Fault Localization Using Neural Networks For Fault Management Based on Alarm Correlation Analysis: The Case of Ethio telecom
No Thumbnail Available
Date
2021-11
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Addis Ababa University
Abstract
Telecommunication sector has considerably large network infrastructure consisting of different
network equipment to provide telecom services to the customers. As a service provider,
telecom operators have to make sure that customers are getting services without interruption.
To do so, they built network operation center to monitor the network status with the help of
a fault management system (FMS). When malfunction or error appears on a piece of network
equipment, they typically produce an alarm and transmits to the FMS, which is then shown to
the monitoring team for trouble ticket creation.
There is also a chance that failures that happened on one network element can cause other network
elements to produce alarms as long as the network equipment’s are functionally dependent.
Thus, during the trouble ticketing, the operators may need to collect further information
such as relationship or dependency between alarms to ultimately distinguish the root source of
the alarms. But the information on FMS lacks explanation and indication of the causes, forcing
monitoring teams to manually figure out the cause and take action, which results in a poor
decision making as well as a longer trouble shooting and service outage time.
This study focuses on studying the association between network equipment alarms using a correlation
analysis technique called Pearson correlation using history alarm data and studying
the feasibility of using artificial neural network to build fault localization model to support the
monitoring team with decision making. Fault localization models using one specific neural
network architecture are built in different literatures. But in this study two neural network architectures
namely, Feed forward neural network (FFNN) and cascade forward neural network
(CFNN), are selected to build a model using the data of correlation analysis output, the microwave
link topology and the history alarms as an input. Data preparation techniques such as
data engineering and feature engineering have been applied to the collected alarm data from
the existing fault management system. After building the models, they are evaluated using
commonly used performance metrics such as accuracy and error measurement. Python is used
as programming tool to perform the correlation analysis as well as to develop and compare the
two neural network models. Experimentation results exhibit that CFNN achieves an accuracy
of 97.7% while FFNN achieves 95.9% overall accuracy.
Description
Keywords
Alarm correlation analysis, Artificial neural network, Fault management system