorithms used for
many classification functions and is based on the Bayes theorem [6]. Gaussian Naïve Bayes is the
extension of naïve Bayes. While other functions are used to estimate data distribution, Gaussian
or normal distribution is the simplest to implement as you will need to calculate the mean and
standard deviation for the training data.
P(A|B)=P(B|A)P(A)/P(B)........................................................................................................... (2)
Here,
 P (A|B) is Posterior probability: Probability of hypothesis A on the observed event B.
 P (B|A) is Likelihood probability: Probability of the evidence given that the probability
of a hypothesis is true
 P (A) is prior probability: Probability of hypothesis before observing the evidence.
 P (B) is marginal probability: Probability of evidence.
3.2 Neural Network
The idea of using neural networks for predicting problems was first expressed by Hu in 1964
which was used for weather forecasting [7]. . Neural network or artificial neural network is same
as the human brain; consist of a network of connected nodes neurons. Neurons are the
corresponding unit of neural networks, and they are connected just like neurons and synapse in a
biological brain.

Page | 869
Figure 1: A neural network
4. Data Collection & Data Preprocessing
4.1 Collecting
The most challenging part of the whole process is to collect significant data relevant to the
project. Sorry to say that because of covid we couldn’t work on field that’s why we collect the
dataset from online source. We collect the dataset from kaggle. There are 112458 data of 15
companies of the world.
4.2 Pre-processing
Data cleaning is the process of fixing or removing incorrect, corrupted, incorrectly formatted,
duplicate, or incomplete data within a dataset. It has main 4 steps
 Step 1- Remove duplicate or irrelevant observations.
 Step 2- Fix structural errors.
 Step 3- Handle missing data.
 Step 4- Replacing With Mean/Median/Mode.
5. Result and Discussions
We performed three types of machine learning classifiers. Here the discussion about the
performance of the three machine learning classifiers.
5.1 Neural Network
Here, the train score and the test score are shown where the classifier is used on the X_train and
Y_train data (figure 2).

Figure 2: Train and Test score of neural network.

Page | 870
We can see that the MSE value of train score is 0.19 MSE which is good but the train score is
1.07 MSE which is very high. MSE stands for “Mean Squared Error”. So the ideal value of MSE
is 0. But the value of MSE in neural network test data is high. The value isn’t negligible and not
acceptable for prediction. So we can’t call it an ideal model for our prediction.
5.2 Comparison between Naive Bayes and Linear regression
Following table shows the comparison between linear regression and Naïve Bayes.
Table 2: Comparison
Test_Size Accuracy for Linear regression Accuracy for Naïve Bayes
2% 95.86% 55.56%
5% 94.56% 60.35%
10% 94.50% 57.15%
15% 95.55% 56.65%
20% 94.06% 55.85%
25% 95.66% 54.10%

Figure 3: Comparison between Naive Bayes and Linear regression.


From figure 3 we can see that, linear regression classifier has the highest accuracy level than naive
bayes at every test split. For efficiency comparison we did not include neural network classifier
model because it did not prove to be a good enough for prediction system as it had huge error
level while testing.
We can say that linear regression model is the best suitable approach to predict the future trends
of any stock market as it has the highest accuracy level.
6. Conclusion
In this research, we focused on manipulating and preparing the dataset collected from internet and
tried to track the best classifier to predict the stock market. We use three machine learning
classifiers linear regression, neural network, and naive bayes. We tried to get the best accuracy
for the prediction for our dataset after performing different levels of data engineering. The
percentage of accuracy is 95.86% in linear regression which acceptable and optimal. In this
research we worked with naive bayes machine learning classifier which is a unique work for stock

Page | 871
market prediction. We used GaussianNB to predict the stock market. The highest accuracy is
60.35%. For prediction with Neural Network the Train score 0.19 and Test score 1.07- which is
not acceptable for real life implementation. We can conclude that to predict the business data
from any given data set linear regression is the best choice that offers minimum error and
maximum efficiency.
Acknowledgment
The authors would like to thank all the staffs of AI lab of the department of Information and
Communication Engineering, Bangladesh Army University of Engineering and Technology,
Qadirabad Cantonment, Natore.
References
Ou, P., Wang, H., 2009. Prediction of stock market index movement by ten data mining
techniques. Modern Applied Science 3, 28–42. URL:
https://pdfs.semanticscholar.org/0e6f/f761862c0b8a2217aa298c5d963a387163f9.pdf.
Patel, J., Shah, S., Thakkar, P., Kotecha, K., 2015. Predicting stock and stock price index
movement using trend deterministic data preparation and machine learning techniques.
Expert Systems with Applications 42, 259–268. doi:10.1016/j.eswa.2014.07.040.
Shah, V. H. (2007). Machine learning techniques for stock prediction. Foundations ofMachine
Learning| Spring, 1(1), 6-12.
Gharehchopogh, F. S., & Khalifehlou, Z. A. (2012). A New Approach in Software Cost
Estimation Using Regression Based Classifier. AWERProcedia Information Technology
and Computer Science, Vol: 2, pp. 252-256.
Gharehchopogh, F. S., & Khalifehlou, Z. A. (2012). A New Approach in Software Cost
Estimation Using Regression Based Classifier. AWERProcedia Information Technology
and Computer Science, Vol: 2, pp. 252-256.] Draper, N. R., Smith, H., & Pownell, E.
(1966). Appliedregression analysis (Vol. 3). New York: Wiley.
J. Li, H. Bu, J. Wu. (2017, June). Sentiment-aware stock market prediction: A deep learning
method. In 2017 international conference on service systems and service management (pp.
1-6).IEEE.
M.J.C. Hu, “Application of the Adaline System to Weather Forecasting,” Master Thesis,
Technical Report 6775-1, Stanford Electronic Laboratories, Stanford, CA, June 1964.

Page | 872
7th International Conference on Engineering Research, Innovation and Education
January 12-14, 2023, Sylhet, Bangladesh

PAPER ID: 215

Studying and Analyzing the Cellular Network Signal Strength


Variations for GSM Networks in a University Campus

Md Sabir Hossain, Abu Sajed Noman, Md. Golam Rashed*,


Mirza A. F. M. Rashidul Hasan, and Rubaiyat Yasmin
Department of Information and Communication Engineering,
University of Rajshahi, Rajshahi-6205, Bangladesh
Email: hossainsabir123@gmail.com; nomanice15@gmail.com; golamrashed@ru.ac.bd;
mirza_ice@ru.ac.bd; sheba@ru.ac.bd
* Corresponding author Md. Golam Rashed
Abstract: The distance and direction of the nearest base transceiver station
Keywords: (BTS), transmission power, network congestion, antenna efficiency, and
barriers are all factors that affect signals received by the user equipment (UE)
● Received Signal in cellular communication. The UEs are devices like mobile phones, tablets,
Strength Indicator etc. When signal strength (ST) is measured by UE in various locations which
are far away from the BTS, there exist many obstacles which cause the value
(RSSI); of ST to be less. Thus, it is important to observe and analyze the ST received
by the UE from the BTS to improve the connectivity services. In this study,
● GSM; we observed the GSM ST received by UE and analyzed it to investigate how
ST varies under different scenarios in popular public spaces. In addition, an
● Line of Sight; identical measurement grid is used in different observation areas to
characterize the change in ST under different weather conditions and different
● Radio Waves; population densities. Here, we considered the campus premises of Rajshahi
University (RU) as a testbed to study and analyze the same. We observed and
● University campus. analyzed the received ST by UE from different academic buildings, residential
halls, and other buildings inside RU campus premises under different
scenarios. It is found that ST received by UE at any particular location(s) from
BTSs of different GSM networks inside the RU campus premises varied
significantly which led to varied connectivity performance to initiate different
GSM network-based services. Further, some recommendation is provided on
how to improve the received ST by UEs from BTSs around the RU campus.

1. Introduction and Motivation


The number of mobile phone subscribers has been increasing remarkably over the last decade
(BTRC, 2018). General people have become at least partially dependent on their cell phones for
more than just convenience. Cell phones are now a primary way of communicating and keeping
one’s social life intact. To keep these facilities for the people, cellular mobile operators must
evaluate network performance and service quality because revenue and customer satisfaction
are directly linked to network performance and quality. Mobile network service satisfaction
assessment is required to analyze network performance and maintain service quality
requirements, most notably from the consumer's perspective.

Page | 873
The most mobile operator operates in areas where there is no direct line-of-sight path between the
transmitter and the receiver antennas (Basiran Nesa, 2020). Due to multiple reflections from
various objects, the electromagnetic waves travel along different paths of varying lengths to the
receiver simultaneously (Abraham U. Usman, 2015). The interaction between these waves,
therefore causes multipath fading, resulting in increased path loss as the distance between the
transmitter and receiver increases.
Many researchers have previously focused on various aspects of cellular networks, particularly
service quality and received signal strength. In Bangladesh, Bangladesh Telecommunication
Regulatory Commission (BTRC) occasionally conduct drive test all over the country to observe
the service quality provided by different cellular operators and signal strength received by the
UEs from BTS of different cellular operators. Besides, those types of observations, oftenexisting
cellular mobile operators also conduct drive tests if needed. Very few researches have been carried
out by academicians for the same due to lack of technical supportive equipment. The authors (N.
Nkordeh, 2016) analyzed the signal strength of mobile networks for GSM networks. In their
investigation, drive tests were utilized to measure the signal strength for GSMnetworks. And also
a data collection of received ST by UEs for an urban public university campus in Bangladesh is
carried out (Basiran Nesa, 2020). That work aimed to analyze and provide a better understanding
of the various GSM signal network at Jahangir Nagar University, BD. In our study, we have
considered Rajshahi University campus premises as a testbed to analyze the received ST by the
cellular mobile station-called UE from the provided signal radiated by the BTSs of different
cellular mobile operators. There are 58 departments at RU. This university has 18 residential halls
for students and six residential halls for female students. The university now has more than twelve
hundred teachers, 750 officers, and about 1800 support staff. The academic activities of the
university continue about 9 am to 5 pm which is termed as the peak hour of the day except for the
two days of the week – Friday and Saturday. Thus, it is important to study the quality of network
coverage and received ST around the RU campus premises. The objectives of this study include:
 To compare the received ST at different locations (academic, administration, residential
building) of RU campus from various operators’ BTS, and compare it with referenceST.
 To identify the best received ST with the most optimal performance from any particular
operator (s).
It is hoped that this work is useful for a given idea or clue to improve the performance of various
mobile operators inside RU campus premises.

2. Received ST Measurement by UE
The transmitting power output as received by a reference antenna at a UE and/or cellular data
device is referred to as ST in cellular communication. Received signal strength indicator (RSSI)
is a measurement of the power present in a received radio signal radio (Martin Sauter, 2010).
Received ST is measured in dBm which is the most accurate measurement available. The dBm
stands for decibel milliwatt. The formulae for calculating dBm measurement is 10× log (signal
strength per milliwatts).
The BTS towers emit signal. As the signal travels away from the tower, it degrades or weakens.
The signal is also affected by any obstacles in the way, such as buildings or trees. When the UE

Page | 874
is close to the BTS, the signal is quite strong. By the time it reaches UE inside building or
obstacles, it has a lower dBm. Measured in dBm, a signal of greater than -70 dBm is considered
an excellent signal in all networks. A poor signal will be -100 dBm or worse in 3G networks and
-110 dBm or worse in 4G networks. It is important to take measurements in several areas to
determine where you have the strongest signal strength. The ST measurement scale as
recommended by BTRC, Bangladesh in the city corporation area is given in Table 1.
Table 1: Signal strength measurement scale in 4G LTE
RSSI Signal strength Description
>= -80 dBm Excellent Strong signal with maximum data speeds
-81 dBm to - 105 dBm Good Strong signal with good speeds
Fair but useful, fast and reliable data speeds may be
-106 dBm to - 110 dBm Fair
attained, but marginal data with drop-outs is possible
-111 dBm to - 119 dBm Poor Performance will drop drastically
-120 dBm No signal Disconnection

2.1 Signal Strength Calculation


To calculate the ST received by the antenna of UEs, it's enough to know the RMS voltage or
current because this is coming from a regulated impedance antenna and transmission line. Let's
say we have a 50-ohm antenna setup and the RMS voltage is 2mV. Then the received signal
strength by the UE can be measured as:
2
P =E RMS ⁄R
= (0.002V)2⁄50Ω
= 8.10 − 8W
Typically, this is converted to decibels relative to 1mW, dBm:
PdBm = 10log10(𝑃 × 0.001𝑊)
= 10 log10(8.10 − 8𝑊 × 0.001𝑊)
=10 log10(8.10 − 5)
= −40.97
Therefore, our signal strength is about -41 dBm. Signals that are closer to 0 are more powerful.
The signal strength might vary between-51 dBm and -113 dBm. Stronger values are those that are
closer to -51 dBm. A signal that exceeds -118dBm is deemed insufficient for proper operation
(L.XU, F. Yang, 2011).

3. Methodology
In this study, ST received by the UE at various places inside RU are collected through a survey.
Smartphones as UE are used to gather received ST data from various locations of Rajshahi
University. For instance, academic buildings, residential buildings, and administrative building
are considered interest points to record received ST data. Signals from different mobile operators
such as Grameenphone, Teletalk, Banglalink, and Robi are recorded for our comparative analysis
during the mid of 2022. The collected received STs are under different situations. For example,
data are recorded in different traffic densities (peak hour, non-peak hour).

Page | 875
In general, all Android phone have the “bars” graphic that displays on our home screen. This is
supposed to show what network strength is on Android phone. The problem with the bar’s
representation is that there are no standards. The accuracy of the bars graphic can vary widely
between phone models, devices from different manufacturers, and operating systems. Thus, in
this study, we used our smartphone in Android Field Test Mode to check the ST. Checking the
dBm reading on the Android phone helped us to test the actual ST received in our smartphone.
Finding the dBm menu screen varies across device manufacturers, phone models, and versions of
Android OS. A typical navigation sequence for our used android smartphone is:
Settings > About Phone > Status Information > SIM card status > Signal Strength.
The whole procedure to measure the received ST in our used android smartphone is
illustrated in Figure 1.

Figure 1: Android Field Test Mode to Check ST.


The data were collected by measuring the varying signal strength in a repeated pattern to have
accurate result. Four different BTSs location were picked within the study area. The BTS of
Grameenphone, Teletalk, Airtell/Robi and Banglalink are located at various location inside and
outside of RU campus.

4. Findings and Comparative Analysis


In this study, we have collected signal strength data received by our UEs at various important
locations inside RU campus premises. The locations include all six academic buildings, four
residential halls, an administrative building, a central library, and a railway station within the RU
campus. Data were collected for two times-one is peak time (office time), and non-peak time, at
every location of interest. We have analyzed and compared the collected data and decided on
whether the signal strength is strong enough at a satisfactory level with the reference of signal
strength measurement sale stated in Table 1. The analysis is presented in the preceding
subsections.
4.1 Received Signal Strength Around Academic Buildings of RU

Page | 876
Figure 2 illustrated the ST data received by used UEs for different mobile operators at various
academic buildings of RU. Figure 2. (left) and (right) show the received signal by UEs during
peak time, and non-peak time, respectively. With reference to Figure 2, we can conclude that none
of the mobile operators provide excellent coverage at any academic buildings at RU. Draftly good
mobile coverage was provided by Grameenphone and Teletalk mobile operators inany case (peak,
non-peak hour) within Sir Jagadish Chandra Bose academic building whereas worst coverage was
provided by Banglalink and Airtel at Dr. M. A. Wazed Mia academic building.
Figure 2. Signal strength received by UEs at RU Academic Buildings.

4.2 Signal Strength Around Residential Halls of RU


The bar chart illustrated in Figure 3 shows received signal variation in different important
residential halls during peak hour and non-peak hours. According to the measurement scale,
Grameenphone provides good network coverage around almost all residence halls. During peak
hour, received signal strength is more satisfactory than non-peak hours in all residential halls.
This is because during peak hours most of the residence halls were less crowdy than during non-
peak hours.

Figure 3. Data variants of received signal strength at RU residential halls.


4.3 Signal Strength Around Other Important Locations at RU
Figure 4 illustrated the received ST at UEs around some important locations inside the RU campus
where often there is a gathering of students moderately. It is seen from the figures thatin almost
all considered places, received ST from different operators are good enough during peak and non-
peak hours in general.

Figure 4. Data variants of received signal strength around important location of RU

Page | 877
From the above discussions, we concluded that though in a few places RU UEs received good
signal strength in almost all places UEs did not receive satisfactory signal strength. For these
reasons, inside the RU campus, call drop is high, call success rate is low, and call quality is low.
To get rid of this problem, it is necessary to set up more BTS for every mobile operator inside
the RU that will be approved by BTRC rules and regulations. Setting up multiple operators’
BTS in one place is recommended by BTRC so that people do not have any problems due to
radiation.

5. Conclusion
Radiated ST from BTS is affected by external factors such as location, weather, time of day,
and so on, all of which have a significant impact on the received ST. In public places, ST is
better during non-peak hours than during peak hours. In this study, we revealed from our
collected received ST data by UEs within the RU campus that better STs are available during
non-peak hours than peak hours. Often, excellent or good qualities ST are not available inside
RU premises to provide good quality service to the RU campus community. In order to maintain
good coverage inside the RU campus premises, it is recommended to install more BTSs of
various mobile operators to offer good quality services to the community of RU campus
premises during both peak hours as well as non-peak hours.
References
BTRC, “Public Consultation on Guidelines for Limiting Exposure to Radiation of
Electromagnetic Fields (Up to 300 GHz)” (2018), http://old.btrc.gov.bd/news/public-
consultation-guidelines-limiting-exposure-radiation-electromagnetic-fields-300-ghz.
Accessed July 2022.
Basiran Nesa and M. Mesbahuddin Sarker (2020): “Analysis of Signal Strength Variations for
an Urban Public University Campus in Bangladesh”, International Journal of
Engineering and Technical Research V9(07).
Abraham U. Usman, et al. (2015) “Instantaneous GSM Signal Strength Variation with Weather
and Environmental Factors”, American Journal of Engineering Research (AJER) e-
ISSN : 2320-0847 p-ISSN :2320- 0936 Volume-4, Issue 3, pp-104-115.
Nsikan Nkordeh, et al. (2016), “Analysis of Mobile Networks Signal Strength for GSM
Networks”, Proceedings of the World Congress on Engineering and Computer Science
2016 Vol I WCECS 2016, October 19-21, 2016, San Francisco, USA.
Martin Sauter (2010). "3.7.1 Mobility Management in the Cell-DCH State". From GSMto LTE:
An Introduction to Mobile Networks and Mobile Broadband (eBook). John Wiley &
Sons. p. 160. ISBN 9780470978221. Retrieved 2013-03-24.
L.XU, et al. (2011): Variation of Received Signal Strength in Wireless Sensor Network, 3rd
International Conference on Advanced Computer Control.

Page | 878
7th International Conference on Engineering Research, Innovation and Education
January 12-14, 2023, Sylhet, Bangladesh

PAPER ID: 222

A comparative study of ARIMA and hybrid ARIMA-GARCH models in


forecasting the exchange rate of Bangladesh.
Md Borhan Uddin a*
a
Department of Statistics, Shahjalal University of Science and Technology, Sylhet-3114,
Bangladesh; E-mail: borhan8930@gmail.com
*Corresponding author: Md Borhan Uddin
Abstract: This study aims to establish comparison among ARIMA model and hybrid ARIMA-GARCH models in
terms of their performance to forecast the currency exchange rate of Bangladesh. The study uses the daily data of
exchange rate of Bangladesh Taka (BDT) against U.S. Dollar (USD). The data has been collected from Bangladesh
Bank which is the central bank of Bangladesh. Exchange rate for the period of 1 January, 2020 to 26 July, 2022 have
been used. The period of 1 January, 2020 to 31 May, 2022 was used to train the models and the remaining data were
test dataset which was used to evaluate the forecasting performance of the models. First of all, unit root test showed
us that the exchange rate series was non-stationary. The exchange rate was transformed to exchange rate returns to
make the series stationary. Then, the Box-Jenkins method was employed to determine the appropriate ARIMA
model for the exchange rate returns series. Hybrid ARIMA-SGARCH, ARIMA-GJR, ARIMA-EGARCH and
ARIMA-IGARCH models were estimated to capture the volatility of the return series. The Root Mean Square Error
(RMSE), Mean Absolute Error (MAE) and Mean Error (ME) values were employed to evaluate the forecasting
performance of the models. According to RMSE, MAE and ME, ARMA(1,1)-EGARCH(1,1) is found to be
adequate. This model captures the volatility of the series and its performance to forecast exchange rate is better than
the other models considered.

Keywords: ARIMA, GARCH, Forecast, Volatility, Exchange Rate

1. Introduction
Forecasting many important economic and financial variables has always been highly crucial for effective
policy making. One of the most important economic variables that implies direct influence on
international trade and finance is the exchange rate. Fluctuations in exchange rate may have significant
impact on various macroeconomic variables e.g., interest rates, prices, wages, unemployment rate etc.
That is why, understanding the behavior of exchange rate greatly concerns investors and policy makers.
Fluctuations in exchange rate may happen due to many economic and political factors which are highly
uncertain themselves. ARIMA model has been used to predict exchange rate for decades now (Dunis &
Huang, 2002). Some studies found ARIMA model to be best suited to model the time series of exchange
rate (Nwankwo, 2014).
Though ARIMA model is powerful forecasting tool and has been widely used, it does not perform well
with the data that contains high volatility. Financial variables e.g., usually have high volatile nature. The
autoregressive conditional heteroscedasticity (ARCH) and generalized autoregressive conditional
heteroscedasticity (GARCH) models can be effective to overcome this limitation. Hence, this study
focuses on hybridization of univariate time series model i.e., ARIMA and the volatility model (GARCH
family models). Many studies in various fields have been using ARIMA-GARCH hybrid models for their
better forecasting performance (Wang et al., 2005). GARCH family models are particularly appealing in
financial and economic sectors for their characteristics
This study takes particular interest in Exchange rate of Bangladeshi Taka (BDT) to U.S. Dollar (USD).
This study aims to model high volatile BDT/USD exchange rate series with ARIMA model and GARCH
Page | 879
family models such as SGARCH, EGARCH, IGARCH and GJR. Different model evaluation criteria e.g.,
RMSE, MAE and ME are employed to compare these models and find the best suited model according to
their forecasting performance. This model may contribute to help the policy makers of Bangladesh taking
vital decisions which may have crucial impact on the economy of Bangladesh.

2. Data description
Bangladesh bank, the central bank of Bangladesh, stores daily exchange rate of Bangladeshi Taka (BDT)
to U.S. Dollar (USD). Daily exchange rate of 1 January, 2020 to 26 July, 2022 have been used in this
study. The data from 1 January, 2020 to 31 May, 2022 were used training dataset and the data from 1
June, 2022 to 26 July, 2022 were used as test dataset for evaluating the forecasting performance of the
model.

3. Methodology
In this section, we briefly demonstrate the specifications of the models, conditional distributions as well
as their forecasting criteria which are employed to model the volatility of exchange rate of Bangladeshi
Taka (BDT) to U.S. Dollar (USD).
3.1 The ARIMA model
The model was proposed by Box and Jenkins (Box & Jenkins, 1976). There are three steps to build an
ARIMA model i.e., model identification, parameter estimation and diagnostic checking. Then this model
can be used to forecasts future values. In the first stage of the model building process, we try to examine
the stationarity of the model which is the most important assumption. The series has to have constant
mean and variance. A sharp decline in ACF plot means the series is stationary. A unit root test,
Augmented Dickey-Fuller test formally examines for the stationary. If the series is non-stationary, we
make the series stationary in mean and variance by differencing and Box-Cox transformation
respectively. The model is identified as ARIMA (p, d, q) where p = autoregressive operator, d =
differencing operator, q = moving average operator. This model can be expressed mathematically,

yt  1 yt  1  .....   pyt  p   t   1 t  1  ......   q t  q where εt ~ N(0, σ 2).


t

We use Akaike information criterion (AIC) and Bayesian information criterion (BIC) / Schwarz criterion
(SBC) to select the order of the operators of the model. Then we perform model diagnostics. The
residuals are assumed to be white noise. Financial data is considered to have high volatility, so it is
necessary to test for the ARCH effect by ARCH-LM test.
3.2 The SGARCH models
Auto Regressive Conditional Heteroscedasticity (ARCH) and related models model the volatility itself in
the model (Dhamija & Bhalla, 2010). The Generalized Auto Regressive Conditional Heteroscedasticity
(GARCH) or Standard GARCH (SGARCH) model is the improved version of ARCH model where
conditional variance is considered as ARMA process. The error process:

εt = vt√ℎ𝑡
In a way that 𝜎v2 = 1 and
ℎt = 𝛼0 + ∑ 𝛼i𝜀𝑡̂-𝑖 2+ ∑ 𝛽jℎt-j
Since 𝜈t is a white noise, the key point here is that the conditional variance of 𝜀t is as the following:
𝐸t-1 𝜀t = ℎt

Page | 880
So, the εt conditional variance complies with an ARMA process. Such models are called GARCH (p, q)
where q is the number of moving average (MA) terms and p is the number of autoregressive (AR) terms.

3.3 The IGARCH Model


Integrated Generalized Autoregressive Conditional Heteroscedasticity (IGARCH) is a restricted version
of the GARCH model, where the sum of the persistent parameters sum up to one, and therefore there is a
unit root in the GARCH process. The constraints for an IGARCH (p,q) model can be written:
∑ 𝛼i+ ∑ 𝛽j = 0 𝑎𝑛𝑑 0 < 𝛽j < 1

3.4 The Exponential GARCH (EGARCH) Model


Standard GARCH model has one limitation which is that there is an assumption of the positiveness for
the coefficients. Exponential GARCH model deals with this limitation where this assumption doesn’t
hold.
𝑙𝑛(ℎt) = 𝛼0 + 𝛼1 (𝜀t-1⁄ℎ t-10.5) + 𝜆1 |𝜀t-1⁄ℎ t-10.5| + 𝛽1 ln(ℎt-1)

3.5 The Threshold GARCH (TGARCH) Model / GJR Model


The leverage effect is expressed in the exponential form in the EGARCH model but it is expressed in the
quadratic form in GJR model or TGARCH (p, q).
So the GJRGARCH model is written by:
ℎt = 𝛼0 + ∑ 𝛼i𝜀 𝑡̂-2𝑖 + ∑ 𝛽jℎt-j + ∑ 𝛾i 𝑢t-i 𝜀 𝑡-̂ 2𝑖
Three evaluation measures have been used in this study namely Mean Error (ME), Mean Absolute Error
(MAE) and Root Mean Square Error to evaluate the performance of the models.

