A Stochastic Neural Network Model for Missile Reliability Prognostics

"The proposed project will demonstrate the feasibility of developing a Stochastic Neural Network for Missile Reliability Prognostics. A neural network model will be developed to predict missile reliability based on the environmental history of particularmissiles. Prediction training will be employed to predict missile reliability at future times, based only on training data available at the current time. Principal components metrics will be used to quantify generic modeling uncertainty based on pastexperience, for purposes of evaluating the predictive accuracy of future reliability estimates. The innovation proposed here will combine the benefits of prediction training and generic uncertainty modeling to quantify the accuracy of reliabilitypredictions based on a statistical comparison of earlier predictions with subsequent observations. Both will be correlated with measurements of the ambient environment recorded by a data logger. Statistical analyses of the environmental data will beperformed to coincide with the update of reliability information, reducing the environmental data to time dependent statistical parameters. In addition to providing the Army with a tool for missile reliability prognostics, an immediate spin-off might verywell be the application to Sandia's maintenance of the Nation's nuclear weapons stockpile, where radiation is an important environmental factor. Other potential applications involve various types o