Process identification and discrimination in the environmental dose rate time series of a radiopharmaceutical facility using machine learning techniques

Multi-facility nuclear sites with research reactors have several environmental area gamma monitors in a network as a part of their surveillance capability. However, the routine release of low levels of 41Ar gas from the reactor is prone to interfere with the recorded gamma dose rate and mask the genuine processes being monitored at the network's central control room. As a potential solution, machine learning techniques have been used in this study to autonomously identify and discriminate the genuine processes, viz., the radioactive consignment loading and its movement, at an interim radiopharmaceutical facility located close to a research reactor. To increase the richness of the recorded univariate dose rate time series, several additional features were created. A labelled dataset of process and non-process dose rate sub-sequences or segments was generated by subject matter experts, based on practical knowledge of the facility, and aided by k-means clustering algorithm. The labelled dataset was used to train several supervised classification models and the random forest class of models gave superior performance. The optimised random forest model was able to identify process sub-sequences with a precision of 82.35% and a specificity of 97.11%. The overall balanced accuracy of the model was 91% with a f1 score of 82%. This machine learning approach proved useful to autonomously identify genuine process driven sub-sequences in the univariate dose rate time series. It has an application in reducing the false alarms at exit portal monitors, especially at those sites where there is a potential for external interference in the monitored dose rate.