Probabilistic techniques for multi-component system diagnostics and surveillance.

Abstract

A new approach at multi-component system fault diagnostics is developed and demonstrated. This approach utilizes detailed system simulation models, uncertain system observation data, statistical knowledge of system parameters, expert opinion, and component reliability data in an effort to identify incipient component performance degradations of arbitrary number and magnitude. A complete probabilistic framework for the procedure is first developed using Bayesian, Chapman-Kolmogoroff, Master and Fokker-Planck equations to represent the stochastic system behavior. Based on these mathematical principles, algorithms for the practical solution to the diagnostic problem are then developed. The technique involves the use of multiple adaptive Kalman filters for fault estimation, the results of which are screened using statistical hypothesis testing procedures to define a set of component events that could have transpired. Latin Hypercube sampling is then used to determine the likelihood of each of these feasible component events in terms of uncertain component reliability data and the density functions characterizing system and component states obtained from the adaptive filters. This information is then combined to yield marginal density functions descriptive of the operational state of the components of interest. This diagnostic framework allows one to transform sometimes disparate pieces of uncertain system information into statistically meaningful diagnostic knowledge. The capabilities of the procedure are demonstrated through the analysis of a simulated small-magnitude binary component fault in a boiling water reactor balance of plant. The probabilistic procedure is applied to first detect the presence of the anomaly and then determine a feasible subset of component states that could have resulted in the observed system behavior. A likelihood is then determined for each feasible component state and marginal probability density functions characteristic of each components operation are constructed. The results show that the procedure has the potential to be a very effective tool for the off-line analysis of system data for small-magnitude, multiple fault diagnosis. Of particular interest is the potential for utilizing the resultant marginal probability density functions to probabilistically monitor the trend in component performance. Such an analysis could be an extremely useful aid in monitoring system safety margins and in the optimum scheduling of maintenance activities.