College of Engineering & Technology

Department of Computer Science

Master of Science Degree
In
Computer Science and Networking
Proposal On
“Cell Outage Detection Using MLP-RNN
Machine Learning Algorithm”
By
Tezera Gale ID No. 013/22

Advisor Mesay A. (PhD)

Feb, 2024

Dilla University

Ethiopia

1
Table of Contents Page
1. Introduction 1
1.1. Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2. Statement of the Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3. Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.4. Objective of the Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.4.1. General Objective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.4.2. Specific Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.5. Research Question . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.6. Scope and Limitation of the Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.7. Significance of the Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.8. Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.9. Work plan and budget break down schedule . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.9.1.Work plan schedule . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.9.2. Budget break down . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
Reference 16

I
List of Figures Page

Figure 1. 1: Self-healing functions.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

Figure 1.8.1. Performance Evaluation Parameters / metrics . . . . . . . . . . . . . . . . . . . . . . 15

Figure1.8.2. Research Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .16

II
 List of Tables page

Table 1: Schedule of tasks to be undertaken . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

Table 2: Estimated cost of proposal work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

III
Abstract
Mobile traffic growth increases exponentially over the years. To gratify the growing traffic,
which requires capacity and coverage, densification of a network is a key solution. As mobile
network becomes larger and larger, it is difficult to manage the network manually rather it
requires automated network management. Self-healing is one of self-organizing network
(SON’s) functionalities that implements automatic fault management in radio access network
(RAN). In practice, mobile cell outage is the major problem in the radio access network and
leads to the lack of network service. The automated and timely detection of a malfunctioning cell
is one of the crucial challenges for network operators. In this thesis, data mining model has been
introduced to detect cell outage automatically using MLP-RNN algorithm, which is a decisive
part of the model, has been adopted and implemented using incoming handover statistical data to
detect cell outage and sleeping cells in self-organizing manner. For this purpose, statistical
handover data has been collected from real UMTS network and then pre-processed using
filtering, aggregation, normalization and then profiling. Moreover, an improved version of using
MLP-RNN algorithm, fast anomaly detection with duplication (FADD), has also been
implemented to improve the detection capability. Receiver Operating Characteristic (ROC) curve
is used to show the degree of the performance of the algorithms. The study shows that the two
versions of MLP-RNN cell outage detectors have detected most cells in outage and locate their
positions. But, FADD has detected 88% compared to 76% of the original LOF.

Keywords:

Self-healing, Self-organizing network, Local outlier factor, Sleeping cell, Cell out detection,
UMTS, data mining, FADD, ROC, and MLP-RNN.

IV
1. Introduction

This chapter provides background of this thesis and describes the motivation, objectives, scope,
and contribution of the research. Moreover, it discusses the methodology used and briefly
describes reviewed literatures which are related to the proposal. Finally, the thesis structure is
outlined.

1.1. Background
Mobile communication is a rapidly growing technology, which provides ease of access to
different services backed by different technologies and seamless connectivity anywhere and
anytime [1]. Due to different services offered by telecom operators, the public interest has
become higher to get the new services. The number of mobile subscription globally reaches
about 5.8 billion in 2023 and will reach 6.9 billion by 2025 [1]. This indicates that
communication has become an integral part of the daily life. Hence, for every operator it is a
challenge to provide adequate coverage and capacity to its customers that effectively increases
an operator's subscriber base and generates more revenues.

To provide adequate capacity and coverage, operators need to increase their mobile network
capacity by increasing the number of base stations and deploy the latest technology to
accommodate a large volume of traffic, while keeping capital expenditures (CAPEX) and
operational expenditures (OPEX) at minimum. Subsequently as the number of base stations
becomes larger and new technologies are introduced, network operation and maintenance
(O&M) activities become difficult and its OPEX increases significantly. Hence, it is necessary
to develop new approach in which mobile system fault and performance management becomes
more effective and automated [2]. Self-organizing network (SON) has been identified for this
purpose by 3rd Generation Partnership Project (3GPP) based on the business requirement
provided by Next Generation Mobile Networks (NGMN) operators Alliance [3, 4]. Later, the
3GPP also included SON as an important element of the new standard of mobile
communications. SON solution is categorized by its functionality as self configuration, self-
optimizations and self-healing in which use cases are developed and organized [4]. R. Barco et
al. in [4] described that, SON mainly targets to reduce operational expenses, improve

-1-
operational efficiency, and enhance and maintain a gratifying user experience, by means of
automating tasks that are currently manually performed by highly experienced staff. SON also
contributes to CAPEX reduction by a more efficient use of network elements and resources.
The SON concept was first introduced by 3GPP as a fundamental element for Long Term
Evolution (LTE) deployment in Release 8 [3]. Later, 3GPP further developed Releases 9 and
10[5].

