pakistan journal of computer and information systemspastic.gov.pk/downloads/pjcis/pjcis_v1_1.pdf ·...

Pakistan Journal of Computer and

Information Systems

Volume. 1, No. 1

2016

Published by:

Pakistan Scientific and Technological Information Centre (PASTIC), Islamabad, Pakistan

Tel: +92-51-9248103 & 9248104; Fax: +92-51-9248113 Website: www.pastic.gov.pk

http://www.pastic.gov.pk/

ITLAb-21

Typewritten text

ISSN (Print): 2519-5395

ITLAb-21

Typewritten text

ISSN (Online): 2519-5409

Pakistan Journal of Computer and Information Systems

This is the first issue of Pakistan Journal of Computer and Information Systems (PJCIS)

published by Pakistan Scientific and Technological Information Centre (PASTIC). Pakistan

Journal of Computer and Information Systems is a peer reviewed journal of Computer Science

and Engineering. It is an open access online journal published twice a year. The mission of the

journal is to provide a forum for researchers to debate and discuss interdisciplinary issues in

Computer Science, Computer System Engineering and Information Systems. It invites and

welcomes contributions in all areas of Computer Science, Computer System Engineering and

Information Systems.

Acknowledgement

Pakistan Scientific and Technological Information Centre (PASTIC) is grateful to the following

researchers/subject experts for carrying out review of the manuscripts and providing their

valuable and critical suggestions on the research papers published in this issue of Pakistan Journal

of Computer and Information Systems (PJCIS). PASTIC also acknowledges their contributions in

bringing out this issue.

Dr. Anne Laurent University of Montpellier-LIRMM,

Montpellier, France

Dr. Mohsin Islam Tiwana College of Electrical and Mechanical

Engineering, (NUST), Rawalpindi, Pakistan

Dr. Nasrullah Memon

University of Southern Denmark,

Denmark

Dr. Muhammad Yasir Qadri National Engineering and Scientific

Commission, Islamabad, Pakistan

Dr. Nisar Ahmad Memon

College of Computer Science and Information

Technology, King Faisal University, Al Ahsa,

Kingdom of Saudi Arabia

Dr. Laiq Hasan

University of Engineering and Technology,

Peshawar, Pakistan

Dr. Mohammad Yakoob Siyal

School of Electrical & Electronic

Engineering, Nanyang Technological

University, Singapore

Dr. Ali Daud

International Islamic University,

Islamabad, Pakistan

Dr. Muhammad Akbar

Effat University, Jeddah,

Kingdom of Saudi Arabia

Dr. Nadia Nawaz

COMSATS Institute of Information

Technology, Wah Cantt, Pakistan

Pakistan Journal of Computer and Information SystemsPatron

Dr. Muhammad AshrafChairman, Pakistan Science Foundation

Editor-in-ChiefDr. Muhammad Akram Shaikh

Director General, PASTIC

EditorsNageen Ainuddin

Muhammad Aqil Khan

Managing EditorsSaifullah Azim

Dr. Maryum Ibrar Shinwari

EDITORIAL ADVISORY BOARD

Dr. Muhammad IlyasFlorida Atlantic University, Florida, USA

Dr. Khalid MehmoodUniversity of Dammam, Dammam, Kingdom of Saudi Arabia

Dr. Muhammad Yakoob SiyalNanyang Technological University, Singapore

Dr. Khalid SaleemQuaid-i-Azam University, Islamabad, Pakistan

Dr. Kanwal AmeenUniversity of the Punjab, Lahore, Pakistan

Dr. Waseem ShahzadNational University of Computer and Emerging Sciences,

Islamabad, Pakistan.Dr. Farooque Azam

National University of Sciences & Technology, Islamabad, Pakistan

Dr. Laiq HassanUniversity of Engineering & Technology, Peshawar,

Pakistan

Dr. Muhammad YounasOxford Brookes University, Oxford, United KingdomDr. Anne LaurentUniversity of Montpellier-LIRMM, Montpellier, FranceDr. M. Shaheen MajidNanyang Technological University, SingaporeDr. Nadeem AhmadSEECS, National University of Sciences & Technology,Islamabad, PakistanDr. Amjad FarooqUniversity of Engineering & Technology, Lahore, PakistanDr. B.S. ChowdhryMehran University of Engineering & Technology, Jamshoro, PakistanDr. Jan MuhammadBaluchistan University of Information Technology, Engineering & Management Sciences, Quetta, PakistanDr. Muhammad Riaz MughalMirpur University of Science & Technology, Mirpur (AJ & K)

Pakistan Scientific and Technological Information Centre (PASTIC), Islamabad, PakistanTel: +92-51-9248103 & 9248104 ; Fax: +92-51-9248113

www.pastic.gov.pk : e-mail: [email protected]

SubscriptionAnnual: Rs. 1000/- US$ 50.00 Single Issue: Rs. 500/- US$ 25.00

Articles published in Pakistan Journal of Computer and Information Systems (PJCIS) do not represent the views of the Editors or Advisory Board. Authors are solely responsible for the opinion expressed and accuracy of data.

Aims and Scope

Pakistan Journal of Computer and Information Systems, a bi-annual

publication of Pakistan Scientific and Technological Information Centre

(PASTIC), Pakistan Science Foundation is aimed at providing a platform

for researchers and professionals of Computer Science & Engineering,

Information and Communication Technologies (ICTs), Library &

Information Science and Information Systems for sharing and

disseminating their research findings and to publish their original and

cutting edge research. This is a peer reviewed open access journal

intended to publish high quality papers on theoretical development as well

as practical applications in all fields of Computer Science and

Engineering, ICT and Information Science. The journal also aims to

publish new attempts on emerging topics /areas, reviews and short

communications. The scope encompasses research areas dealing with all

aspects of computing and networking hardware, numerical analysis and

mathematics, software programming, databases, solid and geometric

modeling, computational geometry, reverse engineering, virtual

environments and hepatics, tolerance modeling and computational

metrology, rapid prototyping, internet-aided design, information models,

use of information and communication technologies for managing

information, study of information systems, information processing, storage

and retrieval, management information systems etc. Review articles are

also published in the journal.

Pakistan Journal of Computer and Information Systems

Contents

Volume 1 Number 1 2016

1. Detection of Ventricular Arrhythmia from ECG

M. Saeed Rattar, Syed. M. Shehram Shah, B.S. Chowdhry, Syed M. Zaigham Abbas Shah………..

1

2. Vertebra Localization Using Shape Based Analysis and Unsupervised

Clustering from X-Ray Images

Anum Mehmood, M. Usman Akram, Mahmood Akhtar………………………………………………….

13

3. Numerical Solution of Initial Boundary Value Problems by MOL-RBF

Method

Fazal-i-Haq, Farah Jaffar, Siraj-ul-Islam…………………………………………………………

25

4. A Comprehensive Review of Intrusion Detection Systems for Wireless

Sensor Networks

Warda Ahmad, Sidra Anwar, Yasir Saleem, Junaid Arshad, Khawar Bashir, Hafsa Tahir……..…

41

5. Modeling of Safety Parameters for Low Level Radioactive Waste

Repository Using Machine Learning Approaches

Sardar Ali Shah, Abdul Majid, Safdar Ali…………………………………………………………..….….

57

PJCIS (2016), Vol. 1, No. 1 : 1-11 Detection of Ventricular

1

Detection of Ventricular Arrhythmia

from ECG

M. SAEED RATTAR*, SYED M. SHEHRAM SHAH, B.S. CHOWDHRY,

SYED M. ZAIGHAM ABBAS SHAH

Institute of Information & Communication Technologies (IICT), Mehran University of Engineering and

Technology, Jamshoro, Pakistan

*Corresponding author’s email: [email protected]

(Received on 28th April, 2016 Accepted on 8th August, 2016)

Abstract

The Ventricular Arrhythmias particularly Ventricular Tachycardia (VT) and

Ventricular Flutter (VF) are life threatening arrhythmias and can lead to heart attacks if

not detected and treated timely. In this paper, a method has been proposed that can

differentiate between normal Electrocardiograms (ECGs) and two abnormal ECGs of VT

and VF. The classification is performed by means of Artificial Neural Networks (ANN).

Reflection coefficients of the Auto-Regressive Models of extractions from the ECG

recordings are computed and used as features for input to the ANN. The ANN is trained

using ECG samples that are characteristic of Non-diseased (normal), VT and VF. After

suitable training and validation, the proposed algorithm has been found to have an

accuracy of 100%, 97% and 94% for classification of Normal ECG, VF and VT

respectively.

Keywords: Ventricular Arrhythmia detection and classification,

Electrocardiogram, AR Modeling, Artificial Neural Networks

INTRODUCTION

The number of people dying from heart diseases is increasing day by day [1]. One

of the main causes of heart related deaths is Ventricular Arrhythmia. Arrhythmia can be

defined as the irregular rhythm of heart, that is, the heart beats may be either too fast or

too slow. Four chambers make up the human heart, with the upper two chambers called

the Atria and the lower two called Ventricles. The arrhythmia which originates from the

lower chambers of the heart is known as Ventricular Arrhythmia [2]. The two main types

of Ventricular Arrhythmia are Ventricular Tachycardia (VT) and Ventricular Flutter

(VF). Ventricular Tachycardia is characterized by faster than normal heartbeats and


2

Ventricular Flutter is characterized by the cardiac muscles contracting without

coordination [3]. These arrhythmias can lead to inevitable death if left untreated.

Therefore, it is of utmost importance that these arrhythmias are detected timely and

action taken timely in order to save lives.

The paper is organized into different sections. After introduction, the previous

work carried out to detect abnormality from ECG signals is described and the proposed

algorithm is explained thereafter. Subsequently, the results are presented and discussed.

Finally, the conclusion is drawn on the basis of results and discussion.

Previous Work

Cardiac Arrhythmias can be detected by ECG signal analysis. Therefore, the topic

of cardiac arrhythmia detection and classification has been of focus in Computer Aided

Diagnosis area. A literature survey indicates various methods through which these

detections have been performed. Lee et al [4] proposes an algorithm for the detection of

different types of Cardiac Arrhythmias using Support Vector Machines[5] based on

Morphological Features. He extracted three morphological features of an ECG based on

the QRS complex which are RR interval, QRS slope, and QRS shape similarity. Their

method has been shown to work for a number of cases. However, not all morphological

features are always present in abnormal ECGs. T. Conde et al. [6] suggests an algorithm

for detecting cardiac arrhythmia by finding a vector representation of the ECG signal and

classifying QRS complex through the use of neural networks [7, 8]. Vuksanovic and M.

Alhamdi [9] developed an algorithm to classify three types of ECG's. The three types

ECGs were Normal, Arrhythmia and Ventricular Arrhythmia. In the proposed method

signals were classified using K-nn [10] rule with classification being performed using

Linear Discriminant Analysis (LDA) [11] and Quadratic Discriminant Analysis (QDA).

Noh et al. [12] proposes an algorithm in which abnormal ECG detection is carried out

with the help of template matching where a normal ECG is first generated and then the

real time ECG input is compared with the generated ECG to determine abnormality. An

algorithm based upon bi-spectral analysis techniques, obtained using autoregressive

model [3] estimation, is proposed by Khadra et al [13] to classify arrhythmia types. The

extracted features are frequency support of the bi-spectrum that serves as the quantitative

measure for classification. In an algorithm proposed by Jung and Tompkins [14], wavelet

decomposition [15] is used for the detection and classification of Ventricular

Tachycardia, Ventricular Fibrillation, Ventricular Flutter and Supraventricular

tachycardia in ECGs.

It is clear from the above discussion that different methods have been proposed to


3

detect and classify ventricular arrhythmia. However, most of them include lengthy

computations or do not provide reasonable detection rates. It is therefore necessary to

develop methods that may perform this detection more accurately and are less

computationally intensive. This paper proposes a novel method that improves upon the

shortcomings of previous work that has been discussed.

Detection of Ventricular Arrhythmia from ECG

The proposed method follows a five step procedure to detect ventricular

arrhythmia from ECG as shown in Figure 1. The first three processes form the

preprocessing phase of the signal which is followed by feature extraction. Lastly, the

extracted features are passed on to the trained Neural Network for classification of the

ECG.

Re-sample to 256

Hz

Bandpass filter

Segmentation to

extract cardiac

cycle

AR modeling

Classify using

Neural Network

Start

Stop

Figure 1: Methodology


4

The database of PhysioNet, specifically Massachusetts Institute of Technology -

Beth Israel Hospital (MIT BIH) Normal Sinus Rhythm Database, Creighton University

(CU) Ventricular Tachyarrhythmia Database and Malignant Ventricular Ectopy database

[16], is used as the source of ECG signals.

A. Resampling

The sampling frequencies of signals used from the PhysioNet database for the

considered arrhythmias were different and shown in Table 1.

Table 1: Sampling Frequencies of Signals in the PhysioNet Database

Parameter Normal Sinus

Rhythm

Ventricular

Tachycardia

Ventricular

Flutter

Sampling

Frequency

128 Hz 250 Hz 250 Hz

By default, the ECG recordings of MIT-BIH NSR, MIT-BIH CU Ventricular

Tachyarrhythmia and MIT-BIH Malignant Ventricular Ectopy were sampled at 128 Hz,

250 Hz and 250 Hz respectively. Since all three recordings needed to be processed by the

same system, it was necessary to resample them at the sample frequency. Therefore, all

these recordings were resampled at 256 Hz. A Linear interpolation kernel is used for this

purpose.

B. Filtering

After resampling process, all three recordings have the same sampling frequency

but contain baseline drift [9] and motion artifacts which are a common artifact in

practical ECG recording. In order to minimize this, a Bandpass filter was applied on the

ECG recordings. The normalized upper and lower cut-off frequency of the Bandpass

filter was chosen to be 0.0062 and 0.7591.

C. Segmentation

Since the ECG recording are of non-uniform duration and also since the

conditions to be detected are characterized by specific parts of the ECG signal,

segmentation has been performed to extract the parts of the signal which characterize the

desired conditions. To do so, all ECG recordings was divided into 200 segments with


5

each segment containing 300 samples. Each segment is of 1.2 seconds as it is sufficient to

capture information of a single cardiac cycle. It is necessary to extract each cardiac cycle

as every cycle forms the basis for feature extraction individually.

D. AR Modeling

The extraction of discriminative features is one of the most important steps for

detection and classification of signals either one dimensional or multidimensional. In this

paper, features are extracted through Auto Regressive Modeling. An Auto Regressive

(AR) model is a linear prediction formula that predicts an output yt of a system based on

its previous inputs. An AR Model of order ‘m’ can be written mathematically as given in

Equation:

𝑦(𝑡) = ∑ 𝑎(𝑖

𝑚

𝑖=1

). 𝑦(𝑡 − 𝑖) + 𝜀 (𝑡) (1)

where a(i) represent coefficients of AR model, y(t-i) is series under investigation, 𝜀(t) is

output of uncorrelated errors and ‘m’ is the order of model and it indicates the number of

past samples which have been used to estimate the present value of signal. When

performing AR modeling, it is important to determine a suitable order of AR model so

that estimated signal is modeled accurately.

In this method, an AR model was estimated for one segment of an ECG signal.

Each segment contained 300 samples and had time duration of 1.2 seconds. Since single

ECG recording is divided into 200 segments with every segment being estimated using 4

coefficients, the total number of coefficients for a single ECG recording becomes 800.

Reflection Coefficients of the AR model were calculated using the Burg’s method [17].

In order to determine that the AR model has suitable accuracy in estimating the

signal, the Periodogram Power Spectral Density [18] was used a metric. The Periodogram

PSD is defined as the modulus squared of the Discrete Fourier Transform of the signal. It

can be expressed mathematically as given in Equation:

𝑆(𝑓) =∆𝑡

𝑁|∑ 𝑥𝑛 𝑒−𝑖2𝜋𝑛𝑓

𝑁−1

𝑛=0

|

2

−1

2∆t < 𝑓 ≤

12∆t

(2)

The Periodogram PSD was computed for AR mode estimations of order 4 and 5

and the comparison of the PPSD of the original ECG and the signal estimated by AR

models has been shown in Figure 2 and Figure 3.


