An adaptive smoother for counting measurements
Abstract
Counting measurements associated with nuclear instruments are tricky to carry out due to the stochastic process of the radioactivity. Indeed events counting have to be processed and filtered in order to display a stable count rate value and to allow variations monitoring in the measured activity. Smoothers (as the moving average) are adjusted by a time constant defined as a compromise between stability and response time. A new approach has been developed and consists in improving the response time while maintaining count rate stability. It uses the combination of a smoother together with a detection filter. A memory of counting data is processed to calculate several count rate estimates using several integration times. These estimates are then sorted into the memory from short to long integration times. A measurement position, in terms of integration time, is then chosen into this memory after a detection test. An inhomogeneity into the Poisson counting process is detected by comparison between current position estimate and the other estimates contained into the memory in respect with the associated statistical variance calculated with homogeneous assumption. The measurement position (historical time) and the ability to forget an obsolete data or to keep in memory a useful data are managed using the detection test result. The proposed smoother is then an adaptive and a learning algorithm allowing an optimization of the response time while maintaining measurement counting stability and converging efficiently to the best counting estimate after an effective change in activity. This algorithm has also the specificity to be low recursive and thus easily embedded into DSP electronics based on FPGA or micro-controllers meeting "real life" time requirements.