Self-adaptive Pulse Shape Identification by Using Gaussian Mixture Model

Identifying the pulse shape is a crucial step in detecting the contamination of surfaces by α/β radioactive isotopes. In this study, the authors propose a self-adaptive method to identify the pulse shape by applying the Gaussian mixture model and the constant-fraction timing algorithm to a ZnS(Ag) and plastic scintillator to improve its accuracy of detection of low-energy particles, and realize automatic calibration to reduce the operational difficulty of the process and improve its stability. The proposed method comprises a feature extraction algorithm running on an FPGA as well as a classifier and a training machine running on an embedded processor. The feature extraction algorithm forms the core of the proposed method, is independent of the amplitude of the pulse signal, and can extract important information (pulse width and rise time) from the pulse signals to simplify the probabilistic model. The proposed method was tested on a ZnS(Ag) plastic scintillator, and the results showed that the feature extraction algorithm was more robust against errors in the amplitude of the pulse signal than the method that involves using a fixed threshold. The error in its results compared with those of measurements could be controlled to within 4% during mixed radiation field tests involving a single source of radiation. Moreover, we have deployed a complete program to train the model on the platform to support onboard model training and self-adaptive parameter estimation without requiring human intervention.