Cognitive R-Tree for stabilizing temperature and load induced gain shifts of scintillation detectors

A stabilization concept based on a self-learning R-Tree index method is presented and demonstrated with measurements from a 1.5×1.5 cerium bromide detector. The concept uses a cognitive filter, a digital filter for nuclear signals that continuously updates itself to the current temperature by adjusting the filter components. The R-Tree combines the information from this cognitive filter together with (a) data about the temperature gradient, (b) the current load on the detector in terms of counts per second and (c) the current gain shift, which is determined from the spectrum. This technique consequently belongs to the so-called supervised learning algorithms, because the source is known in advance. The method is characterised by two operational phases. First a training in an industrial grade climate chamber and with selected strong radiation fields are conducted, which is a common procedure for producing spectroscopic equipment, building a base set of data points in the R-Tree. Second, the R-Tree learning does not stop here. It continues during the whole instrument lifetime. Each time a manual calibration is launched with a known (pre-selected) source, all data for adding new training information is available and the R-Tree is updated. The instrument learns while being in the field. Tests with a cerium bromide and a sodium iodide detector are shown for a prototype system and for a complete commercial radio-isotope identification device. Limits of the stabilization are determined.