4. Result and Discussions


First, we observe time series plot of exchange rate to get an idea about the series (Figure 1 (a)). The latter
part of the series shows very sharp upward movement of the curve which implies USD values recently
increased rapidly against BDT. Autocorrelation plot (Figure 1 (b)) shows a slow decay which suggests
that the series is non-stationary. To confirm the case, a formal Augmented Dickey-Fuller test is
performed. ADF test also suggests that the series is not stationary with p-value = 0.99.

(a) (b)
Figure 1: Time series plot (a) and ACF plot (b) of the daily data of BDT to USD exchange rate for the period of 1
January, 2020 to 26 July, 2022.
We calculate the exchange rate returns as follows to make the series stationary:
𝐸𝑋𝑅 = log(𝐸𝑋𝑡) - log(𝐸𝑋𝑡-1) = log(𝐸𝑋𝑡⁄𝐸𝑋𝑡-1)

Page | 881
Time series plot of the transformed data is shown in Figure 2 (a). The plot displays that the mean of the
series is constant over time. But the variance is heavily fluctuated, specially in the latter part of the series.
This high variance allows us to apply the ARCH family models. However, ACF plot (Figure 2(b)) and
formal ADF test confirm the stationarity of the data with p-value = 0.01.

(a) (b)
Figure 2: Time series plot (a) and ACF plot (b) the daily data of BDT to USD exchange rate returns for the period
of 1 January, 2020 to 26 July, 2022.

4.1 Model fitting and forecasting using ARIMA model


At first, a univariate model is developed to forecast exchange rate of BDT to USD using Box-Jenkins
approach. Then, Akaike information criterion (AIC) and Bayesian information criterion (BIC) is
employed to select the best model among them.

Table 1: AIC and BIC values for corresponding selected ARIMA models.
ARIMA (p, d, q) AIC BIC
(1, 0, 1) -12102.2 -12083.07
(2, 0, 1) -12100.2 -12076.29
(3, 0, 1) -12098.21 -12069.51
(1, 0, 2) -12100.2 -12076.29
(2, 0, 2) -12098.2 -12069.51
(3, 0, 2) -12096.21 -12062.73
(1, 0, 3) -12098.21 -12069.52
(2, 0, 3) -12096.21 -12062.73
(3, 0, 3) -12094.21 -12055.95
ARIMA (1, 0, 1) model has AIC value of -12102.2 and BIC value of -12083.07. ARIMA (1, 0, 1) model
is the most appropriate model among all the selected models as it has the lowest value of AIC and BIC
(Table 1). The estimated model is,
EXRt = 0.0072 EXRt-1 + 0.0071 εt-1
Model diagnostics are initiated after fitting the model. There is no autocorrelation found in the model’s
residuals (Figure 3). This model is used to forecast exchange rate of Bangladesh to USD from 1 June,
2022 to 26 July, 2022. These forecasted values are used to evaluate the forecasting performance of the
model on the test dataset.

Page | 882
Figure 3: Model diagnostics of ARIMA (1, 0, 1).
The result of Lagrange multiplier test (ARCH-LM) indicates that the variance of the residuals is not
constant that means, there is an evidence of the presence of ARCH effect (p-value: 0.002195323). Hence,
we can apply the GARCH family models to capture the volatility of the exchange rate returns series.

4.2 ARIMA-SGARCH model


ARIMA-SGARCH model is applied to the data to capture the volatility. The estimated ARIMA-sGARCH
(1,1) model is as follows:
Mean equation:
EXRt = 0.007384 EXRt-1 + 0.007352 εt-1 + εt
Variance equation:
ht = 0.050006 εt-12 + 0.899980 ht-1
AIC and BIC values of the model are -19.457 and -19.419 respectively.

4.3 ARIMA-IGARCH model


The joint estimation of ARIMA-IGARCH (1,1) model gives the following
Mean equation:
EXRt = 0.007575 EXRt-1 + 0.007540 εt-1 + εt
Variance equation:
ht = 0.050027 εt-12 + 0.949973 ht-1
AIC and BIC values are -18.068 and -18.035 respectively.

4.4 ARIMA-GJR model


Glosten, Jagannathan and Runkle (ARIMA-GJRGARCH (1,1)) Model are presented followingly:
Mean equation:
EXRt = 0.007494 EXRt-1 + 0.007460 εt-1 + εt
Variance equation:
ht = 0.049997 εt-12 + 0.899941 ht-1 + 0.050122 v t-1ε t-12
AIC value for the model is -19.406 and BIC value for the model is -19.363.

4.5 ARIMA-EGARCH model


Jointly estimated ARIMA-EGARCH (1,1) model is as follows:
Mean equation:
EXRt = 0.013369 EXRt-1 + -0.015544 εt-1 + εt
Variance equation:

Page | 883
𝑙𝑛(ℎt) = -2.127157 + -0.018241 t-1 ⁄ℎ 0.5) + 0.018710 |𝜀t-1⁄ℎt-10.5| + 0.933090 ln(ℎt-1)
(𝜀 t-1
AIC and BIC values for the model are -21.737 and -21.693 respectively.

4.6 Model Comparison


Root mean square error (RMSE), Mean absolute error (MAE) and Mean error (ME)are
employed toevaluate the forecasting performance of the model. Comparative RMSE, MAE
and ME values are presented in the Table 2 for the models applied to the exchange rate
returns data.
Table 2: RMSE, MAE and ME values for each model applied to the data
Model RMSE MAE ME
ARIMA 0.001288233 0.0005731449 0.0004550209
ARIMA-SGARCH 0.001296631 0.0005733631 0.0004782215
ARIMA-EGARCH 0.001276251* 0.0005721422* 0.0004188912*
ARIMA-IGARCH 0.001296541 0.0005733724 0.0004779335
ARIMA-GJR 0.001296632 0.0005733844 0.0004781999
* minimum value
According to the result presented in Table 6, the best values for RMSE,MAE and ME are
achieved fromARIMA-EGARCH model which are 0.001276251, 0.0005721422 and
0.0004188912 respectively.
Therefore, ARIMA-EGARCH model captures the volatility of the exchange rate returns series better
thanany other model applied in this study.
5. Conclusion
This study focuses to model the Bangladeshi taka (BDT) to U.S. Dollar (USD) exchange rate
series andmake reliable forecasts with the model. While examining the series, it appears to be
non-stationary.
Exchange rate returns are calculated to make it stationary. Then, Univariate ARIMA model and
hybridARIMA-GARCH family models like ARIMA-SGARCH, ARIMA-EGARCH, ARIMA-
IGARCH,
ARIMA-GJR models were applied to the exchange rate series. According to the three model
evaluation measures (RMSE, MAE and ME), ARIMA-EGARCH model captures the high
volatility of exchange ratereturns better than any other models considered. This model is adequate
to fit the data and has fairly reliable forecasting performance. In further research, the models which
are applied to BDT/USD exchange rate, can also be applied to any other exchange rate as well.

Reference
Box, G. E. P., & Jenkins, G. M. (1976). Time series analysis. Forecasting and control. Tsaf.
https://ui.adsabs.harvard.edu/abs/1976tsaf.conf B/abstract
Dhamija, A. K., & Bhalla, V. K. (2010). Financial Time Series Forecasting : Comparison of various
architectures of Neural Networks and ARCH models ajay dhamija Financial Time Series
Forecasting: Comparison of Neural Networks and ARCH Models. International Research
Journal of Finance and Economics, 49. http://www.eurojournals.com/finance.htm
Dunis, C. L., & Huang, X. (2002). Forecasting and trading currency volatility: an application of
recurrent neural regression and model combination. Journal of Forecasting, 21(5), 317–354.
https://doi.org/10.1002/for.833
Nwankwo, S. C. (2014). Autoregressive Integrated Moving Average (ARIMA) Model for Exchange
Rate (Naira to Dollar). Academic Journal of Interdisciplinary Studies.
https://doi.org/10.5901/ajis 2014.v3n4p429
Wang, W., van Gelder, P. H. A. J. M., Vrijling, J. K., & Ma, J. (2005). Testing and modelling
autoregressive conditional heteroskedasticity of streamflow processes. Nonlinear
Processes in Geophysics, 12(1), 55–66. https://doi.org/10.5194/npg-12-55-2005

Page | 884
7th International Conference on Engineering Research, Innovation and Education
January 12-14, 2023, Sylhet, Bangladesh

PAPER ID: 223

Finite Element Analysis of Different Coil Topologies for


Wireless Power Transfer Based Applications

Ifte Khairul Amina*, Md. Mahfuzur Rahman Chyb, Md. Nahid Islamc,
Azam Jamand and Md. Atiqur Rahaman Shantoe
Department of Electrical and Electronic Engineering, Shahjalal University of Science
& Technology, Sylhet-3114, Bangladesh.
Email: aiftekhar-eee@sust.edu, bmahfuzur22@student.sust.edu,
cnahid39@student.sust.edu, dazam02@student.sust.edu and
eatiqur.rahaman.shanto.bd@gmail.com.

* Corresponding author : Ifte Khairul Amin

Keywords: Abstract: Over recent years, Wireless Power Transfer (WPT) has gained
massive interest in the fields of autonomous operation and dynamic charging
● Magnetic with greater flexibility, reliability, and robustness. One of the major
Resonance (MR);
challenges in wireless power transfer is the loss of power with the increase of
● Wireless Power axial separation between the transmitter and receiver coil. Over the last
Transfer (WPT); decade, a number of researches have been conducted to increase the
● Finite Element efficiency and range of power transfer with no physical link. Coil design plays
Analysis (FEA); an important role when power transfer is required within a short distance. In
this study, a finite element analysis (FEA) will be conducted on different 3D
● Power transfer coil designs such as circular, square, octagonal, and helical to
efficiency (PTE);
find the optimized coil dimension and spacing between transmitter and
● Coupling receiver. Several coil parameters, such as mutual inductance, coupling
coefficient; coefficient, and transformation efficiency have been analyzed in simulation to
● Leakage flux; guarantee overall optimal power output. The design and simulations are
performed in Ansys Maxwell platform, and the optimal coil design is
● Wireless charging. proposed by FEA based on the performance of the magnetic resonance-based
WPT structure targeting higher PTE and reduced leakage flux.

1. Introduction
Wireless Power Transfer (WPT) is a promising and rapidly evolving technology with broad
commercialization potential in the fields of electronic devices, healthcare, automotive, and mobile
cellular communication. Despite being one of the most active research topics, the historyof WPT
technology is over 100 years old, as proposed by Sir Nikola Tesla in the late 1800s. There have
been a number of significant advancements in the WPT domain in recent years, credit goes to the
University of Auckland, Massachusetts Institute of Technology, University of Tokyo, Intel, and
others. The majority of researchers concentrate on near-field WPT due to the

Page | 885
wide range of applications available today, such as electric vehicles (EV), implantable medical
devices (IMD), smartphones, watches, even in mini-scale power transmission.
Many such WPT technologies have already been used, including Inductive WPT, Magnetic
Resonance WPT, Capacitive WPT, Strongly Coupled Magnetic Resonance WPT, Microwave
Power Transfer, and Optical WPT, each with distinct advantages and disadvantages. However,
these WPT technologies face difficulty due to the inherent characteristics, such as high magnetic
leakage, low power transfer efficiency and lower misalignment tolerance at increased axial
separation between the transmitter and receiver coils. Such WPT inconveniences can be reduced
by using an optimized coil design with a strong magnetic coupling arrangement which ensures
maximized power transfer with increasing distance between TX and RX [1].
In this study, a Finite Element Analysis of typical 3D coil designs such as circular, square, helical,
and octagonal shape has been performed emphasizing on wireless charging of EV applications.
The coils were designed using the FEA tool called Ansys Maxwell 3D, and the performance of
various coil shapes were compared [2]. To enhance the magnetic coupling between the coils, the
dimension of the coil was optimized alongside using power pads of ferrite core which increases
the coupling coefficient, mutual inductance, power transfer efficiency and misalignment
tolerance. Finally, the performance analysis of the proposed coil designs reveals the ideal coil
structure for implementation in dynamic wireless charging of electric vehicles (EV) [1-6].
2. Methodology
2.1 Magnetic Resonance Wireless Power Transfer Model

Figure 1: Equivalent Circuit of Magnetic Resonance WPT

In this study, the magnetic resonance based WPT model has been utilized to operate under
resonant conditions. The mono-resonant compensation topology employed is the series-series
compensation topology, which is safe, convenient, and reliable to use. The circuit diagram of
the compensation network in Fig. 1 reveals the magnetic linking between the transmitting and
receiving side with no physical or electrical contact [1]. The inductive, capacitive, and resistive
element in the compensation network refers to the self-inductance, capacitance and the ohmic
resistance of the coil, respectively. The self-inductance and ohmic resistance depend upon the
material and dimensions of the coil and the capacitance can be calculated using the equation:
1
CTX = 2
(1)
(2 f) LTx
1
CRX = (2 (2)
f)2LRX

Page | 886
Where, LTX and LRX are the self-inductance of transmitter and receiver, respectively and f is the
resonant frequency.
As transmitter coil is supplied with alternating current excitation, an alternating magnetic field
is created which links with the receiver coil positioned at its vicinity. Due to rate of change of
magnetic flux linkage, emf is induced in the receiver coil owing to Faraday’s law of
electromagnetic induction [1,2]. The governing equations related to electromagnetic induction are
given below: dф d
C = − RX,TX = − ∬ B ⃗⃗⃗⃗⃗⃗⃗→. dA
⃗⃗⃗⃗⃗⃗⃗→ (3)
RX,TX dt dt RX TX RX

Where, CRX,TX is the induced emf, фRX,TX is the magnetic flux linking receiver coil due to current
⃗⃗⃗⃗T⃗⃗⃗X→ is the magnetic flux density of transmitter coil and A
excitation in transmitter coil, B ⃗⃗⃗⃗2→ is the
cross-sectional area of receiver coil.
The rate of change of magnetic flux dфRX,TX can be written as follows:
dt
dфRX,TX d ITX
=M (4)

dt dt

Where, the constant of proportionality M is the mutual inductance between the coils.
фRX,TX ∬RX ⃗B⃗⃗⃗T
⃗⃗⃗X
⃗→.d⃗A⃗⃗⃗R
⃗⃗⃗X
⃗→
M= = (5)

ITX ITX

In order to calculate the maximum power transfer efficiency, two important parameters known as
Quality factor, Q and Coupling coefficient, k must be calculated. The equations related to Q and
k are:
2 fL
Q= R
(6)
M
and k = (7)
√LTXLRX

Hence, the maximum power transfer efficiency (PTEmax) is [6]:


k2QTXQRX
PTEmax = 2 (8)
(1+ √k2QTXQRX)

Page | 887
2.2 Finite Element Analysis Model
The FEA-based Ansys Maxwell 3D physical model of four different coil topologies is presentedin this
subsection. For a legitimate comparison, each coil has been designed using 6m long, 1mm radius copper
wire. Maximum mesh element size was set at 50mm, 100mm and 250mm for coil, bounded region, and
core, respectively. The bounded region was filled with air and sized accordingly to satisfy the boundary
conditions at all instances of magnetostatic simulation. Figure 2 depicts the proposed circular, square,
helical, and octagonal 3-D models [1-3].
A current excitation of 5A was set for TX while the RX was left unexcited. Selecting the iterative solving
method, the percentage error allowance was set to 2% for circular, square, and octagonal shapes and to
10% for just helical shape, with a total of 10 passes for all the coil models. Losses due to skin effect is
minimized by using stranded copper wire. Precisionmodeling, selection of material and application of
PT are to be specified before designing and loss calculations such as eddy current loss and dielectric
loss will be done in future to measure precise efficiency. Table 1 provides the summary of the
dimensions and design characteristics of the proposed models.

(a) (b) (c) (d)


Figure 2: 3-D model of WPT TX-RX coils (a) Circular, (b) Square, (c) Octagonal and (d) Helical Design

Table 1: Coil Design Parameters


Parameters Circular Square Octagonal Helical
Outer Radius (mm) 65 69 69 89
Pitch (mm) 0 0 0 4.3
No. of Turns 23 25 24 11
Start Helix Radius (mm) 18 18 18 87
Radius Change (mm) 2.04 2.04 2.04 0
Core Thickness (mm) 1 1 1 50
Core diameter (mm) 148 148 140 160

3. Result and Discussions


3.1 The influence of ferrite core on coil performance
The coil performance can be significantly improved by using power pads of ferrite core [3]. Use of ferrite
core between the coils enhances the magnetic coupling and reduces the overall leakageflux by directing
the flux from TX to RX. The mutual inductance between the TX and RX for all coil designs has
increased around 4-5 times since the introduction of the ferrite core power pad, as illustrated in Figs. 3a
and 3b. Similarly, as shown in Figs. 3c and 3d, the coupling coefficient has increased more than 1.5
times in all cases. As the vertical axial separation between the coils increases, both mutual inductance
and coupling coefficient exhibit an exponential decrease profile. Even though circular, octagonal, and
square shapes have much higher mutual inductance and coupling coefficients within 75mm separation
between TX and RX, helical coil outperforms the rest at increased axial distance, as highlighted in the
inset of Figs. 3a-3d.

Page | 888
(a) (b) (c) (d)
Figure 3: Variation of inductance: (a) with ferrite core (b) with air core; Variation of coupling coefficient: (c)with
ferrite core (d) with air core.

The core losses such as hysteresis loss and dynamic magnetic losses can become significant at high
frequency applications. The hysteresis loss of the ferrite core used in FEA modeling can estimated
by the equation (9):
𝑃𝑐𝑜𝑟𝑒 = 𝐶𝑚ƒ𝛼𝐵𝛽 𝑚𝑎𝑥 (9)

Where 𝐶𝑚 = . 1 , α = 1.418, β = 2.755 are called


Steinmetz exponents that depends on core properties
while ƒ = 1 MHz and 𝐵𝑚𝑎𝑥 are operating frequency and
maximum magnetic flux density, respectively. The
hysteresis loss calculated with respect to frequency for
each coil-core model is shown in Fig. 4 [4].
Figure 4: Hysteresis loss vs Frequency

3.2 Effect of axial and angular misalignment on coil performance


Coil misalignment is a common problem that causes significant loss of flux in WPT systems, making
them less efficient. As a result, the effect of misalignment has been investigated in this study with the
goal of developing a misalignment tolerant coil model. The vertical axial separation was maintained
fixed at 75 mm between the TX and RX. To analyze the axial misalignment, the receiver was swept at
a range of 0-140mm with 25mm step size. The resultsin Figs. 5a and 5b reveal that helical coil has
much higher mutual inductance and coupling coefficient than any other coil design at higher horizontal
misalignment. For angular misalignment, the receiver coil was rotated from 0 degree to 20 degree with
a step size of 5 degree. Figs. 5c and 5d demonstrate the performance of coil with angular misalignment.
In this case, the circular and octagonal coil designs show better angular misalignment tolerance than
helical and square shape [5].

(a) (b) (c) (d)


Figure 5: Effect of axial misalignment on: (a) mutual inductance (b) coupling coefficientEffect of
angular misalignment on: (c) mutual inductance (d) coupling coefficient.
3.3 Analysis of Power Transfer Efficiency (PTE)
In this subsection, the power transfer efficiency of all the coil designs have been investigated through
parametric sweep analyses between the axial distance between TX and RX and the operating source
frequency.

Page | 889
Using the series-series topology, the voltage source (Esource) was set 320V, frequency was 1MHz, ohmic
resistance (RTX and RRX) was calculated as 0.001Ω, inductance was the self-inductance (LTX and LRX)
of the respective coil designs, load resistance (Rload) was chosen to be 20Ω and the compensation
capacitance (CTX and CRX) has been calculated using Equation 1 and 2. The three-dimensional surface
plots obtained from the above analysis is depicted in Figs. 6a-6d below.

(a) (b) (c) (d)


Figure 6: 3-D PTE surface plot of (a) Circular, (b) Square, (c) Octagonal and (d) Helical Design

From the Fig. 6 we can observe that, the PTE is maximized at the source frequency of 1 MHz exhibiting
the resonance effect and the splitting of peaks when TX and RX is relatively close is the Frequency
Splitting Phenomenon (FSP), which is dominant in over-coupled region. The coil power transfer
efficiencies are more than 99% for circular, helical, and octagonal shape and more than 90% for square
shape at resonant frequency of 1 MHz with tuned compensation capacitance of 363.81 pF, 499.62 pF,
339.30 pF and 360.34 pF, respectively [6].

4. Conclusion
A comparative study of finite element modeling for 4 different coil shapes such as circular, helical,
octagonal, and square has been analyzed in this paper. The effectiveness of the coil designs has been
investigated for various wireless power transfer-based applications. The performance of the coil models
deemed suitable for utilizing in WPT systems showing high mutual inductance, coupling coefficient and
very high PTE at resonant frequency. The helical coil is found to outperform the other proposed models
at higher axial separation with better misalignment tolerance and hence, can be used efficiently in
dynamic wireless charging systemsfor electric vehicle. The coil dimensions can be scaled accordingly
for various applications of WPT systems based on the required transmission level of power and
directionality.
References
M. Zellagui (Ed.), 2021.Wireless Power Transfer – Recent Development, Applications and New
Perspectives. London, United Kingdom, IntechOpen.
Bouanou T, El Fadil H, Lassioui A, Assaddiki O, Njili S., Analysis of Coil Parameters and Comparison
of Circular, Rectangular, and Hexagonal Coils Used in WPT System forElectric Vehicle Charging,
World Electric Vehicle Journal. 2021, 12(1), 45
Elnail KE, Huang X, Xiao C, Tan L, Haozhe X. Core structure and electromagnetic field evaluation in
wpt systems for charging electric vehicles, Energies. 2018, 11(7), 1734.
M. Mohammad, S. Choi, Z. Islam, S. Kwak and J. Baek. Core Design and Optimization for Better
Misalignment Tolerance and Higher Range of Wireless Charging of PHEV," in IEEE Transactions
on Transportation Electrification, vol. 3, no. 2, pp. 445-453, June 2017.
Wang, J., Li, J., Ho, S.L., Fu, W.N., Li, Y., Yu, H. and Sun, M. Lateral and angular misalignments
analysis of a new PCB circular spiral resonant wireless charger, IEEE Transactions on Magnetics.
2012, 48(11); 4522-4525.
Aboualalaa, M., Elsadek, H., & Pokharel, R. K. (2021). WPT, Recent Techniques for Improving System
Efficiency. IntechOpen.

Page | 890
7th International Conference on Engineering Research, Innovation and Education
January 12-14, 2023, Sylhet, Bangladesh

PAPER ID: 226

Modeling Monthly Rainfall in Noakhali Coastal region of Bangladesh:


A Box-Jenkins Approach
Md Borhan Uddin a*
a
Department of Statistics, Shahjalal University of Science and Technology, Sylhet-3114,
Bangladesh; E-mail: borhan8930@gmail.com
*Corresponding author: Md Borhan Uddin
Abstract: Assessing the impact of climate change on rainfall variability draws a great concern around the world.
Time series analysis has been an important tool for modeling different meteorological phenomena e.g., rainfall. One
of the most effective approaches for analyzing time series data is Box-Jenkins approach. This study aims to model the
monthly rainfall in Noakhali coastal region of Bangladesh. The monthly data of rainfall at M. Court, Noakhali station
were collected from Bangladesh Meteorological Department (BMD) for the period from January, 1980 to April, 2015.
The period of January, 1980 to December, 2010 had been used to train the model and the remaining data were test
data which were used to evaluate the forecasting performance of the model. Identification, estimation and diagnostic
checking; these three steps recommended in Box-Jenkins approach were followed in the development of the model.
The model is selected by Akaike information criterion (AIC). After completing the analysis, ARIMA (0,0,1) (0,1,3)
[12] model was found to be adequate to forecast the rainfall in this particular region. This forecasting model will help
the decision maker in flood prediction, agriculture as well as making different plans based on water demands which
is vital in this coastal region.
Keywords: ARIMA, Box-Jenkins, Forecast, Modeling, Rainfall
1. Introduction
Climate change has become one of the alarming issues for today’s world and has more cumulative
negative impact than any other issues currently (Hengade et al., 2018). The impact is worse on developing
countries due to the lack of proper adaptation as well as timely preparedness. Climate change refers to the
long-term impact on meteorological phenomena and the rainfall is very significant among them. The
forecast of rainfall has particular relevance in agricultural sector which has tremendous impact on
economy. Moreover, the impact of climate change on rainfall causes unprecedented fluctuations that may
result in drought or flood like situations. Therefore, the modeling of rainfall and forecasting its future
values is of great interest.
The methods used for rainfall forecasting are broadly divided into two categories. One is dynamic
approach which is mainly physics based e.g., numerical weather prediction model. Another one is
empirical e.g., Regression, Autoregressive Integrated Moving Average (ARIMA), Artificial Neural
Network (ANN) etc. The study concentrates on employing ARIMA to model monthly rainfall data. These
models were proposed by Box and Jenkins (Box & Jenkins, 1976). The Box-Jenkins approach can build a
model only depending on the past values of the variable e.g., rainfall. It also extends its scope to
accommodate any other correlated variables into the modeling if there appears any cause for this
correlation. ARIMA models are based on three operators i.e., autoregressive operator (AR), moving
average operator (MA) and integration part (I) which is regarded as differencing operator (Valipour et al.,
2013). Different studies concentrated on forecasting rainfall in different regions of Bangladesh (Akhter &
Mahsin, n.d.; (PDF) Forecasting Monthly Precipitation in Sylhet City Using ARIMA Model, n.d.). But there
has been no study to model rainfall in this particular region. Agriculture heavily depends upon rainfall as
ground water harvesting is not feasible in every area of this coastal region. Rain water could possibly be a
sustainable alternative. Floods are more frequent, mostly in Meghna river basin areas. Hence, forecasting
rainfall in Noakhali region holds a particular significance. Therefore, this study focuses on developing a
model which will effectively and reliably forecast monthly rainfall in Noakhali coastal region.