6

Figure 2: Periodogram PSD of order 4 AR Model

Figure 3: Periodogram PSD of order 5 AR Model

It can be seen in Figure 2 and Figure 3, the AR model estimate of 4 coefficients is

more accurate in estimating the ECG signal as compared to signal estimate using 5

coefficients of the AR model.


7

E. Classification

Once the ECG signals have been preprocessed and feature extraction has been

performed, the next step is to perform classification of the signals. To perform

classification, a two layer feed forward Neural Network is used that is shown in Figure 4.

Figure 4: Feed Forward Network

The Hidden layer of the network comprises a Log-Sigmoid [19] transfer function

whose output is given by Equation:

𝑎 =1

1 + 𝑒−𝑛 (3)

Where ‘a’ is the output and ‘n’ is the input matrix. The output layer consists of a

Softmax [20] transfer function whose behavior is described by Equation 4.

µk =𝑒(ηk)

1 + ∑j𝑒(ηk)

(4)

Where ηk is a vector of k values, µk is a vector of k values between the range of 0

and 1 which sum up to 1. For training purposes, the AR model reflection coefficients of

samples of each category of ECG (Normal, VT and VF) were used. As mentioned earlier,

the order of AR Model is selected as 4 which mean a single segment of signal will be

represented by four reflection coefficients and there are 200 segments of an ECG.

Therefore, a single ECG recording will have 800 (200x4) inputs. The details of feed

forward neural network are shown in Table 2. Table 3 shows the numbers of samples

used for training of the network for NSR, VF and VT respectively.

Table 2: Details of Neural Network

Input Layer Hidden (Log

Sigmoid)

Output

(Softmax)

800 I/p Lines 16 Neurons 3 Neurons


8

RESULTS AND DISCUSSION

The proposed method uses reflection coefficients, obtained from AR Model,

which serve as distinct features. The ECG signals to be classified are provided as input to

the ANN after suitable preprocessing and computation of the AR model reflection

coefficients of order 4 for each signal.

The total number of samples and trained number of samples are shown in Table 3.

An ECG recording has 200 segments and each segment is considered separate sample for

training.

Table 3: Total and trained number of samples

Total Number of

Samples

Trained Number of

Samples

NSR 3600 2600

VF 7000 1000

VT 4400 1000

The number of correctly and incorrectly classified ECGs is shown in Table 4.

Table 4: Number of Correctly and Incorrectly Classified ECGs

Number of trained samples NSR VF VT

NSR 1000 18 4

VF X 5820 3400

VT X 162 200

The percentage of correctly and incorrectly classified ECGs is shown in Table 5.

Table 5: Percentage of Correct and Incorrect Detections

Classes NSR VF VT

Correctly Classified ECGs 100% 97% 94%

Incorrectly Classified ECGs 00% 03% 06%


9

It can be observed from Table 5 that the proposed algorithm is able to detect

normal ECGs with very high accuracy (100%), ECGs which are characteristics of VF are

classified with a correctness percentage of 97% whereas these are incorrectly detected as

VT 3%. For VT, the correctly detected percentage is 94%. Table 6 shows the comparison

of our classification with results by other authors who have used same dataset i.e.

PhysioNet (specifically Massachusetts Institute of Technology - Beth Israel Hospital

(MIT BIH) Normal Sinus Rhythm Database, Creighton University (CU) Ventricular

Tachyarrhythmia Database and Malignant Ventricular Ectopy database).

Table 6: Sensitivity (Se %) comparison with the methods cited in the Literature

Algorithms Se (%) of NSR Se (%) of VF Se (%) of VT

SVM based Morphological

Features, Lee et al. [4] 100 92 74.07

Quantitative Analysis, Khadra

et al. [13] 100 91.7 81.8

Wavelet Decomposition, Jung

et al. [14] X 86.7 93.9

PNN Neural Network, Darouei

et al. [21] 98.15 93.3 90

Algorithm proposed in this

paper 100 97 94

It can be observed from Table 6 that the method described in this paper

outperforms previous approaches proposed for the task of classification of the three types

of ECGs; Normal, VT and VF. This is especially true for the case of VF where an

improvement of three percent can be observed. In the case of VT, similar accuracy is

achieved by [14].

The feature extraction process used in the proposed method has been found to

perform reasonably better for differentiating among the three types of ECGs. The

classification process in the proposed method is relatively cheap to implement two layer

feed forward Neural Network which provides satisfactory performance.


10

CONCLUSION

In this paper, three cardiac arrhythmias Normal Sinus Rhythm (NSR), Ventricular

Tachycardia (VT) and Ventricular Flutter (VF) are classified using Artificial Neural

Networks (ANN). The features used are reflection coefficients of AR model which serve

as the inputs to the Feed Forward Neural Network. The features have been shown to be

robust enough to allow classification with a two layer network. Recommendation for

future work in this direction is to look at the classification of arrhythmias of both upper

and lower chambers of the heart using similar techniques. Furthermore, instead of

detecting VT and VF individually, more complicated ECGs could be classified which

may exhibit multiple defects in one recording.

ACKNOWLEDGMENT

This research was supported by Mehran University of Engineering and

Technology, Jamshoro.

REFERENCES

[1] D. Mozaffarian, et al., "Executive Summary: Heart Disease and Stroke Statistics--2016

Update: A Report From the American Heart Association," Circulation, vol. 133, no. 4,

pp. 447-54, Jan 26 2016.

[2] J. Ciabattoni, Doctor C's Medical Guide: What You Need To Know: Xlibris Corporation,

2009.

[3] U. R. Acharya, Advances in Cardiac Signal Processing, 1st ed.: Berlin: Springer, 2007.

[4] S. H. Lee, et al., "Classification of Ventricular arrhythmia using a support vector machine

based on morphological features," in 35th Annu. Int. Conf. IEEE Eng. Med. Biol. Soc.

(EMBC), 2013, pp. 5785-5788.

[5] C. Cortes and V. Vapnik, "Support-vector networks," Machine Learning, vol. 20, no. 3,

pp. 273-297, 1995.

[6] T. Conde, "Automatic neural detection of anomalies in electrocardiogram (ECG)

signals," in Proc. IEEE Int. Conf. Neural Netw., Orlando, FL, 1994, pp. 3552-3558.

[7] T. Heskes, "Energy functions for self-organizing maps BT," in Kohonen Maps, E. Oja

and S. Kaski, Eds., Amsterdam: Elsevier Science B.V., 1999.

[8] J. Hertz, et al., Introduction to the Theory of Neural Computation: Addison-Wesley,

1991.

[9] B. Vuksanovic and M. Alhamdi, "ECG based system for arrhythmia detection and patient

identification," in Proc. 35th Int. Conf. Inform. Tech. Interfaces (ITI), 2013, pp. 315-320.

[10] F. V. d. Heijden, et al., Classification Parameter Estimation and State Estimation: An

Eng. Approach using MATLAB: John Wiley & Sons., 2004.


11

[11] A. J. Izenman, "Linear Discriminant Analysis," in Modern Multivar. Stat. Tech., Springer

New York, 2008, pp. 237-280.

[12] Y. H. Noh, et al., "Implementation of real-time abnormal ECG detection algorithm for

wearable healthcare," presented at the 6th Int. Conf. Comput. Sci. Converg. Inform. Tech.

(ICCIT) 2011.

[13] L. Khadra, et al., "A quantitative analysis approach for cardiac arrhythmia classification

using higher order spectral techniques," IEEE Trans. Biomed. Eng., vol. 52, no. 11, pp.

1840-5, Nov 2005.

[14] Y. K. Jung and W. J. Tompkins, "Detecting and classifying life-threatening ECG

ventricular arrythmias using wavelet decomposition," in Proc. 25th Annu. Int. Conf. IEEE

Eng. Med. Biol. Soc., 2003, pp. 2390-2393.

[15] R. M. Rao and A. S. Bopardikar, Wavelet transforms: Introduction to theory and

applications: Addison Wesley Longman, Inc., 1998.

[16] A. L. Goldberger, et al., "PhysioBank, PhysioToolkit, and PhysioNet: components of a

new research resource for complex physiologic signals," Circulation, vol. 101, no. 23, pp.

E215-E220, Jun 13 2000.

[17] P. M. T. Broersen, Automatic Autocorrelation and Spectral Analysis, 1st ed.: Springer

Science & Business Media, 2006.

[18] S. A. Tretter, Communication System Design Using DSP Algorithms with Laboratory

Experiments for the TMS320C6713TM DSK: Boston, MA: Springer-Verlag US, 2008.

[19] G. Hanrahan, Artificial Neural Networks in Biological and Environmental Analysis: Boca

Raton, FL: CRC Press, 2011.

[20] C. M. Bishop, Pattern Recognition and Machine Learning: Springer-Verlag New York,

Inc., 2006.

[21] A. Darouei and A. Ayatollahi, "Classification of cardiac arrhythmias using PNN neural

network based on ECG feature extraction," in IFMBE Proc. World Congr. Med. Phys.

Biomed. Eng., Munich, Germany, 2009, pp. 309-311.

This page is intentionally blank.

PJCIS (2016), Vol. 1, No. 1 : 13-24 Vertebra Localization Using Shape

13

Vertebra Localization Using Shape Based Analysis

and Unsupervised Clustering from X-ray Images

ANUM MEHMOOD1, M. USMAN AKRAM1*, MAHMOOD AKHTAR1, 2

1 College of Electrical and Mechanical Engineering, (NUST), Pakistan

2 The University of New South Wales (UNSW), Australia


(Received on 1st May, 2016 Accepted on 24th August, 2016)

Abstract

Accurate vertebral detection in X-ray images is a challenging task mainly

due to low contrast and noisy set of image data. For the diagnosis of spinal disorders

such as cervical spine trauma and whiplash, the detection and segmentation of

vertebra are the fundamental tasks. The first step in detection process is the vertebra

localization. In this paper, we propose a new method for the cervical vertebra

localization problem. The proposed method contributes a novel composition of a

mean model matching using the Generalized Hough Transform (GHT) and

unsupervised clustering technique. To detect edges and enhance image contrast,

preprocessing is performed on the input X-ray images. After manually selecting

region of the interest (ROI), we use a separately generated geometric mean model as

a template. A modified GHT is then used for the localization of vertebra followed by

Fuzzy c-Means (FCM) clustering technique to obtain centroids of targeted five

vertebras (C3 − C7). The proposed method secured localization accuracy of 96.88%

when tested on 50 X-ray images of publically available database ‘NHANESII’.

Keywords: Generalized Hough transform, Fuzzy c-Means, Vertebra

localization, Shape based analysis, Unsupervised clustering.

INTRODUCTION

During the past few years, medical imaging has become one of the most useful

technology for diagnosis of different diseases such as brain tumor, lung cancer, liver and

kidney problems, etc. [1] . This revolution has also helped to obtain information which is

useful in many clinical applications and diagnoses of disorders such as osteoporosis,

spinal ruptures and cervical spine trauma [2]. The spinal column, along with sacral region

and coccyx, consists of seven ‘cervical’, twelve ‘thoracic’, and five ‘lumbar’ vertebrae

[3] . The radiographic anatomy of cervical spine is shown in Figure 1.

mailto:[email protected]


14

Figure 1.Radiographic anatomy - cervical spine lateral

Each vertebra has the vertebral body for load-bearing, the vertebra larch to protect

the spinal cord, and transverse processes for ligament attachment. The inter vertebral

discs separate the individual bones providing additional weight-bearing support to these

discs and function like shock absorbing springs. The cervical injuries may affect arms,

legs, and middle parts of the body [4] . The vertebrae column serves as a support to the

other organs of a human body. The localization of vertebra performed accurately can play

an important role in the detection of cervical spine disorders. It is of great significance in

many orthopedics and neurological applications that the right vertebra be treated. In low

contrast X-ray images, the localization of vertebra is quite challenging and a tiring task.

Therefore, accurate vertebra localization with high accuracy would be of great interest to

the radiologists’ community.

Many techniques for vertebra localization and segmentation have been proposed

such as active shape model [5] , GHT [6] , etc. The medical imaging modalities such as

computed tomography (CT) and magnetic resonance imaging (MRI) are greatly benefited

from these contributions.

A model based on two phases for targeting inter-vertebral discs using MRI was

proposed by Alomari et al. [7] . Their work (i.e., two experiments using 50 and 55 MRI

cases) led them attain an accuracy of 87% and 89.1% respectively. Larhmam et al. [8]

built a mean model of the vertebra intended for template matching. They used GHT for

vertebra localization and k-means clustering to obtain targeted vertebra’s clusters.

Klinder et al. [9] proposed a computerized model for the vertebra localization and

segmentation using CT scans. The GHT-based detection by employing adapted

triangulation shape to localize and segment different vertebrae reports around 70%

identification accuracy. Benjelloun et al. [10] presented a comparative study of two

algorithms for segmentation of X-ray images in their semiautomatic implementation for


15

analysis and estimation of vertebra. Based on ASM segmentation, Lecron et al. [11]

developed an implementation which enables the localization of vertebrae using X-ray

images. The authors have reported reduced computing time through synchronized

manipulation of more than one CPUs and GPUs. Larhman et al. [12] presented three step

automatic diagnosis of spinal column based on off-line training of cervical vertebra and

vertebra centroids localization using GHT and adaptive filter post-processing. They have

reported accuracy of 89% using 200 cervical vertebrae. Yao et al. [13] developed a

methodology for extracting and partitioning of the spinal cord using watershed algorithm

on CT scan images. Benjelloun et al. [14] developed a method for the identification of

vertebra boundaries by the use of segmentation along with polygonal regions. This

technique was applied for the analysis of vertebra mobility. Recently, they have

introduced method based on active shape model for vertebra identification and

segmentation [15].

A graphical-model based solution was presented by Dong et al. [16] reporting an

automatic calculation of the visible vertebrae in images using radiographs. Casciaro et al.

[17] used local phase measure for the detection of vertebra and obtained 83% accuracy.

Lecron et al. [18] proposed method based on multiclass SVM. They trained SVM model

using SIFT descriptor and 81.6% localization accuracy was reported using 50 X-ray

images.

We have focused on development of a method for detection of cervical vertebrae

in cervical spine X-rays. This method is used to recognize arbitrary shapes using

generalized Hough transform and fuzzy c-means clustering. This paper is organized as

follows. First of all different existing techniques are reviewed, followed by our proposed

methodology. The next section presents results and a discussion. The last section

concludes our work.

Existing Techniques A. Generalized Hough Transform(GHT)

The Hough transform is a widely used technique in many image processing and

computer vision related applications. It was originally proposed for the detection of lines,

parabolas, circles, etc. but Ballard [6] generalized it to detect arbitrary shapes as well.

Therefore, the generalized Hough transform (GHT) became a technique capable to be

utilized for pattern recognition. The modification in scale, orientation and translation has

no impact on the process of detection. The GHT uses a voting scheme which basically

locates the area of candidate X-ray image matching with the mean model, also named as


16

template. The process of detection using GHT comprises two phases: R-table and

Accumulator.

The R-table basically represents the template image. It holds the information

about the ‘position’ and ‘orientation’ of reference image. During training phase, R-table

is constructed, as follows. Firstly edges are detected then we find a reference point

c = (cx , cy) in the candidate X-ray image by using a mean model of required shape. Then

the ‘distance’ and ‘angle’ are computed using the borderline and reference point of the

image. For every landmark point find ‘orientation’ Φ and the ‘relative position’

r = (rx,ry). We then keep this info in R-table as f(Φ). The R-table utilizes few

parameters representing the template. The voting scheme known as Accumulator is

built as follows: For each landmark point p finds the ‘gradient direction Φp’. Then,

for all the entries representing the location ‘p−ri’ in the accumulator, voting is

carried out. Here ri shows the location (ri, βi) in index Φ = Φp in the R-table. In

this process of voting, local maxima is worked out which helps in shape detection.

