This paper introduces a Hierarchical Uncertainty Quantification (H-UQ) framework to model cascading error effects in structural engineering systems. The framework integrates Bayesian neural networks with graph-based network uncertainty propagation to capture dependencies between hierarchical components. A Cascading Uncertainty Decomposition (CUD) metric is introduced to quantify the contribution of each uncertainty source to the total predictive variance. The framework is validated on benchmark datasets, demonstrating significant improvements in uncertainty calibration and computational efficiency.
Key findings
Proposes a novel H-UQ framework to model cascading error effects in structural engineering systems.
Introduces Cascading Uncertainty Decomposition (CUD) metric to attribute total uncertainty to specific hierarchical sources.
Validates the framework on benchmark datasets, showing up to 34% reduction in total uncertainty estimates compared to traditional methods.
Limitations & open questions
The framework's scalability to larger, more complex structural systems needs further investigation.
The assumptions made in the Bayesian neural networks may limit the applicability in certain scenarios.