Page | 891
2. Materials and Methodology
2.1 Study Location
Noakhali is a district in the southern coastal region of Bangladesh. The latitude & longitude of Noakhali
are 22.75° and 91.17° respectively. The climate of Noakhali is similar to the climate of typical tropical
coastal area. July has the highest precipitation which has an average of 427mm and January has the lowest
with an average of 5mm.
2.2 Data Collection and Missing Data Handle
Bangladesh Meteorological Department (BMD) has 35 weather stations all over Bangladesh to collect
rainfall data. This study uses rainfall data collected in M.court, Noakhali station. This station covers
Noakhali district and nearby areas. Monthly rainfall data from January, 1980 to April, 2015 has been used
in this study. The data from January, 1980 to December, 2010 has been used to develop the model and
rest of the data has been used as test data for measuring forecasting accuracy of the model. The data has
some missing values which have been calculated by linear interpolation technique.
2.3 Model Development
There are three main stages to build an ARIMA model. The stages are model specification
(identification), model fitting and model diagnostics to examine the appropriateness of the model. And
finally forecasting the future values with the model.
The first stage while building an ARIMA model is to identify the model. The most vital and common
assumption in time series analysis is stationarity. That means the series has constant mean and variance. If
the series is non-stationary, it is needed to make it stationary in mean and variance by differencing and
Box-cox transformation respectively. A formal Augmented Dickey-Fuller test can be performed to
confirm stationarity which is a unit root test for stationarity.
The general model introduced by Box and Jenkins has autoregressive (‘p’), moving average (‘q’) and
differencing (‘d’) parameters. The model is expressed in the form of ARIMA (p, d, q) (P, D, Q). The
latter operators are corresponding seasonal parameters.
The general form of above model describing the current value Xt defined by its past values is:
(1 B)(1 1B12)(1 B)(1 B12) Xt  (1 1B)(1  1B12) t (1)

Here, (1 B) = non seasonal autoregressive of order 1; (1 1B12) = seasonal autoregressive of order

1; Xt = The current value of the time series; B = The backshift operator; BXt = Xt-1 and B12Xt= Xt-12;
(1 B) = 1st order non seasonal difference; (1 B12) = 1st order seasonal difference; (1  1B) = non
seasonal moving average of order 1; (1  1B12) = seasonal moving average of order 1;  t = Error term.
We choose the order of the parameters by Akaike information criterion (AIC). After fitting the model, the
last stage of building an ARIMA model is model diagnostics. We need to check whether the residuals are
white noise or not. A version of portmanteau test, Box-Ljung test which is an improved version of Box-
Pierce test is performed to check if the residuals are independent (LJUNG & BOX, 1978). If the residuals
are white noise, we can say that the model fits our data significantly well and this model is adequate for
our data and future forecasts. Two evaluation measures have been used in this study namely Mean
Absolute Error (MAE) and Root Mean Square Error to evaluate the performance of the model on the test
data.
3. Result and Discussions
The first step of Box-Jenkins approach is to check whether the observed data is stationary or not. It is also
important to identify if there exists seasonality in the data. Time series plot (Figure 1)

Page | 892
Figure 1: Time series plot of the observed rainfall data in Noakhali.
shows us that the observed data is not stationary and has clear seasonal pattern of one year (12 months).
To confirm the above notion, ACF (Autocorrelation Function) plots of every lag as well as multiples of
lag 12 have been observed. The autocorrelation functions are significant at identifying stationarity in the
data set. ACF plot of lag 12 (Figure 2 (b)) shows slow decay which is a clear sign of non-stationarity.
Hence, it is concluded that data is not stationary.
It is necessary to meet stationarity assumption while building an ARIMA model. The data needs to be
stationary in both mean and variance. Box-cox transformation is performed to make data stationary in
variance. Box-cox transformation is performed with lambda (λ) value which is determined at 1.395462.
Seasonal first difference (D = 1) is taken to attain mean stationarity in the data. The ACF and PACF plots
(Figure 3) show that first order seasonal differencing is adequate to achieve stationarity. Augmented
Dickey-Fuller Test is performed to confirm stationarity after differencing. ADF test suggests that the data
has become stationary with p-value = 0.01. A preliminary ARIMA (p, 0, q) (P, 1, Q) [12] is selected.

(a) (b)
Figure 2: ACF plot of observed rainfall data (a) and multiples of lag 12 (b).
The order of p, P and q, Q need to be identified. After careful observation of ACF and PACF, a series of
ARIMA models with different combinations of operator orders have been chosen. ARIMA (0,0,1) (0,1,3)

Page | 893
[12] is selected among them based on lowest Akaike information criterion (AIC)

(a) (b)
Figure 3: ACF (a) and PACF (b) plot for Box-cox transformed and first order seasonal differenced data.
Table 1: Akaike information criterion for the corresponding selected models.

Model Akaike information Model Akaike information


criterion criterion
(0, 0, 1) (0, 1, 1) [12] 2428.091 (2, 0, 3) (1, 1, 2) [12] 2328.326
(0, 0, 2) (0, 1, 1) [12] 2429.23 (1, 0, 1) (1, 1, 3) [12] 2324.667
(0, 0, 3) (0, 1, 1) [12] 2430.714 (1, 0, 2) (1, 1, 3) [12] 2325.981
(0, 0, 1) (0, 1, 2) [12] 2333.241 (1, 0, 3) (1, 1, 3) [12] 2328
(0, 0, 2) (0, 1, 2) [12] 2333.841 (1, 0, 1) (2, 1, 1) [12] 2350.786
(0, 0, 3) (0, 1, 2) [12] 2335.828 (1, 0, 2) (2, 1, 1) [12] 2352.541
(0, 0, 1) (0, 1, 3) [12] 2323.754 (1, 0, 3) (2, 1, 1) [12] 2353.493
(0, 0, 2) (0, 1, 3) [12] 2324.373 (2, 0, 3) (2, 1, 2) [12] 2329.418
(0, 0, 3) (0, 1, 3) [12] 2326.367 (1, 0, 1) (2, 1, 3) [12] 2328.012
(1, 0, 1) (1, 1, 1) [12] 2374.414 (1, 0, 2) (2, 1, 3) [12] 2329.993
(2, 0, 1) (1, 1, 2) [12] 2324.551 (1, 0, 3) (2, 1, 3) [12] 2330.265
(2, 0, 2) (1, 1, 2) [12] 2326.298
Once the model has been fitted to the data, number of diagnostic checks are initialized to check how well
the model fits to the data. The residuals are supposed to be uncorrelated with constant variance if the
model fits the data well. Another assumption is residuals are normally distributed.
Histogram of residuals (Figure 4) demonstrates that the residuals are approximately normally distributed.
ACF and PACF plots (Figure 5) show that residuals are uncorrelated. Residuals also have approximately
constant variance.

Page | 894
Figure 4: Distribution of Residuals.

Figure 5: Residual analysis of the data.


Furthermore, to test the independence of the residual, Box-Ljung test is performed. Box-Ljung test
significantly demonstrates that the residuals are independent with p-value = 0.9887.
The model ARIMA (0,0,1) (0,1,3) [12] provides significant result. This model is employed to forecast the
monthly rainfall data from January 2011 to April 2015. This data is used to compare observed and
forecasted values.
Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) values which are 2.303254 and
3.337803 respectively, show that the model fits the data significantly well. Therefore, this model will
produce reliable forecasts of the rainfall data.
Forecasted time series is plotted in Figure 6.
A limitation of this study is that though this study uses long term monthly rainfall data, it only uses data
up to April 2015 due to the unavailability of the data.

Page | 895
Figure 6: Forecasted time series for monthly rainfall data in Noakhali.

4. Conclusion
Time series analysis has been an important statistical tool for modeling and forecasting. ARIMA (0,0,1)
(0,1,3) [12] model has been shown adequate to fit the monthly rainfall data of Noakhali and forecast future
rainfalls. This can help relevant authorities to establish strategies and ensure efficient use of water resources.
Like any statistical model, this model will not forecast future rainfall perfectly. Therefore, thereis a scope to
further research which will focus on improving forecasting accuracy of the rainfall of Noakhali.
References
Akhter, Y., & Mahsin, Md. (n.d.). Modeling Rainfall in Dhaka Division of Bangladesh Using Time Series
Analysis. Retrieved October 2, 2022, from
https://www.academia.edu/26073441/Modeling_Rainfall_in_Dhaka_Division_of_Bangladesh_Usin
g_Time_Series_Analysis
Box, G. E. P., & Jenkins, G. M. (1976). Time series analysis. Forecasting and control. Tsaf.
https://ui.adsabs.harvard.edu/abs/1976tsaf.conf . B/abstract
Hengade, N., Eldho, T. I., & Ghosh, S. (2018). Climate change impact assessment of a river basin usingCMIP5
climate models and the VIC hydrological model. Hydrological Sciences Journal, 63(4), 596–614.
https://doi.org/10.1080/02626667.2018.1441531
LJUNG, G. M., & BOX, G. E. P. (1978). On a measure of lack of fit in time series models. Biometrika,
65(2), 297–303. https://doi.org/10.1093/biomet/65.2.297
(PDF) Forecasting Monthly Precipitation in Sylhet City Using ARIMA Model. (n.d.). Retrieved October 2,
2022, from
https://www.researchgate.net/publication/272744442_Forecasting_Monthly_Precipitation_in_Sylhet
_City_Using_ARIMA_Model
Valipour, M., Banihabib, M. E., & Behbahani, S. M. R. (2013). Comparison of the ARMA, ARIMA, andthe
autoregressive artificial neural network models in forecasting the monthly inflow of Dez dam reservoir.
Journal of Hydrology, 476, 433–441. https://doi.org/10.1016/j.jhydrol.2012.11.017

Page | 896
7th International Conference on Engineering Research, Innovation and Education
January 12-14, 2023, Sylhet, Bangladesh

PAPER ID: 233

A Combined Computational Virtual Approach for Screening of


Mexican Natural Compound Using Machine Learning Based QSAR,
Molecular Docking and Molecular Dynamics Simulation Against
Cancer Target GTPase KRAS

Md Asaduzzaman, Md. Tamzid Hossain Molla and Mohammad Taufiq Alam*


Department of Applied Chemistry & Chemical Engineering, University of Rajshahi, Rajshahi
6205, Bangladesh.
Email: ace.ru.asad01@gmail.com, thmolla@ru.ac.bd and talam@ru.ac.bd*
*Corresponding author [Mohammad Taufiq Alam]

Abstract: 85% of RAS-driven malignancies are caused by Guanosine triphosphate


Keywords: bound Kirsten rat sarcoma virus oncogene (GTPase KRAS), which also acts in various
signaling cascades and regulates cell proliferation, differentiation, and metabolism,
● GTPase KRAS; and it still remains undruggable. A combined computational approach was used to find
● Cancer therapy; out promising natural inhibitors of GTPase KRAS. A study on the quantitative
structure-activity relationship (QSAR) with machine learning was conducted using a
● QSAR;
dataset of 304 bioactive compounds with reported IC50 against GTPase KRAS from
● Molecular Docking the ChEMBL database. The best predicted model, which had an R2 value of 0.663895,
was created using the 882 different PubChem fingerprints of these molecules in
simulation;
conjunction with screening the 423 chemicals in the Mexican Natural Compound
● Molecular Dynamics Database (BIOFACQUIM). Top 35 substances were screened with pIC50 predictions
Simulation greater than 6.0 (IC50 = 1×10-6 nM). For the molecular docking simulation and
ADMET analysis, 9 compounds with lower molecular weights were chosen among
them. The 3'-O-Methylbatatasin III was chosen as the best candidate for molecular
interaction analysis since it had the highest binding affinity and had no violations of
Lipinski's rule of five. A total of 10 interactions were discovered, with 5 of them being
binding site interactions. Then, a stiffness study using Molecular Dynamics Simulation was
performed on the molecule. The study suggests 3'-O-Methylbatatasin III be a contender
for wet lab analysis and an efficient inhibitor for RAS-driven cancer therapy.

1. Introduction
More than 20% of human malignancies have a mutation in Ras, a human oncogene that was discovered
and defined more than 30 years ago.(Maurer et al., 2012) RAS proteins (HRAS, KRAS 4A, KRAS 4B,
and NRAS) are essential signal transduction molecules that function as molecular switches by cycling
between an active, GTP-bound state and an inactive, GDP-bound state. The GDP/GTP exchange and GTP
hydrolysis that are fundamental to RAS proteins are very sluggish. While GTPase Activating Proteins
(GAPs) inhibit RAS signaling by inducing the GTP hydrolysis reaction, Guanine Nucleotide Exchange

Page | 897
Factors (GEFs) speed up the exchange of bound GDP for the plentiful GTP found in cells.(Gremer et al.,
2011)
KRAS is the most often altered Ras isoform out of the three (N, H, and K). Mutant Ras's decreased GTPase
activity causes it to remain in an activated conformation longer, enhancing Ras-dependent signaling and
the survival or proliferation of cancer cells. Cancer patients with Ras mutations have a bad prognosis.
Tumor size decreases in mice after oncogenic Ras inactivation. Ras is hence frequently regarded as an
oncology target of special importance.(Sun et al., 2012)
The remarkable approximation capabilities of in silico virtual screening of potential therapeutic
compounds are drawing a lot of interest. In order to treat cancer, this study attempts to suggest using
natural substances that are typically present in the South American subcontinent area. A total of 423
natural bioactive chemicals from Mexico were docked against the human GTPase KRAS in this
investigation. The chemical known as "3'-O-Methylbatatasin III" was identified using a variety of virtual
screening technologies, including QSAR, molecular docking, ADME, and molecular dynamics
simulations. According to the research, this chemical is a potent therapeutic inhibitor of GTPase KRAS
and has the ability to pass in-vitro, in-vivo, and clinical laboratory testing.
2. Methods and Materials
2.1 Machine Learning based QSAR
According to OECD requirements, a large-scale QSAR modeling is done to find potential natural
inhibitors.(Simeon et al., 2016)
From the ChEMBL Database, 426 active drugs representing a dataset of active GTPase KRAS inhibitors
were gathered. In order to make future computations easier, compounds with reported IC50 were further
screened and their activity was converted to pIC50. 304 bioactive compounds with standard IC50 values
and standard units (nM) are compiled into a dataset. For each molecule in the dataset, the PaDEL-Py
python library computed a total of 881 molecular characteristics known as Pubchem fingerprints as a
binary representation. A 304×881-dimensional data frame was obtained.
In this study, a Random Forest model was created using the scikit-learn package in Python to estimate the
pIC50 value of unknown therapeutic candidate compounds as a function of molecular descriptors. The
test train split method of the scikit-learn package in the python programming language was used to divide
the dataset of 308 compounds into a training set (80%; 246 compounds) and a test set (20%; 62
compounds) prior to the creation of models. The Pearson's correlation coefficient (r), root mean squared
error (RMSE), cross validation, and external validation methods were used to evaluate the statistical
significance of the model to categorize the credibility of prophecy.
The QSAR model was then used to assess a dataset of 423 compounds from the Mexican Natural
Compound Database (BIOFACQUIM) in order to forecast their bioactivity against the GTPase KRAS
protein (Pilón-Jiménez et al., 2019). Compounds selected for further study with a pIC50 value greater
than 5.50 (IC50≥10^-5.5 M).

2.2 Molecular Docking Simulation


2.2.1 Receptor Protein and Ligand Preparation
With an amino acid residue sequence from 1-168 and a total structural weight of 22.5 kDa, the protein
structure was retrieved from Protein Data Bank (www.rcsb.org) (PDB id: 5MBL, resolution: 3.22
Å).(Burley et al., 2021) Using the Swiss PDB Viewer and Gromos96 forcefield, the protein was further
put through energy minimization. Polar hydrogen atoms were added to Pymol to finish off the preparation

Page | 898
of the protein structure. The ligand structures were retrieved from the Pubchem service, and energy
reduction utilizing a universal force field was done in Pyrx software.
2.2.2 Binding Site Prediction and Molecular Docking Simulation
The Pyrx software's Auto Dock Vina was used to simulate the molecular docking interactions between
the receptor protein and potential ligands. From information published in a variety of papers about GTPase
KRAS, the active areas of the protein were categorized. The grid box has dimensions of 36.1024508476,
45.4761164093 and 44.9666104889 (in Angstrom) in the X, Y, and Z axes, respectively, with a center of
40.457 Angstrom, -67.6315 Angstrom, and -6.3626Angstrom.(Trott and Olson, 2010)

2.2.3 ADME and Molecular Interactions Analysis


SWISS-ADME was used to predict a number of characteristics related to the adsorption, distribution,
metabolism, and excretion of certain chemicals.(Daina et al., 2017) The molecular interactions between
the receptor and the chosen substance (3'-O-Methylbatatasin III) were then examined using the BIOVIA
DISCOVERY STUDIO VISUALISER.

2.3 Molecular Dynamics Simulation


In order to investigate the stability of protein-ligand docked complexes, the Desmond Maestro program
(Academic version) was used in a Linux environment to run a 5000 ps (5 ns) molecular dynamics
simulation. (Bharadwaj et al., 2020) A cubic box was created and solvated using the SPC water model at
a distance of 1.0 Angstrom from the complex. Na+ and Cl− ions were used to balance the system. The
system was equilibrated and the pressure was stabilized using NPT ensemble under standard ambient
conditions. The MD simulation was finally done using the Optimized Potentials for Liquid Simulations
(OPLS) forcefield. Several protein-ligand stiffness characteristics, such as the RMSD, RMSF, Radius of
Gyration, Solvent Accessible Surface Area (SASA), and Protein Ligand Interactions were calculated
based on the MD simulation trajectories.

3. RESULTS
3.1. Machine Learning based QSAR
In order to create a Random Forrest (RF) QSAR model, 304 molecules from a dataset of 881 Pubchem
descriptors were divided in an 80/20 ratio, with 80% of the data being used as a training set and 20%
being used as a test set. The model's performance characteristics were listed in Table 01.
Table 1. Performance summary of RF model for predicting pIC50
Training Set R2 0.66389
RMSE 0.44652
External Validation Q2 0.59260
RMSE 0.41915
Q 2
10-Fold Cross Validation 0.36738
RMSE 0.72304

A total of 423 compounds from the Mexican Natural Compound Database (BIOFACQUIM) were
evaluated to determine the pIC50 value against the GTPase KRAS protein. The forecasted pIC50 values
for the compounds vary from 6.39 to 4.57, and 35 molecules with pIC50 values more than 6.0 are selected
for further screening to identify the most promising therapeutic candidates.

Page | 899
3.2 Molecular Docking Simulation
3.2.1 Receptor Protein Preparation
Guanosine triphosphate bound Kirsten rat sarcoma virus oncogene (GTPase KRAS) is found to have 1-
165 amino acid residues, to be co-crystallized with DARPin K27 of 178 residues, a GDP, magnesium ion
and water molecules in structure PDB file. The file is missing a few residues, but solely in the protein's C
terminal, which have no real impact on molecular docking or subsequent molecular dynamics simulation.
Protein structure was protonated by adding polar hydrogen atoms after the protein was reduced to -
8622.507 KJ/mol. The final GTPase KRAS is seen in Figure 1.

Figure 01: GTPase KRAS Protein

3.2.2 Binding Site Prediction and Molecular Docking Simulation


GTPase KRAS has four active sites in total, according to literature. It has been reported that the protein's
residues of 10–18, 29–35, 59–60, and 116–119 positions may function as active sites. (Sun et al., 2012,
Maurer et al., 2012) Only nine compounds were docked in virtually against the protein as out of 35
compounds that were screened in QSAR, 26 of them being discarded due to their high molecular weights.
The interactions between drugs and receptors are better with the higher the binding free energy. All
compounds have a binding free energy between -9.0 kJ/mol and -5.8 kJ/mol, with the top 3 molecules
falling between -9.0 kJ/mol and -8.4 kJ/mol (TABLE 2).
Table 2. Protein- Ligand binding affinities
Ligand Binding Affinity
Pubchem Compound CID-162918353 -9.0
Pubchem Compound CID-14134097 -8.9
3'-O-Methylbatatasin III -8.4
3.2.3 ADME and Molecular Interactions Analysis
The body's ability to metabolize and excrete drug compounds is parameterized by a process called ADME.
TABLE 3 lists a number of the top ligands' drug-like characteristics. The 3'-O-Methylbatatasin III
molecule, which exhibited the best ADME characteristics, was excluded from the examination of
molecular interactions. Five of the 11 interactions, including one hydrogen bond and ten hydrophobic
binds, involved active site residues. Figure 02. provided an illustration of the molecular interactions.
Table 3. ADME Properties
Ligand Molecular weight Log P Log S Lipinski Violation
Pubchem Compound CID-162918353 580.49 -1.33 -1.93 3
Pubchem Compound CID-14134097 594.52 -1.07 -2.16 3
3'-O-Methylbatatasin III 258.31 3.28 -3.93 0

(a) (b)

Figure 02. (a) 3D molecular interactions of protein-ligand complex. b) 2D molecular interactions protein-ligand
complex.

Page | 900
3.3 Molecular Dynamics Simulation
The flexibility of the protein-ligand complex in an artificially solvated state is investigated using
molecular dynamics modeling. 5000 ps (5 ns) of MD modeling of the protein-ligand complex is employed
in this investigation. Based on RMSD, RMSF, Radius of gyration, Protein ligand interactions, and SASA,
the stability of the complex was examined. The protein-ligand complex's molecular dynamics simulation
trajectory analysis is shown in Figure 03.

(a) (b) (c)

(d) (e)

Figure 03. Molecular dynamics simulation trajectory analysis (a) RMSD (b) RMS fluctuation (c) Protein ligand
interactions d) Radius of gyration e) SASA.
RMSD distinguishes atomic displacements from a certain reference frame, while RMSF measures the
displacement of protein residues brought about by ligand interactions, where a typical displacement of
0.1 to 0.3 nm is considered acceptable.(Opo et al., 2021) The system was shown to be properly
equilibrated during the whole simulation based on analysis of the RMSD and RMSF of the protein
backbone and ligand.
Throughout the 5ns simulation, a number of significant interactions with the active site residues were
found. Throughout the whole simulation, the complex displayed a maximum of 9 interactions.
SASA is a measure of the solvent accessibility, and Rg is the total area occupied by the complex. Rg value
for the complex averaged 4.115 nm, and its solvent accessibility was respectable at 81.08219 nm2, on
average.
4. Discussions
the work uses in silico drug discovery techniques to identify possible natural drug candidates from
Mexican Natural Compounds against the cancer target Guanosine triphosphate bound Kirsten rat sarcoma
virus oncogene (GTPase KRAS) protein. Mexican herbal remedies have been utilized therapeutically for
a very long time, but no in vitro or in silico studies against GTPase KRAS have yet been done on them.
A single molecule with an effective anticipated inhibitory capacity against the target was found using a
variety of contemporary screening methodologies, including Machine Learning Based QSAR, Molecular
Docking, and Molecular Dynamics Simulation, as shown in Figure 04.

Figure 04. 3’-O-Methylbatatasin III


This compound is a viable candidate for inhibiting the protein because, according to a molecular
interactions analysis, it has five active site interactions out of a total of eleven interactions. The remaining

Page | 901
interactions give the complex the strength it needs to be densely packed. The results of the molecular
dynamics simulation also confirmed the molecule's ability to act as an inhibitor. Throughout the whole
simulation, the complex produced fantastic outcomes. The equilibrium of the protein-ligand system in the
solvated system is shown by the RMSD, RMSF, Protein-Ligand Contacts, Rg, and SASA plots.
The whole extent of the validations confirms the drug-like potential of 3'-O-Methylbatatasin III, indicating
that the molecule is quite capable of being tested in wet laboratories for additional validation before being
put into commercialization.(Asaduzzaman et al., 2021)
5. Conclusion
As a result of their lower toxicity and increased safety, natural inhibitors are favored over synthetic ones.
The goal of the study was to test a group of natural inhibitors in order to find a prospective herbal candidate
for a cancer target. The results of this study indicate that 3'-O-Methylbatatasin III is a potent treatment
option for cancers that are governed by the GTPase KRAS protein and that it should be chosen for future
testing.

References
ASADUZZAMAN, M., ALOM, J., ALAM, M. T. & KARIM, M. A. In-silico Screening of Arjun Plant (Terminalia arjuna)
Compounds Against Cardiovascular Disease Targeting GRK2 Protein. 2021 International Conference on Computer,
Communication, Chemical, Materials and Electronic Engineering (IC4ME2), 2021. IEEE, 1-4.
BHARADWAJ, S., LEE, K. E., DWIVEDI, V. D. & KANG, S. G. 2020. Computational insights into tetracyclines as inhibitors
against SARS-CoV-2 Mpro via combinatorial molecular simulation calculations. Life Sciences, 257, 118080.
BURLEY, S. K., BHIKADIYA, C., BI, C., BITTRICH, S., CHEN, L., CRICHLOW, G. V., CHRISTIE, C. H., DALENBERG,
K., DI COSTANZO, L. & DUARTE, J. M. 2021. RCSB Protein Data Bank: powerful new tools for exploring 3D
structures of biological macromolecules for basic and applied research and education in fundamental biology,
biomedicine, biotechnology, bioengineering and energy sciences. Nucleic acids research, 49, D437-D451.
DAINA, A., MICHIELIN, O. & ZOETE, V. 2017. SwissADME: a free web tool to evaluate pharmacokinetics, drug-likeness
and medicinal chemistry friendliness of small molecules. Scientific reports, 7, 1-13.
GREMER, L., MERBITZ‐ZAHRADNIK, T., DVORSKY, R., CIRSTEA, I. C., KRATZ, C. P., ZENKER, M.,
WITTINGHOFER, A. & AHMADIAN, M. R. 2011. Germline KRAS mutations cause aberrant biochemical and
physical properties leading to developmental disorders. Human mutation, 32, 33-43.
MAURER, T., GARRENTON, L. S., OH, A., PITTS, K., ANDERSON, D. J., SKELTON, N. J., FAUBER, B. P., PAN, B.,
MALEK, S. & STOKOE, D. 2012. Small-molecule ligands bind to a distinct pocket in Ras and inhibit SOS-mediated
nucleotide exchange activity. Proceedings of the National Academy of Sciences, 109, 5299-5304.
OPO, F. A., RAHMAN, M. M., AHAMMAD, F., AHMED, I., BHUIYAN, M. A. & ASIRI, A. M. 2021. Structure based
pharmacophore modeling, virtual screening, molecular docking and ADMET approaches for identification of natural
anti-cancer agents targeting XIAP protein. Scientific reports, 11, 1-17.
PILÓN-JIMÉNEZ, B. A., SALDÍVAR-GONZÁLEZ, F. I., DÍAZ-EUFRACIO, B. I. & MEDINA-FRANCO, J. L. 2019.
BIOFACQUIM: a Mexican compound database of natural products. Biomolecules, 9, 31.
SIMEON, S., ANUWONGCHAROEN, N., SHOOMBUATONG, W., MALIK, A. A., PRACHAYASITTIKUL, V.,
WIKBERG, J. E. & NANTASENAMAT, C. 2016. Probing the origins of human acetylcholinesterase inhibition via
QSAR modeling and molecular docking. PeerJ, 4, e2322.
SUN, Q., BURKE, J. P., PHAN, J., BURNS, M. C., OLEJNICZAK, E. T., WATERSON, A. G., LEE, T., ROSSANESE, O.
W. & FESIK, S. W. 2012. Discovery of small molecules that bind to K‐Ras and inhibit Sos‐mediated activation.
Angewandte Chemie International Edition, 51, 6140-6143.
TROTT, O. & OLSON, A. J. 2010. AutoDock Vina: improving the speed and accuracy of docking with a new scoring
function,efficient optimization, and multithreading. Journal of computational chemistry, 31, 455-461

Page | 902
7th International Conference on Engineering Research, Innovation and Education
January 12-14, 2023, Sylhet, Bangladesh

PAPER ID: 237

Q-factor and Band Gap Analysis of One-Dimensional


Photonic Crystal Nanobeam Cavity for Sensing Applications
Syed Samiul Alam1,2, Md. Jahirul Islam1,2,*, Md. Rejvi Kaysir1,3 and Sheikh
Shahriar Hossain1
1
Photonics Research Group, EEE, KUET
2
Department of EEE, Khulna University of Engineering & Technology, Khulna-9203,
Bangladesh
3
Department of Electrical and Computer Engineering, University of Waterloo, 200 University
Ave. W, Waterloo, ON N2L 3G1, Canada
Email: dipto.samiul.syed@gmail.com, *jahirul@eee.kuet.ac.bd, mrkaysir@uwaterloo.ca,
shahriar.nwu@gmail.com
Abstract: Optical resonators like photonic crystal nanobeam cavities
Keywords: (PCNCs) are constructed by using photonic crystal Bragg mirrors in the
longitudinal direction and comprehensive internal reflectors in the other
● Photonic Crystal; direction. They are essentially characterized by high quality (Q) factor and
● Nanobeam Cavity; small mode volume (V). The PCNC, on the order of optical wavelengths
which can strongly constrain light, is critical for a wide range of scientific
● FDTD; and engineering applications for example ultra-small filters, low threshold
● Dispersion Relation; nanolasers, high-resolution sensors, quantum information processing,
photonic chips, and group delay as well as biomedical detection, food
● Bandgap safety, and environmental monitoring. In this work, we aimed at
investigating the structural and optical properties of nanobeam cavities on
the performance of a proposed PCNC resonator using MEEP open-source
software integrated with Python programming language. To investigate the
performance of the PCNC resonator, the cavity length, periodicity, and
band gap-dependency of periodicity have been extensively studied. The
overall comprehensive analysis shows that the highest Q-factor of 2.6×106
is found at 0.145 μm cavity length. This analysis could be helpful to design
and fabricate highly sensitive sensors.