B. Fuzzy C Clustering

The fuzzy c-means (FCM) is a technique which permits one sample of data to fit

in to two or more clusters. This algorithm has widely been used and is a popular

technique of clustering introduced by Dunn in 1973 which was later improved by Bezdek

[19, 20] . The FCM partitions a set of vectors Xi, i = 1, 2, 3...n into ‘ Cdivisions and

defines the center of cluster in every group in such a way that dissimilarity cost parameter

is reduced. The technique performs grouping such that an input data can be part of

multiple groups with specific degree of membership indicated by values from 0 to 1. It is

an iterative algorithm. In a group operation, the algorithm finds out the centers‘𝐶𝑗′and the

matrix of belongingness ‘𝑀𝑏′ making use of the phases.

METHODOLOGY

The proposed method is based on offline training of vertebra mean image. The

vertebra localization is performed using generalized Hough transform (GHT) and after

candidate vertebra localization fuzzy c-means clustering is applied to get centroids of

the targeted vertebras (C3−C7). Figure 2 shows an overview of our proposed method.

1) Training

The reference image of GHT stages are used to build a mean image

representation of the vertebra. Therefore, in order to reduce the work overload of the


17

Figure 2: Flow diagram of the proposed system

template matching, a mean shape of the vertebra body is constructed. This mean

model is constructed using a set of 50 vertebrae to represent the average shape. The

average shape is generated through manual contour selection. The mean model of

vertebra body is constructed using the Eq (1) where ‘Nʹ shows the total number of

vertebra used for the mean model (N = 50) and ‘I’ represents one vertebra body

image.

Meanm

N

iiIN 0

1

(1)

2) Vertebra Localization

For the accurate localization of vertebra following steps are to be performed

on input X-ray image:

1) Pre-processing: The input X-ray images are of low contrast and need to be

enhanced for vertebra detection. For this purpose the adaptive histogram equalization can

be applied. It changes the contrast of X-ray image areas by computing the local histogram

of that specific area. This technique reduces the noise amplification of the input image.

The canny edge detector gives the edge points of the enhanced image. Next, the region of

interest (ROI) is manually selected to decide feasible vertebra candidate in the input X-

ray. The ROI covers an area of cervical vertebrae including C3 to C7.


18

2) Generalized Hough Transform (GHT): After preprocessing of input X-ray image

which includes contrast enhancement, edge detection and selection of ROI, GHT is

applied on the resultant image. In GHT training phase, R-table (Table 1) is built which

basically represents the vertebra mean model. It involves ‘position’ and ‘direction’ of

landmark points which we get in pre-processing step. R-table is built by following steps:

• Find out the edges of the input X-ray images using canny edge detector.

• Select a point of reference say (xr,yr).

• Join the point of reference & boundary with a straight line.

• Calculate Φ

• In R-table, enter the point of reference (xr,yr) as f(Φ).

Containing edge points pe(xe, ye), e = 1, 2, ., n where n represents the total

number of points and Φi gives gradient w.r.t.i. An angle formed by the reference point

c=(xr,yr) with horizontal direction is measured by Equation (2).

𝑐 = 1

𝑛∑ 𝑝𝑒 (2)

Table 1: General R-Table Form

Orientation Φ Position (r,β)

0 (rx,βx)/Φx =0

Φ (rx,βx)/Φx =∆Φ

2∆Φ (rx,βx)/Φx =2∆Φ

3∆Φ (rx,βx)/Φx =3∆Φ

... ...

The distance rx and βx are calculated using Equation 3 and Equation 4.

𝑟𝑥 = √(𝑥𝑟 − 𝑥𝑒)2 + (𝑦𝑟 − 𝑦𝑒)2 (3)

𝛽𝑥 = tan−1 𝑦𝑒−𝑦𝑟

𝑥𝑒−𝑥𝑟 (4)


19

Table 1 [6] represents the general ‘R-table’ In second half of GHT,

accumulator is constructed in which for each edge point p, Φp is calculated

representing the gradient direction. Then voting is performed for all the positions. In

the voting technique of GHT, a local maximum is calculated for the identification of

the vertebra shape. After voting procedure, we get the final output of GHT as shown

in Figure 3 associated with ‘Candidate Vertebra Localization’ in which there are

points covering the area of vertebra (C3−C7).

3) Fuzzy C Means clustering: Next, Fuzzy C Mean algorithm is applied on the

points we get as candidate vertebra localization from GHT. Total five clusters (C3 −

C7) are formed with final centroids after several iterations of Fuzzy C Means. After

each iteration membership of data points change and the process continue still we get

same membership results in consecutive two iterations, making algorithm converge.

As a result, five clusters are formed representing targeted C3 to C7 vertebrae.

Figure 3: Pixel wise distance between annotated and automated calculated centers for each

vertebra along with mean pixel distance (Green Horizontal Line)


20

Experimental Evaluation

The publically available database ‘NHANES II’ containing 17000 X-ray images

[21] has been selected for evaluation of our proposed method. It includes both cervical

and lumbar X-ray images of different patients under several conditions and orientation. In

literature, experiments are performed using different number of images of same dataset

but none of them has mentioned on which bases they have selected specific image or

number of images. We have used 50 randomly selected cervical spine images for testing

of our proposed methodology. The mean model is created manually using 50 vertebra

body images of NHANES II dataset. All the images are visualized and center points for

each vertebra from C3to C7 are manually annotated. These annotated centroids are used as

a ground truth for vertebra localization validation. Figure 3 shows the distance calculated

between these annotated centers and the centroids we obtained as a final FCM output.

Each graph show the distance of each cervical vertebra for all the 50 images. Further, this

distance helps to calculate mean and standard deviation, using which we can draw mean

error bar for each cervical vertebra.

Experimental evaluation is performed on two levels using 50 testing images

(Figure 4). In visual examination, we analyze the resultant images individually and count

the correct results. If the centroid obtained is within the body of vertebra, we would

consider it as correctly localized center. We get localization accuracies of 100% for C3,

98% for C4, 98% for C5, 92% for C6 and 92% for C7 with overall accuracy of 96%. Our

results are shown in Figure 4a. The mean error values of C3, C4, C5, C6 and C7 are 8.1378,

7.6241, 7.5284, 8.4043 and 10.8478, respectively as shown by green horizontal line in

Figure 3. In second level, Receiver Operating Characteristics (ROC) curves are generated

using pixel distance as shown in Figure 4b. The ROC curves [22] curves basically

illustrates the performance of the system by varying threshold. They are used to evaluate

the response of system in different conditions. These curves show that proposed system

achieved 96.88% accuracy at a threshold of 14 pixels. This threshold is selected after

analyzing the average size of vertebra in the dataset. Figure 5 shows visual results of the

proposed vertebra localization system for some randomly selected X-ray images. The

first, second, third and fourth columns show original radiograph, GHT, clustering, and

centroid of each cluster representing C3C7 results, respectively. It is observed that C6 and

C7 is not located accurately in many cases as compared to other vertebrae due to the

misleading results of edge detection. The detection becomes more problematic task due

to the noise covering the cervical area. Moreover, contrast enhancement of the input

images has a significant importance in edge detection. Therefore, adaptive histogram

equalization is used to enhance the detection of edges and efficient computation of

gradient. Table II gives a comparison between different techniques. Results are shown for


21

selected images taken from the same NHANES II dataset as used by other researchers.

The algorithm proposed by Larhman et al. [8] achieved an accuracy of 97.5% which is a

bit higher than accuracy achieved by our proposed technique i.e. 96.88%. This is mainly

due to our testing of proposed method on randomly selected subset of images.

Figure 4. Vertebra localization accuracies a) Visual Examination based accuracies b) ROC

curves for each vertebra using pixels distances

Table 2: Comparison of different Vertebra Localization Techniques

Paper Year No. of

Images

No. of

vertebrae

Accuracy

%

Larhman [12]

2012

40

200

89

Benjelloun [23] 2012 40 200 Automatic: 64.5

Semi-Automatic: 89

Lecron [18] 2012 50 250 81.60

Larhman [8] 2013 66 330 97.5

Proposed Method 2016 50 250 96.88


22

Figure 5.Visual results of the proposed vertebra localization system using randomly selected

X-ray images.


23

CONCLUSION

This work has proposed a novel combination of mean model matching using

generalized Hough transform and unsupervised clustering technique to locate five

cervical vertebrae from C3 to C7. The performance of the proposed method has been

tested using the cervical X-ray images of publically available database ’NHANES II’.

The proposed method has been shown to give satisfactory results with a prediction

accuracy of 96.88% on 50 cervical radiographs. The future work is focused towards use

of localized centroids for automatic segmentation of vertebra.

REFERENCES

[1] J. Levy and D. Zieve. (April 06,2016). Imaging and Radiology. Available:

http://slu.adam.com/content.aspx?productId=617&pid=1&gid=007451

[2] L. A. Nadalo, et al. (April 06, 2016). Thoracic Spinal Trauma Imaging. Available:

http://emedicine.medscape.com/article/397896-overview#a1

[3] A. Ngrdi and G. Vrbov. (June 24, 2016)."Anatomy and Physiology of the Spinal Cord,"

in Transplantation of Neural Tissue Into the Spinal Cord, A. Nogradi, Ed., 1st ed:

Springer. Available: https://books.google.com.pk/books?id=E3AkpCqhHP8C

[4] L. Jasmin. (April 24, 2016). Spinal Cord Trauma. Available:

http://www.nytimes.com/health/guides/disease/spinal-cord-trauma/overview.html

[5] T. F. Cootes and C. J. Taylor, "Statistical Models of Appearance for Computer Vision,"

University of Manchester, Manchester, UK, 2004.

[6] D. H. Ballard, "Generalizing the hough transform to detect arbitrary shapes," Pattern

Recognition, vol. 13, no. 2, pp. 183-194, 1981.

[7] R. S. Alomari, et al., "Labeling of lumbar discs using both pixel-and object-level features

with a two-level probabilistic model," IEEE Trans. Med. Imag., vol. 30, no. 1, pp. 1-10,

2011.

[8] M. A. Larhmam, et al., "Vertebra identification using template matching modelmp and

K-means clustering," Int. J. Comp. Assisted Radiol. Surg., vol. 9, no. 2, pp. 177-187,

2014.

[9] T. Klinder, et al., "Automated model-based vertebra detection, identification, and

segmentation in CT images," Medi. Image Anal., vol. 13, no. 3, pp. 471-482, 2009.

[10] M. Benjelloun and S. Mahmoudi, "X-ray image segmentation for vertebral mobility

analysis," Int. J. Comput. assisted Radiol. Surg., vol. 2, no. 6, pp. 371-383, 2008.

[11] F. Lecron, et al., "Heterogeneous computing for vertebra detection and segmentation in

X-ray images," J. Biomed. Imag., vol. 2011, no. 5, pp. 5:1-5:12, 2011.

[12] M. A. Larhmam, et al., "Semi-automatic detection of cervical vertebrae in X-ray images

using generalized hough transform," in Proc. 2012 3rd Int. Conf. Image Process. Theory,

Tools and Applicat. (IPTA), 2012, pp. 396-401.


24

[13] J. Yao, et al., "Automated spinal column extraction and partitioning," in Proc. 3rd IEEE

Int. Symp. Biomed. Imag.: Nano to Macro, 2006, pp. 390-393.

[14] M. Benjelloun and S. Mahmoudi, "Spine localization in X-ray images using interest point

detection," J. Digit. Imag., vol. 22, no. 3, pp. 309-318, 2009.

[15] M. Benjelloun, et al., "A framework of vertebra segmentation using the active shape

model-based approach," J. Biomed. Imag., vol. 2011, no. 9, pp. 1-14, 2011.

[16] X. Dong and G. Zheng, "Automated vertebra identification from X-ray images," in Proc.

Int. Conf. Image Anal. Recog., 2010, pp. 1-9.

[17] S. Casciaro and L. Massoptier, "Automatic vertebral morphometry assessment," in Proc.

29th Annu. Int. Conf. IEEE Eng. Med. Biol. Soc., 2007, pp. 5571-5574.

[18] F. Lecron, et al., "Fully automatic vertebra detection in X-ray images based on multi-

class SVM," in SPIE Medical Imaging, Int. Soc. Optics Photonics, 2012, pp. 83142D-

83142D.

[19] J. C. Dunn, "A Fuzzy Relative of the ISODATA Process and Its Use in Detecting

Compact Well Separated Clusters'," Cybernetics Sys., vol. 3, no. 3, pp. 32 - 57, 1973.

[20] J. C. Bezdek, Pattern recognition with fuzzy objective function algorithms: Springer

Science & Business Media, 2013.

[21] NIH. (Jan 27th, 2016). NHANES II X-ray Images. Available:

https://archive.nlm.nih.gov/proj/ftp/ftp.php

[22] T. Fawcett, "An introduction to ROC analysis," Pattern Recog. Lett., vol. 27, no. 8, pp.

861-874, 2006.

[23] M. Benjelloun, et al., "Template Matching Approaches Applied to Vertebra Detection,"

in Advances in Image Segmentation, P.-G. P. Ho, Ed., 1st ed: InTech, 2012.

PJCIS (2016), Vol. 1, No. 1 : 25-39 Numerical Solution of Initial

25

Numerical Solution of Initial Boundary Value

Problems by MOL-RBF Method

FAZAL-I-HAQ1, FARAH JAFFAR2*, SIRAJ-UL-ISLAM3,

1Department of Mathematics, Stats and Computer Science, KPK Agricultural University, Peshawar, Pakistan

2Department of Mathematics, Shaheed Benazir Bhutto Women University, Peshawar, Pakistan

3Department of Basic Sciences, NWFP University of Engineering and Technology, Peshawar, Pakistan

*Corresponding author’s e-mail: [email protected]

(Received on 27hApril, 2016 Accepted on 17th June 2016)

Abstract

A recently developed meshless method, namely the radial basis function (RBF)

combined with method of lines (MOL), i.e RBFMOL is used to find numerical

solution of initial boundary value problems. Numerical examples are given to illustrate the

practical usefulness of this approach. 2, LL

norms and RMS are used for error

estimation. Superiority of the proposed method is shown as compared with two existing

numerical methods such as variational iteration method and finite difference method.

Keywords: Method of lines (MOL), Korteweg-de-Vries (KdV)equation, Radial

basis function (RBF), Multiquadric (MQ), Inverse Multiquardic (IMQ), Guassian (GA),

Fourthorder Runge Kutta (RK4), Finite difference method(FDM)

INTRODUCTION

In 1895, a partial differential equation was introduced by Korteweg and de Vries to

model the height of surface of shallow water in the presence of long gravity waves which is

called Korteweg-de-Vries (KdV) equation [1]. After 1960 Zabusky and Kruskal [2]

numerically discovered the elastic collision between the KdV solitary waves and then

Gardner et al.[3, 4]find the inverse scattering transform method and also solved the KdV

equation analytically. That was a pioneering work which initiated many research activities

on nonlinear waves. Since non linear phenomena plays crucial role in a variety of scientific

fields, especially in fluid mechanics, solid state physics, thermodynamic etc [5].

Various researchers applied different numerical methods such as bilinear method of

Hirota [6], the homogenous balance method [7], Sine-cosine method [8], Backlund



26

Transformation [9, 10], the Tanh method [11], the variational iteration method (VIM) [12,

13]Darboux transformation [14] etc to find the solution of nonlinear PDEs with

appropriate initial or initial and boundary conditions.

We consider a nonlinear partial differential equation of the form;

0,=3

3

x

u

x

uu

t

u m

(1)

where and are constants and 0,1,2.=m

For 0=m the Eq . (1) becomes linear KdV equation.

For 1=m the Eq . (1) becomes nonlinear KdV equation [15].

For 2=m the Eq . (1) becomes nonlinear Modified KdV equation [16].

KdV equation has some physical applications such as shallow water gravity

waves, internal waves in the atmosphere and ocean etc. In fact it shows a combined effects

of nonlinearity )(x

uu

and the simplest longwave dispersion )(

3

3

x

u

[4].

