Research and Development of Source Term Activity Reconstruction System Based on Deep Learning

Abstract

The radiation damage weakens the efficiency of decommissioning nuclear facilities, and the source term is hard to be located and measured because it often distributes in or inside objectives. In practice, the hot spot and source intensity are measured by the gamma camera then the radiation field is estimated by particle transport algorithm as a reference in scheme planning. This technology is expensive in measuring and simulating. Hence, this paper presents a source term activity reconstruction (hereinafter referred to as STAR) method based on deep learning to solve this problem. Firstly, a UNET liked framework is constructed to establish the correlation between the radiation field and source activity. Secondly, a source activity reconstruction problem is used to validate the framework. Then we deploy it to the Raspberry Pi with a γ-ray detector to verify it can be applied in practical works. The verification results show that Raspberry Pi can complete source term activity reconstruction in a few seconds without consuming too much computing resources. In framework validating, 5000 samples consist of randomly generated activity distribution and its grid dose value which is calculated by the Monte-Carlo program. The results show that the average reconstruction error is less than 15% and the trained framework is performing well in Raspberry Pi. This method reduces the requirements for instruments and the dose detection is parallelable. Therefore, it can be widely used in nuclear facility decommissioning to improve the efficiency of source reconstruction.