A series of repeated nuclear material balances forms a time series
of often autocorrelated observations.
Outliers, deviations from an in-control production process or time
series pattern, indicate an out-of-control
situation relative to the process norm. In this paper various
methods, especially neural networks, will be
examined with respect to their use to detect nuclear material
diversions or losses more rapidly and accurately
than currently used methods. The neural network technique will be
enhanced with the use of a simulation
computer program for creating the training data set. This
simulation approach provides the opportunity of
including outliers of various types in a data set for training the
neural network because an actual process
data set used for training possibly may not have outliers. In this
paper, the methods will be compared on
their ability to identify outliers and reduce false alarms. These
methods were tested on data sets of nuclear
material balances with known removals, and the results are tabulated
and described. Based on these results,
we believe the algorithms used will assist the nuclear industry in
process control, provide a new approach
to nuclear material safeguards, and also provide a new approach to
training neural networks for process
control applications.