Learning Dynamic Fault Trees from Data for Interpretable Failure Prediction

Fault Tree Analysis (FTA) is a standard method in reliability engineering to model causes of system failures. While Static Fault Trees (SFTs) capture logical dependencies, they cannot represent time-dependent or sequential failure behavior. Dynamic Fault Trees (DFTs) overcome this limitation but are still constructed manually using expert knowledge, limiting their use in data-driven applications. This paper presents a method for automatically learning DFTs from temporal event data. We extend the LIFT algorithm with support for dynamic gates (PAND, SEQ) and use the PAMH test to evaluate statistical associations between event combinations and their parent, while a Chi-square test distinguishes between SEQ and PAND gates based on the frequency of order violations. The result is an interpretable model that captures both structural and temporal failure logic. We evaluate the method on synthetic datasets generated from a manually constructed DFT designed by a domain expert, with varying levels of noise and event frequencies. The learned DFTs show high predictive and structural accuracy under ideal conditions and remain robust under noise with appropriate parameter tuning. This demonstrates the potential of the approach for interpretable failure prediction in intelligent manufacturing systems.