Since there is a class of nonlinear wave equations (soliton equations) which have

the property of complete integrability such as Sine Gordon equation, the modified

Korteweg-de-Vries equation and the cubic nonlinear Schrodinger equation [17].

We will study modify KdV equation.

0,=3

32

x

u

x

uu

t

u

(2)

we can also write this equation as

0,=2

xxxxt uuuu (3)

mKdV equation describes the motion of waves in nonlinear optics, plasma or fluids [18]. It

occurs in many nonlinear fields like acoustic waves in certain anharmonic lattices, schottky

barriers of transmission lines, traffic congestion models, ion-acoustic solitons, Alfven

waves in a collisionless plasma etc [7, 19]. It advocates several characteristic such as

Conservation laws, N-solitons, Muira transformation, Inverse scattering transformation,

Darboux transformation and Bilinear transformation [7].The modified KdV (mKdV)

equation was solved by various numerical techniques both analytically and numerically

such as Homotopy perturbation method (HPM) [19], Reduced differential transform

method [5], Exp-function method [3], Backlund transformation [20], Adomian

decomposition method (ADM) [15, 21]. The Extended Cosine-function is used to obtain

the solution of traveling wave of the mKdV equation [22] , Homotopy analysis method


27

[18], Yuanxi [23] found series of explicit and exact solutions by Trial function method,

Gesztesy et al.[24] applied Commutation method, Turabi and Dogan[16] used ADM and a

‘lumped’ Galerkin method with quadratic B-spline Finite Element method and obtained

the numerical solution, Ibrahim and Kalaawy [22] used Extended Cosine-function to find

the traveling wave solution of mKdV equation.

Meshless method is used in this study due to its remarkable properties of computing

high dimension data, bringing changes in the domain of interest (like free surfaces and

large deformations in the field of geometry) and consuming lesser time due to its

independence from mesh. There are different meshless methods like Diffuse element

method (DEM), Reproducing kernal particle method (RKPM), Moving Least square

method, Radial basis function (RBF) [25].

The radial basis function (RBF) combined with method of lines (MOL), or

RBFMOL is used to find numerical solution of initial boundary value problems i.e

mKdV equations.

This paper is organized as follows.

In the introduction authors introduce the KdV and mKdV equations. Authors

discuss RBF in detail that how they use the RBFs interpolation to approximate the solution

in the section named Radial Basis Function. Method of lines for mKdV equation using

RBFs is applied in next section. The section named Implementation and Results shows the

results of numerical examples and comparison through tables and figures. After this

convergence analysis is given and the last section includes conclusion.

Radial Basis Function

When a function to be approximated there are three cases to be considered

1. It depends on many variables or parameters,

2. It is defined by possibly many data,

3. The data are scattered in their domain.

The RBF approach is suitable for these cases [26].

It is one of the most advanced meshless method. It was first introduced in 1990 by

Kansa[27] for the solution of PDEs. RBF has an edge over the other numerical methods

like Finite volume method (FVM), Finite difference method (FDM) and Finite element

method (FEM) because of its easy implementation, fast convergence [2], independent of

mesh and accurate results.


28

Due to its properties, RBFs have been applied successfully to obtain numerical

solution of many types of PDEs and ODEs including heat transfer equation, shallow water

equation for tide and currents simulation, the nonlinear Burger equation upto two

dimension [25]. Because of its meshless property it is useful for 3D problems as well as

those problems that require re-meshing such as in nonlinear analysis [28]. It has many

applications in science and mathematics such as mapping of two or three dimensional

images like portraits or underwater sonar scan into other images for comparison [26].

RBF can be globally supporting, infinitely differentiable and consist of a shape

parameter which needs to be selected from selected or random region for obtaining

required accuracy [29].

Commonly available RBFs are

(1) Multiquadric (MQ ):22=)( crr jj ,

(2) Inverse Multiquadric (IMQ ):22

1=)(

crr

j

j

,

(3) Gaussian (GA): 22

=)( jrc

j er

,

(4) Quadric (Q): )(=)( 22 crr jj ,

(5) Inverse quadric (IQ):)(

1=)(

22 crr

j

j

,

where c is the shape parameter.

Interpolation of RBF

Using )(xu N as the approximate function, RBF method gives [30]

,)(=)(=)(1=

cxrcxu T

kk

N

k

N (4)

where N is the number of data point, is any form of RBF , kk xxr which

represents the Euclidean norm between collocation points x and kx in the interval ],[ ba

T

N xxxx )]()...(),([=)( 21 , T

Ncccc ],...,,[= 21 ,

Let kk

N uxu =)( , then


29

T

NuuuuuAc ],...,,[=where,= 21 (5)

)(

...

)(

)(

=

2

1

N

T

T

T

x

x

x

A

)()()(

)()()(

)()()(

=

21

22221

11211

NNNN

N

N

xxx

xxx

xxx

from Eq. (4) and Eq. (5), we get

,)(=)(=)( 1 uxMuAxxu TN (6)

where )](),(),...,(),([=)(=)( 121

1 xMxMxMxMAxxM NN

T

,

Method of Lines for mKdV Equation Using RBFs

In this section we solve mKdV equation (3) in a finite domain

0,=2

xxxxt uuuu (7)

],[ bax

the initial condition is

)(=),( 0

0 xutxu (8)

and boundary conditions are

)(=),(,)(=),( tgtbutftau (9)

here and are real constant and )(0 xu , )(tf , )(tg are known functions.

Now we discretize the spatial derivatives according to the first step of MOL by using RBF

interpolation . So that we choose N nodes in ],[ ba

such as bxxxxa NN =<<<<= 121

by RBF interpolation the following results are obtained

uxMuAxctxutxu T

kk

N

k

N )(=)(==),(),( 1

1=

(10)

when we apply Eq. (10) to Eq. (7) , and collocate on the node kx , we obtain

NkuxMxMudt

dukxxxkxk

k 1,2=0,=)()((2 (11)

as )(tuk is abbreviated to ku

as we have

)]()()()([=)( 1)(21 kNxkxNkxkxkx xMxMxMxMxM


30

)(=)( klklx xMx

xM

, Nl ,1,2=

and also we have

)]()()()([=)( 1)(21 kNxxxkxxxNkxxxkxxxkxxx xMxMxMxMxM

)(=)(3

3

klklxxx xMx

xM

, Nl ,1,2,=

We write this system in term of the column vectors.

Let ,][= 121

T

NN uuuuU

NNklxx xM )]([=M ,

NNklxxxxxx xM )]([=M ,

So that above equation (11) can be written as

0,=M)M(*2 UUUdt

dUxxxx (12)

Where * is the multiplication of two vectors component-by-component.

We can also write the above equation as

),(= UFdt

dU (13)

Where

UUUUF xxxx M)M(*=)( 2 ,

as the initial condition is

,)]()()()([=)( 0

1

0

2

0

1

0

0

T

NN xuxuxuxutU

(14)

and the boundary conditions are

)(=)(),(=)(1 tgtutftu N (15)

After completion of first step of RBFMOL method we apply an ODE solver to solve

equations . (13)-(15).

In this study fourth order Runge Kutta (RK4) is used as an ODE solver which is given by

6

)22(= 43211 KKKKt

UU nn

,

)(=1

nUFK , )2

(= 12 Kt

UFK n ,

)2

(= 23 Kt

UFK n , )(= 34 tKUFK n [30]


31

Implementation and Results

Example 1: Consider the mKdV equation along with the initial condition [15]

,14

4=,0)(0,=6

22

2

xa

aaxuuuuu xxxxt

(16)

where “a” is a real constant , by taking 6= and 1= .

The exact solution is

.1))6((4

4=),(

222

taxa

aatxu

The boundary conditions are extracted from the exact solution.

The computational domain is [0.1,0.5] , 0.50.1= t , 0.0001=t , 0.1=5,= hN .

We used the following error norms.

|,|max==1

k

N

kNkmax

N uuuuL

,)(== 2

1=22 k

N

k

N

kL

N uuhuuL

,/)(=(RMS)squaremeanRoot 2

1=

Nuu k

N

k

N

k

where Nu is approximate solution and u represents the exact solution.

In Table 1 RBFs shows higher accuracy than FDM as a result of use of 2, LL norms and

Root mean square (RMS) for error estimation, which shows the accuracy of 510 and 610 by different values of shape parameter of RBFs while FDM shows 110 accuracy .

The achievements of 5th and 6th order accuracy has been shown in Table 1. The Figures 1-2

show better graphical and 3D representation of approximate solution of RBFs with the

exact solution of IBVP.


32

Figure 1: Approximate solution of RBFs

of example 1

Figure 2: 3D graph of approximate solution

of RBFs of example 1

Table 1: Computational domain is [0.1, 0.5] × [0.1, 0.5]

t MQ

Max-error 2L error

RMS

0.1 2.2311E-5 1.0303E-5 1.2315E-5

0.2 4.8340E-5 2.2140E-5 2.6463E-5

0.3 7.5163E-5 3.4350E-5 4.1056E-5 0.4 1.0279E-4 4.6904E-5 5.6061E-5

0.5 1.3122E-4 5.9806E-5 7.1482E-5

IMQ

0.1 3.4265E-5 1.7730E-5 2.1192E-5 0.2 6.2099E-5 3.1012E-5 3.7066E-5

0.3 8.9720E-5 4.4179E-5 5.2804E-5

0.4 1.1711E-4 5.7212E-5 6.8382E-5 0.5 1.4425E-4 7.0102E-5 8.3788E-5

GA

0.1 2.4163E-6 1.1231E-6 1.3424E-6

0.2 4.8627E-6 2.2281E-6 2.6631E-6 0.3 7.3371E-6 3.3148E-6 3.9619E-6

0.4 9.8375E-6 4.3831E-6 5.2389E-6

0.5 1.2361E-5 5.4332E-6 6.4940E-6

FDM

0.1 1.0124E-1 5.1572E-2 6.1641E-2

0.2 2.2261E-1 1.0025E-1 1.1983E-1

0.3 3.6342E-1 1.5022E-1 1.7955E-1 0.4 5.0851E-1 2.0883E-1 2.4960E-1

0.5 6.3184E-1 2.7034E-1 3.2312E-1


33

Example 2: From the following mKdV equation the traveling wave solution was obtained

[15].

),(sec=,0)(0,=6 2 xbkhbxuuuuu xxxxt (17)

for all 0b , and k is an arbitrary constant. 0.001=0.1,=5,= thN .

We have used RBFs such as MQ, IMQ, GA and the results are compared with

FDM.

In Table 2 comparison shows that RBFMOL has better results and higher order

accuracy as compared to FDM. In the case of RBF, GA shows high accuracy than MQ and

IMQ. The graphical results of approximate solution of RBFs and exact solution of IBVP

are shown in Figure 3-4.

Figure 3: Approximate solution of RBFs

of example 2




34

Table 2: Numerical result for the traveling wave solution with c = 1.0 , k = 7

t MQ

Max-error 2L error

RMS

0.1 5.4013E-6 2.8995E-6 3.4655E-6

0.2 8.6746E-6 4.5079E-6 5.3879E-6 0.3 1.2434E-5 6.3992E-6 7.6485E-6

0.4 1.6751E-5 8.5572E-6 1.0227E-5

0.5 2.1484E-5 1.0969E-5 1.3110E-5

IMQ

0.1 3.2521E-6 1.5303E-6 1.8291E-6

0.2 1.0060E-5 5.2381E-6 6.2608E-6

0.3 1.7582E-5 9.4075E-6 1.1244E-5 0.4 2.5892E-5 1.4023E-5 1.6761E-5

0.5 3.5074E-5 1.9126E-5 2.286E-5

GA

0.1 2.5514E-8 1.6556E-8 1.9789E-8 0.2 6.0699E-8 3.4416E-8 4.1136E-8

0.3 1.0944E-7 5.3298E-8 6.3703E-8

0.4 1.6736E-7 7.2970E-8 8.7216E-8 0.5 2.2994E-7 9.3323E-8 1.1154E-8

FDM

0.1 1.4932E-2 7.4600E-3 8.9164E-3 0.2 3.0645E-2 1.5362E-2 1.8361E-2

0.3 4.8201E-2 2.4098E-2 2.8803E-2

0.4 6.7360E-2 3.3758E-2 4.0348E-2 0.5 8.8381E-2 4.4442E-2 5.3118E-2

Example 3: Consider mKdV equation [5]

0,=6 2

xxxxt uuuu (18)

with initial condition

)(sec=,0)( xhxu

and exact solution given by

),(sec=),( txhtxu (19)

In this example we have compared results with modified variational iteration method [31]

and RBFs. The pointwise results are presented in Table 3 which demonstrates that at each

point both MQ and GA produce better results. The Figure 5 shows graphical results of

approximate solution of RBFs and VIM with exact solution and proved that RBFs are

better. Figure 7 displayed pointwise absolute error of RBFs and VIM graphically, in which

VIM shows varying error after nodal point 0.3 and at the point 0.8 the error shoot

extensively.


35

Figure 5: Approximate solution of RBFs and

VIM of example 3



Figure 7: Comparative result of absolute error of RBFs and VIM of example 3


36

Table 3: Comparison of the absolute error of RBFs with VIM corresponding to example 3

x-t MQ GA VIM

00 0 0 0

0.10.1 2.736E-4 1.649E-5 1.512E-3

0.20.2 4.879E-3 3.278E-5 9.902E-3

0.30.3 1.333E-2 4.815E-5 3.152E-2

0.40.4 2.342E-2 5.469E-5 2.818E-1

0.50.5 2.717E-2 1.767E-4 8.197E-1

0.60.6 2.363E-2 3.832E-4 1.3742

0.70.7 1.394E-2 1.311E-3 1.3184

0.80.8 3.273E-3 1.595E-3 6.357E-2

0.90.9 1.172E-4 1.638E-3 2.3936

1.01.0 1.110E-16 1.110E-16 5.1783

Convergence Analysis

The pointwise rate of convergence in space is computed by using the following

formula

)/(

)/(

110

110

ii

hihi

hhlog

UuUulog [32]

In this formula u is the exact solution and hiU is the numerical solution with

space step size ih . For the computation of rate of convergence in space for each MQ, IMQ

and GA, the space step size h over the domain [0.1,0.5] was kept fixed as 0.1=h

while keeping the value of shape parameter constant at different collocation points N .

Table 4: 𝑳∞ error norms and space rate of convergence at time t=0.1 of example 1

𝑵𝒕𝒊𝒎𝒆𝒔 𝑳∞ order 𝑳∞ order 𝑳∞ order

MQ IMQ GA

1000 2.2311E-5 3.4265E-5 2.4163E-6 2000 1.8648E-5 0.25 2.5135E-5 0.44 6.6369E-8 5.18 4000 1.8110E-5 0.04 2.3320E-5 0.10 3.2585E-8 1.02 8000 1.7863E-5 0.01 2.2562E-5 0.04 1.5693E-8 1.05 16000 1.7744E-5 0.009 2.2215E-5 0.02 7.2465E-9 -0.48


37

Table 5: 𝑳∞ error norms and space rate of convergence at time t = 0.1 of example 2

𝑵𝒕𝒊𝒎𝒆𝒔 𝑳∞ order 𝑳∞ order 𝑳∞ order

MQ IMQ GA

100 5.4013E-6 3.2521E-6 2.5514E-8 200 9.3815E-7 2.52 1.3529E-6 1.26 3.6123E-7 -3.82 400 9.3811E-7 6.15 1.3333E-6 0.02 3.1026E-7 0.21 800 9.3809E-7 3.07 1.3273E-6 6.50 2.8619E-7 0.11 1600 9.3808E-7 0.000 1.3249E-6 2.61 2.8035E-7 0.02

In the Table 4 various collocation point N are used and the time step size

0.0001=t had been kept constant. The result shows that rate of convergence in RBFs

increases along with the improvement of accuracy L due to increase of collocation

points N . In the Table 5 it is observed that the situation was same as it was noticed in

Table 4, with fixed time step 0.001=t , but in the case of GA the accuracy reduced with

the higher collocation points.

