System Health Management (SHM) systems have found their way into many safety-critical aerospace and industrial applications. A SHM system processes readings from sensors throughout the system and uses a Health Management (HM) model to detect and identify potential faults (diagnosis) and to predict possible failures in the near future (prognosis). It is essential that a SHM system, which monitors a safety-critical component, must be at least as reliable and safe as the component itself—false alarms or missed adverse events can potentially result in catastrophic failures. The SHM system including the HM model, a piece of software, must therefore undergo rigorous Verification and Validation (V&V).

In this paper, we will describe an advanced technique for the analysis and V&V of Health Management models. Although our technique is generally applicable, we investigate in this paper HM models in the form of Bayesian networks (BNs). BNs are a powerful modeling paradigm to express notions of cause and effect, probability, and reliability. A BN model typically contains many parameters (e.g., thresholds for discretization and conditional probability tables); they need to be set carefully for reliable and accurate HM reasoning. We are investigating the use of Parametric Testing (PT), which uses a combination of n-factor and Monte Carlo methods, to exercise our HM model with variations of perturbed parameters. Multivariate clustering on the analysis is used to automatically find structure in the data set and to support visualization. Our approach can yield valuable insights regarding the sensitivity of parameters and helps to detect safety margins and boundaries.

As a case study we use HM models from the NASA Advanced Diagnostics and Prognostics Testbed (ADAPT), which is a realistic hardware setup for a distributed power system as found in spacecraft or aircraft.