1. Introduction
Photonic crystal nanobeam cavity (PCNC) is a high-quality factor, small mode volume optical
resonator made of photonic crystal Bragg mirrors (Villeneuve et al., 1995). These cavities with
dimensions on the order of optical wavelengths that can strongly confine light are critical for a
wide range of scientific and engineering applications, including ultra-small filters, high-resolution
sensors, low-threshold nanolasers, quantum information processing, group delay, and photonic
chips. It can also be used for environmental monitoring, biomedical detection, photon tunneling
microscopy, sub-wavelength imaging, optical switches, ultrahigh-sensitivity phase shifters, and
food safety (Erdil et al., 2019). Deeming the escalating demand for new detection technologies in
various fields, substantial research has been committed to developing copious types of refractive
index (RI) sensing devices for sensing and detection exertions for the past several years.
Several research groups around the globe have coordinated various research on PCNCs. High
Q-factor resonators constitute an important technology in the matter of attainment of sensing

Page | 903
which can increase performance in terahertz (THz) bands. These types of resonators can attain a
supreme level of sensitivity for the dete ction of analytes permittivity which is complex through
emphasized strengthening interaction amidst resonant mode and analyte. In that range, high Q-
factor like 5000 to 8700 can be achieved (Otter et al., 2014). A disruption in the structure of PCNC
could be endorsed to amplify the sensitivity for a dielectric mode of a photonic crystal nanobeam.
There is a tradeoff between the quality factor and sensitivity when designing a dielectric PCNC
RI sensor (Qiao et al., 2018). A figure of merit (FOM) has been proposed by Leif et al. (Sherry et
al., 2005) to assess the performance of RI sensors. Considering the dielectricmode of PCNC, the
Figure of Merit (FOM) is restricted by the tradeoff between Q-factor and sensitivity. High
sensitivity can be attained from Mach Zehnder Interoferometer as it uses a ring resonator, the
complication is that these types of sensors are limited to the low Q-factor. On the contrary, the
PCNCs compose a high Q-factor conserving high sensitivity to some extent simultaneously.
It is noted that one-dimensional (1D) PCNCs have materialized as one of the most auspicious
scaffolding for RI sensing recently due to high Q-factor, sensitivity, and mode volume (V) (Deng
et al., 2017). Notable, high Q-factor cavities are coveted for the low detection limit of RI sensing
as it leads to narrow resonance linewidth, therefore, resulting in small resonance shift resolution.
This also requires staunch constraint of the optical mode in the medium of the waveguide. Over
the past few years, imposing results have been attained in terms of the Q/V (i.e., FOM), by
shrinking the cavity dimensions as well as enhancing the reflectivity of the mirrors. To acquire
high Q factors with small mode volumes is one of the essential features of PCNC. These cavities
are mainly fabricated by introducing an optical defect in a structure with a periodic change in the
refractive index (Saber et al., 2019). If the optical frequency of the electromagnetic field is in the
band gap of the photonic crystal, light is trapped in the defected region. The high Q-factor greatly
relies on the band gap and the variation of the band gap impacts Q-factor.
In this work, we aimed at investigating the structural and optical properties of nanobeam
cavities on the performance of a proposed PCNC resonator. Firstly, the modeling of a 1D PCNC
has been done using MEEP open-source platform with the incorporation of python programming
language. Thereafter, the relation between the band gap and other factors was observed
extensively through dispersion calculations using MPB (MIT photonics bands), which is another
open-source software package for computing the band structures or dispersion relations, and
electromagnetic modes of periodic dielectric structures.

2. Cavity Design and Modeling


2.1 Structure of Cavity
The structure of the cavity of the PCNC depends on several parameters such as cavity radius, the
number of cavities, the distance between the adjacent cavities, the number of tapered holes, cavity
length, the number of defected periods, and waveguide height and width. Apart from these design
parameters, it is also required to specify the air gap and minimum and maximum wavelengths and
frequencies to design the nanobeam cavity. For this work, we intend to create a resonant cavity by
linearly tapering the lattice periodicity from a certain section of the 1D photonic crystal (see Fig.
1).
In addition, band gap modulation is an important component of photonic crystal design since
the Q-factor of a PCNC directly depends on the band gap size of the crystal. By varying the band

Page | 904
gap, it is feasible to get a variety of Q factors. The photonic crystal’s figure of merit (FOM) is
determined using its Q-factor and resonant wavelength with the following relationship
𝑆×𝑄 (1)
𝐹𝑂𝑀 =
𝜆𝑟𝑒𝑠
Where, S is sensitivity and 𝜆𝑟𝑒𝑠 is the resonant wavelength.

Fig. 1. Cavity structure of a 1D PCNC with tapered holes and air padding.
2.2 Simulation Procedure
To investigate the performance of the resonator, the cavity length, periodicity, band gap-
dependency of periodicity, etc. have been extensively studied using MEEP software which
incorporates Finite Difference Time Domain (FDTD) method. A time-domain electromagnetic
simulation simply evolves Maxwell’s equations over time within some finite computational
volume, essentially performing a kind of numerical experiment. This can be used to calculate a
wide variety of useful quantities (MEEP Documentation, n.d.). To find out band gap wavelengths,
the height and width of the waveguide were considered as 0.5 μm and 0.22 μm, respectively
(Deotare et al., 2009). The overall simulation process is shown in Fig. 2.

Band gap
Selection of cavity MPB simulation
wavelength for MEEP simulation
radius and for dispersion
from dispersion for Q-factor
periodicity relations relations

Fig. 2. The overview of the Q-factor analysis process accumulated in this work.
Table I Structural and optical parameters for Q-factor analysis

Parameters Values
Starting periodicity 0.43 µm
Ending Periodicity 0.33 µm

Hole Radius 0.28 µm

Number of Defected Periods 5 and 3


Minimum Source Wavelength 0.13 µm
Maximum Source Wavelength 0.17 m

Page | 905
3. Results and Discussion
3.1 Variation of Band gap with Periodicity
Using MIT Photonic Bands (MPB), a free and open-source software package for computing the
band structures or dispersion relations, we calculated the dispersion relation inside the irreducible
Brillouin zone, which covers axial wave vectors along the x-direction from 0 to π , provided the
𝑎
1D periodicity. As is shown in Fig. 3(a) a band gap of 0.36 μm is found for a = 0.33 μm, radius r
= 0.12 μm.

Fig. 3. Variation of band gap with the varying of (a) periodicity a = 0.33 µm (b) periodicity a = 0.63 µm
(c) periodicity a = 0.63 µm. For all the cases radius r = 0.12 µm is considered.

To further examine the periodicity-dependent band gap variations, periodicity was progressively
increased while keeping the radius constant, and a proportionate trend was discovered. First, the
periodicity was set to 0.33 µm and the radius to 0.12 µm. Periodicity was subsequently increased
from there while maintaining the radius constant at 0.12 µm. There was a considerable increase in
the band gap and also the band gap was getting steeper as the periodicity increased. At 0.33 µm
the minimum and maximum wavelengths were 1.03 µm and 0.1.39 µm which increased up to 1.94
µm and 2.01 µm, respectively for periodicity variations from 0.33 µm to 0.75 µm. Later, the
periodicity-dependent band gap is plotted in Fig. 4 keeping the radius constant. This is evident that
the band gap decreases with an increase in periodicity.

Fig. 4. Variation of band gap with the change of periodicity by keeping the cavity radius constant.

Page | 906
3.2 Variation of Q-factor with Cavity Length
For this segment, we have considered the periodicity of the lattice as 0.43 µm, while the waveguide
width and height are 0.50 and 0.22 µm, respectively. The hole radius is 0.12 µm (Deotare et al.,
2009). An FDTD simulation was done to find out the highest Q- factor for the band gap between
1.30 and 1.70 μm. The cavity is considered by 3 tapered holes and is surrounded by the waveguide.
The entire computational cell is surrounded by perfectly matched layers (PML). With a band gap
between 1.30 μm and 1.70 μm and 3 tapered holes, the cavity length was varied from 0.11 μm and
0.18 μm. In this range, the highest Q-factor found at 0.145 μm is 2.6 × 106 [see Fig. 5(a)]. With
a cavity length of 0.51 μm to 0.59 μm and there exist 3 tapered holes surrounded by the waveguide.
Investigations show that the highest Q-factor was achieved at 0.575 μm which was 869.58 [see
Fig. 5(b)]. Then the number of tapered periods was decreased from 3 to 5 which resulted in a lower
Q-factor of 135.115 at 0.550 μm. It is noted that a high Q-factor is the core requirement to attain
an extreme level of sensitivity. Finally, the periodicity-dependent wavelength variation is shown
in Fig. 6. It is found that both the λmax and λmin increase with the increased periodicity.

Cavity Length (μm)


Fig. 5. Q-factor vs. cavity length within (a) 0.11 μm to 0.17 μm (b) 0.51 μm to 0.59 μm. Highest Q-factor of
2.6 × 106 is found at 0.145 μm.

Fig. 6. Variation of wavelength with the change of periodicity.

Page | 907
4. Conclusion
In this work, numerical investigations have been performed to find out the highest Q-factor in
several wavelength ranges and periodicity-dependent band gaps for one-dimensional photonic
crystal nanobeam resonators. A silicon-based structure was considered, and several design
parameters were studied to observe their effects on the performance. An extensive study shows
that by varying the periodicity the band gap width as well as the resonant wavelength changes
significantly. A maximum Q-factor of 2.6 × 106 is found at a cavity length of 0.145 μm, which
is significantly higher than the previously obtained value of 7.5 × 105 in this range. It is noted
that Q-factor has a predominant impact on the figure of merit of a photonic crystal and a high Q-
factor is essential in sensing applications.

References
Deng, C.-S., Li, M.-J., Peng, J., Liu, W.-L., & Zhong, J.-X. (2017). Simultaneously high-Q and
high-sensitivity slotted photonic crystal nanofiber cavity for complex refractive index
sensing. Journal of the Optical Society of America B, 34(8), 1624.
https://doi.org/10.1364/josab.34.001624
Deotare, P. B., McCutcheon, M. W., Frank, I. W., Khan, M., & Lončar, M. (2009). High quality
factor photonic crystal nanobeam cavities. Applied Physics Letters, 94(12).
https://doi.org/10.1063/1.3107263
Erdil, M., Ozer, Y., & Kocaman, S. (2019). High-Q Slot-Mode Photonic Crystal Nanobeam
Cavity Biosensor With Optomechanically Enhanced Sensitivity. IEEE Journal of Selected
Topicsin Quantum Electronics, 25(2), 1–6. https://doi.org/10.1109/JSTQE.2018.2880592
MEEP Documentation. (n.d.). https://meep.readthedocs.io/en/latest/Introduction/
Otter, W. J., Hanham, S. M., Ridler, N. M., Marino, G., Klein, N., & Lucyszyn, S. (2014). 100
GHz ultra-high Q-factor photonic crystal resonators. Sensors and Actuators, A: Physical,
217, 151–159. https://doi.org/10.1016/j.sna.2014.06.022
Qiao, Q., Xia, J., Lee, C., & Zhou, G. (2018). Applications of Photonic Crystal Nanobeam
Cavities for Sensing. Micromachines, 9(11). https://doi.org/10.3390/mi9110541
Saber, I., Boddeda, R., Raineri, F., Sanchez, D., Beaudoin, G., Sagnes, I., Glorieux, Q., Bramati,
A., Levenson, J. A., & Bencheikh, K. (2019). Photonic crystal nanobeam cavities with
optical resonances around 800 nm. In arXiv (Vol. 36, Issue 7, pp. 1823–1828). arXiv.
https://doi.org/10.1364/josab.36.001823
Sherry, L. J., Chang, S.-H., Schatz, G. C., Van Duyne, R. P., Wiley, B. J., & Xia, Y. (2005).
Localized Surface Plasmon Resonance Spectroscopy of Single Silver Nanocubes. Nano
Letters, 5(10), 2034–2038. https://doi.org/10.1021/nl0515753
Villeneuve, P. R., Fan, S., Joannopoulos, J. D., Lim, K. Y., Petrich, G. S., Kolodziejski, L. A., &
Reif, R. (1995). Air-bridge microcavities. Applied Physics Letters, 67, 167.
https://doi.org/10.1063/1.114655

Page | 908
7th International Conference on Engineering Research, Innovation and Education
January 12-14, 2023, Sylhet, Bangladesh

PAPER ID: 257

Sentiment Analysis of Online Public Comments in


Bangla on BD Government Decisions
Md. Abdus Samad1*, Md Shamihul Islam Khan2 , Mohammed Jahirul Islam3, Mohammad
Abdullah Al Mumin4
1,3
Institute of Information and Communication Technology, Shahjalal University of Science
and Technology, Sylhet, Bangladesh
2,4
Department of Computer Science and Engineering, Shahjalal University of Science and
Technology, Sylhet, Bangladesh

Email: 1abdus-iict@student.sust.edu, 2shamihul78@gmail.com, 3jahir-cse@sust.edu,


4
mumin-cse@sust.edu
● *Corresponding author [Md. Abdus Samad]
Abstract:
Keywords:
As social media and online news portals gain popularity among all levels of
● Sentiment people, more investigation is required to ascertain how the platforms can be
Analysis; used to foster meaningful citizen-government partnerships. Bangladesh has a
total of 26.0 million people using social media actively, thus expressing their
● Bangla Sentiment opinion on a variety of topics, including governmental activities. As there is
Analysis; no other direct and easy way to gain people's opinions towards governmental
activities, online platforms might be a potential source of obtaining people's
● Natural Language opinions. Although some research has already been done in other languages
Processing; regarding people's opinions on government decisions, there is no such research
in the Bangla language on this topic. The purpose of this study is to assess the
● Machine Learning; sentiment (tone) of people toward various governmental decisions, which
therefore assists the government in assessing how people think about any
● Governmental
particular decision. In this study, a total of 7645 comments were collected and
Decision; pre-possessed, and labels were given in three polarities (positive, negative, and
● Deep Learning. neutral). Seven machine learning algorithms and a deep learning algorithm
were trained and evaluated in this study. The deep learning-based model
produces better classification accuracy (73.60%) compared with machine
learning models. We found different ML models performed better in terms of
accuracy in different n-gram features. For unigram features, Logistic
Regression outperformed the other six algorithms with an accuracy score of
66.8%. On the other hand, Multinomial Naive Bayes outperforms Bigram and
Trigram features. The algorithm achieved an accuracy of 65.7% and 63.61%
on bigram and trigram features, respectively.
1. Introduction

Sentiment Analysis (SA) is a field of research that examines how people’s opinions, judgments,
sentiments, feelings, and attitudes are expressed in written language (Osimo & Mureddu, 2012).
According to Aydoan and Akcayol (2016), it is the computational handling of emotions,opinions,
and subjectivity in any writing. Because it combines artificial intelligence, machine learning,
computational linguistics, retrieval of information, natural language processing, and semantics,
SA is an interdisciplinary field (Ain et al., 2017). People have gradually gotten more

Page | 909
and more vocal about their opinions on various social media platforms and the web in general
since the introduction of Web 2.0 (Liebeskind et al., 2017). That makes it a crucial source of
information retrieval.
Bengali is the native language of close to 230 million people (out of which 163.9 million are
Bangladeshi). Among Bangladeshis, 63.3 million utilize the internet. Among internet users,
26.0 million use social media frequently. Besides, the capital of Bangladesh, Dhaka, has 1.1 per
cent of all active Facebook (FB) users, placing it second globally (Murad, 2017). So, the public’s
opinions about governmental decisions could be accessed in great detail on social media and in
the newspaper. As there is no other direct and easy way to gain people’s opinions towards
governmental activities, and online platforms are a direct and cost effective way of collecting
opinions, the mentioned platforms might be a potential source in this case.
In this research, utilizing Machines Learning (ML) and Deep Learning (DL) algorithms, we
covered some aspects of analyzing the attitude toward governmental decisions that are generated
on social media and in the online press. The goal of this study is to assist public decision-makers
(the government or its other organizations) in the decision-making process by analyzing public
sentiments towards their decisions.
The organization of this paper is as follows. In section 2, we present the existing work. In section
3, we discuss the experimental methodology for this study. The results and discussion are
presented in section 4. At the end, we presented concluding remarks and future directions in
section 5.
2. Related works
Over time, researchers use the SA techniques in different fields with different techniques and
improve the quality of the algorithm (Mäntylä et al., 2018; Vinodhini and Chandrasekaran, 2012).
Prior to the availability of the internet and the popularity of social media, opinions were collected
from some written documents or other means. Thereafter, as social media got popular over time,
with millions of users starting to use social media platforms, researchers focused on capitalizing
on those platforms.
Many sentiment analysis studies have been done using social media platforms within the last few
years. In a study, Stylios et al. (2010) proposed text and data mining mechanisms to capture public
sentiment from online sources on governmental decisions. Their goal is to automatically detect
the public’s standpoint on government decisions and thereafter be capable of deducing how
citizens’ opinions may influence future government activities. A very similar study was
conducted by Adeyemo and Ojo (2014) later as well, in terms of goals and objectives. Milakovich
(2010) covers the beginnings of participatory politics. A study was conducted by Liebeskind et
al.(2017) on 5.3 million public messages on incumbent politicians’ Facebook pages to find the
best method for automatically classifying such a large corpus of comments. Later, Saragih and
Girsang (2017) intend to look at customer engagement in the transportation industry by analyzing
data generated in social media. Pérez-Granados et al. (2012) and Rahab et al. (2017) utilized
online newspaper comments to analyze sentiments.
Scholars have been conducting sentiment analysis on Bangla texts as well. Vinodhini and
Chandrasekaran (2012) conducted a study to quantify the positive and negative aspects of a
sentence. The research was conducted by Chowdhury and Chowdhury (2014) and Asimuzzaman
et al. (2017) on a different period on microblog posts in Bangla. Chowdhury andChowdhury
(2014) intended to use it to automatically extract users’ thoughts from microblog

Page | 910
posts in Bangla and afterwards classify the total polarity of the texts as negative or positive. On
the other hand, in their work, Asimuzzaman et al. (2017) used Adaptive Neuro-Fuzzy System to
predict the polarity by applying fuzzy rules to Bangla tweets.
In this era of technology, online portals have a huge impact on governmental decision-making
(Hongquan, 2012). Therefore, in resourceful languages, much sentiment analysis research is done
on government decisions utilizing online portals. In this notion, in the Bangla language, there is
still no such research on governmental decisions despite having a huge number of users. In this
study, we aimed to cover this particular research gap.

3. Methodology
In this part of the paper we discuss the experimental methods for our study. We brieflydescribed
the methodology of data collection, processing, and labeling. We also discuss the model training,
testing and evaluation methods as well.

Figure 1: Methodology

3.1 Data collection and pre-processing


A total of 7645 comments were collected manually from online newspapers and social media
platforms. Among the total collected comments, 2661 comments were filtered out due to
irrelevancy. Collected comments were pre-processed in different steps, demonstrated below:

Figure 2: Pre-processing steps

3.1.1 Punctuation marks, whitespace, numeric values, and emoji removal


The majority of the comments were noisy, and some of them contained extraneous URLs, emojis,
dots, and images. We trimmed all of them.
3.1.2 Conversion of Romanized Bangla texts
We found some comments were written in Romanized Bangla (Bangla words written in English
alphabet). As such, a Romanized Bangla “Khub khushi holam” became “খব খশিু " হলু ম"
3.1.3 Stop-word Removal

Despite the fact that they appear frequently within the text, certain word categories are typically
unimportant for sentiment analysis. We trimmed them for this study.
3.1.4 Feature extraction (N-Gram and TF-IDF)
In this research, we used TF-IDF to extract features. One dimensional IDF and two dimensional
TF is multiplied to calculate the TF-IDF as in equation 1 below:

Page | 911
TF-IDF = TF * IDF ............................. (1)

Comment সরক র সশি ক শিস দ্ধ ু ন্ত শি-য়ে/য়ে2


TF-IDF value 0.0948500964042655 0.154390639646518 0.220542164544889 0.134390639646518

Table 1: TF-IDF representation of words

Let a corpus {C1, C2, C3…., Cn} with a total of N no. of documents. Let Ci denotes arbitrary
document for 1 ≤ i ≤ x. Let Cj = {s1, s2, ..., sq} be a document having total z terms. Also suppose
si denotes an arbitrary feature word. If s(i,j) indicates the number of total appearancesof si in a
document Cj then the term frequency of si concerning Cj, denoted by TF(i,j), can be expressed by
the following terms.
s (i , j )
TF(i,j) = (2)
s (1 , j )+ s (2 , j )+ ....+ s ( z , j )............................

The ratio of total documents to the number of total documents that contain feature words is termed
IDF(i). Let si stand for the total number of documents that mention ti at least once. With regard
to the corpus, IDFi, of ti, may thus be defined as:

IDFi = log ( Ns i) (3)

A group of n subsequent tokens from a certain text, voice, or blog are known as ngrams.
There are terms for n = 1, 2, and 3. If n=1 it is unigram, if n= and n=3 they are bigram,
and trigram, respectively. We tested unigram, bigram, and trigram separately.
Label encoding and dataset splitting
In this study, we classified sentiments into three polarity classes, namely positive, negative, and
neutral. Before training the model with the collected data, we encode the label. Thereafter, we
divide training and testing data into an 8:2 ratio.
Model Training
For training, seven ML models were selected, namely Random Forest (RF), K-nearest Neighbour
(KNN), Decision Tree (DT), Logistic Regression (LR), Stochastic Gradient Descend (SGD),
Multinomial Naive Bayes (MNB), and Linear Support Vector Machine (LSVM). A deep learning
model, namely Convolutional Neural Network (CNN), is also usedfor this study.

4. Result and Discussions


We found different ML models performed better in terms of accuracy in different n-gram
features. MNB performed well on bigram and trigram features, whereas LR performed better on
unigram features in terms of accuracy and F1 scores. On the other hand, the Convolutional Neural
Network outperformed ML algorithms in terms of accuracy.
For unigram features, Logistic Regression outperformed the other six algorithms with an
accuracy score of 66.7% (Table: 5). On the other hand, Multinomial Naive Bayes outperforms

Page | 912
on bigram and trigram features. The algorithm achieved an accuracy of 65.7% and 63.61% on
bigram and trigram features, respectively. CNN with multiple layers produces better validation
accuracy (73.60%) compared with machine learning algorithms.
CNN produces 99.97% training accuracy as well. Chowdhury and Chowdhury (2014) performed
a study on Bengali Twitter data using SVM and MaxEnt. They gained 66% and 65% accuracy for
the algorithms, respectively, which is very similar to our results. Rabeya et al. (2019) performed
a Bangla song review study in a lexicon-based backtracking approach which gave them 71.23%
accuracy, which is slightly higher than our gained accuracy by the ML algorithms.
Algorithm Accuracy Precision F1-Score

Unigram Bigram Trigram Unigram Bigram Trigram Unigram Bigram Trigram


LR 66.7 63.11 61.81 67.66 66.66 67.95 65.41 59.15 56.29
DT 58.03 59.52 59.82 57.52 59.09 59.53 57.48 59.05 59.45
RF 64.01 63.31 63.41 65.2 64.51 65.76 60.85 59.15 58.54
MNB 66 65.8 63.81 65.63 65.08 62.86 64.96 64.88 62.94
KNN 61.12 62.61 61.32 60 61.45 60.01 59.89 61.38 59.86
LSVM 57.33 54.84 53.34 65.18 62.59 76.54 47.95 44.14 42.09
SGD 60.42 57.03 55.33 66.28 72.37 62.48 52.62 47.35 44.83

Table 2: Model accuracy and Evaluation

This probably happened because their polarity was just to accept or reject the song (only 2 polarity
classes). Alam et al. (2017) studied Bangla comments from an unknown source andsize with
the deep learning algorithm CNN. Their classification accuracy was 99.87%, which was slightly
better than ours. As they only used positive and negative classes, their accuracy is higher than
ours. In terms of precision, trigram outperformed the other two feature categories byachieving
76.54% precision in linear SVM. The highest precision in unigram is achieved by Logistic
Regression (67.66%), whereas Stochastic Gradient Descend has achieved 72.37%, which is the
highest among bigram features.