CONCLUSION

The different values obtained by the shape parameter which were selected

randomly by proposed method gave the most accurate results. Significant differences are

observed in the accuracy among RBFs at different shape parameters. It is also observed

that the number of nodes and accuracy are inversely proportional to each other. The

proposed method shows better numerical results for solving IBVPs. All the results indicate

that Gaussian (GA) has the best accuracy as compared to MQ and IMQ of RBFs. The

numerical results confirm that RBFMOL method has better accuracy over the two

numerical methods i.e FDM and VIM. The efficient results of numerical solution of

nonlinear PDEs are obtained due to its two major advantages i.e the meshless property and

use of the best ODE solver.

REFERENCES

[1] R. Knobel, An Introduction to the Mathematical Theory of Waves vol. 3: American

Mathematical Society Institute of Advance Study, 2000.

[2] M. Dehghan and A. Shokri, "A numerical method for solution of the two

dimentional sine gorden equation using the radial basis functional," Math. Comput.

Simul., vol. 79, pp. 700-715, 2008.


38

[3] K. Raslan, "The application of He’s exp-function method for mKdV and Burgers

equations with variable coefficients," Int. J. Nonlinear Sci., vol. 7, no. 2, pp.

174-181, 2009.

[4] R. Grimshaw, "Solitary Waves in Fluids," in Advances in Fluid Mechanics. vol. 47,

1st ed: WIT Press, 2007.

[5] Y. Keskin and G. Oturanc, "Reduced differential transform method for generalized

KdV equation," Appl. Math. Comput., vol. 15, no. 3, pp. 380-393, 2010.

[6] R. Hirota, "Exact solution of the Korteweg-de-vries equation for multiple collisions

of solutions," Phys. Rev. Lett., vol. 27, pp. 1192-1194, 1971.

[7] Z. Yan, "The modified KdV equation with variable coefficients: Exact

uni/bi-variable travelling wave-like solutions," Appl. Math. Comput., vol. 203, no.

1, pp. 106 - 112, 2008.

[8] Z. Y. Yan and H. Q. Zhang, "Auto-darboux transformation and exact solutions of

the brusselator reaction diffusion model," Appl. Math. Mech., vol. 22, pp. 541-546,

2000.

[9] M. J. Ablowitz and P. A. Clarkson, "Solitons, Nonlinear Evolution Equations and

Inverse Scattering," in Cambridge Studies in Social and Cultural Anthropology, ed:

Cambridge University Press, New York, 1991.

[10] A. A. Coaly, Backlund and Darboux transformation: The Geometry of Solitons, 1st

ed.: American Mathematical Society Providence, Rhode Island, 2001.

[11] N. Malfeit, "Solitary wave solutions of nonlinear wave equation," Amer. J. Phys.,

vol. 60, pp. 650-654, 1992.

[12] M. A. Abdou and A. A. Soliman, "Variational iterational method for solving

burgers and coupled burgers equations," Comput. Appl. Math., vol. 181, no. 2, pp.

245-251, 2005.

[13] M. T. Darvishi, et al., "The numerical simulation for stiff systems of ordinary

differential equations," Comput. Math. Application, vol. 54, no. 7-8, pp.

1055-1063, 2007.

[14] M. Wadati, et al., "Relationships among inverse method, Backlund transformation

and an infinite number of conservation laws," Prog. Theor. Phys., vol. 53, pp.

419-436, 1975.

[15] D. Kaya and I. E. Inan, "A convergence analysis of the ADM and an application,"

Appl. Math. Comput., vol. 161, pp. 1015-1025, 2005.

[16] T. Geyikli and D. Kaya, "An application for a modified Kdv equation by the

decomposition method and finite element method," Appl. Math. Comput., vol. 169,

no. 2, pp. 971-981, 2005.

[17] V. M. Roythos, "Homoclinic orbits for a perturbed lattice modified KdV equation,"

Theor. Math. Phys., vol. 134, no. 1, pp. 117-127, 2003.


39

[18] C. Wang, et al., "Solving the nonlinear periodic wave problems with the homotopy

analysis method," Wave Motion, vol. 41, no. 4, pp. 329-337, 2005.

[19] F. Babolian, et al., "Application of homotopy perturbation method to some

nonlinear problems," Appl. Math. Sci., vol. 3, no. 45, pp. 2215-2226, 2009.

[20] J. B. Bi, "Novel solution of mkdv equation with the modified backlund

transformation," J. Shangai university (English edition), vol. 8, no. 3, pp. 286-288,

2004.

[21] W. Lei, et al., "Adomian decomposition method for nonlinear differential

difference equation," Commun. Theor. Phys., vol. 48, pp. 983-986, 2007.

[22] R. S. Ibrahim and O. H. El-Kalaawy, "Traveling wave solution for a modified

ion-acoustic waves in a collisionless," Adv. Studies Theor. Phys., vol. 2, no. 19, pp.

919-928, 2008.

[23] Y. Xie, "New explicit and exact solutions to the Mkdv equation," Int. J. Nonlinear

Sci., vol. 6, no. 2, pp. 124-128, 2008.

[24] F. Gesztesy, et al., "Commutation methods applied to the mKdV equation," Trans.

Amer. Math. Soc., vol. 324, no. 2, pp. 465-525, 1991.

[25] J. Li and Y. C. Hon, "Domain decomposition for radial basis meshless methods,"

Numer. Methods Partial Diff. Equ., vol. 20, pp. 450-462, 2004.

[26] M. Buhmann, Radial Basis Function, Theory and Implementation: Cambrige

University Press, 2003.

[27] E. J. Kansa, "Multiquardics-a scattered data approximation scheme with

applications to computational fluid dynamics i:surface approximation and partial

derivative estimates," Comput. Math. Applicat., vol. 19, no. 8/9, pp. 127-145, 1990.

[28] R. B. Platte, "Accuracy and stability of Global Radial Basis Function method for

the numerical solution of Parial differential equations," Ph D. Dissertation, Dept.

Math. Sci., Univ. Delaware, Newark, DE, United States, 2005.

[29] I. Dag and Y. Dereli, "Numerical solution of KdV equation using radial basis

function," Appl. Math. Modelling, vol. 32, no. 4, pp. 535-546, 2008.

[30] Q. Shen, "A meshless method of lines for the numerical solution of kdv equation

using radial basis functions," Eng. Anal. Bound. Elem., vol. 33, pp. 1171-1180,

2009.

[31] A. Abassy and M. A. E. Zoheiry, "Solving nonlinear partial equations using the

modified variational iteration pade technique," Comput. Appl. Math., vol. 1, pp.

73-91, 2007.

[32] S. Islam, F. Haq, and I. A. Tirmizi, "Collocation method using quartic b-spline for

numerical solution of the modified equal width wave equation," Appl. Math.

Inform., vol. 28, no. 3-4, pp. 611-624, 2010.

This page is intentionally blank.

PJCIS (2016), Vol. 1, No. 1: 41-56 A Comprehensive Review of Intrusion

41

A Comprehensive Review of Intrusion Detection

Systems for Wireless Sensor Networks

WARDA AHMAD, SIDRA ANWAR, YASIR SALEEM*, JUNAID ARSHAD,

KHAWAR BASHIR, HAFSA TAHIR

Department of Computer Science and Engineering, University of Engineering and Technology Lahore, Pakistan

*Corresponding author’s e-mail: [email protected]

(Received on 7th June, 2016 Accepted on 1st August, 2016)

Abstract

Wireless Sensor Networks (WSNs) are gaining wide spread acceptance in the

regimes where close communication with physical world are important. Due to

importance in almost all fields of practical life WSN are vulnerable to wide range of

attacks. Securing the network from those attacks is a vital task in order to achieve the

required performance. Intrusion Detection Systems (IDS) are security mechanisms

against network vulnerabilities. The purpose of this paper is to compare and evaluate

most recently proposed IDS methods in WSNs and identifying their strengths and

weaknesses.

Keywords: Intrusion detection system, Wireless sensor network, Energy

consumption, Detection rate, Review

INTRODUCTION

Wireless sensor networks (WSNs) are types of networks which consist of

hundreds of thousands of tiny nodes. These nodes are sensing devices also called motes.

These motes generate data as well as they act as network relays. Each of these

motes/nodes consists of a transceiver, sensor(s) and a microprocessor. These motes are

designed to consume low power to extend their life-time up to several months and years.

Their processing capability is also low because they only sense data like temperature,

heat, humidity etc from the physical world and send this information to the main system

for further processing. Some motes have on-board microprocessors which initially

process data before sending it to macro processor. They also have low cost. They have a

testing issue in making productive self-organized WSN, on the grounds that sensor nodes

are circulated in wide area [1].


42

One of the important things is that WSN should be adaptable, dependable, self-

organized and secure and have flaw resilience. Due to the low cost and straightforward

proliferation attributes of wireless sensor networks, they have wide spread applications in

many fields of science, military and health. Some of their uses include sensing and

accumulating information regarding different exercises, case investigation of war zone

(e.g. Boomerang Sniper Identifying System), distinguishing NBC (Nuclear, Biological,

Chemical) assaults, observing parkway movement, learning natural life and seas (Great

Duck Island - GDI Project), fire alert framework, home automation systems, agriculture,

transportation and space investigation to name a few [2].

Due to these vast applications of WSNs they are more vulnerable to intrusion

attacks. The security of WSNs is very critical task because of their self-organizing nature,

dependency on the other nodes, limited bandwidth, low battery power [3]. There are two

types of mechanisms which give security to the network. These are prevention based and

detection based. Prevention based mechanisms provide confidentiality, integrity and

authentication security which include cryptography, secure routing, key management and

so on. All these are known as first line security mechanisms. The second is detection

based mechanisms which includes IDS. This is said to be second line security mechanism

for the network [4].

Intrusion is a process of sending malicious software to the network or hijacking a

network. Intrusion detection is a system which is used to detect the intrusion attacks on

wireless sensor networks. In 2015 many intrusion attacks were done on very popular

networks .The most common and devastating attack is cyber-attack like the one on

French television network TV5Monde on 8th April, 2015. Another attack was on US

power grid on Oct 21, 2015 and many others. Due to this, an Intrusion detection system

plays very important role.

One of the major issues in the implementation of IDS to wireless sensor network

is to choose the type of strategy that consumes less power in WSN. The reason being that

each node of WSN has low powered battery, limited computation capability and very less

storage available. Intrusion detection systems are computationally expensive and they are

made for wired and ad hoc network so they are not directly applied to these wireless

sensor networks.

In this paper, we present a comprehensive review of recent intrusion detection

system models for wireless sensor networks. The section named Intrusion Detection

System (IDS) in Wireless Sensor Networks gives an overview of IDS in wireless sensor

networks. In this section we have described some basic intrusion detection methodologies


43

The next section presents the related work done in this field. After this the review of

some recent IDSs for wireless sensor network has been presented. Analysis and

comparison of these IDSs is presented in the next section. This section also discusses

some strengths, weaknesses and future work of the analyzed IDSs. Finally, the last

section closes the paper with a conclusion.

Intrusion Detection System (IDS) in Wireless Sensor Networks

In the network or a system intrusion is the process of gaining unauthorized access

and then performing unauthorized activity. Intrusion is done by two methods; first one is

known as passive intrusion which includes eavesdropping and Information gathering the

second one is known as active intrusion like packet dropping, malicious packet

forwarding, hole attacks etc. For these types of intrusion attacks first line of security

system i.e. Intrusion prevention is not sufficient. So there must be the second line of

security system i.e. Intrusion detection [5].

IDS is said to be the second line of security in any security system. It means it

detects the intrusion activity (active or passive), type of intrusion (warm hole, black hole,

sink hole etc) and protocol layer at which intrusion occurs. It also detects the intruder i.e.

location of intruder. All of this information is very useful for mitigating the intrusion by

placing appropriate controls. Hence Intrusion detection system (IDS) is a hardware or

software system, used for detection of attacks whether they are internal or external attack.

IDS have four main components: sensor, detector, knowledge base and response

component. Sensor collects the data. Then this collected data is analyzed by the detector

with the help of knowledge base because knowledge base contains all the signatures of

serving attacks. And then response component manages the response given by the

detector on the basis of information in knowledge base [2].

IDSs are classified on many bases which are given below:

A. Source of audit data

It means that from which location the data is to be analyzed. On this base IDSs

are classified into three types.

Network Base Intrusion Detection System (NIDS)

Host Based Intrusion Detection System (HIDS)

Hybrid Intrusion Detection System


44

NIDS is systems which captures network traffic on the specific network with the

help of sensors and then analyze this traffic and detect the intrusion attacks like DoS

attack, Port scans etc. It analyze the packets and their headers to find out any malicious

signature present in them [6]. HIDS is a system which detects those attacks which are

inside the system. It maintains the log information of the system then by analyzing this

find out attack. Hybrid intrusion detection system is the combination of both above

systems.

B. Detection methodologies

There are three basic detection methodologies used by intrusion detection system.

1) Anomaly based detection: In this type of detection technique IDS tries to analyze the

normal operation of the system. The normal performance of the system is profiled and

any deviation from the normal is said to be anomaly. This is a runtime detection

technique. Anomaly base detection is classified in three categories on the basis of

nature of their processing. Statistical based (Uni-variant, multi variant, time series

model), Knowledge based (Expert system, UML, FSM), Machine learning based

(Markov model, Fuzzy logic, Neural Network) [5,8].

2) Specification based intrusion detection: In this technique manually some

specifications and limitations are designed and then the behavior of the system is

monitored. It is similar to anomaly based detection but the difference is that its

specifications are manually designed [5].

3) Misuse based (ruled based) detection: In this detection type profile of known attacks

are managed. And on the basis of this log attacks are detected. This approach is very

useful and it has lowest false positive rate. Advantage of this technique is that it can

easily find out the known attacks. The main disadvantage of this attack is that if the

attack occur whose signature is not present in profile than this attack will be difficult

to detect [7].

C. Executing location of the gathered data

On the bases of executing location IDSs are divided into four types: Centralized

IDS, Stand-alone IDS, Distributive and Cooperative IDS and Hierarchal IDS. Centralized

IDS are those systems which monitor all the activities in the network and find intrusions

in the monitored data. Stand-alone IDS are the systems which run independently on each

node, collect their own data then make decisions. Distributive and Cooperative IDS is

proposed for flat networks. Each node collect data individually and if node detects the

intrusion it sends report to the cooperate network. Hierarchal IDS is developed for

multilayer architecture networks [2].


45

Related Work

Recently, much work has been done on intrusion detection systems in wireless

sensor networks. A few surveys have already been published in order to review the

existing work. Some of which are presented here.

An overview of security attacks in wireless sensor network has been presented in

an article [8] and also the models and architectures of some of the intrusion detection

system (IDS) in wireless sensor network is studied in this article. A comparison and

characteristics of different IDS models have also been presented in this paper.

Another study conducted by Robert Mitchell and Ing-Ray Chen provides a survey

of intrusion detection in wireless sensor networks. This study classifies the existing

wireless intrusion detection system (IDS) techniques based on collection process,

analysis technique, detection technique, and trust model and target wireless network. It

then summarizes the advantages and disadvantages of these wireless intrusion detection

system techniques with respect to specific parameters of target wireless networks and

finally suggests some future research areas [7].

A survey of Intrusion Detection Systems (IDSs) for wireless sensor network is

conducted in an other article[5]. This paper first provides the detailed information about

IDSs, and then provides a survey of IDSs proposed for Mobile Ad-Hoc Networks. After

that IDSs proposed for wireless sensor networks are discussed. Further in this study, the

analysis, comparison and strengths and weaknesses of each system are discussed in

detail.

Similarly another study presents a survey of IDSs in wireless sensor networks

(WSNs). This study also presents the cyber-attacks occurring in wireless sensor networks

in detail. As the features of wireless sensor networks are different from wired networks

and non-energy constrained wireless networks, so intrusion detection systems in wireless

sensor networks behave differently and also take different approaches to detect the

intrusions. In this study, these approaches are discussed in detail [2].

