Induction of Fault Trees through Bayesian Networks

Cyber-physical systems have increasingly intricate architectures and failure modes, which is due to an explosion of their complexity, size, and failure criticality. While expert knowledge of individual components exists, their interaction is complex. For these reasons, obtaining accurate system reliability models is a hard task. At the same time, systems tend to be continuously monitored via advanced sensor systems. This data describes the components' failure behavior and can be exploited for failure diagnosis and learning of reliability models. This paper presents an effective algorithm for the learning of Fault Trees from data. Fault trees (FTs) are a widespread formalism in reliability engineering. They capture the failure behavior of components and their propagation through an entire system. To that end, we first use machine learning to compute a Bayesian Network (BN) highlighting probabilistic relationships between the failures of components and root causes. Then, we apply a set of rules to translate a BN into an FT, based on the Conditional Probability Tables to decide, amongst others, the nature of gates in the FT. We evaluate our method on synthetic data and a benchmark set of FTs.