Tiny machine learning with convolutional neural network for intelligent radiation monitoring in nuclear installation

Abstract

This study focuses on developing an intelligent radiation monitoring system capable of operating on a low-power single-board computer (Raspberry Pi) for deployment in remote monitoring stations within nuclear facility environments. The proposed system utilizes a radionuclide identification method based on tiny machine learning (TinyML) with a convolutional neural network (CNN) architecture. The radionuclide dataset was acquired through measurements of standard radiation sources, with variations in distance, exposure time, and combinations of multiple sources-including Cs-137, Co-60, Cs-134, and Eu-152. The radiation intensity data from detector measurements were structured into a response matrix and subsequently converted into a grayscale image dataset for model training. Keras is used to design and train machine learning models, while Tensor Flow Lite is used to model size reduction. Experimental results demonstrate that the developed model achieves an accuracy of 99.338% for Keras model trained on computer and 84.568% after deployment on the Raspberry Pi. Furthermore, this study successfully designed and embedded the TinyML model into an environment radiation monitoring system at the PUSPIPTEK nuclear installation.