Gamma ray spectrum inversion based on master-secondary encoder-decoder network

Gamma-ray diagnosis can detect the energy and spatial distribution of fast ions, as well as identify disruption signs. The detector's response to the gamma-ray spectrum involves complex mappings, requiring a fast and accurate spectrum reconstruction method. The challenge lies in the ill-conditioned nature of spectrum inversion, where errors in measurement can significantly amplify the uncertainties of the inversion results. To solve this, additional information is needed, introducing non-linearity into the problem. Traditional approaches typically rely on iterative algorithms, such as linear regularization, maximum likelihood estimation method (ML-EM), and Gold deconvolution (Gold). Recently, neural networks have gained traction due to their strong capability in handling non-linear and highly ill-posed problems. In this paper, we present a method leveraging a master-secondary network structure that splits the spectrum inversion into two simpler sub-problems, improving outcomes beyond those of a single network. This network structure is verified suitable for solving highly ill-posed inversion problems and applying to gamma-ray spectrum reconstruction. Our method's accuracy is compared to ML-EM and Gold, demonstrating superior stability and effectiveness, particularly under high noise conditions, achieving a level suitable for practical applications. This method has been successfully applied to gamma-ray spectrum detection in the EAST tokamak facility.