Bayesian Shapelet Learning for Improving the Efficiency of Mobile Radiation Detection Systems

IEEE Nuclear Science Symposium and Medical Imaging Conference (NSS/MIC), pp. 1–6

Abstract

Mobile radiation detection requires large volume detector equipment and high sensitivity. The paper presents a shapelet extraction method that is applied to the time series spectral data of such a mobile search system. The detection capability of a system is often limited by varying background effects. With the technique, time pattern are analysed to find those shapelets of greatest discriminative potential for differentiating background and source. Previous literature results already indicated the advantage of taking the temporal behaviour of spectra into account. Several influencing factors are identified, including the velocity of the vehicle to be of importance for the selection of the shapelet size. A physics-informed autoencoder is used to decide for the presence of the shapelet. The actual discriminator build on top of the shapelet transform is a Bayesian network classifier. It takes the shapelet distance metrics as input and links detection candidates to the respective conditional probability that indicates whether a real source was present or not. The algorithm is build in an online applicable way, in order to provide quick reaction times especially when the system is in motion. Results are presented based on the acquisition with two large volume 2x4x8 NaI:Tl detectors. The detectors were mounted in a car to determine a varying background test profile. On top of this background different sources were tested including a Ra-226 pass, yielding promising results in improving the detection efficiency of the system.

Authors 3

1. Marcus J. Neuer Aachen Data Mining and Machine Learning

   RWTH Aachen University

   Affiliation as printed

   InnoRIID Gmbh,Dsseldorf,40225

   Data Mining and Machine Learning, RWTH Aachen University, Aachen

   FHDW Mettmann, Mettmann

2. Elmar Jacobs
   Affiliation as printed

   InnoRIID Gmbh,Dsseldorf,40225

3. Christian Henke
   Affiliation as printed

   InnoRIID Gmbh,Dsseldorf,40225