Sensitivity Analysis with Reliability Analysis

Birnbaum importance is a local, structural measure: it quantifies how much system reliability would change if a specific component were made perfectly reliable or failed with certainty, assuming all other components are at their nominal reliability. Sobol indices are global variance-based measures: they quantify what fraction of the total variance in system reliability is attributable to each uncertain input parameter across its full probability distribution, including interaction effects. For systems with independent components and simple architectures, both give similar rankings; for systems with strong interactions or highly uncertain inputs, Sobol indices are more informative.

There is no universal minimum, but component failure rate estimates should have coefficient of variation below roughly 0.5 for the sensitivity rankings to be stable. If failure data are very sparse (fewer than 5–10 observed failures per component), Bayesian reliability analysis combined with sensitivity analysis is recommended so that prior information can compensate for sparse data and the sensitivity of results to prior choice can itself be examined.

Yes, but with significant adaptation. Software reliability models (e.g., Jelinski-Moranda, NHPP models) express failure rate as a function of residual defects and testing time rather than physical failure mechanisms. Sensitivity analysis can rank input parameters of those models — initial fault count, fault detection rate, operating profile — by their influence on predicted software reliability. The interpretation differs from hardware reliability but the methodological pipeline is the same.

Analytical importance measures (Birnbaum, Fussell-Vesely) are computationally efficient and exact for systems representable as fault trees or reliability block diagrams with independent components. Monte Carlo simulation is necessary when the system reliability function is implicit (e.g., defined by a finite-element model), when inputs are correlated, or when global variance-based sensitivity indices are required. For large or complex systems, Latin hypercube sampling and quasi-Monte Carlo methods reduce the number of simulations needed.

They are closely related but distinct. Uncertainty analysis quantifies the overall uncertainty in the reliability estimate given uncertain inputs — producing a confidence interval around the system reliability figure. Sensitivity analysis goes further and decomposes that uncertainty to attribute it to specific input parameters. In practice, both are run together: uncertainty analysis shows how uncertain the reliability estimate is; sensitivity analysis shows which inputs are responsible for that uncertainty.