Modeling Dependable Systems with Continuous Time Bayesian Networks

In the domain of information systems modeling for dependability is an established method. Most approaches dealing with structural or probabilistic modeling do not consider time information and handle only discrete data. But in reality systems have a time varying behavior and numerous measures are continuous. In this paper an approach for modeling dependable information systems for fault prediction is presented. The method considers time behavior and continuous variables. The technique is based on continuous time Bayesian networks (CTBN) which make assumptions for time to transition or time to failure feasible. A drawback of CTBN is that only discrete data is processed, thus continuous variables have to be discretized. This is carried out by grouping measures with distributions which are similar with the restriction that values from continuous range are nearby. Furthermore this technique is capable of performing a data reduction such that subsequent computations can be done with moderate hardware resources. Based on such preprocessed data the structure of the Bayesian network (BN) is learned by a Max-Min Hill-Climbing (MMHC) algorithm. Known misbehavior (e.g. faults) is incorporated into the BN by introduction of auxiliary variables. A structural model generated in this way forms the backbone of a continuous time Bayesian network. Henceforth CTBN parameter estimation (e.g. time characteristic) is doable by established learning methods.