Self-configuration functions focus on defining the configuration parameters of network

elements automatically in the planning or deployment stage. Self-optimization functions adjust
network parameters, which improve network performance, based on the situation during
operation. Self-healing, in simple terms, is an automated fault management which automatically
detects any fault occurring in the network, diagnoses those fault to avoid any service breakdown
as well as maintaining Service Level Agreements (SLAs), and reduces operational costs of the
network [1]. Figure 1.1 shows the self-healing functions. One of the major mobile network
faults is mobile cell outage. Traditional troubleshooting is a manual process which is carried out
through alarms and user complaints. Such practice costs much time and effort for cell outage
detection (COD) [6]. In order to implement the detection functionality accurately and timely, a
number of COD frameworks have been adopted by different researches based on detection
indicators such as key performance, cell level statistical data and location information which
can be collected from base stations and user equipment (UE). COD mechanisms generally
involved data collection, preprocessing and analyzing it to extract relevant information so that it
can judge if a cell outage occurs. For such process data mining approaches are used to
preprocessing and extract knowledge from collected data that comprise of huge amount of
information.

-2-
 Figure 1: Self-healing functions.

SON concept consists of three solutions, namely self-configuration, self optimization, and self-
healing. Self-configuration refers to the automated configuration of the newly established
networks. Self-optimization mechanism automatically adjusts network parameters in order to
guarantee high quality of service and reduce maintenance costs. The purpose of the self-healing
mechanism is to automatically detect and address network operation problems, avoiding
significant impact on subscribers’ experience and reducing operational expenses.

Later in (NGMN, 2010) NGMN formulates top 10 operational efficiency recommendations for
SON. The list includes standardized Minimization of Drive Tests (MDT) method and a cell
outage detection function, which is part of the self-healing mechanism. MDT is a method that
enables mobile operators to collect measurements from a regular cell phone. A mobile operator
can employ these measurements to improve capacity, coverage, and performance characteristics
of a radio network. One of the use cases is the identification of malfunctioning cells. The
automation of this process allows fast and effective cell outage detection, while reducing the

-3-
risk of human error to a minimum. Despite its benefits, the implementation of MDT
functionality is a challenging task for network operators. It requires an intelligent management
of data flow, data storage, and data processing mechanisms.

Data mining is the process of discovering interesting patterns and knowledge from large
amounts of data [5]. It is also popularly referred to as knowledge discovery from database
(KDD) which is the automated or convenient extraction of patterns representing knowledge
implicitly stored or captured in large databases, data warehouses, the Web, other massive
information repositories, or data streams [5]. The abandoned growth of databases in recent
years, for example global backbone telecommunication networks carry tens of petabytes (250
bytes) of data traffic every day [6] , brings data mining to the forefront of new business
technologies.

The motivation of this thesis work is to adapt mobile cell outage detection model using LOF
data mining technique based on incoming handover statistical data of Universal Mobile
Telecommunication System (UMTS). The thesis work also focuses on how handover (HO) data
is collected and preprocessed to detect and localize outage cells. Moreover, it evaluates the
performance of the detection system.

The experiments have demonstrated that there is no single machine learning algorithm which
can handle efficiently all the types of attacks. The decision table (rule base classifiers) achieved
the lowest false negative value, but it was far from the highest accuracy rate detection. On the
other hand, Random Forest classifier performs significantly better than MLP ANN but, MLP-
RNN classifier registered the highest accuracy rate 99.79%, with the smallest root mean square
error (RMSE. The experimental result shows that for the real dataset, MLP-RNN performs
better than the two other techniques in terms of the execution time with evaluation Parameters
in terms of accuracy, sensitivity, precision, and F-score compared to the several existing has
been recommended for cell outage detection (COD) methods [7]

-4-
1.2. Statement of the Problem
MDT is a data collection process and an auxiliary mechanism for SON functions, such as self-
healing. MDT concerns radio and service quality measurements from a regular user device.
Since the measurements are collected from all devices, MDT database may include an enormous
amount of records. The management of this huge database, especially processing, is a very
challenging task for network operators. Knowledge Discovery in Databases (KDD) is known to
be an efficient tool for the processing and extraction of information from large volumes of data.
KDD is an umbrella term for machine learning and data mining techniques, which address
anomaly detection, clustering, and other types of information retrieval problems. Therefore,
network operators can employ KDD methods to analyze huge MDT databases and improve a
network’s performance.

