BP neural network based reconstruction method for radiation field applications

Abstract

In order to achieve optimal radiation protection, rapid and accurate reconstruction of radiation field has vital significance in the selection of working paths during the overhaul of nuclear power plants and the decommissioning of nuclear facilities. The radiation field is usually reconstructed by various interpolation methods, but the reconstruction accuracy of such methods is insufficient, With the improvement of AI technology, neural networks have great potential in radiation field reconstruction, but conventional neural networks is prone to local minima and vanishing grandient problem. This paper aims to develop a radiation field reconstruction method based on an adaptive Back-propagation (BP) neural network neural network method with learning rate decay and a corresponding sampling method for multisampling in places where flux gradient changes drastically, and verify its accuracy and feasibility. The proposed method achieves global optimality and avoids vanishing grandient problem by virtue of adaptive algorithm and learning rate decay, ensuring that the radiation field is reconstructed with the smallest relative average error when the sampling point is determined, moreover, the proposed sampling method can greatly improve the accuracy of radiation field reconstruction. The accuracy of the proposed method was tested with three MC simulated radiation fields with simpler cases, and the feasibility of the proposed method was further validated with two MC simulated, more complex and realistic scenes. The results of the proposed method show that the errors of the three test cases are 1.7%, 6.8%, and 7.8%, and the errors of the two validated cases are 8.8% and 7.7%, respectively. The merit of this method was preliminarily verified, further validation is underway to validate its application in real world scenarios.