Another article related to this study first provides a review of some of the existing

intrusion detection systems for wireless sensor networks. Then in this study, a new

intrusion detection system is proposed named as Insomnia Mitigating Intrusion Detection

System (IMIDS). IMIDS is a cluster based layered model that can efficiently reduce sleep

deprivation attack in wireless sensor networks [9].


46

Recently Proposed IDSs for WSNs

Since security threats for WSN are different from wired networks due to the

limited energy and storage constraints, therefore IDS for WSN are designed accordingly.

In this section we have briefly discussed the most recent IDSs for WSNs based systems.

The systems included for review are as follows:

A Global Hybrid Intrusion Detection System for Wireless Sensor Networks

Distributed Detection of Flooding and Gray Hole Attacks in Wireless Sensor

Network

An Improvised Hierarchical Black Hole Detection Algorithm in Wireless Sensor

Networks

Policy and Network-based Intrusion Detection System for IPv6-enabled Wireless

Sensor Networks

Insomnia Mitigating Intrusion Detection System (IMIDS)

A. A Global Hybrid Intrusion Detection System for Wireless Sensor Networks.

A Hybrid Intrusion Detection System (HIDS) using Support Vector Machines

(SVM) as learning algorithm and attack signatures for detection is introduced in Figure 1

[10]. A cluster based topology is used, where one known node is designated as the

Cluster Head (CH) which collects data from all other sensors in the cluster and sends the

aggregated data to the base station. The network lifetime can be increased by the use of

CH, by lowering the energy consumption of network. The architecture of the system is

given in Figure 1.

A Cluster Head is selected on the basis of its energy. The residual energy of CH is

calculated by the following formula:

Vi(t) = [Initial – Ei(t)] / r

Where Initial = initial energy, Ei(t) = residual energy, and r = current round of CH

selection. The process of CH selection is announced by the Base Station (BS), which then

calculates the average deviation and values of the energy data. The old CH announces the

end of its authority while the new CH sends alert messages to the nodes. All members of

the cluster are authenticated by the CH, which is in turn authenticated by the BS. Due to

energy constrains in the network nodes, agents are activated only when required.


47

SVM is used as a method of anomaly detection, which is suitable for detection of

small sample data. Since IDS data is large in size, therefore during the training phase the

data collected from all layers of system is sorted and pre-processed offsite, where enough

resources are available. The training data then goes through a data reduction phase, which

reduces the size of data so that it can be further processed by the SVM. A maximum

margin linear hyper plane is defined by the SVM after mapping the training data. Given

the training set for the sample set (xi, yi):

i = 1……., n, x € Rd y € ||+1, -1||

Where ||+1|| is normal, and ||-1|| is abnormal, the classify hyper plane equation is given as

w . x + b = 0

Here w represents a normal vector and offset is given by the parameter b. The

Support Vectors are the training samples on the hyper plane. In context of this scheme,

the vectors of each node are sent to its one-hop neighbor, and the final hyper plane is

calculated discriminator for all nodes separating data into two classes.

The signature based model uses discovery protocol employing signatures for the

detection of harmful nodes and protect network against attacks by these nodes. Based on

a set of rules, this protocol defines the behavior of the target as either normal or

abnormal. This system has used four types of rules for detecting following attacks:

Selective forwarding attack

Black hole attack

Hello flood attack

Wormhole attack

The decision making model uses the following set of rules to take proper action

for a given situation:

If SVM detects an attack and signature model does not detect the attack, then it is

classified as an error or false alarm

If both SVM and signature model detect an attack, then it is classified as an

attack.


48

Figure 1: A Hybrid IDS for Wireless Sensor Networks

B. Distributed Detection of Flooding & Gray Hole Attacks in Wireless Sensor Network

In [11], authors have discussed the main types of Denial of Service (DoS) attacks

i.e. flooding, gray hole and black hole, with respect to the energy consumption and have

given a methodology to prevent these types of attacks in the WSN. Energy consumption

is a critical issue to be addressed for WSN and appropriate measures should be taken to

conserve energy.

The proposed system uses cluster heads (CH) to perform the intrusion detection

procedure for identification and isolation of attacker nodes responsible for the DoS

attack. A predicted energy value for all the nodes in the cluster is given by the CH and

the value of actual energy consumed by the node is obtained from all the nodes. Anomaly

in the predicted and actual value determines an attack.

The system is assumed to consist of homogenous WSN, in which all nodes have

the same energy initially. The residual energy of all nodes is sent to the CH after regular

intervals, which is used to calculate actual energy. The formulas for actual energy and

predicted energy are given as follows:

Actual energy (E1(v))=Initial energy-Residual energy

Predicted Energy (Ek +1(v))=ek+Ø(Ek(v) –Ek-1(v))

Whenever the selection of new cluster head takes place, the newly elected CH get

the routing table from the previous one, which contains the energy information about the

nodes in the cluster. The actual energy consumption can be determined from the

difference of energy levels between two intervals. The mismatch of predicted and

consumed energy of a node is considered as a malicious node. If calculated energy is

greater than predicted energy, then the node is launching a flooding attack because

sending large number of packets requires abnormally high amount of energy. Gray hole


49

attack is detected in the opposite case i.e. when predicted energy is greater than

calculated energy as selective number of packets are dropped by the attacker node; hence

decreasing the transmission. By using the above system, the malicious nodes causing

flooding and gray hole attacks are detected efficiently. The system is consuming less

energy, and is thus lightweight.

C. An Improvised Hierarchical Black Hole Detection Algorithm in WSNs.

Black hole attacks are one of the most harmful types of routing attacks on the

network layer. These attacks aim for the Cluster Heads (CH), by designating a malicious

node as CH and absorbing all the data from other nodes in the cluster. The malicious

node presents itself as the shortest route, and thus absorbs all the received messages and

performs selective forwarding. The system proposed in [12] prevents black hole attack by

implementing a simple strategy where each node sends a control packet to CH and one of

the agents at the end of each transmission. The control packet consists of identity of the

node and the number of packets received by the CH (Nbpkr) [12]. The flow chart of the

proposed algorithm is given in Figure 2 [12].

Figure 2: Flow Chart for Black Hole Detection Algorithm

BS Compares No. of packets in

agent and CH

Start

All nodes send control packet to

CH and Agent

Agent and CH send control packet

to BS

BS broadcasts and alarm message

to sensor nodes

Sensor nodes maintain their

black-hole table to omit detected

CH and selection of new CH

Not Equal


50

The difference in the Nbpkr and number of packets sent by the agent and CH to

the base station determines the presence of black hole attack. In case of attack presence,

an alarm packet containing the identity of the malicious node is sent to all nodes by the

base station. A black hole table is maintained by each node which helps in the selection

of new CH by preventing malicious node from getting selected as CH again.

D. Policy and Network-based Intrusion Detection System for IPv6-enabled Wireless

Sensor Networks

The system proposed in [13] defines an IDS based on abnormal behaviors and

traffic signatures to define attacks on the network. The system uses a network based

approach suitable for WSN. There are two types of network nodes in the system. Those

nodes which are deployed with network based IDS (NIDS) act as watchdogs for

identification of possible attacks by eavesdropping on the packets exchanged between

neighbor nodes acting as host based IDS (HIDS). A set of rules is organized on each

NIDS which matches the monitored messages with rules, and if the match occurs it

generates an alarm and sends it to the Event Management System (EMS). The nodes with

maximum number of alarms is detached from network and designated as a compromised

node.

The authors in this paper have considered a heterogeneous WSN system;

therefore each NIDS has different set of rules based on the neighborhood nodes. A policy

programming approach is adopted by the authors for this purpose. A predefined role

group is required in order to classify different rules for each of the traffic type in WSN.

Each rule is transmitted via configuration channel to NIDS nodes.

The Event Management System (EMS) runs on the sink node, with no restraints

on energy and high processing capabilities. Its function is collection of data from NIDS

and its comparison in order to identify the compromised nodes or intruders. The system

consists of the following three modules:

1) Packet Monitoring Module: It collects the communication data from nodes within

range of NIDS. Due to memory constraints, not all the eavesdropped packets are

stored. Each packet is stored in the temporary buffer for applying rules, after which it

is discarded.

2) Detection Module: this module is responsible for the alarm triggering by storing,

management and application of rules specified by the administrator. The analysis of

network traffic at discrete locations helps improving the performance.


51

3) Action Module: It sends an alert to EMS whenever an intrusion is detected by NIDS

in the neighborhood. The administrator can then compare the alert with alert

messages sent by other NIDS for the same node and takes action accordingly.

E. Insomnia Mitigating Intrusion Detection System (IMIDS)

Insomnia Mitigating Intrusion Detection System (IMIDS) is proposed to detect

sleep deprivation attack in heterogeneous wireless sensor network. A sleep deprivation

attack is an attack in which intruder forcefully awake the sensor nodes until they can

consume their energy. After that sensor nodes stop working and their sleep cycles are

disturbed. In this case, the lifetime of sensor node is minimized. IMIDS uses cluster

based mechanism in which each sensor network is first divided into clusters which are

further subdivided into sectors. The objective of using the cluster based mechanism is to

reduce the network energy consumption in an efficient manner. The block diagram of

sensor network is shown in Figure 3 [9].

It consists of five layers. The description of each layer is given below.

1) Layer 1 is the lowest layer of the sensor network. It consists of the Leaf sensor

nodes that detect the data and send it above to the layer 2.

Figure 3: IMIDS Layered Model

Sink

Node

Cluster Coordinator Cluster

Coordinator

Sector

Monitor Sector

Monitor

Sector

Monitor Sector

Monit

or

Forwarding

Sector Head Forwarding

Sector Head

Sector

Coordinat

or

Sector

Coordinat

or Leaf

Node Leaf

Node

Leaf

Node

Leaf

Node


52

2) Every sector in layer 2 has a sector coordinator that receives the data sent by the

layer 1. Layer 2 also has a capacity to detect anomaly. The purpose of sector

coordinator is to keep the list of all leaf nodes in a sector. Sector coordinator also

separates the suspected nodes from the valid nodes. The detail of suspected nodes

is saved in suspected list and forwarded to the sector monitor while valid nodes

are sent to the forwarding sector head of the above layer.

3) Elements of layer 3 are forwarding sector head and sector monitor. Forwarding

sector head adds the detail of valid packet to the forwarding table and sends the

valid data to the cluster coordinator. While sector monitor takes the data of the

suspected nodes from layer 2, detects the intruders and compromised nodes and

adds their details in quarantine list. Finally it sends the data to the cluster

coordinator of layer 4.

4) Every cluster in layer 4 has a cluster coordinator which is used to monitor the

forwarding sector head and sector monitor of each sector in a cluster. Cluster

coordinator adds the detail of valid packets into the valid list and sends valid data

to the sink node. Two or more cluster coordinators also interact with each other to

form a global intrusion detection system.

5) Layer 5 is the upper Layer of the IMIDS layered model. It has a sink node that

takes data from layer 4 and acts as an access point or a gateway between sensor

networks and other networks. Sink node also saves the backup data of all clusters

[9].

Comparison and Analysis of Recently Proposed IDSs

In this section, we have compared and analyzed the intrusion detection system

models discussed in previous section. The analysis is done on the basis of some

parameters. These parameters are detection rate, false positive rate, technique used and

energy consumption. Detection rate and false positive rate is defined as follows:

Detection Rate: The number of the detected attacks divided by the total number of

attacks.

False Positive Rate: The number of normal connections classified as an anomaly

divided by the total number of normal connections (patterns).

An IDS model should have high detection rate, low false positive rate and should

also have a low energy consumption to perform efficiently and to maximize its lifetime.


53

Comparison and analysis of our discussed IDSs is shown in Table1 while Table 2 shows

the pros and cons and future work of these analyzed IDSs.

Table 1: Analysis and Comparison of IDSs

Intrusion

Detection

Systems

(IDSs)

Parameters

Detection

Rate

False Positive

Rate

Technique

used Energy Consumption

Global Hybrid

IDS

Almost

98%

Near 2% Anomaly

detection with

SVM and

signature

based

Low, energy

consumption is reduced

by using cluster head

Distributed

Detection of

Flooding and

Gray Hole

Attacks

High No evaluation

regarding

false positive

rate is found

Learning

based energy

prediction

algorithm

Low, Energy

consumption is

determined by

calculating the

difference of energy

levels between two time

intervals

An Improvised

Hierarchical

Black Hole

Detection

Algorithm

High No evaluation

regarding

false positive

rate is found

Rule based

technique

Energy Consumption is

50mj for 20 nodes and it

is less than as compared

to other existing

algorithms

Policy and

Network-based

IDS

High No evaluation

regarding

false positive

rate is found

Rule based

technique

No evaluation is found

Insomnia

Mitigating IDS

Detection

Accuracy

is 100% for

20 monitor

nodes

No evaluation

regarding

false positive

rate is found

Anomaly

detection

technique

Energy consumption

with clustering is less as

compared to without

clustering


54

Table 2 Pros, Cons and Future Work of Intrusion Detection Systems

IDS Pros Cons Future Work

Global

Hybrid IDS

High Detection Rate,

Low False Positive Rate,

Low Communication

Cost,

Lift time of network is

increased

None Detailed simulation

of different attacks

needs to be

performed

Distributed

Detection

of Flooding

and Gray

Hole

Attacks

High detection ratio,

Life time of network is

increased,

Low computation

complexity,

Faster intrusion

detection

None Detailed analysis of

algorithm needs to

be performed

An

Improvised

Hierarchical

Black Hole

Detection

Algorithm

Efficient algorithm,

Save the network from

black hole attack,

Improves the node

security

It mostly emphases on

black hole attack,

The model used in this

algorithm is

comprehensive and

complex hence

computation complexity

is increased.

Simulation of

sensor nodes as

black hole nodes

along with cluster

head can be

conducted.

Policy and

Network-

based IDS

Detect and reports the

security attacks

It cannot detect the

wireless channel

dynamically

Optimization of

process is needed to

store the new

detection rules

Insomnia

Mitigating

IDS

Low energy

consumption,

Maximize the life time

of network,

High detection rate

It only detects the sleep

deprivation attack

No future work is

found


55

CONCLUSION

Wireless Sensor Networks (WSNs) are the large scale networks which consist of

the hundreds of thousands of small nodes. The security of WSNs is very critical issue

because of their self-organizing nature, dependency on the other nodes, limited

bandwidth, and low battery power. To secure and prevent the network from intrusions,

different intrusion detection systems (IDSs) are used. In this paper, we have presented the

survey of recent intrusion detection systems (IDSs) for wireless sensor networks. We

have first discussed the intrusion detection systems (IDSs), their classification and some

detection methodologies. Then we have presented the review of some recent intrusion

detection systems for wireless sensor network. Finally, we have summarized the analysis,

comparison and pros and cons of these IDSs in tabular form.

REFERENCES

[1] I. F. Akyildiz and M. C. Vuran, Wireless Sensor Networks vol. 4: John Wiley & Sons,

2010.

[2] O. Can and O. K. Sahingoz, "A survey of intrusion detection systems in wireless sensor

networks," in Proc. 6th Int. Conf. Modeling, Simulation Appl. Optimization (ICMSAO),

2015, pp. 1-6.

[3] N. A. Alrajeh, et al., "Intrusion detection systems in wireless sensor networks: A review,"

Int. J. Distributed Sensor Networks, vol. 2013, pp. 1-7, 2013.

[4] S. Shen, et al., "Signaling game based strategy of intrusion detection in wireless sensor

networks," Comput. Math. Applicat., vol. 62, pp. 2404-2416, 2011.

[5] I. Butun, et al., "A survey of intrusion detection systems in wireless sensor networks,"

IEEE Commun. Surveys Tutorials, vol. 16, pp. 266-282, 2014.