SON includes a self-healing mechanism, which can leverage anomaly detection methods for the
automated detection of malfunctioning cells. A cell failure is a highly important problem,
because it directly affect on customer satisfaction of wireless services. Unsatisfied clients may
change a telecommunication provider, leading to the loss of the provider’s revenue. Therefore,
effective and timely detection of cell failures is a vital problem for mobile operators. Sleeping
cell is a special case of cell outage, which makes mobile service unavailable for subscribers even
though the cell still appears to be operable from the networks point of view. Usually, sleeping
cell outage lasts for days until it becomes unveiled by multiple user complaints or after a detailed
analysis of expensive drive tests measurements. For these reasons the timely detection of
sleeping cells is of high importance for network operators.

One of the possible mobile radio access network problems for an operator is a cell outage.
Several researches about COD which are based on the analysis of the performance of the
problematic cell or its neighboring cells have been presented. However, with these methods only
the most severe cases can be detected but many other outage situations may not be identified. For
instance, sleeping cell, which is a specific type of cell outage that can occur in the network, is not
identified by the fault management system. This is because sleeping cell is invisible for network
operators via traditional alarms [8]. This means that FM system does not generate alarms based
on network traffic. Sleeping cell problem can only be identified from customers complains or

-5-
network performance evaluation. This thesis focuses on automatic detection of cell outage that is
the major type of fault in the radio access part of the UMTS mobile network. The detection is
based on HO statistics data to reduce resolution time in fault handling process.

1.3. Literature Review

SON is often used to categorize a mobile network for which the activities of configuring,
operating, fault handling and optimizing are largely automated. It is a collection of functions for
automatic configuration, optimization, and healing of mobile networks. Self healing is one of the
functionalities of SON that enables the mobile network to automatically perform the task of
troubleshooting which includes detection, diagnosis, and correction of faults [9]. Cell outage
Management (COM) is the main task of self-healing and it contains COD and cell outage
compensation (COC).
There are a lot of researches conducted on the detection and diagnosis of faults in mobile
networks, especially on radio access part of the network. Many detection methods use input data
that collected from minimize drive testing (MDT), UE or network management system (NMS).

Min Liu et al., [10] enveloped prediction of Congestion Degree for Optical Networks Based on
BP Artificial Neural Network. By creating BP-ANN, I proposed and implemented an approach
for predicting the degree of congestion in optical networks. The fundamentals are discussed, and
simulation results show that their proposed method is capable of predicting network congestion
levels. They choose four critical request attribute values as input for BP-ANN learning and
training. The volume and quality of example data, as well as the correctness of the network
model, affect prediction accuracy. By improving them, we will be able to make more accurate
forecasts. BP-ANN. Because the trained BP-ANN has a strong nonlinear conversion mapping
ability, it can make the most of huge data in the real-world network environment.

Khan et al., [11] study early detection of congestion using specific location or area and usually
call attempts made in the location. In this study Multilayer Perceptron Neural Network is used to
detect congestion. The trained network was used to anticipate traffic congestion in a specific
location using daily traffic data and a multilayer network model. They investigated applying this

-6-
on the data set, which resulted in output prediction of Stand-alone Dedicated Control Channel
(SDCCH) failure attempts in particular area using Levenberg-Marquardt as a training method in
neural networks for output prediction that takes less time in training.

Raheem and Okeene [12] study a neural network approach to GSM Traffic Congestion
Prediction. Using twelve months of real traffic data, they present a GSM congestion prediction
model based on multilayer Perceptron neural networks (MLP-NNs) with sigmoid activation
functions and Levenberg-Marquardt [9] Algorithms (LMA). The trained network model was
used to forecast traffic congestion on a particular route. The correlation coefficient between
expected traffic congestion volumes and actual traffic congestion volumes is 0.986, according to
regression analysis. This result demonstrates the utility of artificial neural networks in traffic
congestion prediction and control.

