Simulations play an important role in the continuing effort to reduce development time and risks. However, large and complex models are necessary to accurately simulate the dynamic behavior of complex engineering systems. In recent years, the use of data-driven models based on machine learning (ML) algorithms has become popular for predicting the structural dynamic behavior of mechanical systems. Due to their advantages in capturing non-linear behavior and efficient calculation, data-driven models are used in a variety of fields like uncertainty quantification, optimization problems, and structural health monitoring. However, the black box structure of ML models reduces the interpretability of the results and complicates the decision-making process. Hierarchical Bayesian Networks (HBNs) offer a framework to combine expert knowledge with the advantages of ML algorithms. In general, Bayesian Networks (BNs) allow connecting inputs, parameters, outputs, and experimental data of various models to predict the overall system-level dynamic behavior. This characteristic of BNs enables a divide and conquer approach. Hence, complex engineering systems can be split into more easily describable subsystems. HBNs are an extension of BNs that can use knowledge about the structure of the data to introduce a bias that can contribute to improving the modelling result. In this work, an approach to design a HBN is presented where each model in the network can be a parametric reduced finite-element models. The influence of the hierarchical approach is evaluated by comparing a HBN and a BN of the model from the Sandia structural dynamics challenge.