[6] L. Koc, et al., "A network intrusion detection system based on a Hidden Naïve Bayes

multiclass classifier," Expert Syst. Applicat., vol. 39, no. 18, pp. 13492-13500, 2012.

[7] R. Mitchell and R. Chen, "A survey of intrusion detection in wireless network

applications," Comput. Commun., vol. 42, pp. 1-23, 2014.

[8] Y. Maleh and A. Ezzati, "A review of security attacks and Intrusion Detection Schemes

in Wireless Sensor Networks," CoRR, vol. abs/1401.1982, no. 2014.

[9] T. Bhattasali and R. Chaki, "A survey of recent intrusion detection systems for wireless

sensor network," in Advances Network Security Applicat., D. C. Wyld, et al., Eds., ed:

Springer Berlin Heidelberg, 2011, pp. 268-280.

[10] Y. Maleh, et al., "A Global Hybrid Intrusion Detection System for Wireless Sensor

Networks," Procedia Comput. Sci., vol. 52, pp. 1047-1052, 2015.

[11] N. Dharini, et al., "Distributed detection of flooding and gray hole attacks in Wireless

Sensor Network," presented at the 2015 Int. Conf. Smart Techn. Management Comput.,

Commun. Controls, Energy Materials (ICSTM), Avadi, Chennai, India, 2015.


56

[12] A. B. Karuppiah, et al., "An improvised hierarchical black hole detection algorithm in

Wireless Sensor Networks," presented at the 2015 Int. Conf. Innovation Inform. Comput.

Techn. (ICIICT), 2015.

[13] J. P. Amaral, et al., "Policy and network-based intrusion detection system for IPv6-

enabled wireless sensor networks," presented at the 2014 IEEE Int. Conf. Commun.

(ICC), 2014.

PJCIS (2016), Vol. 1, No. 1 : 57-71 Modeling of Safety Parameters

57

Modeling of Safety Parameters for Low Level

Radioactive Waste Repository Using Machine

Learning Approaches

SARDAR ALI SHAH*1, ABDUL MAJID2, SAFDAR ALI3

1 Pakistan Institute of Nuclear Science and Technology, Islamabad, Pakistan 2 Pakistan Institute of Engineering and Applied Sciences, Islamabad, Pakistan

3 Pakistan Atomic Energy Commission, Islamabad, Pakistan


(Received on 20th April, 2016 Accepted on 6th September, 2016)

Abstract

The paper is aimed at the safety assessment of intermediate and low level

radioactive waste (LLRW) disposal facility. Now a days extensive research is going on to

develop safety assessment methodologies for radioactive waste disposal facilities. For

disposal of low level radioactive waste, near surface disposal facility is assumed the

preferred option. Safety assessment is helpful to get public confidence. The main

objective of disposal of radioactive waste is to protect the human health and the

environment from its worse effects. Therefore, it is necessary to manage the radioactive

waste safely. In this work, machine learning (ML) approaches of support vector

regression (SVR), generalized regression neural network (GRNN), artificial neural

network (ANN) and multiple linear regressions (MLR) have been applied for the

modeling of different safety parameters of LLRW disposal facility. Simulations have

been performed to model the distribution coefficients (Kd), leaching rates (𝜆𝑙), and

retardation factors (Rf) of radionuclide present in the RW. Experimentations are

conducted in Matlab environment. Percentage absolute difference is used to evaluate the

performance of the proposed models. The best results have been achieved by SVR and

GRNN models with correlation coefficients R=0.99812 and 0.94773 for Kd, respectively.

The performance of ML models is compared with conventional linear regression (LR)

methods. Experiments highlights that the proposed ML models provide better results

compared to conventional LR methods. This study is useful for the development and

safety assessments of our national future assessment of low level radioactive waste

disposal facility.

Keywords: Radioactive waste, Near surface disposal facility, Distribution

coefficient, Retardation factor



58

INTRODUCTION The proper disposal of radioactive waste is a problem which needs to be

addressed properly to avoid any serious consequences. One of the main reasons is to

protect humans and environment from the hazardous radiations emitted from waste. We

are dealing with low level radioactive waste (LLRW) generated from nuclear power

plants, medicine and research[1]. For the disposal of LLRW near surface disposal

facilities are used contrary to high level radioactive waste (HLRW) which is disposed

deep within the ground.

Theoretical methods use mathematical predictive models and these models are

easy to simulate on computers. In this work, MLR, GRNN, ANN and SVR were

simulated and the results are compared.

Since we are concerned mainly to near surface disposal facilities in this paper, so

there is a possibility that water can enter into the disposal area in accidental scenario[2].

Due to which radio nuclides will be released either into rocks or into cover soil. In case

they are released into rocks, radio nuclides released from the facility will become part of

the underground fresh water. Now this water can affect human beings either directly or

indirectly. This contaminated water may be used for farming which will affect us

directly. Indirectly this water used for farming will affect us by using eatables like milk,

meat etc of animals. In case water is released to cover soil, local residents will obviously

be affected directly[3].

Distribution coefficient or partition coefficient is the ratio of the concentration of

an element on a solid and the concentration in the liquid phase (water etc.) [4].

Adsorbed Concentration

Distribution coefficientDissolved Concentration

(1)

Retardation of the contaminant can be estimated using distribution coefficient Kd.

Kd model is the simplest and robust model. Kd values are empirical and represent a very

simplest model of sorption or attenuation on soil. In general all isotopes of an element

have the same Kd value, because sorption is a chemical property which is not affected by

atomic mass or nuclear emissions. Also Kd values are highly dependent on environmental

factors such as pH, particle size distribution and temperature etc[5].

The retardation factor (Rf) is commonly used in transport models. It describes the

chemical interaction between the contaminant and geological materials (such as soil,


59

sediments, rocks, simply referred to as soils). Specifically, retardation factor indicates the

rate at which contaminant is transported from one place to other [6].

Higher the bulk density, higher is the retardation factor. For high adsorption

coefficient, retardation coefficient is high. Similarly if water content is low, retardation

factor is also low.

𝑅𝑓 = 1 + 𝜌𝑏𝐾𝑑

𝜃 (2)

where

ρb= bulk density of medium (Kg/m3),

Kd= distribution or partition coefficient (m3/Kg),

θ= Total porosity of the medium.

The ratio of the amount of radio nuclides released from a given facility to the

amount remaining in the waste is called leaching rate. We assume that leaching of radio

nuclides (partitioned into the pore water) is due to the steady-state infiltration or drainage

through the waste. Equation (3) can be used for leaching rate.

( )

l

w b d

q

z K

(3)

where

q = the rate of drainage of water through the waste forms (m/y),

θw = the water filled porosity of the waste form ,

ρb = the bulk density of the waste form (kg/m3),

Kd = the waste form distribution coefficient (m3/ kg),

z = the height of the waste form (m).

Materials and Methods

Reduction and assessment of migration of radio nuclides present in waste is done

by safety assessment of radioactive waste disposal facility. As safety assessment of

radioactive waste is a very vast field, some safety parameters like distribution coefficient

(Kd), retardation factor (Rf) and leaching rate (𝜆𝑙) are modelled in this paper. Different

machine learning (ML) approaches like multiple linear regression (MLR), support vector

regression (SVR), generalized regression neural networks (GRNN) & artificial neural

networks (ANN) are used for future prediction of RW disposal facility.


60

Modeling of different safety parameters of low level or intermediate level waste is

a problem because in nuclear power plants, medical and research centers, huge amount of

radioactive waste (RW) is produced every year. It is essential to safely dispose this RW

so that workers and other people can be protected. Disposal of RW is a very vast and

multi-dimensional field. There are huge amount of radio nuclides present in the RW

dumped in near surface disposal facilities. However in this paper radionuclide of our

interest will be cesium (Cs) due to its concentration in the waste. Here in this paper a few

safety parameters are modeled using different techniques. For modeling safety

parameters like distribution coefficient Kd, retardation factor Rf and leaching rate 𝜆𝑙,

techniques used in this paper are MLR, SVR, GRNN and ANN. These models will help

in safety assessment and to assess sorption of radio nuclides from near surface disposal

facilities.

Our proposed model is of the following form:

( , , , )Model f pH CEC SA Aqueous cesium (4)

A set of known Kd values for cesium taken from the literature are analyzed

considering it to be output parameter while other four parameters are considered to be

input parameters. This cesium data set included 176 cesium Kd values. Two separate data

sets were compiled. The first one included both soils and pure mineral phases. It contains

176 Kd values. The lowest Kd value is 0.6 ml/g while the largest Kd value is 52,000 ml/g.

The average cesium Kd value is 2635 ± 530 ml/g. The second data set shown contain data

which include only soil studies, that is, data from pure mineral phases and rocks were

eliminated from the data set. This second data set contains 57 Kd values [7]. Statistics

were also applied on this data set. These four methods were applied to predict output

parameter Kd from four input parameters pH, Cation exchange capacity (CEC) in

meq/100g, surface area (SA) in m2/g aqueous cesium in µM. Matlab codes were written

in Matlab 2013 [8] to model these parameters.

In this paper machine learning approaches like MLR, GRNN, ANN and SVR are

used to model safety parameters [9]. The safety parameters are distribution coefficient Kd,

retardation factor Rf, leaching rate 𝜆𝑙. Researchers use ML approaches to make decisions

for their problems. In the literature, there are many techniques but machine learning

techniques are significant in a way that they can be applied to solve a large variety of

problems. ML approaches learn from experience [10].

Matlab and matlab toolboxes were used for modeling Kd values [8]. A set of

known Kd values for cesium taken from the literature are analyzed considering it to be

output parameter, while other four parameters (i.e. pH, CEC, SA, aqueous cesium) are


61

considered to be input parameters. First the four techniques were used to model Kd values

taken from the literature. Then these modeled values of Kds were used to find values of

retardation factor Rf and leaching rate 𝜆𝑙 using Equations 2 and 3. Other values of

parameters like bulk density of medium (Kg/m3) ρb, total porosity of the medium θ, the

rate of drainage of water through the waste forms (m/y) q, the water filled porosity of the

waste form θw, the height of the waste form (m) z, were taken from literature [2].

In this paper, our ML based methodology comprises three steps. First step is data

set partitioning, in the second step models are constructed and in the last step models are

tested. Performance is measured by finding percentage absolute difference (PAD) as:

exp

exp

% 100predd d

d

K KPAD

K

(5)

In our case, out of 55 samples, for training 39 samples were selected, while for validation

and testing 8 samples were selected respectively. Any machine learning training process

consists of learning domain, training set, learning system and finally testing. Keeping

these factors in mind, machine learning models can be developed [11].The four

techniques used are discussed as below:

Multiple Linear Regression (MLR)

In this paper, MLR is used to predict the value of distribution coefficient Kd from

a set of four predictors i.e. pH, Cation exchange capacity (CEC) in meq/100g, surface

area (SA) in m2/g, aqueous cesium in µM. In other words the four input parameters are

pH, CEC, SA and aqueous cesium and the output parameter is distribution coefficient Kd.

A matlab code was written for the prediction of Kd out of these four parameters.

Generalized Regression Neural Network (GRNN)

This is basically a neural network based function approximation or function

estimation algorithm. It predicts the output of a given input data. In principle, neural

network needs a training data to train itself. Training data should contain input-output

mapping. Now if the network is trained with the training data set and a new testing data

set is fed, it will accordingly give the output or predict the result.

The main advantage of GRNN models is that it estimates the appropriate

regression model from the given data. The network is over-trained for low value of

Gaussian function parameter σ, and the prediction error would be higher for evaluation


62

data. In a similar way, for large value of σ, the network remains under-trained and the

prediction error would increase for training data.

Training procedure involves determination of the optimum value of spread

constant σ. Best practice is that find the position where the Mean Squared Error (MSE) is

minimum. First divide the whole training sample into two parts, training sample and test

sample. Apply GRNN on the test data based on training data and find out the MSE for

different [11, 12].

Artificial Neural Network (ANN)

ANN uses generalized mathematical models of human or neural biology. A

general mathematical model of simple neuron is shown in figure 1. A simple neuron has

many inputs but a single output. ANN model also needs a training sample set with

desired output values. Back-propagation technique has been employed. Neural Network

Toolbox of Matlab is employed for the training[8]. This toolbox is a built in set of

MATLAB functions that provide an environment to develop not only feed forward neural

networks but also recurrent neural networks. For optimal network performance, ten

neurons are used in the hidden layer. Levenberg–Marquardt algorithm is used for weights

adjustment. Initial weights and bias values are randomly selected. Tansig and pure linear

activation functions are used for hidden and output layers. The values of four parameters

and the desired Kd value of each sample point are fed as the network input [13].

Support Vector Regression (SVR)

Support vector machine is a practical learning method based on statistical learning

theory. Like ANN, SVR is also first trained on training data samples and then it is tested

for those data samples not present in the training dataset[12]. In using SVR methodology

first data set is generated. After data set generation, parameters are optimized on training

dataset. Then optimized model is validated using validation dataset and finally optimized

model is validated using verification dataset. For developing SVR machine the

parameters initialized are type and width of the kernel, Trade–off parameter, C and ε–

insensitivity zone

Results

Correlation coefficients for the parameters are determined for both types of data

sets. The parameter which have largest correlation coefficient with cesium Kd was

CEC (R2=0.3364). Similarly the correlation coefficients of other parameters i.e Ph and


63

aqueous Cs with cesium Kd values are 0.0356 and 0.0989 respectively. The correlation

coefficient between parameter aqueous Cs and Cs Kd values is poor. This is due to the

fact that this parameter contains concentration of the solution before and after contact

with the soils. The correlation coefficients between the chosen input parameter CEC and

Kd is shown in Figure 1. The correlation coefficient R values and corresponding equations

of all four input parameters and Kd is shown in Table 1.

Figure 1: Relationship between Cs Kd values and CEC.

Table 1: Input parameter correlation coefficient values and corresponding Kd equations.

Sr. No. Input Parameter R2 value Equation

1 CEC (meq/100g) 0.336 Kd = 0.5051 CEC + 2.457

2 SA (m2/g) 0.176 Kd = 0.5679 SA + 1.841

3 Aqueous cesium (µm) 0.102 Kd= -0.1354 AqCs + 2.0512

4 pH 0.036 Kd= 1.6236 pH + 1.0251

Kd = 0.5051CEC + 2.457

R=0.58

0

0.5

1

1.5

2

2.5

3

3.5

4

4.5

5

-4 -3 -2 -1 0 1 2 3

log K

d(m

l/g)

log CEC (meq/100g)


64

Kd Prediction Using MLR

Four independent variables (pH, CEC, SA, Aq Cs) are the input parameters and

Kd is the corresponding output variable. The coefficients of independent variables are

computed under ordinary least squares criterion, using training data, as follows:

1 2 3 4 8107.9 910.7 129.8 3.1 3.2 MLR

dK X X X X (6)

where

X1=pH, X2= Cation exchange capacity (CEC), X3= surface area (SA) and X4= Aqueous

cesium (Aq Cs). i.e. intercept b0= 8107.9

Regression coefficients are b1= -910.7, b2= 129.8, b3= 3.1, b4= -3.2

Linear correlation between theoretical and predicted Kd is shown in Figure 2.

Figure 2: Linear correlation graph between experimental and predicted Kd for MLR

Kd Prediction Using GRNN

For GRNN, Gaussian function parameter σ is tuned and the best results of this

network are obtained by finding the optimal spread values of σ. In our case σ =0.1025,

0.0275 for four parameters, respectively. Figure 3 shows the linear correlation graph

between the two variables.

0.5 1 1.5 2 2.5

x 104

0.5

1

1.5

2

2.5x 10

4

Experimental Kd

Pred

icte

d K

d ~=

1*T

arge

t + -

3.3e

-12

R=0.76158

Data

Fit

Y = T


65

. Figure 3: Linear correlation graph between experimental and predicted Kd for GRNN.

Kd Prediction Using ANN

Neural Network Toolbox of Matlab [8] is employed for the training. For optimal

network performance, ten neurons are used in the hidden layer. Predicted Kd is shown in

Figure 4.