5. Conclusion and future works


For this study, we used seven ML algorithms and a DL algorithm. The accuracy we got is quite
satisfactory compared with other SA studies (with three polarity classes) in Bangla, but compared
with resourceful language, there is still room for improvements. Also, the DL algorithm (73.60%
accuracy) outperformed the ML algorithms (66.7%). As, we did not find anyexisting datasets
available publicly. Hence, we developed a golden dataset with over five thousand comments prior
to training the models with the data. As the study topic is quite generalized and people use various
types of words to express their thoughts, we think the dataset is still not large enough.
In the future, we will expand the dataset. In addition, as we explored DL models that produced
better accuracy, in the future we will use other DL algorithms.
Acknowledgment
This work is supported by Shahjalal University of Science and Technology (SUST) research center.
References
Adeyemo, A. B., & Ojo, A. K. Classification of Social Blogs Comments Using Text Mining.

Page | 913
IJCSI International Journal of Computer Science. 2014, 11(6); 54–59.
Ain, Q. T., Ali, M., Riaz, A., Noureen, A., Kamran, M., Hayat, B., & Rehman, A. Sentiment
analysis using deep learning techniques: a review. International Journal of Advanced
Computer Science and Applications. 2017, 8(6).
Alam, M. H., Rahoman, M.-M., & Azad, M. A. K. Sentiment analysis for Bangla sentences
using convolutional neural network. 2017 20th International Conference of Computer and
Information Technology (ICCIT). 2017, 1–6.
Asimuzzaman, M., Nath, P. D., Hossain, F., Hossain, A., & Rahman, R. M. Sentiment analysis
of bangla microblogs using adaptive neuro fuzzy system. 2017 13th International
Conference on Natural Computation, Fuzzy Systems and Knowledge Discovery (ICNC-
FSKD). 2017, 1631–1638.
Aydoğan, E., & Akcayol, M. A. A comprehensive survey for sentiment analysis tasks using
machine learning techniques. 2016 International Symposium on INnovations in Intelligent
SysTems and Applications (INISTA). (2016), 1–7.
Chowdhury, S., & Chowdhury, W. Performing sentiment analysis in Bangla microblog posts.
2014 International Conference on Informatics, Electronics & Vision (ICIEV). 2014, 1–6.
Hongquan, L. The Internet Public Opinion in Crisis Management: Analysis from the
Perspective of Chinese Government. 2012 International Conference on Public
Management (ICPM-2012). 2012, 197–203.
Liebeskind, C., Nahon, K., HaCohen-Kerner, Y., & Manor, Y. Comparing sentiment analysis
models to classify attitudes of political comments on facebook (november 2016). Polibits
2017, 55; 17–23.
Mäntylä, M. V, Graziotin, D., & Kuutila, M. The evolution of sentiment analysis—A review of
research topics, venues, and top cited papers. Computer Science Review. 2018, 27; 16–32.
Milakovich, M. E. The Internet and increased citizen participation in government. JeDEM-
eJournal of eDemocracy and open government. 2010, 2(1); 1-9.
Mahmud Murad. (BDNEWS24. Com, Bangladesh), 2017. Dhaka ranked second in number of
active Facebook users”, Available in: https://bdnews24.com/bangladesh/dhaka-ranked-second-
in-number-of-active-facebook-users, (accessed 13 January 2022)
Osimo, D., & Mureddu, F. Research challenge on opinion mining and sentiment analysis.
Universite de Paris-Sud, Laboratoire LIMSI-CNRS, Bâtiment. 2012, 508.
Pérez-Granados, D., Lozano-Garzón, C., López-Urueña, A., & Jiménez-Guarín, C. Sentiment
analysis in Colombian online newspaper comments. In Recent progress in data
engineering and internet technology. Springer. 2012, 113–119.
Rabeya, T., Chakraborty, N. R., Ferdous, S., Dash, M., & Al Marouf, A. Sentiment analysis of
bangla song review-a lexicon based backtracking approach. 2019 IEEE International
Conference on Electrical, Computer and Communication Technologies (ICECCT). 2019,
1–7.
Rahab, H., Zitouni, A., & Djoudi, M. Siaac: Sentiment polarity identification on arabic algerian
newspaper comments. Proceedings of the Computational Methods in Systems and
Software. 2017, 139–149.
Saragih, M. H., & Girsang, A. S. Sentiment analysis of customer engagement on social media in
transport online. 2017 International Conference on Sustainable Information Engineering
and Technology (SIET). 2017, 24–29.
Stylios, G., Christodoulakis, D., Besharat, J., Vonitsanou, M.-A., Kotrotsos, I., Koumpouri, A.,
& Stamou, S. Public opinion mining for governmental decisions. Electronic Journal of E-
Government . 2010, 8(2); pp202-213.
Vinodhini, G., & Chandrasekaran, R. M. Sentiment analysis and opinion mining: a survey.
International Journal. 2012, 2(6); 282–292.

Page | 914
7th International Conference on Engineering Research, Innovation and Education
January 12-14, 2023, Sylhet, Bangladesh

PAPER ID: 283

Enhancement in Efficiency of ZnTe Based Solar Cell by Using


Intermediate Band Absorber Layer

Md. Mazharul Islam, Md. Sohel Rana and M. Julkarnain*


Department of Electrical and Electronic Engineering, University of Rajshahi,
Rajshahi-6205, Bangladesh
E-mail: M M Islam: shakil.ru455@gmail.com; M S Rana: sohel4709@gmail.com;
and M. Julkarnain: jnain.apee@ru.ac.bd
*Corresponding author [M. Julkarnain]

Abstract: Zinc Telluride (ZnTe) thin film solar cell have attracted great
Keywords: attention of researchers due to its optimum direct energy gap, higher
absorption coefficient and good electronic properties. Moreover, ZnTe with
● ZnTe; impurity band or intermediate band is one of the promising candidates to
● ZnTe:O; enhance the performance of solar cells due to drastically improve in absorption
characteristics. In this work, a basic CdS/ZnTe and an intermediate band solar
● Solar Cells; cell structure (IBSC) (CdS/ZnTe:O/ZnTe) has been investigated by numerical
● Intermediate Band; analysis using a one-dimensional Solar Cell Capacitance Simulator (SCAPS-
1D). Firstly, a basic conventional structure (Al/ZnO/CdS/ZnTe/Pt) analyzed
● IBSC; and optimized with respect to different input parameters like thickness, carrier
● SCAPS-1D; concentration, and defect densities of absorber. Secondly, an intermediate
band layer (ZnTe:O) has been introduced in between the ZnTe and CdS, and
● PCE. the effect of this layer on device performance is analyzed. The input physical
parameters of this new structure have been optimized depending on optimum
performance of the cell. The structure with ZnTe:O IB layer
(Al/ZnO/CdS/ZnTe:O/ZnTe/Pt) gives an overall conversion efficiency of
57.85% (Voc=1.697V, Jsc = 47.26mA/cm2 , FF= 72.19%) at optimum
absorber thickness of 1μm. The efficiency improved more than 200% by
introducing intermediate band (IB) in ZnTe based solar cells whereas optimum
thickness of absorber of this structure Al/ZnO/CdS/ZnTe:O/ZnTe/Pt became
half compare to basic structure. The reduction in thickness reduces the material
consumption thus the cost of the devices. These simulation results could be an
important guideline to a researcher for fabricating cost effective and highly
efficient ZnTe solar cells.

1. Introduction
Zinc Telluride (ZnTe) thin film solar cell attracts a great attention for solar cell fabricating
community due to its wide direct energy gap (2.23-2.28eV) ( Kaneta et al., 2000, Skhouni et al.,
2016), higher absorption coefficient, α >105 cm−1, low electron affinity 3.73 eV and good
electronic properties (Pistone et al., 1998; Skhouni et al., 2016). ZnTe is one of the most

Page | 915
promising II–VI semiconductor candidates for the development of low-cost and high-efficiency
thin film solar cells. However, ZnTe cannot absorb enough photons with energy lies in visible
region due to its high band gap, thus limiting the conversion efficiency. To remedy this II–VI
semiconductors with an intermediate band (IB) located inside the semiconductor bandgap seem
to be the appropriated solution for increasing the conversion efficiency. Indeed, the intermediate
band solar cells (IB-SCs) offer another concept of absorption based on multi-photon absorption
with assistance of an IB situated in the bandgap of an active semiconductor (Luque et al., 1997;
Skhouni et al., 2016)., the IB must be partially filled to allow absorption of photons with energy
below the energy bandgap. It is possible for ZnTe to generate an IB by doping with oxygen
forming ZnTe:O alloy which is an attractive material for photovoltaic devices (Wang et al., 2009).
By introducing the intermediate band, i. e. for IBSCs, the conversion efficiency of single gap
semiconductor based solar cells can be increased from 40.7% to 63.2% (Luque et al., 2001; Luque
et al., 2010; Skhouni et al., 2016).
In depth understanding in device operation mechanism is crucial for further optimizing device
performance. Device simulation can provide deep insight into the device physics and forecast the
effect of material properties on device performance. In this work, the ZnTe:O/CdS based solar
cell structure has been investigated by numerical analysis using a one-dimensional Solar Cell
Capacitance Simulator (SCAPS-1D). Firstly, a basic conventional structure
(Al/ZnO/CdS/ZnTe/Pt) analyzed and optimized. Secondly, an intermediate band layer (ZnTe:O)
has been introduced in between the ZnTe and CdS, and the effect of this layer on device
performance is analyzed with respect to different input parameters like thickness and carrier
concentration of different layers.
2. Modeling and physical parameters
This work has been numerically simulated by a one-dimensional Solar Cell Capacitance
Simulator (SCAPS-1D) developed by a group at the Department of Electronics and Information
Systems (ELIS), University of Gent, Belgium (Burgelman et al., 2000). The physical parameters
including the bandgap, electron affinity, mobility, doping, dielectric permittivity, effectivedensity
of states, etc. for different layers can be modified which helps to obtain a desired structure.The
conventional layered structure of ZnTe/CdS thin-film solar cell used in this work is as
Al/ZnO/CdS/ZnTe/Pt. Here, first layer Platinum (Pt) on glass substrate used as back contact. The
subsequent p-ZnTe semiconductor act as an absorber layer, n-CdS as buffer layer, and n-ZnO as
window layer, and finally, the Al used as front contact of the cell. To study the effect of an
intermediate band (IB) layer on device performance, ZnTe:O has been inserted in between the p-
ZnTe and n-CdS layer. Hereafter, the basic device structure (Al/ZnO/CdS/ZnTe/Pt) and proposed
device structure (Al/ZnO/CdS/ZnTe:O/ZnTe/Pt) is mentioned as (a) without IB layer, and (b) with
IB layer, respectively.
In the SCAPS software, physical parameters of different layers need to be considered to analyze
the device characteristics. The physical parameters of different layers which used in this
simulation have been taken from previous report (Skhouni et al., 2016; Skhouni et al., 2017) and
presented in Table-1.
The key and crucial task of a simulation is to determine variable physical parameters of a solar
cell. For this purpose, the performance of the cells has been analyzed systematically by varying
different parameters; like thickness and carrier concentrations of different layers. The effect of
working temperature and back contact metals work function has been investigated. In this
simulation, the thickness and carrier concentration of window layer (ZnO) remain fixed to 50nm

Page | 916
and 1x1018 cm-3, as these parameters have less effect on cell performance. All simulations have
been conducted under global AM 1.5 spectral illumination with a light power of one sun (1000
W/m2).

Table 1: Physical parameters of different layers of proposed ZnTe/CdS solar cell. (Skhouni et
al., 2016; Skhouni et al., 2017)
n-ZnO n-CdS p-ZnTe
Parameters p-ZnTe:O
(Window) (Buffer) (Absorber)
Thickness, W (m) 0.05 0.1 0.1–2 1.0
Band gap, Eg (eV) 3.37 2.43 2.19 2.30
Electron affinity, (eV) 4.45 4.0 3.73 3.73
Dielectric permittivity, r 9.00 9.35 10.3 10.3
Effective Density of states, NC (cm-3) 2.95 ×1018 1.75 × 1018 1.18 ×1018 1.176×1018
Effective Density of states, NV (cm-3) 1.15 ×1019 1.46 × 1019 1.16 × 1019 1.166×1019
Electron thermal velocity (cm/s) 2.38 × 107 2.83 × 107 3.24 × 107 3.24 × 107
Hole thermal velocity (cm/s) 1.52 × 107 1.39 × 107 1.51 × 107 1.51 × 107
Electron mobility, e (cm2/Vs) 100 100 330 330
Hole mobility, p (cm2/Vs) 25 25 80 80
Donor Concentration, n (cm-3) 1.0 × 1018 1.0 × 1018 0 0
Acceptor Concentration, p (cm-3) 0 0 2.16 ×1019 1013-1016
3. Result and Discussions
3.1 Optimization of PV performance by varying layer parameters
The effect of thickness of buffer layer on device output parameters for both cells has been
investigated. The thickness of buffer layer varied from 0.01 to 0.2µm and its corresponding output
parameters are shown in Fig. 1. Figure 1 illustrates that; all the performance parameters but VOC
of both cells increase with increasing the thickness of buffer layer. The J SC, FF, and PCE shows
qualitatively similar behavior with increasing thickness of buffer layer. The VOC slightly increases
from 1.7144 to 1.7148 V for cell without IB layer, whereas it slightly decreases from 1.735 to
1.695 V for cell with IB layer. Though, the enrichment of performance is insignificant in response
to thickness variation, it comparatively rises quickly up to 0.08µm after that it rises slowly. This
type of response also reported by other researchers (Khattak et al., 2018). The small improvement
in performance can be explained by the fact that by increasing CdS buffer layer thickness, more
atoms can be ionized so as to photo-generate more electrons (Bouarissa et al., 2021). Though, the
performance increases with increasing thickness we prefer 0.1µm to avoid higher series
resistance, as the series resistance increases with increasing thickness of buffer layer.
The variation of device performance with donor concentration of buffer layer has been studied
and illustrated in Fig. 2. Buffer layer’s donor concentration has been varied from 1×10 16 to
1×1019cm−3. The Figure 2 represent that, with increasing the doping concentration from 1×1016 to
5×1017cm−3 Jsc, FF and PCE is increased monotonically, whereas the value of Voc is decreased
smoothly. Above the concentration of 5×1017cm−3 the performance parameters seem almost
constant. When the doping concentrations of buffer layer is higher than the absorber layer, higher
band bending observed on the absorber side, which limits the density of interfacial holes (Song et
al., 2016). Again, the interface recombination rate controlled by both interface recombination
velocity and the electron and hole densities at the interface (Song et al., 2016). As the density of
interfacial holes’ decreases, the recombination rate also decreases with increasing doping

Page | 917
concentration of buffer layer. From this analysis, the optimal doping concentration of buffer layer
set as 1 × 1018cm−3.
CdS/ZnTe CdS/ZnTe:O/ZnTe CdS/ZnTe CdS/ZnTe:O/ZnTe
91.0 17.2
91 75 17.1 58 80
60
90 56 90.5
70 17.0 75 17.0
54
89 56
90.0

PCE (%)
52 70

PCE (%)
PCE (%)
PCE (%)
65 16.9 16.8

FF (%)
FF (%)
88

FF (%)
50 89.5
FF (%)

65 52
87 60 48
16.8 89.0 60 16.6 48
86 46
55 16.7 44 88.5 55
85 16.4 44
42
84 50 16.6 40 88.0 50

1E16 1E17 1E18 1E19 1E16 1E17 1E18 1E19


0.00 0.04 0.08 0.12 0.16 0.20 0.00 0.04 0.08 0.12 0.16 0.20

1.720 46.2
1.74 11.0
46.4 1.72 10.96
1.719
1.73

Jsc (mA/cm2)
1.7148 46.2 10.92

Jsc (mA/cm2)
Jsc (mA/cm2)

Voc (V)
Voc (V)

Jsc (mA/cm2)

1.718 1.70 46.0


1.72
46.0
Voc (V)

Voc (V)
10.88
1.717
1.71 10.9 1.68
45.8
1.716 10.84
45.8
1.70 45.6
1.7144 1.66
1.715 10.80
1.69 45.4
1.68 10.8 45.2 1.714 1.64 10.76 45.6

1E16 1E17 1E18 1E19 1E16 1E17 1E18 1E19


0.00 0.04 0.08 0.12 0.16 0.20 0.00 0.04 0.08 0.12 0.16 0.20

Thickness of CdS Layer (m) Thickness of CdS Layer (m) Donor Concentration of CdS (cm-3) Donor Concentration of CdS (cm-3)

Fig. 1. Effect of the buffer layer thickness on Fig. 2. Effect of the donor concentration of buffer
device performance of ZnTe cells. layer on device performance of ZnTe cells.

To investigate the effect of absorber layer on device performance parameters (JSC, VOC, FF, and
PCE), the thickness of the absorber layer for both structures has been varied from 0.2 to 2μm and
illustrated in Fig. 3. These curves revealed that, JSC, VOC and PCE increases for both structures
with increasing thickness of ZnTe layer. Fast increasing tendency of these parameters is observed
up to the absorber thickness of 1µm. After that, very slow rise is taken place. The FF increases
with increasing thickness for cell without IB layer (CdS/ZnTe) wheres it decreases for cell with
IB layer (CdS/ZnTe:O/ZnTe). With increasing thickness of absorber layer, the number of
absorbed photons increases, and hence more electron-hole pairs are generated. In addition,
spectral response of absorber layer increases with increasing thickness (Bouarissa et al., 2021).
Thus, the JSC and finally PCE enhanced for thicker absorber. On the other hand, the depletion
layer gets very nearer to the back contact as the absorber layer thickness is reduced, and more
electrons are collected and recombined there. As a result, fewer electrons participating in the
generation process, resulting in a lower fill factor and lower power conversion efficiency
(Adewoyin et al., 2017; Biplab et al., 2020).
CdS/ZnTe CdS/ZnTe:O/ZnTe CdS/ZnTe CdS/ZnTe:O/ZnTe
92 74 17.8 58
72.6 18 60
90 17 17.6
91 73
72.4 16 17.4
57
PCE (%)

88
PCE (%)

15
PCE (%)
PCE (%)

17.2
FF (%)
FF (%)
FF (%)
FF (%)

90 72
72.2 14 58
86 17.0
13 56
89 71 16.8
84 72.0 12
16.6
11
82 10 56 88 70 16.4 55
71.8
0.0 0.5 1.0 1.5 2.0 0.0 0.5 1.0 1.5 2.0 1E16 1E17 1E18 1E16 1E17 1E18

1.78 1.700 11.3 46.4


1.72 1.700 12 47.6
1.76 11.2
1.70 1.698
11 46.2
Jsc (mA/cm2)

1.74
Jsc (mA/cm2)

Jsc (mA/cm2)

47.2
Jsc (mA/cm2)

1.68 11.1
Voc (V)

1.696
Voc (V)

Voc (V)

10 1.72
Voc (V)

1.66 1.695 11.0 46.0


46.8 1.70 1.694
9
1.64 1.68 10.9
45.8
8 1.692
1.62 46.4 1.66 10.8

1.60 1.690 7 1.64 1.690 45.6


0.0 0.5 1.0 1.5 2.0 0.0 0.5 1.0 1.5 2.0 1E16 1E17 1E18 1E16 1E17 1E18

Thickness of Absorber (ZnTe) Layer (m ) Thickness of Absorber (ZnTe) Layer (m) Acceptor Concentration of ZnTe (cm -3) Acceptor Concentration of ZnTe (cm -3)

Page | 918
Fig. 3. Effect of the absorber layer thickness on Fig. 4. Effect of the acceptor concentration of
device performance of ZnTe cells. absorber layer on performance of ZnTe cells.

The variation of VOC, JSC, FF, and PCE with acceptor concentration of absorber layer for all both
structures are elucidate in Figure 4. The acceptor concentration is varied from 1×1016 to 1×1018
cm-3 for the absorber layer, whereas the donor concentration of buffer layer set as 1×1018 cm-3.
The VOC, FF and PCE increases steadily, while the J SC decreases with increasing acceptor
concentration. The rising of VOC with acceptor concentration can be attributed to enhance
saturation current of the cells (Khattak et al., 2018). The reason behind the declining of J SC at
higher concentration can be explained by boosting recombination process and carrier scattering.
The probability of the collection of the photo-generated electrons reduces by boosting the
recombination process at higher carrier densities (Atowar Rahman, 2021; Khattak et al., 2018).
Moreover, at higher carrier concentrations, carrier scattering enhanced, which in turn increased
the resistivity of the absorber layer and hence reduced JSC (Atowar Rahman, 2021). These analyses
lead to set the optimum thickness as >1.5μm and 1μm for CdS/ZnTe and CdS/ZnTe:O/ZnTe
structure, respectively and acceptor concentration as 1×1017 cm-3 for both structures.
To investigate the effect of thickness of intermediate band (IB) layer on device performance
parameters, the thickness of the IB layer has been varied from 0.2 to 2μm and illustrated in Fig.
5. These curves revealed that, JSC and PCE increases with increasing thickness of ZnTe:O layer
whereas VOC and FF decreases with increasing the thickness. Fast increasing/decreasing tendency
of these parameters is observed up to the IB layer thickness of 1µm. After that, very slow change
is taken place. From these study, intermediate band (IB) layer thickness is set to1µm. The
variation of performance parameters with doping density of IB layer has also been studied (not
shown here) and set the doping density of 1×1016 cm-3 as optimum value.

CdS/ZnTe:O/ZnTe CdS/ZnTe CdS/ZnTe:O/ZnTe


90 60

54 60
85

80 48 100
PCE (%)

75 42 50
FF (%)

70 36
80
65 30
40
24
60
Jsc (mA/cm2)

18
55
QE (%)

60

0.0 0.5 1.0 1.5 2.0 0.0 0.5 1.0 1.5 2.0 30
1.76 60

50 40
1.74
20
JSC(mA/cm2)

40
V OC(V)

1.72 20
10
30

1.70
20
0 0
1.68 10
0.0 0.4 0.8 1.2 1.6 300 400 500 600 700 800

0.0 0.5 1.0 1.5 2.0 0.0 0.5 1.0 1.5 2.0
Voltage(V) Wavelength(nm)
Thickness of IB (ZnTe:O) layer (m) Thickness of IB (ZnTe:O) layer (m)

Fig. 5. Variation of Voc, Jsc, FF, and PCE with Fig. 6. (a) J-V curves and (b) QE curves of ZnTesolar cells
thickness of IB (ZnTe:O) layer. with and without IB (ZnTe:O) layer.

Page | 919
3.2 Enhancement in performance

Figure 6 shows the J-V and QE curves of optimized ZnTe solar cells with and without IB layer.
From Fig. 6(a), it is clear that the output performance of cell enhances significantly after adding
the IB layer. The PCE drastically enhances from ~15.3 to ~57.85% (about 225%) with
improvement in VOC from 1.693 to 1.697V, JSC from 10.15 to 47.26mA/cm2. The FF slightly
deteriorate from 89 to 73%. The key reasons of this performance enhancement can be explained
by improvement in photogenerated carriers. After incorporating IB layer in solar cells, both the
ZnTe:O and ZnTe layer act as absorber. In our case, ZnTe:O layer helps to improve the absorption
of lower photon energy. Thus, a greater number of photons can be captured and more electron-
hole pairs generated. This can be revealed from the QE curve of solar cell as shown in Fig. 6(b).
This figure elucidates that, the quantum efficiency (QE) of solar cell with IB (ZnTe:O) layer is
higher than that of cell without IB layer throughout the range of 400-900 nm. Therefore, the ISC
can be strongly boosted up and hence PCE in the proposed solar cell.
4. Conclusion
The CdS/ZnTe based solar cell with and without intermediate band (IB) layer has been
investigated by numerical analysis using SCAPS-1D. The basic conventional structure
(Al/ZnO/CdS/ZnTe/Pt) analyzed and optimized. After that, an intermediate band layer (ZnTe:O)
has been introduced in between the ZnTe and CdS, and the effect of this layer on device
performance is analyzed. The input physical parameters of this new structure have been optimized
depending on optimum performance of the cell. The structure with ZnTe:O IB layer
(Al/ZnO/CdS/ZnTe:O/ZnTe/Pt) gives an overall conversion efficiency of 57.85% (Voc=1.697V,
Jsc = 47.26mA/cm2 , FF= 72.19%) at optimum absorber thickness of 1μm. The efficiency
improved more than 200% by introducing intermediate band (IB) in ZnTe based solar cells. The
findings of this simulations could be an important guideline to a researcher for fabricating cost
effective and highly efficient ZnTe based intermediate band solar cells.
Acknowledgment
The authors gratefully acknowledge Dr. Marc Burgelman and his team, University of Gent,
Belgium, for providing SCAPS-1D software.
References
Adewoyin, A.D., Olopade, M.A., Chendo, M., Enhancement of the conversion efficiency of
Cu2ZnSnS4 thin film solar cell through the optimization of some device parameters. Optik
2017, 133; 122–131.
Atowar Rahman, M., Enhancing the photovoltaic performance of Cd-free Cu2ZnSnS4
heterojunction solar cells using SnS HTL and TiO2 ETL. Solar Energy. 2021, 215; 64–76.
Biplab, S.R.I., Ali, M.H., Moon, M.M.A., et al., Performance enhancement of CIGS-based solar
cells by incorporating an ultrathin BaSi2 BSF layer. Journal of Computational Electronics.
2020, 19; 342–352.
Bouarissa, A., Gueddim, A., Bouarissa, N., et. al., Modeling of ZnO/MoS2/CZTS photovoltaic
solar cell through window, buffer and absorber layers optimization. Materials Science and
Engineering: B. 2021, 263; 114816.
Burgelman, M., Decock, K., Niemegeers, A., Verschraegen, J., Degrave, S., SCAPS manual.
2021, Version: 8-4-2021, University of Gent, Belgium.
Kaneta, A., Adachi, S., Photoreflectance study in the E-1 and E1+Delta (1) transition regions of
ZnTe, J. Phys. D: Appl. Phys. 2000, 33; 901.

