Comparative study of analytical and machine learning models for gamma source localization in waste drums

초록

In this study, we compared analytical and machine learning models for two-dimensional localization of radioactive sources within a simulated waste drum. A reflector-coated liquid light guide (LLG) was wound around a drum, and a 60Co gamma source was located at 72 different positions within that drum. Time-difference signals corresponding to each source position were obtained and used to generate time-difference distributions. These distributions revealed that time differences of the source peaks correlated linearly with angular position, while their amplitudes exhibited complex dependencies on both radial and angular coordinates. For 24 arbitrary source positions, an analytical localization model achieved mean absolute errors (MAEs) of 2.2 cm for radial distance and 6.6 degrees for angular position, with an overall position error of 2.8 cm. However, model performance degraded significantly at the LLG intersection region and at small radial distances where signal intensity decreased. To address these limitations, a multilayer perceptron (MLP) neural network was implemented, and MAEs were reduced to 0.6 cm in radial distance, 4.7 degrees in angular position, and 1.1 cm in position error. The use of the MLP neural network leads to a relative reduction of 72.5%, 29.2%, and 60.4% in MAEs, respectively. Additionally, the mean expanded uncertainties (k = 2) were reduced by 76.4%, 73.4%, and 75.6%, respectively. These findings demonstrate that integrating machine learning enhances radioactive source localization accuracy.

키워드