Title: BRACES: Bayesian Reliability Analysis for Complex Engineered Systems

Speaker: Paul Boggs, SNL Org. 08961

Date/Time: Tuesday, November 11, 2008, 2:00-3:00 (NM), 1:00-2:00 (CA)

Location: CSRI Building/Room 90 (Sandia NM), Building 915/Room S145 (CA)

Brief Abstract: Nuclear weapons testing is a key application area at Sandia.  Weapons are complex systems of interconnected components for which reliability is of paramount importance, and the processes for estimating this reliability quantitatively are extraordinarily complicated.  While these estimation processes use analytical methods that incorporate data acquired through an array of testing procedures, there is significant pressure to reduce the costs of testing while maintaining the integrity of the stockpile.  Moreover, a reliability estimate is just that, an estimate, particularly for systems for which it is impractical, or impossible, to perform sufficient classical tests, to ensure a specified level of confidence.  In these cases, it is relevant to establish a formal process for assessing the uncertainty in the estimate itself.  Both of these aspects motivate our attempt to investigate alternatives to our current practices in weapons testing and reliability assessment.

In this talk, we describe novel mathematical strategies, based on structured probabilistic models, for testing our systems.  We propose to use these strategies to assess the impact of various combinations of subsystem- and component-level tests on overall system reliability, and its associated uncertainty due to epistemological realities.  We will also consider the problem of predicting future reliability given time-dependent data.  The Bayesian approach of Martz-Waller will first be extended to testing regimes where continuous data, e.g., voltage, are collected in addition to that collected in traditional, and simpler, "pass-fail" tests.  This approach will be extended further to accommodate time-dependent systems; our goal is to handle predictions in an environment where aging may be important.  Finally, we will extend this approach to more general network-class applications for which certain components may exist in, or directly affect, multiple systems.  Exact, or approximate, inference algorithms will allow probability distributions characterizing reliability to be updated through and across systems.  This ability is crucial for estimating the uncertainty in an overall system reliability estimate.  Our construction will, for the first time, enable the formulation of optimization questions concerning the best testing strategies and the relative importance of each individual test on the overall reliability and uncertainty estimates.