Page | 920
Khattak, Y.H., Baig, F., Toura, H., et. al., Effect of CZTSe BSF and minority carrier life time on
the efficiency enhancement of CZTS kesterite solar cell. Current Applied Physics. 2018, 18;
633–641.
Luque, A., Marti, A., Increasing the efficiency of ideal solar cells by photon induced transitions
at intermediate levels, Phys. Rev. Lett. 1997, 79; 5014.
Luque, A., Marti, A., Prog. Photovolt. 2001, 9; 173.
Luque, A., Marti, A., Adv. Mater. 2010, 22; 160.
Pistone, A., Arico, A.S., Antonucci, P.L., et. al., Sol. Energy Mater. Sol. Cells. 1998, 53; 255.
Song, T., Kanevce, A., Sites, J.R., Emitter/absorber interface of CdTe solar cells. Journal of
Applied Physics. 2016, 119.
Skhouni, O., Manouni, E., Mari, B., et. al., Numerical study of the influence of ZnTe thickness
on CdS/ZnTe solar cell performance. Eup. Phy. J. Appl. Phy. 2016, 74(2); 24602.
Skhouni, O., Ahmed, E., Manouni, et. al., Boosting the Performance of Solar Cells with
Intermediate Band Absorbers-The Case of ZnTe:O, J. Energy and Power Engg. 2017, 11; 417-
426.
Wang, W., Lin, A. S., Phillips, J. D., Intermediate-band photovoltaic solar cell based on ZnTe:O.
Applied Physics Letters. 2009, 95(1); 011103

Page | 921
7th International Conference on Engineering Research, Innovation and Education
January 12-14, 2023, Sylhet, Bangladesh

PAPER ID: 287

Exposition of the Issues, Difficulties, and Proposed Solutions


in Bangla Text Summarization
Roman Raihan1, M. Ebrahim Sazin2, Enamul Hassan3
● 1,2,3
Dept. Of CSE, Shahjalal University of Science & Technology, Sylhet,
Bangladesh. Email: romanraihan60@gmail.com1, sazin.me@gmail.com2,
enam-cse@sust.edu3

*Corresponding author [Roman Raihan (romanraihan60@gmail.com)]

Abstract: In this digital world, data is growing at a faster rate than ever
Keywords: before, and in this fast-paced world, it is quite impossible to filter out
necessary information by reading through all the available resources. If there
● Text
existed a way to know the summary of a large text before starting to read it
summarization, thoroughly, it would be a great help. That’s where automated text summarizers
come into play. There are two types of text summarizers: extractive text
Natural language summarizers and abstractive text summarizers. In this important field of NLP,
processing; a lot of work has been done in High Resource Languages like English. But in
Bangla, there is not enough work. In this paper, we try to find out the reasons
● Bangla Language behind this fallback. We have surveyed the existing datasets, tools, and
Abstractive and algorithms for Bangla text summarization. And we have got three major
reasons for falling back. The first one is about datasets. Bangla faces a slow
Extractive Text speed of development and public release of datasets compared to English.
Summarization; There are two reasons for the lower number of datasets in Bangla. First, the
absence of dedicated forums for building good datasets; and second, the lack
● Bangla Data Sets;
of shareable standard datasets in the public domain. The second one is about
● Automated Text tools. For languages with a lot of resources, like English, there are a lot of
summarization tools to choose from. But, it seems difficult to summarize the
Summarization. text in the Bangla language due to the limited availability of NLP tools and
techniques present. And the third one is about morphological characteristics.
Each language has unique traits, and because of its different characteristics
and morphological variations, researchers face different challenges while
doing text summarization of these languages. Though there are extractive and
abstractive text summarization approaches, these approaches do not yield good
results because of their morphological variations and the presence of no norms.
We are going to discuss the current situation of Bangla text summarization,
the drawbacks, how to overcome these bottleneck situations, and further work
for researchers in detail in this paper.

Page | 922
1. Introduction
Text summarization is the technique of shortening text without taking away its meaning. There
are two methods of doing this. The first one is extractive text summarization and the second one
is abstractive text summarization.
Extractive summarization is a method of creating a text summary from the original documents by
compiling the most important sentences and phrases. Sentences often remain in the same order as
the original source.
Abstractive summarization, on the other hand, employs linguistic approaches to grasp and
evaluate the original material to construct summaries. The objective of abstractive summarization
is to offer a comprehensive summary that uses sophisticated language development and
compression techniques to express information precisely.
Most of the work has been done in the extractive fields. But comparatively, very little work has
been done in the abstractive field. But there are still some issues in both the extractive and
abstractive fields. There are three major challenges for Bangla text summarization. The first is
the lack of good datasets, the second is the lack of NLP tools, and the third one is the very own
complex norms of the Bangla language.

2. Background
To understand the summarization issues and difficulties, we need to know some basic things in
this field. Here are some of these things:

Corpus: A group of texts is called a corpus. When we use it to study language, we call it a corpus
(plural: corpora).

Tokenization: Tokenization is the process of converting a string or text into a list of tokens.
Tokens may be thought of as pieces in the same way that a word is a token in a sentence and a
phrase is a token in a paragraph. For example, if we apply tokenization in this sentence, ‘শিম
ভ ল শিছ’, the output will be ‘শিম’, ‘ভ ল’, ‘শিছ’.

Lemmatization: Lemmatization is a Natural Language Processing (NLP) text normalization


approach that converts any kind of word to its basic root mode. Such as, in Bangla, 'কর ' , ' কশরছ',
'কশরছল ম ', 'কশরছছল ', 'করব ', 'শকর ', 'কছরব ' , 'য়কশরছছল ', 'কছরবন' etc have the sameroot ‘কর’.
Lemmatization means converting these all words to their basic word "কর ". This process is used
in the abstractive approach to generate new texts.

Stemming: Stemming is the process of removing suffixes or prefixes from a word to reduce it
to its basic form. But it does not guarantee that the base form will always be a meaningful word.
Such as, 'য়কশরছল ম', 'কছরব ', 'শকর', will be converted as 'য়কর' , 'কর', 'শকর' respectively.

Vectorization: Vectorization, or Word embeddings in natural language processing (NLP), is an


intelligent way to turn text data into numerical data.

Page | 923
3. Existing Works:
3.1 Bangla Extractive Text Summarization
Using document indexing and keyword-based information retrieval, Islam et al. [1] initially
presented Bangla extractive summarization in 2004. Uddin et al. [2] then used English extracted
text summary methods in Bangla. Das et al. [3] employed topic identification, page rank
algorithms, and other techniques for extractive summarization in 2010. Sentence ranking and
stemming process-based Bangla extractive summarization were suggested by Kamal Sarkar et
al. [4] and then enhanced by Efat et al. [5]. For the extractive approach, Haque et al. presented
two strategies. The first was a key phrase-based technique [6], while the second was a pronoun
replacement-based approach to sentence ranking [7]. Munzir et al. [8] created a Bangla extractive
summarization model based on deep neural networks. For Bangla text summarization, Abujar et
al. [8] suggested Word2Vec-based word embedding. In 2017, three approaches for extracting text
summarizing in Bangla gained popularity. The first was Abujar et al.’s [9] heuristic technique.
Akhter et al. [10] offered a K-means clustering approach as the second. Thethird technique was
an LSA (Latent Semantic Analysis) method described by Chowdhury et al. [11]. Ghosh et al. [12]
applied the graph-based phrase scoring feature for the first time in Bangla. Ullah et al. [13] offered
a cosines similarity based extractive summarizing strategy for Bangla, while Sarker et al. [4]
proposed a word frequency-based extractive summary technique for Bangla.

3.2 Bangla Abstractive Text Summarization


We have noticed that Yeasmin et al. [15] have described the various methods concerning
abstractive approaches. An LSTM-RNN-based method appears to be the most widely used
technique for Bangla abstractive text summarization that we have come across to this point. Using
Seq2Seq recurrent neural networks (RNNs), Talukder et al. [16] developed an abstractive
summarization model. They have utilized bi-directional RNNs with LSTM functioning as the
encoder and attention functioning as the decoder. They, too, made use of the Bahdanau [17]
attention mechanism, much like the majority of the others. A Seq2Seq-based long-short-term
memory (LSTM) network model was published by Bhattacharjee et al. [18]. This model included
attention to both the encoder and the decoder. Their proposed method makes use of a model that
is based on local attention and generates a long string of words that are lucid and human-like in
their created sentences, and it includes information that is significant from the original material.
Another LSTM-RNN-based text creation technique has been presented by Abujar et al. [19] for
abstractive text summarization in Bangla. Abujar et al. [20] presented a different model only
recently in the year 2021. They constructed the model with the assistance of bidirectional
recurrent neural networks (RNNs) that contained LSTM cells at both the encoding layer and the
decoding layer. They relied on the attention process to achieve the desired outcome. They chose
a Seq2Seq model, which is typically implemented in machine translation, as their guide for the
structure of the model. In addition to this, they made use of a word embedding file that has already
been trained and was created specifically for Bangla NLP research.

Page | 924
3.3 Existing Bangla Datasets

Abstractive text summarization techniques for every language usually use deep learning models.
And we know that for the training of a deep neural network-based model, we need a large dataset.
Generally, the larger our dataset is, the better the model will perform. For the Bangla dataset,
Talukder et al. [16] have collected data from various sources such as news and Facebook posts
and comments. They have summarized that data manually. Bhattacharjee et al.
[18] have also prepared a dataset of more than 19K data entities from the Bangla edition of
Bdnews24.com, which was the largest dataset on Kaggle¹ till then. As online news portals contain
a lot of advertisements and garbage words, they cleaned their dataset. Abujar et al. [21] have done
an abstractive summarization for which they have collected data from various social media.

4. Issues and Difficulties


4.1 Issues and Difficulties with the Summarization Techniques
The abstractive summarization method is better than the extractive summarization method
because it is more like a summary written by a person. But we need to come up with new sentences
to do that. And generating new sentences depends on a good lemmatizer. But because of the rules
of the Bangla language, it is hard to make such a good lemmatizer for that language. The word
"do" can be written as "do," "did," "done," "does," "doing," etc. in English. But in Bangla, we
know there are a lot more choices. This makes things hard. Again, people aren't sure which one is
the root of a word. There is no guarantee that it is a word or that the root will be the word's
conjugation. Also, there are some difficulties in text normalization. If a numerical value is present
in the text, in the Bangla language, it could have different meanings. For example, “সকু ল ১০টু
- ১২ টু র ময়ুধ7ই শুবদ7ু ৎ চয়ুল িসয়ুব।” in this sentence, “১০-১২” means from 10 to 12. But
“বু ু লু য়ুদি ভু রতয়ুক ৩-১ Cগু য়ুল হু শুরয়ুেয়ুছ।” in this sentence, “৩-১” doesn’t mean 3 to
1. This type of ambiguity hasn't been cleared in Bangla yet.
Another thing is about modeling. We've seen that more work has been done on extractive
summarization, but not much on abstractive summarization. But almost all the abstractive
approaches choose the LSTM-RNN based Seq2Seq model, which is much slower than
transformers models. Though we have Bangla BERT, it is not used by anyone so far. Another
issue is there is no programmatically usable database of ontological interpretations for Bangla
terms. For Bangla, there is no lexical database comparable to WordNet.

4.2 Issues and Difficulties with Datasets


Nowadays, deep learning techniques are frequently used for text summarization, which requires
a large training dataset. But in Bangla, there are not enough large datasets in Bangla.
Bhattacharjee et al.[1] has built a dataset of 19k entities. But they only took a single sentence as
text with a corresponding single sentence as a summary. So there is not a large dataset with multi-
sentence text with a multi-sentence summary. Again, almost every dataset is built with newspaper
content. This makes the datasets one dimensional. Whereas other languages like English have
different types of large datasets. There are some difficulties in building datasets in Bangla. There
are not good enough sources of data from where we can collect data. Most of the

Page | 925
news portals don't have abstractive summaries with their news. There is a lack of shareable
standard datasets in the public domain. Without news portals, there are not enough good sources
for data collection.

5. Research Scopes
Researchers could try Seq2Seq modeling with GRU, sometimes which is better than LSTM-
RNN. LSTM-RNN has a drawback. It is very slow compared to transformers models. So
researchers can use transformers or Bangla-BERT.
Researchers could fix the issues discussed in our issues and difficulties section. They could build
good and large datasets collected from various kinds of sources.

6. Conclusion
In this paper, we have discussed the existing work in Bangla text summarization. We have
discussed the issues and difficulties of this field. Bangla is new in this field compared to
English. So researchers have a lot of work to do. They could build tools and datasets to make
summarization better in Bangla. Bangla is an Indo-Aryan language. So maximum difficulties
are the same in these languages. Researchers could get some ideas to overcome these difficulties
by observing how other researchers have overcome the difficulties for other Indo-Aryan
languages.

References
M. T. Islam and S. M. Al Masum, “Bhasa: A corpus-based information retrieval and
summariser for bengali text,” in Proceedings of the 7th International Conference on
Computer and Information Technology, 2004.

M. N. Uddin and S. A. Khan, “A study on text summarization techniques and implementfew


of them for bangla language,” in 2007 10th international conference on computer and
information technology. IEEE, 2007, pp. 1–4.

Das and S. Bandyopadhyay, “Topic-based bengali opinion summarization,” in Coling 2010:


Posters, 2010, pp. 232–240.

K. Sarkar, “Bengali text summarization by sentence extraction,” arXiv preprint


arXiv:1201.2240, 2012.

M. I. A. Efat, M. Ibrahim, and H. Kayesh, “Automated bangla text summarization by sentence


scoring and ranking,” in 2013 International Conference on Informatics, Electronics and
Vision (ICIEV). IEEE, 2013, pp. 1–5. -20-

Page | 926
M. M. Haque, S. Pervin, and Z. Begum, “Enhancement of keyphrase-based approach of
automatic bangla text summarization,” in 2016 IEEE Region 10 Conference (TENCON).
IEEE, 2016, pp. 42–46.
——, “An innovative approach of bangla text summarization by introducing pronoun
replacement and improved sentence ranking,” Journal of Information Processing
Systems, vol. 13, no. 4, pp. 752–777, 2017.
Al Munzir, M. L. Rahman, S. Abujar, S. A. Hossain et al., “Text analysis for bengali text
summarization using deep learning,” in 2019 10th International Conference on
Computing, Communication and Networking Technologies (ICCCNT). IEEE, 2019, pp.
1–6.
S. Abujar, A. K. M. Masum, M. Mohibullah, S. A. Hossain et al., “An approach for bengalitext
summarization using word2vector,” in 2019 10th International Conference on Computing,
Communication and Networking Technologies (ICCCNT). IEEE, 2019, pp. 1–5.
S. Abujar, M. Hasan, M. Shahin, and S. A. Hossain, “A heuristic approach of text
summarization for bengali documentation,” in 2017 8th International Conference on
Computing, Communication and Networking Technologies (ICCCNT). IEEE, 2017,
pp. 1–8.
S. Akter, A. S. Asa, M. P. Uddin, M. D. Hossain, S. K. Roy, and M. I. Afjal, “An extractivetext
summarization technique for bengali document (s) using k-means clustering algorithm,”
in2017 IEEE International Conference on Imaging, Vision & Pattern Recognition
(icIVPR). IEEE, 2017, pp. 1–6.
S. R. Chowdhury, K. Sarkar, and S. Dam, “An approach to generic bengali text summarization
using latent semantic analysis,” in 2017 International Conference on Information
Technology (ICIT). IEEE, 2017, pp. 11–16.
P. P. Ghosh, R. Shahariar, and M. A. H. Khan, “A rule based extractive text summarization
technique for bangla news documents.” International Journal of Modern Education &
Computer Science, vol. 10, no. 12, 2018. -21-
S. Ullah, S. Hossain, and K. A. Hasan, “Opinion summarization of bangla texts using cosine
simillarity based graph ranking and relevance based approach,” in 2019 International
Conference on Bangla Speech and Language Processing (ICBSLP). IEEE, 2019, pp.
1–6.
S. Yeasmin, P. B. Tumpa, A. M. Nitu, M. P. Uddin, E. Ali, and M. I. Afjal, “Study of
abstractive text summarization techniques,” American Journal of Engineering Research,
vol. 6, no. 8, pp. 253–260, 2017.
M. A. I. Talukder, S. Abujar, A. K. M. Masum, F. Faisal, and S. A. Hossain, “Bengali
abstractive text summarization using sequence to sequence rnns,” in 2019 10th
International

Page | 927
Conference on Computing, Communication and Networking Technologies (ICCCNT). IEEE,
2019, pp. 1–5.

[1] D. Bahdanau, K. Cho, and Y. Bengio, “Neural machine translation by jointly learning to
align and translate,” arXiv preprint arXiv:1409.0473, 2014.

[2] P. Bhattacharjee, A. Mallick, S. Islam et al., “Bengali abstractive news summarization


(bans): a neural attention approach,” in Proceedings of International Conference on Trends in
Computational and Cognitive Engineering. Springer, 2021, pp. 41–51.

[3] S. Abujar, A. K. M. Masum, S. Islam, F. Faisal, S. A. Hossain et al., “A bengali text


generation approach in context of abstractive text summarization using rnn,” in Innovations in
Computer Science and Engineering. Springer, 2020, pp. 509–518.

[4] M. M. Islam, M. Islam, A. K. M. Masum, S. Abujar, and S. A. Hossain, “Abstraction based


bengali text summarization using bi-directional attentive recurrent neural networks,” in
Emerging Technologies in Data Mining and Information Security, A. E. Hassanien, S.
Bhattacharyya, S. Chakrabati, A. Bhattacharya, and S. Dutta, Eds. Singapore: Springer Nature
Singapore, 2021, pp. 317–327. [23] H. P. Luhn, “The automatic creation of literature abstracts,”
IBM Journal of research and development, vol. 2, no. 2, pp. 159–165, 1958.

[5] F. A. Fouzia, M. A. Rahat, M. T. Alie Al Mahdi, A. K. M. Masum, S. Abujar, and S. A.


Hossain, “A bengali text summarization using encoder-decoder based on social media dataset,”
in Emerging Technologies in Data Mining and Information Security, A. E. Hassanien, S.
Bhattacharyya, S. Chakrabati, A. Bhattacharya, and S. Dutta, Eds. Singapore: Springer Nature
Singapore, 2021, pp. 539–549.

Page | 928
7th International Conference on Engineering Research, Innovation and Education
January 12-14, 2023, Sylhet, Bangladesh

PAPER ID: 318

Electronic Structure and Charge Transport Properties of Copper-,


Silver- and Gold- Quinolates utilizing Dispersion-Corrected Density
Functional Theory for Application in Organic Light Emitting Diodes

Md. Rakib Hossain, Ahsan Ullah, Sajib Kumar Mohonta and Nazia Chawdhury*

Department of Physics, Faculty of Physical Science, Shahjalal University of Science and


Technology, Sylhet, Bangladesh.
Author Email Addresses: rakib82hossain@gmail.com; ahsan10@student.sust.edu; skmohonta-
phy@sust.edu, nc-phy@sust.edu
*Corresponding author [Nazia Chawdhury]

Keywords: Abstract: The electronic and charge transport properties of Metal-


Quinolates (Metal = Cu, Ag, Au) compounds are theoretically
● Metal-Quinolate; investigated and compared with Lithium Quinolatesusing Austin-
● Charge Transport; Frisch-Petersson functional with dispersion (APFD) corrected
density functional theory (DFT). The calculated energy gap
● Density Functional
between Highest Occupied Molecular Orbital and Lowest
Theory; Unoccupied Molecular Orbital of LiQ, CuQ,AgQ and AuQ are
● APFD Functional; 3.40 eV. 2.49 eV, 1.91 eV and 2.15 eV. The ionization potential,
● Molecular Orbital. electron affinity and chemical hardness of the compounds are
calculated. We find that the charge transport rate of AgQ is higher
than that of LiQ, CuQ and AuQ. The chemical hardness of all
complexes is recorded and found that AgQ has the higher value
of chemical hardness. Our study suggest that AgQ is the most
efficient charge transport material for OLEDs.

1. Introduction

Tang et al. presented the first study on organic light-emitting diodes (OLEDs) in 1987.1,2 OLEDs
have been actively promoted because of their potential for use in flexible displays, solid-state
lighting, etc.3,4 However, research on improving the material properties of devices has been
extensive. As a result, the charge transfer property is one of the most crucial aspects for grasping
the physical properties of organic compounds.5,6
The feasibility of lithium-quinolate complexes such as 8-hydroxy-quinolinato lithium (LiQ) as
electron injection layers (EILs) and electron transport layers (ETLs) in an organic light-emitting
diode has been investigated by several researchers (OLED).7-12 LiQ has been the most widely
utilized quinolate material.12 It serves as an electron injection layer (EIL) since it reduces the

Page | 929
energy barrier between the cathode and the electron transport layer (ETL), enabling more efficient
electron injection from the cathode. However, LiQ's poor electron carrier mobility may hinder
ETL's electron hopping.13
In order to comprehend the impact of metal ions on electronic and charge transport characteristics
and to address low electron mobility, Jeon et al. recently theoretically explored a series of M-
quinolate complexes (M=singly oxidized metal ions) using the nonlocal Becke's three
parameterized Lee-Yang-Parr functional (B3LYP).14 The B3LYP functional does not account for
the influence of van der Waals interactions with extended range, which are prevalent in nature.
This research evaluates the electrical characteristics of a range of metal quinolate complexes using
dispersion-corrected density functional theory. APFD is a hybrid functional that modifies the
regular Austin-Frisch-Petersson functional by including dispersion adjustments.15 APFD provides
significantly superior performance for a selection of compounds compared to B3LYP.16,17

2. Methodology
DFT simulations of M-quinolate materials (M = Li, Cu, Ag, Au) are executed with unrestrained
APFD at the LanL2MB level of theory in the Gaussian 16 software.18 The charge transport rate
at the molecular level has derived by Marcus formalism.19
4π2 1 2 −[
λ
]

Khole/electron = Jh/e e 4kBT (1)


h
√4πλhole/electron kBT

where h: Planck’s constant, kB: Boltzmann constant, T: temperature, λhole/electron: the reorganization
energy for hole or electron, and Jh/e: the transfer integral for hole or electron.

Figure 1: Potential energy curve for neutral and cationic structure.

The transfer integral for the hole (Jh) and electron (Je) have approximated from Koopman’ theory
with equation (2)20
HOMO−(HOMO−1) (LUMO+1)−LUMO (2)
J = , J =
h 2 e 2
To comprehend the property of charge transfer, the reorganization energy is the primary
determinant. Under a reversible charge transfer process, the sum of relaxation energies (λ1 and λ2

Page | 930
in Figure 1.) may be utilized to estimate the inner reorganization energies for hole and electron.
The equations can be defined as 21
𝜆ℎ = (𝐸𝑐𝑎𝑡𝑖𝑜𝑛 − 𝐸𝑐𝑎𝑡𝑖𝑜𝑛) + (𝐸𝑛𝑒𝑢𝑡𝑟𝑎𝑙 − 𝐸𝑛𝑒𝑢𝑡𝑟𝑎𝑙) (3)
𝑛𝑒𝑢𝑡𝑟𝑎𝑙 𝑐𝑎𝑡𝑖𝑜𝑛
𝜆𝑒 = (𝐸𝑎𝑛𝑖𝑜𝑛 − 𝐸𝑎𝑛𝑖𝑜𝑛) + (𝐸𝑛𝑒𝑢𝑡𝑟𝑎𝑙 − 𝐸𝑛𝑒𝑢𝑡𝑟𝑎𝑙) (4)
𝑛𝑒𝑢𝑡𝑟𝑎𝑙 𝑎𝑛𝑖𝑜𝑛

3. Result and Discussions


3.1 Optimized structures
Figure 2 depicts the DFT/UAPFM molecular structures of M-quinolate (M=Li, Cu, Ag, Au).
Table 1 displays the interatomic Distances between metal-nitrogen (N-M) and metal-oxygen (O-
M) and the Nitrogen- Metal-Oxygen (NMO) bond Angles. The organization of the structure is
determined by the oxidation states of the metal ions. In all complexes, the N-M bond lengths are
somewhat longer than the O-M bond lengths because metal-nitrogen has lower binding energythan
metal-oxygen.

Figure 2: The optimized structures of M-quinolate complexes.

In addition, the estimated bond length of N-M and O-M for group XI transition metals increases
monotonically with the atomic number of metal ions. Furthermore, the binding angle of oxygen-
metal-nitrogen reduced as the atomic number of metal ions increased in the same group.
Table 1: Optimized structural parameters, bond length (Å), and inter-ring angle (⁰) of M-quinolate
materials obtained by DFT/UAPFD/LanL2MB level.

Molecules N-M O-M NMO

LiQ 1.884 1.635 100.53

CuQ 2.017 1.913 94.14

AgQ 2.290 2.136 81.74

AuQ 2.544 2.145 75.33

Page | 931
3.2 Frontier molecular orbitals
From the frontier molecular orbitals, HOMOs,
and LUMOs of all materials, we determine
that the electron density distribution of
HOMO and LUMO is significantly different
for CuQ, AgQ, and AuQ compared to LiQ.
The metal ion dominates the LUMO state of
CuQ, AgQ, and AuQ, whereas the quinolate
moiety dominates the HOMO state. The
difference in electron density between the
HOMO and LUMO levels is frequently
referred to as a charge transfer property,
specifically ligand-to-metal charge transfer
(LMCT).22,23 However, in LiQ we observe
ligand-to-ligand charge transfer character or
π-π*with small contribution from lithium
atom. We find a negative electrostatic
potential between aromatic carbon and
oxygen and a positive one near the metal. For
Figure 3: The molecular orbital distribution and
Group XI metal ions, the electrostatic electrostatic potential of the M-quinolate
potential of electrophilic region rises as the complexes.
atomic number of the metal ion rises.

3.3 Electronic energy levels


The inherent conductivity of materials with a
narrower band gap offers substantial practical
potential in electrical devices. Consequently,
it is essential to regulate the band gap. Figure
4 illustrates the HOMO and LUMO energy
levels and the computed energy gap between
them. LiQ has the highest energy gap among
the four molecules, whereas AgQ has the
lowest.
Yadav et al.24 observed that the energy gap
must be narrower for significant charge
recombination since the LUMO levels
contribute more to the recombination in
organic layers than the HOMO levels. The
calculated LUMO energy levels of CuQ,
AgQ, and AuQ are much more stable than Figure 4: The HOMO and LUMO energy levels of
LiQ, CuQ, AgQ, and AuQ.
those of LiQ. Thus, it is predicted that CuQ,
AgQ, and AuQ will inject electrons from the
cathode more effectively.