Kd Prediction Using SVR

Gaussian Kernel function is employed for this problem. To find the optimal

values of these parameters, selected ranges for these three parameters are as: C=[80,120],

σ=[0.000001,0.1] and ε = [6, 1.8]. The combination for which the average PAD value is

minimum for both the training and validation set is chosen to be optimal. The optimal

values for the trade-off parameter C, the kernel width σ and ε–insensitivity zone are

found to be 100, 1.7255 and 1.0 x 10-6 respectively in our case. The correlation

coefficient in this case is 0.99812.

A comparison of PAD values is given in Table 2. The performance comparison in

terms of correlation coefficient is shown in Table 3.

Now using these predicted Kd values, two other parameters i.e. retardation factor

Rf and leaching rate of radio nuclides are modeled. Using Equation (2) retardation factor

is calculated. The values of bulk density ρb and total porosity θ of the medium are taken

from the literature. The graph between Kd and Rf is shown in Figure 5.

0.5 1 1.5 2 2.5

x 104

0.5

1

1.5

2

2.5x 10

4

Experimental Kd

Pred

icte

d K

d ~=

0.9

*Tar

get +

2.8

e+02

R=0.94773

Data

Fit

Y = T


66

Figure 4: Linear correlation graph between experimental and predicted Kd for ANN

Similarly Equation (3) is used to calculate leaching rate 𝜆𝑙. The values of rate of

drainage of water q, water filled porosity of waste 𝜃𝑤, bulk density of waste 𝜌𝑏 and height

of waste form z is taken from the literature. The graph between Kd and leaching rate 𝜆𝑙 is

shown in Figure 6.

0 0.5 1 1.5 2 2.5

x 104

0

0.5

1

1.5

2

2.5x 10

4

Experimental KdPre

dic

ted

Kd

~=

0.6

6*T

arg

et

+ 8

.5e

+0

2 Training: R=0.83584

Data

Fit

Y = T

0 5000 10000 15000 20000

0

5000

10000

15000

20000

Experimental KdPre

dic

ted

Kd

~=

0.8

7*T

arg

et

+ 9

.3e

+0

2 Validation: R=0.91989

Data

Fit

Y = T

0 5000 10000 15000 20000

0

5000

10000

15000

20000

Experimental KdPre

dic

ted

Kd

~=

0.9

2*T

arg

et

+ 3

.9e

+0

2 Test: R=0.90598

Data

Fit

Y = T

0 0.5 1 1.5 2 2.5

x 104

0

0.5

1

1.5

2

2.5x 10

4

Experimental Kd

Pre

dic

ted

Kd

~=

0.7

*Ta

rge

t +

8.7

e+

02 All: R=0.84916

Data

Fit

Y = T


67

Table 2: Comparison of ML approaches

PAD for SVR (%) PAD for GRNN (%) PAD for ANN (%) PAD for MLR (%)

1.333 9.891 12.916 18.374

12.500 9.903 15.743 16.851

1.818 9.892 14.146 15.607

0.487 9.890 9.090 13.143

1.428 9.891 15.405 15.805

1.818 9.892 9.146 9.324

7.142 9.897 15.609 9.148

8.438 9.899 14.052 15.515

0.243 9.890 13.792 14.590

6.153 9.896 12.209 19.062

0.090 9.890 13.128 11.959

6.079 9.896 11.916 12.871

1.838 9.892 14.542 15.656

1.845 9.892 14.675 12.205

2.103 9.892 7.913 17.346

1.956 9.892 4.913 19.839

Table 3: Performance comparison of ML approaches in terms of correlation coefficient

Input Dataset SVR Model GRNN Model ANN Model SVR Model

Train. Data 0.998 0.947 0.835 0.761

Valid. Data 0.983 0.950 0.919 0.741

Test. Data 0.975 0.942 0.905 0.752

Mean R-value 0.985 0.946 0.887 0.751


68

Figure 5: Graph between Distribution coefficient and Retardation factor

Figure 6: Graph between distribution coefficient (Kd) and leaching rate (λl)

0

20000

40000

60000

80000

100000

120000

140000

160000

1.4

8

1.0

8

1.3

8

0.2

6

8.2

22

.08

0.3

09

1.0

64

0.9

31

1.0

05

0.3

98

0.4

67

0.4

68

0.5

97

0.4

4

1.4

46

0.0

36

0.8

33

0.8

88

1.0

6

9.7

74

0.2

04

5.8

62

8.3

13

1.3

8

0.2

6

Ret

ardat

ion f

acto

r R

f

Distribution coefficient Kd (m3/kg)

Kd vs Rf

0

0.001

0.002

0.003

0.004

0.005

0.006

0.007

0.008

0.009

1.4

8

1.0

8

1.3

8

0.2

6

8.2

22

.08

0.3

09

1.0

64

0.9

31

1.0

05

0.3

98

0.4

67

0.4

68

0.5

97

0.4

4

1.4

46

0.0

36

0.8

33

0.8

88

1.0

6

9.7

74

0.2

04

5.8

62

8.3

13

1.3

8

0.2

6

Lea

chin

g r

ate

y-1

Distribution coefficient Kd (m3/kg)


69

Discussion

The correlation coefficient value of each parameter with Kd is calculated to

determine which parameter has the largest correlation coefficient. The correlation

coefficients of parameters CEC, SA, pH and aqueous cesium with Kd was 0.58, 0.42, 0.19

and 0.32 respectively. Hence the parameter with the largest correlation coefficient with

cesium Kd was CEC (R = 0.58).

Multiple linear regression (MLR), generalized regression neural network

(GRNN), artificial neural network (ANN) and support vector regression (SVR) are used

to predict Kd and results obtained from these methods are compared. Correlation

coefficient for these machine learning approaches MLR, GRNN, ANN and SVR are

0.76158, 0.94773, 0.84916, and 0.99812, respectively. Thus the results obtained from

SVR and GRNN are better than other two approaches. The correlation coefficients for

SVR and GRNN are 0.99812 and 0.94773 respectively.

Similarly Performance is measured by finding percentage absolute difference (PAD) as

follows:

SVR GRNN ANN MLRPAD PAD PAD PAD (7)

After modeling Kd values using different machine learning approaches, the results

of best approach is taken which in our case is that of SVR for which R=0.99812. These

Kd values are plotted against retardation factor (Rf). The results are shown in Figure 5,

which shows a smooth variation between Kd and Rf. The values of Kd are constant for a

given range of Rf values but at some points some peaks are observed which shows that at

these points one must be careful in selecting these Kd values. For example in Figure 5, the

value of retardation factor is constant in the range of 0.308 up to 13.022 but at points

0.307 and 0.204 there is a peak showing abnormal variation of Rf with Kd.

In a similar way, Kd values are also plotted against leaching rate and the results

are shown in Figure 6. These graphs also shows smooth variation between Kd and Rf but

at some particular points there are peaks showing abnormal behavior.


70

Conclusion

Four different models are used for modeling Kd values in this work and their

performance is tested using % PAD and coefficient of performance R. The best results

have been achieved by SVR and GRNN models with correlation coefficients R=0.99812

and 0.94773 for Kd, respectively. Predicted Kd values are then plotted against retardation

factor and leaching rate, which shows smooth variation with Kd except some particular

points. So we should avoid taking those Kd values at those particular points.

Recommendations and Future Work In future distribution coefficient Kd, retardation factor Rf and leaching rate 𝜆𝑙

prediction can be modelled for other radionuclides against different disposal concepts.

This study could be used for development of radionuclides transport models in different

media. This study will also integrate with entire safety assessment activities of near

surface disposal facilities.

REFERENCES

[1] R. H. Little, et al., "Assessment of radioactive and non-radioactive contaminants found in

low level radioactive waste streams," in Proc. Waste Manag. Conf., Tucson, AZ, United

States, 2003.

[2] IAEA, "Safety assessment methodologies for near surface disposal facilities: Results of a

co-ordinated research project," IAEA, Vienna2004.

[3] E. Tolentino and C. C. O. d. Tello, "An overview of technical requirements on durable

concrete production for near surface disposal facilities for radioactive wastes," in Proc.

INAC 2013 Int. Nucl. Atl. Conf., Recife, PE, Brazil, 2013.

[4] S. C. Sheppard, et al., Solid/liquid partition coefficients (Kd) and plant/soil concentration

ratios (CR) for selected soils, tills and sediments at Forsmark, 1st ed.: Swedish Nuclear

Fuel and Waste Management Company, 2011.

[5] M. Bragea, et al., "Influence of distribution coefficients on the transfer of radionuclides

from water to geological formation," Chem. Bull. Politehnica Univ. Timisoara, vol. 53,

pp. 1-2, 2008.

[6] S. A. Adeleye, et al., "Sorption of caesium, strontium and europium ions on clay

minerals," J. Mater. Sci., vol. 29, pp. 954-958, 1994.

[7] K. M. Krupka, et al., "Review of geochemistry and available Kd values, for cadmium,

cesium, chromium, lead, plutonium, radon, strontium, thorium, tritium (3H), and

uranium," in Understanding Variation in Partition Coefficient, Kd, Values vol. II, 1999.

[8] MatLab. Natick, Massachusetts, United States: The MathWorks Inc., 2013.


71

[9] S. Kwon and W. J. Cho, "A sensitivity analysis of design parameters of an underground

radioactive waste repository using a backpropagation neural network," J. Korean Soc.

Rock Mech., vol. 19, pp. 203-212, 2009.

[10] F. L. d. Lemos, et al., "Uncertainty / sensitivity methodologies for safety assessments of

low-level waste disposal facilities," in Waste Manag. Conf., 1999.

[11] A. Majid, et al., "Predicting lattice constant of complex cubic perovskites using

computational intelligence," Comput. Mater. Sci., vol. 50, pp. 1879-1888, 2011.

[12] A. Majid, et al., "Combination of support vector machines using genetic programming,"

Int. J. Hybrid Intell. Syst., vol. 3, pp. 109-125, 2006.

[13] A. Ceguerra and I. Koprinska, "Automatic fingerprint verification using neural

networks," in Proc. ICANN 2002: Int. Conf. Artif. Neural Netw., Berlin, Heidelberg,

2002, pp. 1281-1286.

Guidelines For Authors

Pakistan Journal of Computer Science and Information Systems (PJCIS) is an

official journal of PASTIC meant for students and professionals of Computer Science &

Engineering, Information & Communication Technologies (ICTs), Information Systems,

Library and Information Science. This biannual open access journal is aimed at

publishing high quality papers on theoretical developments as well as practical

applications in all above cited fields. The journal also aims to publish new attempts on

emerging topics/areas, original research papers, reviews and short communications.

Manuscript format: Manuscript should be in English, typed using font Times New

Roman and double spaced throughout with 0.5 inches indention and margins of at least

1.0 inch in Microsoft word format (doc or docx) or LaTex. The manuscripts should be

compiled in following order: Abstract, Keywords, Introduction, Materials and Methods,

Results, Discussion, Conclusion, Recommendations, Acknowledgement and References.

Follow the format guidelines provided in the following table:

Type Characteristics Letter case

Title Font size 16, bold, single line spacing

centered

Sentence case

Author names Font size 12, bold, centered, single line

spacing

Uppercase

Institutional

affiliations

Font size 10, centered, single line

spacing

Sentence case

All headings Font size 16, bold, centralized and

numbered

Uppercase

Body text Font size 12 and justified Sentence case

Captions

(Table/Figures)

Font size 12 and bold Sentence case

Title: This should contain title of the articles by capitalizing initial letter of each main

word. Avoid abbreviations in the title of the manuscript.

Author’s Name: Title should be followed by complete author names, e-mail addresses,

landline, fax and cell numbers. In case of more than one author with different Institutions,

use numbering in superscripts (1, 2, 3 etc) to separate the author's affiliation. Do not use

any other symbol such as *, † or ‡. For multiple authors, use comma as separator.

Author’s Affiliation: Mark the affiliations with the respective numbering in superscripts

(1, 2, 3 etc) used in the authors field. PJCIS encourages the listing of authors’ Open

Researcher and Contributor Identification (ORCID).

Corresponding author: One author should be identified as the corresponding author by

putting asterisk after the name as superscript.

Abstracts: Abstract must be comprehensive and self explanatory stating brief

methodology, important findings and conclusion of the study. It should not exceed 300

words.

Keywords: Three to eight keywords depicting the article should be provided below the

abstracts separated by comma.

Introduction: The introduction should contain all the background information a reader

needs to understand the rest of the author’s paper. This means that all important concepts

should be explained and all important terms defined. Introduction must contain the

comprehensive literature review on the problem to be addressed along with concise

statement of the problem, objectives and valid hypotheses to be tested in the study.

Material and Methods: An adequate account of details about the procedures involved

should be provided in a concise manner.

Results & Discussion: Results should be clear, concise and presented in a logical

sequence in the text along with, tables, figures and other illustrations. If necessary,

subheadings can be used in this section. Discussion should be logical and results must be

discussed in the light of previous relevant studies justifying the findings. If necessary, it

may be split into separate "Results" and "Discussion" sections.

Recommendations: Research gap may be identified and recommendations for future

research should be given here.

Acknowledgement: In a brief statement, acknowledge assistance provided by people,

institutions and financing.

Abbreviations: Abbreviations used should be elaborated at least first time in the text

inside parenthesis e.g., VPN (Virtual Private Network).

Tables/Figures: Caption must have descriptive headings and should be understandable

without reference to the text and should include numbers [e.g., Figure 1: Summary of

quantitative data analysis]. Put citation at the end of the table/figure or in the text, if it has

been copied from already published work. Photographic prints must be of high resolution.

Figures will appear black & white in the printed version however, they will appear

coloured in the full text PDF version available on the website. Figures and tables should

be placed exactly where they are to appear within the text. Figures not correctly sized will

be returned to the author for reformatting.

References: All the references must be complete and accurate. When referring to a

reference in the text of the document, put the number of the reference in square brackets

e.g., [1]. All references in the bibliography should be listed in citation order of the

authors. Always try to give url or doi in case of electronic/digital sources along with

access date. References should follow IEEE style, Details can be downloaded from

Journal’s home page. http://www.pastic.gov.pk/pjcis.aspx

Publication Ethic Policy: For submission to the Pakistan Journal of Computer Science

and Information Systems please follow the publication ethics listed below:

The research work included in the manuscript is original.

All authors have made significant contribution to the conception, design,

execution, analysis or interpretation of the reported study.

All authors have read the manuscript and agree for submission to the Pakistan

Journal of Computer Science and Information Systems.

This manuscript has not been submitted or published elsewhere in any form

except in the form of dissertation, abstract or oral or poster presentation.

All the data has been extracted from the original research study and no part has

been copied from elsewhere. All the copied material has been properly cited.

Journal Review Policy: PJCIS follows double blind peer review policy. All the

manuscripts are initially evaluated by the internal editorial committee of the PASTIC.

Manuscripts of low quality which do not meet journal’s scope or have similarity index >

20% are rejected and returned to the authors. Initially accepted manuscripts are

forwarded to two reviewers (one local and one foreign). Maximum time limit for review

is six weeks. After submission of reviewer’s report final decision on the manuscript is

made by the chief editor which will be communicated to the corresponding author

through email. All dates i.e., submission and acceptance will be clearly printed on the

manuscript.

Submission policy: Before submission make sure that all authors have read publication

ethics policy for PJCIS and agree to follow them. There is no submission or processing

charges for publishing in Pakistan Journal of Computer Science and Information

Systems. Manuscript complete in all respect should be submitted to

[email protected] or [email protected]

Editorial Office: Correspondence should be made at the following address:

Editor in Chief

Prof. Dr Muhammad Akram Shaikh

Director General

Pakistan Scientific & Technological Information Centre (PASTIC)

Quaid-e-Azam University (QAU) Campus

Islamabad

Please send all inquiries to [email protected]




pakistan journal of computer and information systemspastic.gov.pk/downloads/pjcis/pjcis_v1_1.pdf ·...

Documents