Siddiqui and Choudhary [13] developed Telecom voice traffic prediction for GSM using Feed
Forward Neural Network. The study is being carried out to estimate peak-hour voice traffic
congestion. They check QoS reports on a regular basis to verify that voice traffic is being used to
its full potential. This QoS report includes resource parameters such as the number of traffic
channels (TCH), stand-alone dedicated channels (SDCCH), and number of SDCCH seizure
attempts, SDCCH success calls, SDCCH block calls, and SDCCH drop calls, total calls, TCH
assign, TCH success calls, TCH availability rate, TCH drop, Incoming handoff (HO) success
rate, Outgoing handoff (HO) success rate, Half rate (HR), and Mean holding time. Siddiqui and
Choudhary's research focuses on utilizing a neural network to forecast voice traffic using the
above-mentioned QoS criteria on a daily basis and working with real-world data for quality
assessment.

Elisha Didam Markus et al, [14] Studied Predicting Telephone Traffic Congestion Using
Multi-Layer Feed forward Neural Networks. It shows how to estimate traffic congestion in a
telephone network using an artificial NN model. To represent telephone traffic, the design
technique employs a multilayered feed forward NN with back propagation algorithm. All
simulations were conducted in the Mat lab. The correlation coefficient between projected and
real traffic congestion levels was 87 percent in a regression analysis, demonstrating the utility

-7-
and effectiveness of Neural Networks in traffic prediction and control. The simulation findings
show that after the influential variables that influence congestion are determined, the neural
network MLP can forecast network congestion in the short and long run.

Yeshinegus [15] investigate and come up with a prediction modeling for detecting incoming
international calls that are terminated using local mobile numbers or fraud detection in
telecommunication the case of ethio telecom. The derived model can be integrated with the
existing system to detect frauds in telecommunication companies, specifically in ethio telecom.
This is done by implementing derived models from data mining tools, techniques and algorithms.
Yeshinegus this study CRISP-DM (Cross industry Standard process for data mining) model is
used and classification method like J48 and PART from decision tree and multilayer Perceptron
algorithms on data collected from ethio telecom. WEKA data mining tool used to design a model
for predicting fraudulent activities. For this study prepaid sample voice (call detail record) CDR
data has been used along with SMS, GPRS and other data such as pre-paid wallet recharge log
from OCS and CCB data warehouse in ethio telecom. The experiment result showed that the
model from the PART algorithm exhibited 100% accuracy level followed by J48 algorithm with
99.98%

Dereje [16] discovers hidden knowledge from ethio telecom mobile network data specifically
GSM mobile so as to determine call setup success rate. According, to the research to overcome
the drawback of simple statistical method they proposed data mining techniques, methods and
methodology. In order to discover knowledge from the data they used the divisive hierarchical
clustering methods to cluster the data. The K-means algorithm, WEKA tool and CRISP data
mining process model are used. As a result, they shows the knowledge which was discovered
during analysis of each cluster and the relationship between attribute against CSSR. The data
shows the most of the call setup failed and categorized under very poor CSSR category. To
enhance CSSR emphasize should be given to attributes used as KPIs. The study result reveals
which attribute should enhance to improve the call setup success rate. Enhancing CSSR leas
giving QoS to customers and it implies customer satisfaction and increases company revenues.
Finally they recommended ET to apply data mining technique using cluster analysis on GSM
mobile network data to analyze the data, evaluate the performance of the network, to assess the

-8-
quality of the service and to make better decision. Yared developed a predictive model that can
determine mobile call drops from ethio telecom mobile network data using data mining
techniques. They implement data mining process, classification methodology. The data collected
from fault management system, to build the model. They used WEKA for their study and four
classification techniques such as J48, random forest algorithm from decision tree as well as
PART and JRIP algorithm from rule induction are used. As a result, J48 decision tree algorithm
with 10-fold cross validation registered better performance and processing speed 95.43% and
0.06 sec respectively.

