Reasoning about uncertainty in prognosis: a confidence prediction neural network approach

Uncertainty representation and management is the Achilles heel of fault prognosis in condition based management systems. Long-term prediction of the time to failure of critical military and industrial systems entails large-grain uncertainty that must be represented effectively and managed efficiently, i.e. as more data becomes available, means must be devised to narrow or "shrink" the uncertainty bounds. Prediction accuracy and precision are typical performance metrics employed to access the performance of prognostic algorithms. That is, we would like the predicted time to failure to be as close as possible to the real one. Also, the bounds or limits of uncertainty must be as "narrow" as possible. This paper introduces a novel confidence prediction neural network construct with a confidence distribution node based on a Parzen estimate to represent uncertainty and a learning algorithm implemented as a lazy or Q-learning routine that improves online prognostics estimates. The approach is illustrated with test and simulation results obtained from a faulty helicopter planetary gear plate.