Pattern Recognition in Studies that Inject Synthetic Signals into Real Background

摘要

In quantitative nuclear non-proliferation studies, it is common to have modest or even large amounts of real background data, but relatively little background-plus-signal data Therefore, signal effects are often modelled, and the modelled effects are injected into real background data Then, various pattern recognition options can be applied or developed to learn to distinguish background-only from background-plus-signal data. This paper evaluates impacts of estimation error in the modelled signal on signal-detection probabilities in the case of using linear discriminant analysis as the pattern recognition option.