Principal component analysis of low resolution energy spectra to identify gamma sources in moving vehicle traffic

A system intended to detect, classify, and track radioactive sources in moving vehicle traffic is under development at Lawrence Livermore National Laboratory (LLNL). This system will fuse information from a network of sensor suites to provide real time tracking of the location of vehicles emitting gamma and/or neutron radiation. This work examined aspects of the source terms of interest and applicable gamma detection technologies for passive detection of emitted gamma radiation. The severe restriction placed on the length of count due to motion of the source is presented. Legitimate gamma sources expected in traffic are discussed. The requirement to accurately classify and discriminate against these “nuisance” sources and cost restraints dictate the choice of NaI(Tl) detectors for this application. The development of a capability to automatically analyze short duration, low signal-to-noise NaI(Tl) spectra collected from vehicles passing a large, stationary detector is reported. The analysis must reliably discriminate between sources commonly transported in motor vehicles and alert on the presence of sources of interest. A library of NaI(Tl) spectra for 33 gamma emitting sources was generated with MCNP4B Monte Carlo modeling. These simulated spectra were used as parent distributions, from which multiple realizations of short duration spectra were sampled. Principal component analysis (PCA) of this data set yielded eigenvectors that enable the conversion of unknown spectra into principal component space (PCS). An algorithm using least squares fitting of the positions of library sources in PCS as basis functions, capable of identifying library nuclides in unidentified spectra, is reported. Analysis results for experimental spectra are compared against those achieved using simulated spectra. A valuable characteristic of this method is its ability to identify sources despite unknown shielding geometries. The successful identification of radionuclides and false identification rates found were excellent for the signal levels involved. For many of the sources, identification performance against experimental spectra was somewhat poorer than found using simulated spectra. The results demonstrate that the PCA-based algorithm is capable of high success rates in identifying sources in short duration, low signal-to-noise NaI(Tl) spectra.