Abstract/Summary

This thesis concerns the prediction of faults in a telecommunication network, this is done in an expressive way that offers engineers insight into the cause of a fault to help guide mitigating action. The thesis is a detailed account of the preprocessing steps applied to the data and an exploration of Rule Induction techniques resulting in a novel method of Rule induction. A method is developed to enable classification algorithms to forecast events. The result of this work is a system that can forecast critical events with high precision. Telecommunications are a vital part of modern society. They are relied upon for the running of both businesses and personal lives and so minimising the disruption to this network is very important. To this end there is a system of alarms in place that detail faults and warnings that engineers can respond to in a timely manner. A method of forecasting these alarms would allow engineers more time to take action. If these forecasts took the form of feature rich expressive rules then engineers would be offered a greater insight into the cause of an issue and may be able to mitigate the oncoming fault in the future. To create more expressive rules a method is introduced to hybridise event forecasting and event classification, allowing a classifier to produce forecasting rules. This offers a number of benefits over traditional forecasting techniques. This work also contributes an accompanying novel abstaining classifier adapted to producing rules with a broad range of features. This a very desirable feature when dealing with alarm data as other forecasting techniques are not able to capitalise on the potentially valuable data held within each alarm outside of the overarching alarm type. Rules that can incorporate these details will offer an engineer more information to work with.