Channel attention GAN-based method for gamma-ray full spectrum analysis

Abstract Traditional spectrum decomposition methods face significant limitations when dealing with complex, low signal-to-noise ratio, and low count rate gamma-ray spectra. To overcome these challenges, this study proposes a full-spectrum recognition and inversion approach based on a generative adversarial network (GAN) combined with a channel-attention mechanism. The method aims to enable rapid qualitative identification of unknown nuclides and accurate estimation of their radioactivity. A database of typical single-source nuclides was constructed using Monte Carlo simulations, and a mixed gamma-ray spectra dataset containing up to four nuclides was generated through random combinations for model training. Experimental results demonstrate that the proposed model achieves nearly complete recognition for single-nuclide cases and maintains an average recognition rate of over 90% under complex mixed-spectra conditions. For successfully identified peaks, the counting error remains within 4%, demonstrating the strong quantitative analysis capability of the model. In addition, its generalization performance in peak separation, low-intensity peak detection, and spectral drift correction is systematically evaluated. The results demonstrate that the proposed method provides high recognition accuracy and stability in complex spectrum analysis, providing robust technical support for radiation monitoring and nuclide identification.