Learning Classification-DBNs from Data

Abstract

As our societies are becoming ever more digital and reliant on automated systems, it becomes increasingly important to monitor the technologies we depend on using automated systems to guard against failures and downtime. While many fault detection solutions have already been proposed, we found that methods for continuously monitoring the state of a system in an explainable way have not yet been widely researched, while this could provide helpful information to the user. Therefore, we propose C-DBNs, a special case of Dynamic Bayesian networks that have been tailored to classify dynamic processes using existing probabilistic models. We also introduce S-RAD: a novel method for automatically discretizing datasets for usage with C-DBNs to automate the process of learning explainable models even further. Our ?rst results seem promising and provide a reliable alternative to existing methods of discretization without prior knowledge.