Page | 932
3.3 Ionization potential, electron affinity, and chemical hardness
Optoelectronic devices rely on the ability to
inject and transfer charges in a balanced way
for optimal performance.25 Therefore, 5

Electron affinity (EA), a measure of electron 4 IP


EA
transport, is one of the essential elements in 3

assessing whether or not an organic

Energy (eV)
2
semiconductor is acceptable for use in
1
organic electroluminescent devices. Accurate
EA values are necessary to construct effective 0

organic light-emitting diode devices -1

(OLEDs). The total energy gap between -2

anionic and neutral molecules is the unit of LiQ CuQ AgQ AuQ
measurement for electron affinity.26 In
Increasing atomic number of metal ions
addition, it is necessary to understand their
ionization potential (IP) to design the Figure 5: The ionization potentials and electron
structure of organic molecules for organic affinities for LiQ, CuQ, AgQ, and AuQ.
electronics.
We have analyzed IP, EA, and chemical hardness to get insight into the electrical structure of
metal complexes. AgQ has the lowest IP (Figure 5). These findings follow the same trend as the
calculation of Jeon et al.14. Comparing the theoretical conclusions with the practical reality in a
fair quantitative or qualitative way is a crucial step in proving the relevance of this study. For
example, LiQ has the highest chemical hardness among the four compounds examined, as shown
in Table 2. Additionally, the chemical hardness of AgQ and AuQ is almost comparable, and both
are lower than that of LiQ and CuQ. AgQ and AuQ are likely the most electrochemically reactive
materials for electron/hole transport in OLEDs based on these results.
Table 2: The HOMO, LUMO, energy gap (Eg), ionization energy (IP), electron affinity (EA),
and chemical hardness values (𝜂) of LiQ, CuQ, AgQ and AuQ.

Molecules LiQ CuQ AgQ AuQ

LUMO+1(eV) 1.185 -0.101 -0.006 -0.049

LUMO (eV) 0.413 -1.107 -1.609 -2.148


HOMO (eV) -2.989 -3.593 -3.517 -4.303

HOMO-1(eV) -3.971 -4.3110 -4.5096 -5.1017

Eg(eV) 3.4019 2.4862 1.9077 2.1548


IP (eV) 4.7202 4.9744 4.7905 5.3097

EA (eV) -2.1287 -0.2046 0.1856 0.7211

𝜂 3.4245 2.5895 2.3024 2.2943

Page | 933
3.3 Charge transfer properties 70

Kelectron
Figure 6 shows the hole and electron transport 60 Khole

Charge transport rate (X1013)


rates in M-quinolate complexes. We observe 50
that Kelectron of AgQ is more than six times
higher than Kelectron of LiQ that makes it a 40

promising ETL for OLEDs. 30

20

4. Conclusion 10

The effect of metal ions on electrical and 0

charge transport properties in various M-


Figure 6: The comparative electron transfer rate
quinolate compounds (where M=Li, Cu, Ag,
of LiQ, CuQ, AgQ, and AuQ.
and Au) are theoretically investigated using
dispersion-corrected density functional

theory. Detailed data on charge transport characteristics like transfer integrals, reorganization
energies, and intermolecular charge hopping rates are presented. The obtained theoretical results
suggest that AgQ performs better than LiQ as EIL and ETL additives in all areas of organic
electronics. We expect that the experimentalists may verify our calculation using APFD
functional.

Acknowledgment
We are grateful to Ministry of Education, Government of the People’s Republic of Bangladesh
for financial support to this project (ID PS217535).

References
Tang, C.W., VanSlyke, S.A., (1987). Organic electroluminescent diodes. Appl. Phys. Lett. 51,
913–915. https://doi.org/10.1063/1.98799.
Chien, C.-H., Chen, C.-K., Hsu, F.-M., Shu, C.-F., Chou, P.-T., Lai, C.-H., (2009).
Multifunctional Deep-Blue Emitter Comprising an Anthracene Core and Terminal
Triphenylphosphine Oxide Groups. Adv. Funct. Mater. 19, 560–566.
https://doi.org/10.1002/adfm.200801240.
Kang, S., Lee, J.Y., Kim, T., (2019). Unveiling the Root Cause of the Efficiency-Lifetime Trade-
Off in Blue Fluorescent Organic Light-Emitting Diodes. Electron. Mater. Lett. 16, 1– 8.
https://doi.org/10.1007/s13391-019-00180-5.
Van Slyke, S.A., Chen, C.H., Tang, C.W., (1996). Organic electroluminescent devices with
improved stability. Appl. Phys. Lett. 69, 2160–2162. https://doi.org/10.1063/1.117151.
Rohloff, R., Kotadiya, N.B., Crăciun, N.I., Blom, P.W.M., Wetzelaer, G.A.H., (2017). Electron
and hole transport in the organic small molecule α-NPD. Appl. Phys. Lett. 110, 073301.
https://doi.org/10.1063/1.4976205.

Page | 934
Kang, S., Moon, J.H., Kim, T., Lee, J.Y., (2019). Design of efficient non-doped blue emitters:
toward the improvement of charge transport. RSC Adv. 9, 27807–27816.
https://doi.org/10.1039/c9ra04918e.
Singh, D., Bhagwan, S., Saini, R.K., Tanwar, V., Nishal, V., (2016). Optoelectronic Properties of
Color-Tunable Mixed Ligand-Based Light-Emitting Zinc Complexes. Journal of Elec Materi
45, 4865–4874. https://doi.org/10.1007/s11664-016-4721-0.
Kido, J. and Hayashi, T., International Manufacturing and Engineering Services Co Ltd, (2003).
Organic electroluminescent device. U.S. Patent Application 10/393,952.
Schmitz, C., Schmidt, H.-W., Thelakkat, M., (2000). Lithium−Quinolate Complexes as Emitter
and Interface Materials in Organic Light-Emitting Diodes. Chem. Mater. 12, 3012– 3019.
https://doi.org/10.1021/cm0010248.
Zhou, D.Y., Cai, S.D., Gu, W., Liao, L.S., Lee, S.T., (2010). Efficiency dependence on alkalimetal
compound/Al bilayer cathode in organic light-emitting diodes. Appl. Phys. Lett. 97, 223302.
https://doi.org/10.1063/1.3522883.
Zheng, X., Wu, Y., Sun, R., Zhu, W., Jiang, X., Zhang, Z., Xu, S., (2005). Efficiency
improvement of organic light-emitting diodes using 8-hydroxy-quinolinato lithium as an
electron injection layer. Thin Solid Films 478, 252–255.
https://doi.org/10.1016/j.tsf.2004.08.020.
Lee, W., Jung, J.W., (2016). High performance polymer solar cells employing a low-
temperature solution-processed organic–inorganic hybrid electron transport layer. J. Mater.
Chem. A 4, 16612–16618. https://doi.org/10.1039/c6ta06911h.
Kim, J.-H., Park, J.-W., (2017). Designing an electron-transport layer for highly efficient, reliable,
and solution-processed organic light-emitting diodes. J. Mater. Chem. C 5, 3097– 3106.
https://doi.org/10.1039/c7tc00488e.
Jeon, S.H., Cho, Y.M., Kim, T., Kang, S., (2019). Discovering new M-quinolate materials:
theoretical insight into understanding the charge transport, electronic, self-aggregation
properties in M-quinolate materials (M = Li, Na, K, Rb, Cs, Cu, Ag, and Au). J Mater Sci 54,
9523–9532. https://doi.org/10.1007/s10853-019-03584-8.
Austin, A., Petersson, G.A., Frisch, M.J., Dobek, F.J., Scalmani, G., Throssell, K., (2012). A
Density Functional with Spherical Atom Dispersion Terms. J. Chem. Theory Comput. 8,
4989–5007. https://doi.org/10.1021/ct300778e.
Malloum, A., Fifen, J.J., Conradie, J., (2018). Solvation energies of the proton in methanol
revisited and temperature effects. Phys. Chem. Chem. Phys. 20, 29184–29206.
https://doi.org/10.1039/c8cp05823g.
Hansen, P.E., Saeed, B.A., Rutu, R.S., Kupka, T., (2020). One-bond 1 J (15 N,H) coupling
constants at sp2 -hybridized nitrogen of Schiff bases, enaminones and similar compounds: A
theoretical study. Magn Reson Chem 58, 750–762. https://doi.org/10.1002/mrc.5052
Frisch, M.J.E.A., Trucks, G.W., Schlegel, H.B., Scuseria, G.E., Robb, M.A., Cheeseman, J.R.,
Scalmani, G., Barone, V., Mennucci, B., Petersson, G. and Nakatsuji, H., (2009). Gaussian
09, Revision d. 01, Gaussian. Inc., Wallingford CT, 201.

Page | 935
Marcus, R.A., Sutin, N., (1985). Electron transfers in chemistry and biology. Biochimica et
Biophysica Acta (BBA) - Reviews on Bioenergetics 811, 265–322.
https://doi.org/10.1016/0304-4173(85)90014-x.
Koopmans, T., (1934). Über die Zuordnung von Wellenfunktionen und Eigenwerten zu den
Einzelnen Elektronen Eines Atoms. Physica 1, 104–113. https://doi.org/10.1016/s0031-
8914(34)90011-2.
Kang, S., Kim, T., (2020). The nature of electronic and carrier transport properties of M-
(quinolate)2 complexes (M = Be, Mg, Ca, and Sr): A computational study. Organic
Electronics 87, 105980. https://doi.org/10.1016/j.orgel.2020.105980.
Liang, Y., Guo, Y., Wang, Ye, Zhang, S., Pan, Y., Shi, Y., Wang, Yanan, Zhu, L., Jin, B., Sun,
Y., Zhang, B., Feng, X., Yuan, M., Wang, H., Zhao, G., (2020). Combined ultrafast
spectroscopic and TDDFT theoretical studies on dual fluorescence emissions promoted by
ligand-to-metal charge transfer (LMCT) excited states of tungsten-containing organometallic
complexes. Chemical Physics Letters 748, 137396.
https://doi.org/10.1016/j.cplett.2020.137396.
Han, D., Liu, J., Miao, R., Zhao, L., Zhang, G., (2015). Theoretical design study on the electronic
structure and photophysical properties of a series of osmium (II) complexes with different
ancillary ligands. Polyhedron 85, 506–
510.https://doi.org/10.1016/j.poly.2014.09.018.
Yadav, R.A.K., Dubey, D.K., Chen, S.-Z., Liang, T.-W., Jou, J.-H., (2020). Role of Molecular
Orbital Energy Levels in OLED Performance. Sci Rep 10. https://doi.org/10.1038/s41598-
020-66946-2.
Chitpakdee, C., Namuangruk, S., Khongpracha, P., Jungsuttiwong, S., Tarsang, R., Sudyoadsuk,
T., Promarak, V., (2014). Theoretical studies on electronic structures and photophysical
properties of anthracene derivatives as hole-transporting materials for OLEDs.Spectrochimica
Acta Part A: Molecular and Biomolecular Spectroscopy 125, 36–45.
https://doi.org/10.1016/j.saa.2013.12.111.
Yoshida, H., Yoshizaki, K., (2015). Electron affinities of organic materials used for organic light-
emitting diodes: A low-energy inverse photoemission study. Organic Electronics 20, 24–30.
https://doi.org/10.1016/j.orgel.2015.01.037.
Kwak, S.W., Hong, J.H., Lee, S.H., Kim, M., Chung, Y., Lee, K.M., Kim, Y., Park, M.H., (2020).
Synthesis and Photophysical Properties of a Series of Dimeric Indium Quinolinates.
Molecules 26, 34. https://doi.org/10.3390/molecules26010034
Mahmoud, M.E., Haggag, S.S., Abdel-Fattah, T.M., (2009). Surface layer-by-layer chemical
deposition reaction for thin film formation of nano-sized metal 8-hydroxyquinolate
complexes. Polyhedron 28, 181–187. https://doi.org/10.1016/j.poly.2008.09.03

Page | 936
7th International Conference on Engineering Research, Innovation and Education
January 12-14, 2023, Sylhet, Bangladesh

PAPER ID: 325

Cyberbullying detection on teenage users' Bangla comments


from Facebook using machine learning

Srabon Bhowmik Shanto1,*, Mohammed Jahirul Islam2 and Md. Abdus Samad3

1,2,3
Institute of Information and Communication Technology, Shahjalal University of Science
and Technology, Sylhet, Bangladesh
*Corresponding author [Srabon Bhowmik Shanto]

Abstract: The consequence of Covid-19 lockdown resulted in the increased


Keywords: amount of social media use among the large number of teenage population in
Bangladesh. Reckless thoughts sharing are resulted into cyberbullying which
● Cyberbullying should be monitored and controlled. Teenagers likely to pass their leisure time
detection; on such platforms with freedom to post and share anything. There is no
research dedicatedly done with teenage users comments on social media to
● Machine Learning ; detect cyberbullying. The goal of this study is to automatically detect bullied
● Natural Language comments posted by teenage users on Facebook using Machine Learning
algorithms. The dataset has total 7072 comments mainly collected from
Processing;
Facebook. Collected data were pre-processed, tokenized and labelled (e.g.,
● Bangla Sentiment bully and not bully) before training. Thereafter processed data were feed into
nine machine learning models. Result has shown that Support Vector Machine
Analysis; (Radial Basis Function kernel) resulted highest accuracy of 76.2% in unigram.
● Teenage users’ In case of Bigram, Support Vector Machine with linear kernel gained highest
accuracy of 76.4%, whereas for trigram the same algorithm outperformed
sentiment.
other algorithms with an accuracy of 76%. Utilizing the outcome of this study,
researcher may get information about teenage mental health conditions.
Furthermore, appropriate models can be used to automatically detect
cyberbullying to ensure safe online environment for the users.

1. Introduction
Cyberbullying is when someone intentionally, repeatedly, and violently harms or harasses another
person using information and communication technology Slonje et al. (2013). Cyberbullying has
become a significant problem now a days. Online bullying has become a significant social
problem as a result of the widespread use of social media platforms. People will experience severe
psychological harm as a result of this phenomenon, which will limit their capacity for healthy
growth and development Olweus (2012).
Bangladesh only recently began providing fast internet service and most people regularly use the
internet in their daily lives. Furthermore, teenagers of school/college age now prefer to spend
their free time on social networking sites most like Facebook Islam et al. (2020).

Page | 937
Maximum research work on cyberbullying detection is in the English language text. But in Bangla
language, a few work is found to automatically detect cyberbullying in Bangla language text.
Moreover there is no research which is dedicatedly done with the data of teenagers’ comments to
detect the cyberbullying. Detecting Cyberbullying in Bangla language will be useful for
preventing cybercrime such as online harassment and blackmailing which are becoming one of
the major concern in Bangladesh nowadays Rahman et al. (2017). This research goal is to detect
cyberbullying from teenagers’ Bangla comments which are collected from various Facebook
groups/page where teenage users share their opinions, and views in comment.
We trained nine machine learning models with this data and selected the model which performed
the highest accuracy so that it can be used to detect cyberbullying comments automatically.

2. Related works
Many researchers are working hard to identify cyberbullying across many languages. Therefore,
there are some earlier studies in this area that are still available. Ahammed et al. (2019) divided
the process of identifying hate speech in Bangla into three main steps. The authors has compiled
data from Facebook. They incorporated hate speech using machine learning. The maximum F1
score for the SVM algorithm was 0.71, and the maximum F1 score for the Naive Bayes algorithm
was 0.73.
Al- Amin et al. (2017) tried to identify the sentiment or opinion of a full text, whether it was
positive or negative, using posts from Bangla Microblogs. They created the training corpus using
a semi-supervised bootstrapping methodology and used Support Vector Machines (SVM) and
Maximum Entropy for classification (MaxEnt). When the word2vec word co-occurrence score
and the sentiment polarity score of the words were combined, the accuracy for the authors' attempt
to categorize the sentiment of Bengali comments was 75.5%.
To determine the sentiment of Bengali Facebook Status, Islam et al. (2016) used the Naive Bayes
Model. They obtained a 0.72 F-score by combining the Bigram and Naive Bayes models.Using a
new algorithm, Hussain et al. (2018) identified abusive Bangla text. The weights of the test
comments are determined using their suggested algorithm, which gives words a "abusive weight"
and a "non-abusive weight" based on the word's usage. 71.6% of their work was accurate.
The objective of this study is to use machine learning algorithms to automatically identify
bullying comments made by teenage users on Facebook.

3. Methodology
3.1 Dataset Collection and Preparing

Preparing the dataset was one of the challenging tasks of this research. We have selected some
groups and pages and collected teenage user comments with the help of instant data scraper
chrome extension. We have added ⅓ data from two existing datasets to increase the number of
comments in our dataset (Karim et al., 2021; Ahmed et al., 2021). Later we labelled the data as
bully (4504 comments) and not bully (2568 comments). To review teenage bullying comments

Page | 938
we categorized them into 4 subcategories e.g., (personal 1575 comments, geopolitical 1198
comments, religious 664 comments, and political 570 comments).

Figure 1: Number of comments in different subcategories (left) and labels (right)

3.2 Data Processing

For data preprocessing, we have removed emoji’s from the data. In next steps we removed
punctuation and special characters e.g., https, ০-৯, different language alphabets and so on.
Next, we have tokenized data and removed stopwords e.g., রু খু , রেয়ুেয়ু, রকম, শুধ,ু
সেয়ু.
Finally we have joined the rest words and data were ready to feed into the machine learning
models.

3.3 Feature extraction

Machine learning algorithms cannot extract features from dataset by its own. In this research we
have used Tf-idf vectorizer with different N-grams e.g., unigram, bigram, and trigram as
feature extraction technique.
Rahman et. al (2020) said, one of the easiest to measure feature extraction methods in NLP is the
TF-IDF. A word’s frequency of occurrence in a document is indicated by its term frequency,
whereas a word’s frequency across all documents is indicated by its inverse document frequency.
For TF-IDF, the formula is as follows:
𝐾
𝑡𝑓𝑖𝑑𝑓𝑤 = 𝑓𝑤,𝑟 × log ( ) ..................... (1)
𝑟𝑓𝑤

𝑤ℎ𝑒𝑟𝑒, 𝑡𝑓𝑖𝑑𝑓𝑤 = 𝑤𝑒𝑖𝑔ℎ𝑡 𝑜𝑓 𝑡𝑒𝑟𝑚 𝑤, 𝑓𝑤,𝑟 = 𝑓𝑟𝑒𝑞𝑢𝑒𝑛𝑐𝑦 𝑜𝑓 𝑡𝑒𝑟𝑚 𝑤 𝑖𝑛 𝑑𝑜𝑐𝑢𝑚𝑒𝑛𝑡 𝑟, 𝐾 =


𝑡𝑜𝑡𝑎𝑙 𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑑𝑜𝑐𝑢𝑚𝑒𝑛𝑡𝑠, 𝑟𝑓𝑤 = 𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑑𝑜𝑐𝑢𝑚𝑒𝑛𝑡𝑠 𝑐𝑜𝑛𝑡𝑎𝑖𝑛𝑖𝑛𝑔 𝑡ℎ𝑒 𝑡𝑒𝑟𝑚 𝑤.

N-gram is a sequence of the N-words in the modeling of NLP. Let us take an example to explain
N-gram. “ির কত অপমু নশুত হেয়ু তু ই কনরেমর বু সু ছু িস বন্ধ করনব”. A single word
sequence makes up a one-gram or unigram. As for the above example , in one gram or unigram
it can be
i.e. “ির”, “কত”, “অপমু শনত”, “হছে”, “িসু ”, “বন্ধ”. Two-gram or the bi-gram contains two-word
sequence i.e. “ির কত”, “কত অপম শনত”, or “অপম শনত হছে”. For three-gram or the tri-gram, it has
three words sequences i.e.”ির কত অপমু নশুত”, “হছে তু ই কনরেমর” or “িস বন্ধ করনব”.

Page | 939
Figure 2: Flowchart of the methodology

3.4 Machine Learning Model Training

Nine machine learning models has been selected to perform with the preprocessed data e.g.,
Linear SVM, Sigmoid SVM, Polynomial SVM, RBF SVM, MNB, RF, Logistic Regression, DT,
KNN, SGD, Adaboost , XGBoost. 70% of the data used for training and 30% of the data used for
testing and fit machine learning model with these data.
For implementing SVM we have used all of its existing kernel e.g., linear, polynomial with degree
= 2, rbf, sigmoid. On Random Forest implementation we have used random_state=10, 99
estimators. In KNN we divided the clusters into 2. In Stochastic Gradient Descent 12 penalty and
99 max iteration. With Adaboost we have used 50 estimators.

4. Result Calculation and Discussion


After implementing all the machine learning algorithms with tf-idf ngrams we get following
results:

Table 1: Accuracy Matrices

Page | 940
From the above table we can see, for ‘unigram’, RBF SVM performed highest accuracy of 76.2%,
Multinomial Naïve Bayes performed highest precision value of 77.5%, Random Forest performed
highest recall value of 72%, RBF SVM performed highest F1-score of 72%. For ‘bigram’, Linear
SVM performed highest accuracy of 76.4%, Logistic Regression performed highest precision
value of 76.5%, Random Forest performed highest recall value of 71.5% and also performed
highest F1-score of 72%. For ‘trigram’, Linear SVM performed highest accuracy of 76%,
Multinomial Naïve Bayes performed highest precision value of 81.5%, Random Forest performed
highest recall value of 71.5% and also performed highest F1-score of 72%.

In this study, we used 7072 data, but we discovered that machine learning models can perform
more accurately with more data. Faisal et al. (2021) had achieved 85% for accuracy, precision,
and recall and.84% in F1 score with SVM and they had 44001 comments in their dataset.

5. Conclusion and Future work


Bullies on Facebook can be challenging to spot in Bengali Language. This is the result because
of the language's diversity and the way comments are typed in it. We would be increasing the size
of our dataset in future. The results of our research show that the support vector machine had the
highest accuracy. We will be implementing some deep learning models like LSTM, GRU, CNN,
and DNN in the future to raise the evaluation score.

References
Ahammed, S., Rahman, M., Niloy, M. H. et al., Implementation of machine learning to detect
hate speech in Bangla language, 8th International Conference System Modeling and
Advancement in Research Trends (SMART), IEEE. 2019; 317-320.

Ahmed, M. F., Mahmud, Z., Biash, Z. T. et al., Cyberbullying detection using deep neural
network from social media comments in bangla language, arXiv preprint
arXiv:2106.04506.2021.

Al-Amin, M., Islam, M. S., & Uzzal, S. D., Sentiment analysis of Bengali comments with
Word2Vec and sentiment information of words, International conference on
electrical, computer and communication engineering (ECCE), IEEE. 2017;186-
190.

Hussain, M. G., Al Mahmud, T., & Akthar, W., An approach to detect abusive bangla text,
International Conference on Innovation in Engineering and Technology (ICIET), IEEE.
2018;1-5.

Islam, M. S., Islam, M. A., Hossain, M. A et al., Supervised approach of sentimentality


extraction from bengali facebook status,19th international conference on
computer andinformation technology (ICCIT), IEEE. 2016; 383-387.

Islam, A. N., Laato, S., Talukder, S. et al., Misinformation sharing and social media fatigue
during COVID-19: An affordance and cognitive load perspective, Technological
forecastingand social change. 2020; 159, 120201.

Karim, M. R., Dey, S. K., Islam, T. et al., Deephateexplainer: Explainable hate speech detection

Page | 941
in under-resourced bengali language, 8th International Conference on Data Science and
Advanced Analytics (DSAA), IEEE. 2021; 1-10.

Olweus, D., Cyberbullying: An overrated phenomenon? , European journal of developmental


psychology. 2012; 9(5), 520-538.

Rahman, S. S. M. M., Biplob, K. B. M., Rahman, M. et al., An investigation and evaluation of


N-Gram, TF-IDF and ensemble methods in sentiment classification, In International
Conferenceon Cyber Security and Computer Science, Springer. 2020; 391-402.

Rahman, M. R., Prevention of cybercrimes in Bangladesh. Society & Change. 2017;11(4), 7-18.

Sharif, O., Hossain, E., Hoque, M. M., Offensive language detection from multilingual code-
mixed text using transformers, arXiv preprint arXiv: 2103.00455. 2021.

Slonje, R., Smith, P. K., Frisén, A., The nature of cyberbullying, and strategies for prevention,
Computers in human behavior. 2013; 29(1), 26-32.

Page | 942
7th International Conference on Engineering Research, Innovation and Education
January 12-14, 2023, Sylhet, Bangladesh

PAPER ID: 334

Highly Sensitive SF11/Au/PtSe2/graphene Layered-Based


Surface Plasmon Resonance Biosensor for Alcohol Sensing

Md. Masud Rana 1, *, Shahriar Mostufa 1


1
Electrical and Electronic Engineering, Rajshahi University of Engineering &
Technology (RUET), Rajshahi-6204, Bangladesh
* md.masud.rana@eee.ruet.ac.bd [Md. Masud Rana]

Abstract: In this paper, a hybrid SF11 /Au/PtSe2/graphene layer based surface


Keywords: plasmon resonance (SPR) biosensor with enhanced refractive index is
presented for alcohol sensing. The proposed SPR biosensor is designed based
● Biosensor; on the angular analysis of the attenuated total reflection (ATR) method. The
● High index prism; sensor’s surface plasmon polaritons (SPPs) and the sensing region refractive
index (RI) are changed owing to the interaction of various concentrations of
● Surface Plasmon;
alcohol. The refractive index is varied between 1.33 and
● Alcohol; 1.59 for the various concentration of alcohol. The detection of alcohol levels
of 0%, 20%, 40%, and 60% in water and various alcohol groups of methanol,
● Refractive index;
ethanol, propanol, butanol, pentanol, and phenol has been detected with the
● Chemical Sensor. proposed sensor. The proposed sensor exhibits an angular shift sensitivity of
106.92 deg/RIU, the figure of merits (FOM) of 26.73 RIU−1and detection
accuracy (DA) of 6.95. The proposed sensor exhibited the highest sensitivity
with the enhanced refractive index for the detection of alcohol.