Jember[17] develop an application of data mining in fraud detection on mobile communication

service in the case of Ethio telecom mobile data using (Call detail record) CDR data. The CDR
contains a vast amount of data about each call made and is a significant source of data for
research to reveal hidden patterns of calls made by consumers, in addition to its normal use for
bill processing tasks. He performs data mining with artificial neural networks using MATLAB.
As a result, he discovered an accuracy rate of 89 percent and recommends further investigation
into other likely causes of fraud on the switch's pure CDR data.

Gebremeskel [18] Detecting unlawful calls from ETC's CDR Switching machine and enabling
early identification of those calls are important priorities. For their research, they used neural
network techniques and Brain Maker neural network software. CDR was used as the study's data
source, with an emphasis on ETC's Pre-paid mobile phone. As a result, fraudulent calls with an
accuracy level of 88.465 are archived, but non-fraudulent calls with an error rate of 4.19 percent
are not.

Asemelash [19] Research on 3G mobile fault occurrence prediction using Neural Networks was
developed. The research was carried out utilizing a case study of Addis Ababa 3G mobile sites
and a Nonlinear Auto Regressive (NAR) Neural Network time series prediction method. To train
the neural network, they employed the Levenberg-Marquardt method and an iterative strategy of
hidden layer neural number selection. As a result, the optimal model with the lowest mean
square error of prediction is chosen. The model was further evaluated with actual fault

-9-
occurrence time, which was not included in the training, and it achieved a prediction rate of
90.71 percent.

Aipeng Guo and Chunhui Yuan [20] Studied Network Intelligent Control and Traffic
Optimization Based on SDN and Artificial Intelligence designed a network control and solution
mechanism for network intelligent traffic optimization based on SDN and artificial intelligence.
Additionally, they also analyzes the objectives of traffic optimization as well as routing
calculation algorithms and routing optimization algorithms mainly focused on SDN-based
network traffic algorithm optimization and experimental verification. Design a network control
mechanism for network intelligent control as well as solutions for traffic optimization based on
SDN and artificial intelligence. They analyze operators‘ network requirements (e.g., the carrying
of the 5th generation mobile network (5G) service, multi-protocol label switching virtual private
networks optimization, cloud of services and the IP backbone network). The proposed
architecture consists of three modules, including a network status collection/perception module,
an AI intelligent analysis module and an SDN controller module. future will try to apply network
intelligent control and traffic optimization based on SDN and artificial intelligence solutions to
operators' actual networks in order to solve real data problems in future work; at the same time,
more network intelligent control application scenarios will be studied.

In some study Key Performance Indicators (KPIs) are used for Cell outage detection. Ahmed
Zoha et al. in [13] proposed COD framework that adopts a model driven approach that makes
use of mobile terminal-assisted data collection solution based on MDT functionality. They first
collected UE reported MDT measurements and extracted minimal KPI representation by
projecting them to a low-dimensional embedding space. They used two kind of anomaly
detection methods namely principal component analysis recurrent neural network (PCA-RNN)
and one-class support vector machine-based detector (OCSVMD) together with the embedded
measurements. The two learning algorithms were compared and evaluated. Moreover, the geo-
location associated with each measurement of COD framework was used to localize the position
of the faulty cell. Full dynamic LTE simulation tool was utilized to simulate the LTE network
consists of 27 e-Node Bs and to test the detection performance of the OCSVMD and PCA-RNN.

- 10 -
In the study presented by Szilágyi and Nováczki [21], integrated detection and diagnosis
framework is presented that can perform fault classification based on statistical analysis and find
the most probable root cause of problems. For detection, monitored radio measurements and
other KPIs were compared to their usual behavior captured by profiles automatically without
threshold and manual setting. But, diagnosis is depending on previous fault cases. The
abnormality level was used to calculate the likelihood of a failure case. The target with largest
likelihood value is considered to be the diagnosed failure.

Other detection approach is proposed by S. Rezaei & H. Radmanesh [22]. In this paper, an
automatic unified detection and diagnosis framework has been presented using unsupervised
clustering of both traffic and signaling KPIs for diagnosis. Moreover, experts reasoning also
incorporated in the design to enable automatic decision support in an operating mobile
communications environment for fault diagnosis. They considered the real data of a live GSM
network for performance evaluation.

