KALMAN FILTER FOR TRANSIENT EVENT DETECTION IN RADIATION DATA

In unattended radiation monitoring applications, automated real-time data analysis is used to detect anomalous events. For bulk count rate data, this analysis is typically done by comparing the present count rate with some calculated background rate. If the present count rate varies from the background rate by some statistically significant margin, or if it exceeds some set threshold rate, then an event is logged and possibly acted upon. The calculation of an appropriate background rate is often done with a “running average” technique designed to allow for gradual changes in the background (e.g., resulting from diurnal effects) while still catching anomalous radiation events. This paper discusses some of the weaknesses of the running average technique, focusing particularly on the delayed background effect and the reliance on unphysical tuning knobs in the algorithm. An alternative to the running average technique is presented. It uses a pair of constantmodel Kalman filters to approximate the radiation count rate data as well as a local background. The Kalman-based technique is shown to be free of delay and to have no unphysical knobs. This technique is also examined from the perspective of implementation in embedded electronics. Comparisons are made between the Kalman filter system and the running average system on the metrics of memory usage, execution speed, and sensitivity.