1. Introduction
In general, the surface plasmon is a charge density oscillation that is confined to the metal-
dielectric medium interface surface. The refractive index of the sensing medium next to the metal
has a significant impact on the surface plasmon's ability to propagate. Due to the very strong
refractive index sensing capability, SPR based sensor becomes very effective in biosensing as
well as other sensing applications, for instance, chemicals and thermal sensing (Mollah et al.,
2020; Rahman et al., 2017; Hossain and Rana, 2016). When the light wave incident on the metal
surface generates the surface plasmon wave (SPW), and in resonance or SPR conditions, the
reflectance of the output light wave becomes drastically minimum due to the maximum
excitation of the surface plasmon wave (Dai et al., 2019;Bochenkov, Frederiksen and Sutherland,
2013). Therefore, by adopting the angular interrogation method, the resonance or SPR angle is
achieved. For a particular resonance or SPR angle, the output reflectance becomes minimum, and
with an increase of the refractive index (analyte) in the sensing medium, the SPR angle shifts
right. This change in the angle of the SPR is used to measure the sensitivity and detect the
respected analyte in the sensing medium.
At present, the SPR-based sensors sensitivity has significantly increased due to significant
advancements in fabrication technology. The introduction of 2D and transition metal
dichalcogenide (TMDC) materials has significantly improved the sensitivity of the prism-

Page | 943
coupled Kretschmann configuration. Several papers have been reported by researchers focusing
on the enhancement of the sensitivity of the prism-coupled Kretschmann configuration using 2D
and TMDC materials (Rahman et al., 2017; Hasib et al., 2019;Maurya et al., 2015). Although the
overall Kretschmann configuration sensitivity has been greatly improved by many researchers,
the refractive-index detection range of these proposed sensors is not much improved. As a result,
the detection of biological analytes having high refractive index changes remains undetectable by
the existing SPR-based prism-coupled Kretschmann configuration. Therefore, the main focus of
this paper is to introduce a new sensor with significant improvement in the refractive index
sensing range for the detection of alcohol. Since the maximum refractive index sensing range for
alcohol detection is 1.53, it is not possible to detectthis refractive index analyte using the currently
existing proposed prism-coupled Kretschmann configuration sensors. Therefore, in this article,
we have utilized the high refractive index prism materials for gaining the refractive index sensing
range of 1.53 and with the SF11 material as a prism, the refractive index (RI) sensing range of
the SPR prism-coupled Kretschmann configuration sensor improved significantly.
In this article, a novel structure SPR-based sensor of Prism (SF11)/Au/PtSe2/graphene has been
designed and analyzed using the FEM for alcohol sensing. The proposed sensor exhibits a
refractive index range of the analyte from 1.33 to 1.59, which is the highest to the best of the
author's knowledge and the proposed sensor exhibits an overall angular shift sensitivity of
106.92 deg/RIU, detection accuracy (DA) of 6.95 and the figure of merits (FOM) 26.73/RIU for
the various alcohol groups of Methanol, Ethanol, Propanol, Butanol, Pentanol, and Phenol
detection.

2. Design Methodology
Kretschmann configuration provides the foundation for the proposed biosensor. Here, p-
polarized light with a 633 nm wavelength was incident on a prism (SF11) at an acceptance
angle, and the resonance angle was discovered using the angular interrogation method. The output
reflectance intensity decreases as the incident light wave travels across several layer interfaces,
reaching its least value under the resonance state. Using a complementary metal- oxide display
(CMOD) or charge-coupled device (CCD), the output reflectance intensity maybe observed
(CMOS) (Akib et al., 2021). Table. 1 shows the material specifications which have been used to
model the sensor and Figure.1 illustrates the structural design of the sensor. The suggested model
is designed and examined using a numerical simulation that is based on the FEM techniques. We
have simulated the 2D geometry of the suggested sensor and the proposed model using COMSOL
where the top of prism SF11 has been exposed to light from a source with a wavelength of 633
nm shown in Figure. 2. Some crucial parameters of sensors such as angular sensitivities, the figure
of merit (FOM), and the detection accuracy (DA) characterize sensor performance denoted as
Eq.(1-3)(Dey, Islam and Park, 2021).
Δ𝜃spr
S= [deg/RIU] (1)
Δn
∆θspr
DA = (2)
FWHM
S
FOM = FWHM [RIU−1] (3)

Page | 944
Figure. 1. Structural diagram of proposed hybrid Prism (SF11)/Au/PtSe 2/Graphene-based SPR
biosensor.

Table. 1 Materials Specification at 633 nm wavelength


Specification Refractive index Thickness Reference

Prism (SF11) 1.7786 1.5µm (Hossain et al., 2019)


Gold (Au) 0.13783+3.6196*i 50 nm (Mishra, Mishra and Verma, 2016)
PtSe2 2.9029 + 0.8905i 2 nm (Rahman et al., 2020)
Graphene (monolayer) 3+1.1491*i 0.34 nm (Rahman et al., 2020)
Sensing medium or (1.33-1.59) 1.44382 µm -
analyte

Figure. 2. General 2D structure of COMSOL Multiphysics (software view) of proposed Prism


/Au/PtSe2/Graphene-based SPR biosensor.

Page | 945
3. Result and Discussions
3.1 Different groups of alcohol detection
For detecting multiple groups of alcohol, the analyte has to be placed on a sensing medium. As
different groups of alcohol such as Methanol, Ethanol, Propanol, Butanol, Pentanol, and Phenol
exhibit incremental refractive index of 1.33, 1.36, 1.38, 1.40, 1.41, and 1.53 (Ahmed et al., 2019).
Therefore, with this optical property of refractive index increment, the different groups of alcohol
can be identified easily with this proposed sensor. The proposed biosensor exhibits the sensing
range of refractive index from 1.33 to 1.59 and all alcohol refractive indexes are within this range.
Thus, the proposed sensor can easily detect the different groups of alcohol using resonance angle
shifting. Figure. 3 and Table. 2 illustrate that due to the refractive index increment the analyte
SPR angle shifts rightwards. Therefore, taking methanol optical property as a base, the Ethanol,
Propanol, Butanol, Pentanol, and Phenol can be detected with 57.10, 58.70, 60.30, 61.20, and 74.30
SPR angle change, respectively. Furthermore, the shift of resonance angle for Ethanol, Propanol,
Butanol, Pentanol, and Phenol is 2.20, 3.80, 5.40, 6.30,and 19.40, respectively. Thus, by observing
the resonance or SPR angle shift the proposed sensor can easily detect these various alcohol
groups.

Figure. 3. Resonance angle shift respected to the Methanol, Ethanol, Propanol, Butanol,
Pentanol, and Phenol group of alcohol.
Table. 2. Detection of a different group of alcohol with the proposed sensor using the resonance
angle shift

Alcohol group Refractive Refractive SPR Reflectance SPR angle Angular


specification index index angle intensity shift Sensitivity
change (𝜃spr) (%) (∆𝜃spr) (deg/RIU)
(∆n) (deg) (deg) ∆𝜃spr/∆n

Methanol 1.33 Ref 54.9 12.12 Ref Ref


Ethanol 1.36 0.03 57.1 11.35 2.2 73.33
Propanol 1.38 0.05 58.7 12.3 3.8 76
Butanol 1.40 0.07 60.3 14.11 5.4 77.14
Pentanol 1.41 0.08 61.2 14.78 6.3 78.75
Phenol 1.53 0.2 74.3 21.82 19.4 97

Page | 946
Table. 3. Comparison of the extended range of our proposed sensor with some recently reported
SPR sensors

Model Structure Refractive Sensitivity DA FOM Reference


index (deg/RIU) (RIU−1)
Range

Graphene Coating 1.33-1.40 33.98 0.2987 2.78019 (Verma, Prakash and


Tripathi, 2015)
SF11 With graphene- 1.34 - 1.41 85.375 - - (Hossain et al., 2020)
MoS2-TiO2-SiO
SF11/Au/graphene 1.33-1.39 85.00 3.38 - (Hossain et al., 2019)
Monolayer of graphene - 87.8 1.28 17.56 (Hossain et al., 2019)
and MoS2
SF11/ Au-Graphene-MoS2 1.334- 89.29 - - (Rahman et al., 2017)
hybrid 1.40
Graphene coated with WS2 - 95.71 1.763 25.19 (Rahman et al., 2018)
SF11/Au/PtSe2/Graphene 1.33-1.59 106.92 6.95 26.73 Our work

Finally, the comparison Table. 3 has been tabulated to compare the proposed sensor with some
recently reported work. From the comparison, it has been obtained that the refractive index
sensing range of the proposed sensor has been improved enormously with this proposed prism
(SF11)/Au/PtSe2/graphene-based biosensor along with the angular shift sensitivity, DA, and FOM.
4. Conclusion
In this article, a novel prism (SF11 )/Au/PtSe2/Graphene structured SPR based biosensor has
been proposed for an enhanced refractive index sensing range from 1.33 to 1.59 to detect the
alcohol. The detection of Methanol, Ethanol, Propanol, Butanol, Pentanol, and Phenol has been
detected with the proposed SPR-based sensor. To the best of the author's knowledge, thesensing
refractive index range of the proposed sensor is the highest to this date and the detectionof alcohol
has been first shown using the prism-based SPR sensor. This proposed biosensor might be used
greatly in the future for biosensing and chemical sensing applications due to its high sensitivity
and extended refractive index sensing range.
References
Ahmed, K. et al., Tetra-core surface plasmon resonance based biosensor for alcohol sensing,
Physica B: Condensed Matter, 2019, 570(5); 48–52.
Akib, T. B. A. et al., Design and numerical analysis of a graphene-coated spr biosensor forrapid
detection of the novel coronavirus, Sensors, 2021, 21(10); 1–19.
Bochenkov, V. E., Frederiksen, M. and Sutherland, D. S., Enhanced refractive index sensitivity
of elevated short-range ordered nanohole arrays in optically thin plasmonic Au films,
Optics Express, 2013, 21(12); 14763-14773.
Dai, X. et al. ‘Sensitivity enhancement of a surface plasmon resonance with tin selenide (SnSe)
allotropes’, Sensors, 2019, 19(1); 1-9.

Page | 947
Dey, B., Islam, M. S. and Park, J., Numerical design of high-performance
WS2/metal/WS2/graphene heterostructure based surface plasmon resonance refractive
index sensor, Results in Physics, 2021, 23(3); 104021.
Hasib, M. H. H. et al., Improved transition metal dichalcogenides-based surface plasmon
resonance biosensors, Condensed Matter, 2019, 4(2); 1–11.
Hossain, B. and Rana, M., Graphene coated high sensitive surface plasmon resonance biosensor
for sensing DNA hybridization, Sensor Letters, 2016, 14(2); 145–152.
Hossain, M. B. et al., High performance refractive index SPR sensor modeling employing
graphene tri sheets, Results in Physics, 2019, 15(1); 102719.
Hossain, M. B. et al., A Numerical Approach to Design the Kretschmann Configuration Based
Refractive Index Graphene-MoS2 Hybrid Layers With TiO2-SiO2 Nano for Formalin
Detection, Photonic Sensors, 2020, 10(2); 134–146.
Maurya, J. B. et al., Performance of graphene–MoS2 based surface plasmon resonance sensor
using Silicon layer, Optical and Quantum Electronics, 2015, 47(11); 3599–3611.
Mishra, A. K., Mishra, S. K. and Verma, R. K., Graphene and beyond Graphene MoS2: A New
Window in Surface-Plasmon-Resonance-Based Fiber Optic Sensing, Journal of Physical
Chemistry C, 2016, 120(5); 2893–2900.
Mollah, M. A. et al., Plasmonic temperature sensor using D-shaped photonic crystal fiber,
Results in Physics, 2020, 16(1), p. 102966.
Rahman, M. M. et al. ‘Sensitivity enhancement of SPR biosensors employing heterostructure of
PtSe2 and 2D materials’, Optical Materials, 2020, 107(4); 110123.
Rahman, M. S. et al., Design and numerical analysis of highly sensitive Au-MoS2-graphene based
hybrid surface plasmon resonance biosensor, Optics Communications, 2017, 396(2); 36–
43.
Rahman, M. S. et al., A novel graphene coated surface plasmon resonance biosensor with tungsten
disulfide (WS2) for sensing DNA hybridization, Optical Materials, 2018, 75; 567–573.
Verma, A., Prakash, A. and Tripathi, R., Performance analysis of graphene based surface plasmon
resonance biosensors for detection of pseudomonas-like bacteria, Optical and Quantum
Electronics, 2015, 47(5); 1197–1205.

Page | 948
7th International Conference on Engineering Research, Innovation and Education
January 12-14, 2023, Sylhet, Bangladesh

PAPER ID: 335

A Compact Partially Ground Loaded Slotted MPA for Brain


Tumor Detection

1Md. Masud Rana*, 2Maruf Hasan and 3Hasibul Islam


1,2,3
Rajshahi University of Engineering & Technology, Kazla-6204, Rajshahi
1
md.masud.rana@eee.ruet.ac.bd,
*Corresponding author [Md. Masud Rana]

Abstract: In this paper, a new, compact, wideband, biomedical compatible


Keywords: microstrip patch antenna (MPA) is presented to find out brain tumors inside
the brain. The proposed MPA is designed on a low loss, stable, and high-
● Antenna;
frequency compatible considering substrate of Rogers RO3003 (εr = 3). The
● Brain tumor. overall dimension of the MPA is 50 mm × 50 mm × 1.5 mm. The rectangular
patch of the antenna is modified by rectangular slots for improving the
● Imaging ; resonance frequency. Besides, the bandwidth is also enhanced by partial
● Medical; ground. The proposed antenna operates at 3.94 GHz with higher fractional
bandwidth of 62.18%. The maximum directivity of the antenna is close to 5
● MPA;
dBi. The proposed MPA is mimicked with a six-layer head apparition
● SAR; containing skin, fat, skull, Dura, cerebrospinal liquid (CSF), and brain in
CST software to identify tumors. The safety measurement is also tested by
● Wideband; specific absorption rate (SAR). The most popular confocal microwave
imaging (CMI) is used for tumor imaging. From the reproduction results, it is
discovered that the proposed MPA is exceptionally precise and safe for
tumor imaging.

1. Introduction
Many deadly diseases nowadays affect people's lives, economic status and health. The brain
tumor is one of the most threatening, maladies in this modern world. Brain tumor is the
accumulation of mass tissue which tends to the abnormality of cell in the brain. The brain
tumor causes different type of human illness. In 2022 about 96,000 people inUnited States
are diagnosed with brain tumor and 700,000 are living with initial brain tumor. Devastating and
deadly disease brain tumor is in practice detected by a magnetic resonance imaging (MRI),
complete blood count blood protein tests, X-ray. But these established methods are experiencing
some major drawbacks. That’s why microwave imaging (MWI) is gaining more popularity among
all these techniques. (Luo, et al., 2017 ; Tarik, et al., 2020)
For MWI application, various types of microstrip patch antenna (MPA) are using by the
researcher. Because of some unique feature, like small size, trouble-free fabrication, and
lightweight, MPA is getting greater prevalence in the field of biomedical application.
Concentrating to those challenges, a number of MPA is designed. Some of them are the ultra-
wideband (UWB) frequency slot antenna array (Islam, et al.,2019), Electromagnetic band gap

Page | 949
based MPA (Inum , et al., 2018), wearable MPA (Porter, et al., 2016) , UWB pentagon antenna
(Shokry and Allam, et al., 2016), 3D slot-loaded wideband antenna ( Mobashsher , et al., 2016)
have utilized previously, tapered slot UWB antenna (Al-Zuhairi, et al., 2017), rectangularslotted
wideband patch with a partial ground technique (Selvaraj, et al., 2018) were applied for this
application.
In this paper, a new biomedical compatible MPA is designed to detect the brain tumour based on
MWI system. Although the design of the patch antenna is simple; but the antenna is modified
with rectangular slot and partial ground to have gain, larger BW. The six-layer of human head
phantom is designed with skin, fat, skull, Dura, Cerebrospinal fluid (CSF), and brain respectively
with relative dielectric permittivity, electrical conductivity and mass density. The |S11| outputs are
obtained from human head phantom with tumor and without tumor. The proposed slotted MPA
conserved better gain and efficiency with the body phantom. By switching the distance between
antenna and head phantom, the efficiency and gain is fluctuated too. The proposed antenna
performance is also responded heterogeneously for changing the thickness and permittivity of the
tissue layer. The SAR of the antenna also provides the safety range according to IEEE standard.

2. Antenna Design
2.1 Methodology & Antenna Model
The proposed antenna is presented in Fig. 1. In the Fig. 1 front view, back view, and side view
given. On Fig. 1(a) the yellow colored surface is the patch of the antenna. The Fig. 1(b) the partial
ground is also marked with yellow color. The substrate is in the middle between the patch and
partial ground described in Fig. 1(c). The antenna has a partial ground made of
copper and the substrate layer is created with Roger 3003 (𝜀r = 3, tan𝛿 = 0.0001). The ground
plane is optimized for getting better bandwidth, and better impedance matching. The 50 ohm inset
feed line is used to feed power in the antenna.
The rectangular patch is made of copper material modified with slotted shape. The overall
dimension of the patch is 19.5 ×20×0.035 mm3. The slots are made in order to increase the current
flowing path and to improve the resonance frequency. The calculated parameters of the patch,
substrate layer, ground layer and feeding line are stated in Table 1. In Table 1 the ground, feed
line, substrate, and patch dimensions are given.

Fig.1 Proposed MPA (a) front side view (b) backside view (c) side view

Page | 950
Table 1. Geometric dimension of patch antenna
Antenna Part Parameters(symbol) Value(mm) Antenna Part Parameters(symbol) Value(mm)
Length, Lg 16.4 Length, Lb 9.75
Ground Width, Wg 50 Width, Wa 9
Thickness,Tg 0.1 Width, Wb 8.5
Length, Ll 20 Width, Wc 4.25
Feed Line
Width, We 2.5 Patch Width, Wd 2
Width, Ws 50 Width, We 4.9
Substrate Length, Ls 50 Width, Wf 4
Thickness,Ts 1.5 Thickness, Tp 0.035
Patch Length, La 20 Length, Ld 3

3. Numerical Analysis
3.1 Antenna Performance without Head Phantom
The design evolution steps and the following |S11| graph are given in Fig. 2. The antenna design started with
a normal conventional rectangular patch antenna. In this step the antenna had resonance at 3.63 GHz with
fractional bandwidth of 41.87%, covering frequency range 2.91 to 4.43 GHz. In the next step, the
rectangular portion then etched for getting good performance. But the antenna on this step was achieved
only 40.52% fractional bandwidth with the resonance shifts of 3.48 GHz with frequency range of 2.79-
4.20 GHz. Following to the next step, the patch was made staircase on the upper portion. The final antenna
achieved performance with resonance 3.94 GHz with 60.66% fractional bandwidth. The antenna covers the
frequency range was 3-5.39 GHz. In this final step, the fractional bandwidth improved in a very efficient
way. The use of additional two slots in the final antenna on both right and left sides also helped to reduce
the |S11| to -48.09dB.

Fig. 2. Reflection coefficient value (|S11|) for different evaluation steps

3.2 Effect of Partial Ground


In this section, the effect of partial ground in the patch is discussed. The partial ground is not only
useful for better impedance matching but also it is very good for better radiation efficiency. The
partial ground in antenna increases the capacitance which reduces the quality factor (Q). In
addition, the stored energy is also reduced due to partial ground structure in the antenna. For

Page | 951
getting the accurate dimension of partial ground total of five iterations were performed. In Table
2, the analysed results are given.
Table 2. Summarised result of the effect of partial ground
Ground Surface
Minimum Resonance Q-
Length Bandwidth Current,
|S11|, dB (GHz) factor
Lg (mm) (A/m)
50 13.2 MHz -16.3 4.612 52.8 350
40 12.6 MHz -15 4.62 52.9 366.7
30 40 MHz -21.52 5.32 56.2 133
20 90 MHz -21.86 5.32 52.6 59.11
16.4 2.39 GHz -48.09 3.94 37.5 1.65

From Table 2, it is seen that the value of Q-factor is very large at 50 mm ground length. The
observed Q-factor value is larger than 10 which cause very narrowband. From Table 2, it is also
noted that, at 16.4 mm length of partial ground, minimum bandwidth was observed with Q- factor
value of 1.65. For the partial ground, if Q-factor <10, the wideband occurs in the communication
source. So, the value of Q-factor at implemented ground layer provides lowest value. This is why
the wideband occurs in this antenna. Due to high capacitance the bandwidth increased and Q-
factor was reduced thus improving the overall performance.

3.3 Antenna Performance with head phantom


In this paper, the six-layer human head phantom model is presented. The applied layers are skin,
fat, skull, Dura, Cerebrospinal fluid (CSF), and brain. The properties of the human headphantom
are summarized in Table 3. Human head phantom model is given in Fig. 3.

Fig 3. Human head phantom model (a) schematic representation, (b) cross-sectional view in CST
software, and (c) perspective view.
Table 3. Properties of the head phantom
Radius Permittivity Conductivity,σ Density
Tissue
(mm) (εr) (S/m) (Kg/m3)
Brain 80.5 43.22 1.29 1030
CSF 82 70.1 2.3 1005.9
Dura 83.5 46 0.9 1130
Skull 87.6 5.6 0.03 1850
Fat 89 5.54 0.04 910
Skin 90 45 0.73 1090
Tumor 5 55 7 1040

Page | 952
In Fig. 4 three antenna states are discussed by rotating the antenna with a static head phantom
model. The simulation results are given in Fig 5.

Fig. 4 (a) Step-1, (b) Step-2, (c) Step-3

(a) (b)

Fig. 5 (a) The output result of |S11| for the defined steps (b) Enlarged captions

In Fig. 6, the SAR analysis for distance between 20 mm and 25 mm head phantom and antenna
are shown respectively:

Fig.6 SAR value at distance between (a) 20 mm, and (b) 25 mm of head phantom and antenna.

In Fig. 6 (a) and (b), the SAR result for the distance of 20 mm and 25 mm between antenna and
human head phantom model is presented. Here the maximum value of SAR is 1.1 W/kg for the
20 mm distance, which is below 1.6 W/kg so it shows the safety limit of the SAR.

3.3 Tumor Detection


In this paper proposed antenna placed in seven different locations to find the presence of a tumor
on the head phantom. The locations marked with T1, T2, T3, T4, T5, T6, and T7 to find the value
of |S11| for with tumor and without tumor both in Fig. 7 and summarised result is given in Table
4.

Page | 953
Fig. 7. Tumor Detection

Table 4 Result of |S11| on the seven points

|S11| in presence |S11|without Variation of Fractional


Position
of tumor, dB tumor, dB |S11|, dB Bandwidth, %

T1 -46.59 -46.65 1.06 53.63


T2 -46.6 -47.76 1.16 55.69
T3 -46.51 -47.77 1.26 55.7
T4 -46.25 -47.59 1.34 56.22
T5 -46.48 -47.64 1.16 55.96
T6 -46.58 -47.66 1.08 53.89
T7 -46.59 -47.68 1.09 53.62

3.4 Tumor Imaging


The dielectric image from the received signal of the MPA is processed by using some computer
based algorithm such as CMI, MIST, and HMI (Wang, 2018). In Fig. 8 the reconstruction of
tumor dielectric images is given.

Fig.8 Reconstructed tumor image at (a) 20 mm, and (b) 25 mm (c) 30 mm distance between antenna and
head phantom.
It is observed from the Fig. 8 that with the increasing distance between the antenna and head
phantom the intensity of tumour detected region is contracted. The strength of reflected signal
power from the tumor is reduced for the increasing distance that’s why at 30 mm distance the
tumor detected region is waned. However, the proposed antenna can detect tumor and it is can
detect tumor at any distance between antenna and phantom. The effectiveness of the proposed
antenna in human head phantom imaging is evident from all the analysis. In sum, the advantages
of the proposed antenna are-
 The proposed antenna has a high frequency operating range
 The |S11| (< -10 dB) is very low of the proposed antenna

Page | 954
 The directivity gain (dBi), efficiency, and VSWR is excellent
 The SAR of the proposed antenna is very low according to the defined limit.

4. Conclusion
An MPA is designed with a slotted patch and partial ground for a wideband frequency range to
detect the brain tumor. The slots and partial ground in the antenna provide better radiation
efficiency with high fractional bandwidth. The |S11| of the antenna is less than -10 dB. The
maximum bandwidth observed 2.39 GHz with a high frequency operating range between 3-5.39
GHz. Maximum efficiency of the proposed antenna is 94.8% in free space with VSWR below 2.
Besides, the proposed slotted MPA conserved better gain and efficiency with the body phantom.
The SAR of the antenna also provides safety range of IEEE standards. The collected |S 11| in
presence of head phantom with tumor is also scrutinized to find a tumor. The measured |S11| from
the head phantom is post processed in MATLAB with CMI algorithm to detect the tumorat 20
mm, 25 mm, and 30 mm distance away between antenna and head phantom to reconstruct the
image respectively. The 2D tumor image is observed after these steps. The overall results
confirmed that the proposed patch antenna can successfully detect brain tumors.

References
Al-Zuhairi DT, Gahl JM, Al-Azzawi A, Islam NE. Simulation design and testing of a dielectric
embedded tapered slot UWB antenna for breast cancer detection. Progress in
Electromagnetics Research C. 2017, 79; 1-5.
Balanis, C.A.,. Antenna theory, hoboken. New Jersey: John Wiley & Sons, Inc, 2005, 8(1); 21-
31.
Inum R, Rana M, Shushama KN, Quader M. EBG based microstrip patch antenna for brain tumor
detection via scattering parameters in microwave imaging system. International journal of
biomedical imaging. 2018, 12 (2); 1-12.
Islam MT, Mahmud MZ, Islam MT, Kibria S, Samsuzzaman M. A low cost and portable
microwave imaging system for breast tumor detection using UWB directional antenna array.
Scientific reports. 2019, 29 (10); 1-3.
Luo S, Ji Z, Yang S, Xing D. Near-field transmission-type microwave imaging for
noninvasive evaluation of electromagnetic characteristics: Toward early breast tumor
detection. IEEE Photonics Journal. 2017, 9(6); 1-0.
Mobashsher AT, Bialkowski KS, Abbosh AM. Design of compact cross-fed three- dimensional
slot-loaded antenna and its application in wideband head imaging system. IEEE Antennas
and Wireless Propagation Letters. 2016, 9 (3); 56-60.
Porter E, Bahrami H, Santorelli A, Gosselin B, Rusch LA, Popović M. A wearable microwave
antenna array for time-domain breast tumor screening. IEEE transactions onmedical imaging.
2016, 18 (1); 1-9.
Shokry MA, Allam AM. UWB antenna for brain stroke and brain tumour detection. In2016 21st
International Conference on Microwave, Radar and Wireless Communications (MIKON)
2016, 9 (5); 1-3.
Wang L. Microwave sensors for breast cancer detection. Sensors. 2018, 23(2); 655-665

Page | 955

You might also like