Not-So-Na ï ve Bayesian Networks and Unique Identification in Developing Advanced Diagnostics

Problems in accuracy and effectiveness in system diagnosis and prognosis arise from constructing models from design data that do not match implementation, failing to account for inherent uncertainty in test data, and failing to account for characteristics unique to specific units due to variations in usage, environment, or other factors. Large sums of money have been expended by owners of these systems, but little improvement in measures such as retest-OK rate and cannot duplicate rate has been reported. In fact, simply losing track of where specific units are located has resulted in substantial losses of money. In this paper, we study the problem of performing diagnosis and prognosis on systems and describe an approach to building models based on data collected about specific units. We rely on the emerging Department of Defense (DoD) unique identification (UID) program that is focusing on obtaining this data and apply Bayesian methods for constructing such diagnostic models. Specifically, we discuss an alternative class of Bayesian model that we call the "not-so-naive" Bayesian network (NBN). We also discuss the concept of the NBN in the context of the UID program as a means of tracking and deriving probabilities for creating the network. Finally, we focus on the specific problems encountered and lessons learned from working with a large, real-world database for the US Navy's STANDARD missile