In another study in [23] by I. De-La-Bandera et al., HO statistics has been used for cell outage
detection. The proposed Cell Outage Model includes different cases of cell outage. First, cell
outage that does not affect the eNB in which eNB generates KPIs from the cell in outage. Second
is site outage which affects eNB and there is no KPIs available from the Operation support
system (OSS). The third case is when Cell is not in outage, but there is a failure in eNB-OSS
connection. LTE simulator test results indicates that algorithm enables to detect a cell outage
when KPIs from the cell are either available or not. The drawback of the algorithm is that cell in
outage with very low traffic cannot be detected.

There are also researches based on data mining and machine learning techniques. Automated
network troubleshooting using data mining is presented by E. Rozaki [24]. Amonitoring scheme
is proposed for mobile networks based on the use of rules and decision tree data mining
classifiers. The goal of the study is to improve anomaly detection and fault localization based on
a top-down (Bayesian networks) model. The data mining techniques was used to train a system
to learn network fault rules.

- 11 -
T. Zhang et al. in [25] proposed COD architecture based on the handover statistical data to detect
small cell outage in heterogeneous network (HetNets). They use data mining methods and
preprocessed sequential HO data spatially and temporally. Detection algorithms performance is
evaluated using their own designed simulator with some reasonable assumptions. The results of
simulation show that their system is more effective to detect small cell outage in comparison to
the model using MDT measurements.

In a study by Y. Ma, et al. in [26], an unsupervised data mining algorithm called Dynamic
Affinity Propagation (DAP) clustering which uses reference signal received power (RSRP) and
reference signal received quality (RSRQ) as input data from UEs, eNodeBs and OAM to detect
cell outages was introduced. The LTE-Advanced simulation environment is used to test the
proposed algorithm and cell outages are successfully detected.

The experiments have demonstrated that there is no single machine learning algorithm which can
handle efficiently all the types of attacks. The decision table (rule base classifiers) achieved the
lowest false negative value, but it was far from the highest accuracy rate detection. On the other
hand, Random Forest classifier performs significantly better than MLP ANN but, MLP-RNN
classifier registered the highest accuracy rate 99.79%, with the smallest root mean square error
(RMSE. The experimental result shows that for the real dataset, MLP-RNN performs better than
the two other techniques in terms of the execution time with evaluation Parameters in terms of
accuracy, sensitivity, precision, and F-score compared to the several existing has been
recommended for cellular network detection methods [27].

1.4. Objective of the Study

1.4.1. General Objective

The main objective of this thesis is to investigate a COD model using MLP-RNN machine
learning technique based on HO statistical data collected from UMTS network and evaluate the
performance of the detection scheme.

- 12 -
 1.4.2 Specific Objectives

The specific objectives to be accomplished in this thesis are:

• Review related literatures and understand COD methods.
• Study the various machine learning techniques and identify the available HO statistical data to
come up with a COD model for the thesis work.
• Collect incoming HO statistical data and preprocess it for analysis by transforming the data
into a format suitable for the proposed machine learning detection algorithm.
• Develop a Matlab program for COD algorithm and test it by altering the control parameters to
get best performance.
• Evaluate the performance of the detection algorithms using Receiver Operating Characteristic
(ROC) curve.
• Discuss the results and draw recommendations based on the findings.

1.5. Research Question

As the problems mentioned in Section 1.2 still remain as open research issues, this thesis
attempts to address these problems by answering the following research questions:
➢ Which machine learning algorithm can be more efficient for the purpose of detecting
COD?
➢ Which model training approach is better in order to detecting COD?
➢ How to evaluate the performance of cell outage detection?
➢ How to improve network performance for further network communication?

1.6. Scope and Limitation of the Study

This thesis addresses one of the use cases of self-healing which is cell outage detection. The
scope of the thesis is to investigate cell outage detection model based on incoming HO statistical
data of UMTS mobile network using outlier anomaly detection algorithm. Even though cell
outage detection concept is the same for all mobile technologies, the thesis work is restricted to
only UMTS network. UMTS Node-B sites are selected for analysis. Due to time limitation and

- 13 -
complex matters, the thesis is limited to cell outage detection only rather than diagnosis the root
cause of the problem. Hence, further research can be conducted to investigate the cause of the
cell outage

1.7.Significance of the Study

The Significances of this work are as the follows:

• Characterizing attack traffic and innocent traffic in networks: by investigating real traffic
traces, we find out important features that can be used to differentiate normal traffic and SYN
flood attacks traffic in network scenarios.
• Proposing novel machine learning mitigation algorithms: based on these features, a machine-
learning algorithm integrated in the WEKA controller has been developed to detect and drop
attack traffic while innocent traffic is almost not affected;
• An adaptive method to optimize the parameters of mitigation algorithms for accuracy
improvement: A Wireshark is deployed to evaluate the approach in real devices and real time.
Based on experiments conducted in the Wireshark, an adaptive mechanism to improve the
mitigation accuracy is proposed.

