Abstract:
Telecommunications networks are ever growing and rapidly expanding.
Their
management becomes complicated with same kind of equipment purchased from differ-
ent manufacturers and incorporation of newer technologies to accommodate customer
demands. In such a scenario, modeling of an ever changing telecommunications network
becomes complicated. Automatic methods are necessary and modeling of event/alarm
intensity becomes crucial for monitoring of a telecommunication network in these settings.
The framework of Salmenkivi [1, 2] has been extended to incorporate classical Poisson
likelihood and Bayesian integrated likelihoods proposed by Scargle [3].
Scargle has
proposed three Bayesian integrated likelihoods to segment γ−ray bursts coming from the
space. He has used these Bayesian integrated likelihoods with hierarchical algorithm to
segment the data to model intensity of Gamma ray bursts. Two of those three likelihoods
mentioned as Scargle1 and Scargle2 likelihoods are used under both hierarchical and
dynamic programming algorithms to model intensity of event/alarm data collected from
a typical telecommunications network.
Unlike Salmenkivi, this study directly considers the discrete event/alarm data.
Event/alarm data collected from telecommunications networks and a large amount of
synthetic datasets are processed with hierarchical and dynamic programming algorithms
by employing classical Poisson and Bayesian integrated likelihoods.
The same data
has also been processed with hierarchical Bayesian models proposed by Green [4] and
Dobigeon et al., [5, 6]. The results of hierarchical and dynamic programming algorithms
are compared with those obtained from hierarchical Bayesian models.
Finally, the British coal mining disasters dataset is processed with hierarchical and
dynamic programming algorithms in various time resolutions. This is done to focus on
event/alarm thresholds below 1. New results have emerged and a different behavior of
classical Poisson and Bayesian integrated likelihoods has been found and reported. A
novel hierarchical Bayesian model has been proposed and simulated with Gibbs sampler
that models time differences between events/alarms.
