Files
Abstract
The performance of existing methods for the digital and analog discrimination of alpha and beta particles in mixed radiation fields is examined for signals produced using a novel ruggedized scintillator. Leveraging frequency analysis of simulated signals, an experiment has been designed and used to obtain a representative sample of time domain electronic pulse data. The experiment was conducted at the Triangle Universities Nuclear Laboratory, where emissions from radioactive elements Americium 241, Plutonium 239, Cesium 137, Cobalt 60, Polonium 210, Strontium 90, and Thallium 204 were used to collect a diverse set of energies for alpha, beta, and gamma particles. Statistical signal processing techniques are utilized to determine optimal parameters for use in rise-time, pulse gradient, and time over threshold pulse shape discrimination techniques. Support vector machines, and k nearest neighbor classifiers are used to extend the discrimination performance by utilizing rise-time, pulse gradient, and time over threshold coupled with the peak height of the event. Rise-time discrimination is found to be 98.95 percent accurate for beta particles, and 99.29 percent accurate for alpha particles. Pulse gradient discrimination is found to be 99.92 percent accurate for beta particles, and 99.98 percent accurate for alpha particles. Finally, time over threshold discrimination is found to be 94.72 percent accurate for beta particles, and 99.50 percent accurate for alpha particles. Using classification methods and peak height, most methods produce an additional 1-2 percent accuracy in all cases.All of the analysis is repeated after artificially down sampling the available data to mimic 50 MHz, 100 MHz, 200 MHz, 1 GHz, and 2 GHz sampling rates. Under this treatment, the best consistency is obtained with pulse gradient discrimination, which maintains greater than 99 percent accuracy even at 50 MHz, compared to rise-time, which falls to under 50 percent accuracy, and time over threshold which falls to under 80 percent accuracy at 50 MHz. The dataset is made available for more advanced study, and the gamma data is also available for researchers to attempt beta/gamma or alpha/gamma discrimination. While more features can be found for the pulses, and more advanced techniques can be used to obtain higher accuracy metrics, the benefit of the analysis completed here is its transferability to embedded and computationally constrained systems. Each method used here, with the exception of k nearest neighbors is capable of classifying an incoming pulse in a handful of clock cycles, compared to more complex systems that require more power and more computation time.