1.8. Methodology

In order to have a better understanding of this thesis work, different related literatures, journals,
and books on COD and UMTS network have been reviewed. Relevant materials on detection
algorithms of data mining techniques are also referred. Moreover, UMTS process manuals and
network design documents are also consulted to understand the UMTS network management.

This thesis was performed by using the following methods:

• The thesis work started with a literature review to understand more the purpose of the research
and then familiarize with the available related research works.
• Necessary data for the study has been identified.
• COD model has been proposed.
• In ordered to accomplish the research process, COD algorithm WEKA, Matlab program has
been developed.

- 14 -
• A real UMTS network scenario has been considered and the required data has also been
collected in collaboration with domain experts.
• Data has been preprocessed using MS-Excel and detection system analyzed its output to detect
outage cells.
• Finally, the results have been discussed and published in the form of a final thesis paper.

Figure 1.8.1. Performance Evaluation Parameters / metrics

The Accuracy on the testing set (ACC):

ACC = (TP + TN)/(P + N) = (TP + TN)/(TP + TN + FP + FN)

This proposal analyzes a machine learning-based cell outage detection. The evaluation is based
on best-suited metrics such as; recall or sensitivity, specificity, precision, False Positive Rate
(FPR) and computational time rather than stick on generic metrics like accuracy. Finally,
discussed the results and give recommendation based on findings.

- 15 -
Figure 1.8.2. Research Methodology

- 16 -
 1.9. Work plan and budget break down schedule
1.9.1. Work plan schedule
This section will show the estimated time table in-order to complete each specific objectives of
proposal work.

Table 1: Schedule of tasks to be undertaken

No Task /Activities

Month Jan Feb Mar Apr May June July Agus

Year 2024 2024 2024 2024 2024 2024 2024 2024

1 Title Selection

2 Introduction/Background

3 Statement of the Problem

4 Literature Review

5 Objective

6 Methodology

7 Work plan and budget

8 Proposal Submission

9 Paper Completion

10 Final paper Submission

11 Thesis /Defense

- 17 -
 1.9.2. Budget break down
The following table will describe estimation of the costs that are expected to be incurred by this
proposal work.

Table 2: Estimated cost of proposal work

Total Equipment and Price

No Types of Equipment Unit Quantity Unit of price Total price

1 HP computer core i5 No 1 20000 20000

2 Scanner No 1 5000 5000

3 CDMA (Wi-Fi) No 1 1500 1500

4 External Hard disk No 1 1200 1200

5 A4 paper Packet 3 150 450

6 Digital camera (For video No 1 4500 4500

& audio Recording)

7 Pen Packet 1 250 250

8 Software( LOIC WEKA, No 3 300 300

Wireshark)

