Much scientific literature claims that existing digital twins often lack semantics, leaving the plant maintenance team responsible for interpreting and responding to faults. To enable semantics in a digital twin, it must rely not only on the data produced by sensors, but also on a deeper knowledge of the system and the processes taking place within it. This paper proposes a framework for the automated generation of Bayesian Networks (BNs) from knowledge graphs, which should store information from different sources, such as topology, documents originally written in natural language, and domain-specific ontologies based on RDF (Resource Description Framework). BNs will be used to infer failure symptoms and causes, while automated refinement of BNs is expected to address scalability issues. As a first representative demonstrator, a two-room facility was modeled in the Dymola environment and coded according to the Brick Ontology and the Digital Building Ontology. A BN was extracted from it and tested for fault analysis. Finally, the two knowledge graphs were compared to conclude on their efficiency for automated BN generation.