9 To Print Documents No 500 1.50 750

Total Price 33950

- 18 -
Reference
1. Mueller, C. M., Kaschub, M., Blankenhorn, C., & Wanke, S.(2008). A cell outage detection
algorithm using neighbor cell list reports. In K. Hummel & J. Sterbenz (Eds.), Self-organizing
systems (Vol. 5343, pp. 218–229)., Lecture notes in computer science Berlin Heidelberg:
Springer.
2. Ciocarlie, G., Cheng, C. C., Connolly, C., Lindqvist, U., Nitz, K., Novaczki, S., Sanneck, H.,
& Naseer-ul Islam, M. (2014a). Anomaly detection and diagnosis for automatic radio network
verification. In 6th international conference on mobile networks and management, MONAMI
2014.
3. Ciocarlie, G., Lindqvist, U., Nitz, K., Novaczki, S., & Sanneck, H. (2014c). On the feasibility
of deploying cell anomaly detection in operational cellular networks. In Network operations and
management symposium (NOMS), 2014 IEEE, pp 1–6, doi:10. 1109/NOMS.2014.6838305.
4. 3rd Generation Partnership Project. (2009a). Evolved universalterrestrial radio access network
(e-utran); self-configuring and self-optimizing network (SON) use cases and solutions (release
9). Tech. Rep. TR 36.902, 3GPP 28. 3rd Generation Partnership Project. (2009b). Technical
specification group radio access network; study on minimization of drive-tests in next generation
networks (release 9). Tech. Rep. TR 36.805, 3GPP
5. 3rd Generation Partnership Project. (2010). 3GPP; TSG radio access network; further
advancements for e-utra physical layer aspects (release 9). Tech. Rep. TR 36.814, 3GPP
6. 3rd Generation Partnership Project. (2011). Technical specification group radio access
network; evolved universal terrestrial radio access (e-utra); radio resource control (rrc); protocol
specification (release 10). Tech. Rep. TS 36.331, 3GPP
7. 3rd Generation Partnership Project. (2012). Technical report 3rd generation partnership project;
technical specification group radio access network; evolved universal terrestrial radio access (e-
utra); mobility enhancements in heterogeneous networks (release 11). Tech. rep., 3GPP 32. 3rd
Generation Partnership Project. (2014). Self-organizing networks (SON); self-healing concepts
and requirements (release 12). Tech. rep., 3GPP TS 32.541 V12.0.0
8. GSMA Intelligence (2014) Understanding 5g: Perspectives on future technological
advancements in mobile. http://www.cisco. com/c/en/us/solutions/collateral/service-
provider/visual-networking-index-vni/white_paper_c11-520862.html
9. Ha¨ma¨lainen, S., Sanneck, H., & Sartori, C. (2012). LTE selforganising networks (SON):

- 19 -
Network management automation for operational efficiency (1st ed.). Hoboken: Wiley
Publishing.
10. Hapsari, W., Umesh, A., Iwamura, M., Tomala, M., Gyula, B., & Sebire, B. (2012a).
Minimization of drive tests solution in 3GPP. Communications Magazine, IEEE, 50(6), 28–36.
11. Hapsari, W., Umesh, A., Iwamura, M., Tomala, M., Gyula, B., & Sebire, B. (2012b).
Minimization of drive tests solution in 3GPP.Communications Magazine, IEEE, 50(6), 28–36.
12. He, H., & Garcia, E. A. (2009). Learning from imbalanced data. IEEE Transaction on
Knowledge and Data Engineering, 21(9),1263–1284. doi:10.1109/TKDE.2008.239.Transaction
Pattern Analysis Machine Intelligence, 32(11),
2006–2021.
13. Holma, H., & Toskala, A. (2011). LTE for UMTS: Evolution toLTE-advanced (2nd ed.).
Hoboken: Wiley Publishing.
14. Johansson, J., Hapsari, W., Kelley, S., & Bodog, G. (2012). Minimization of drive tests in
3GPP release 11. Communications Magazine, IEEE, 50(11), 36–43.
15. Khanafer, R., Solana, B., Triola, J., Barco, R., Moltsen, L., Altman, Z., et al. (2008).
automated diagnosis for umts networks using bayesian network approach. Vehicular
Technology, IEEE Transactions on, 57(4), 2451–2461. doi:10.1109/TVT.2007. 912610.
16. Kolehmainen, N. (2007). Downlink packet scheduling performance in evolved universal
terrestrial radio access network. Master’s thesis, University of Jyva¨skyla¨, Finland.
17. Laiho, J., Raivio, K., Lehtimaki, P., Hatonen, K., & Simula, O. (2005). Advanced analysis
methods for 3g cellular networks. Wireless Communications, IEEE Transactions on, 4(3), 930–
942. doi:10.1109/TWC.2005.847088.
18. Luo, F. L., Unbehauen, R., & Cichocki, A. (1997). A minor component analysis algorithm.
Neural Networks, 10(2), 291–297.
19.Ciocarlie, G., Lindqvist, U., Novaczki, S., & Sanneck, H. (2013). Detecting anomalies in
cellular networks using an ensemble method. In Network and service management (CNSM),
2013 9th international conference on, pp 171–174, doi:10.1109/CNSM. 2013.6727831.
20. Next Generation Mobile Networks (2008a) Recommendation on SON and O&M
Requirements. Tech. rep., NGMN, URL http://

- 20 -