Attention-Unet based gamma-ray full spectrum qualitative and quantitative analysis method

Abstract

Rapid full-spectrum analysis of gamma-ray spectra is crucial for public radiation safety. Existing analytical algorithms face challenges in low-count, low-resolution, and overlapping gamma spectra. To address these issues, an improved Attention-Unet neural network method for full-spectrum gamma-ray analysis is proposed, aiming to establish a mapping relationship between the gamma-ray spectra and incident spectra based on the principles of gamma-ray spectra formation, for rapid qualitative and quantitative analysis. The results indicate that, with a gamma-ray branching ratio threshold of 10 %, the Attention-Unet model achieves an accuracy of over 95 % in peak position detection for more than 80 % of the characteristic peaks in the incident spectra output. For correctly identified characteristic peaks, the positive relative error in peak counts is less than 5 %, and the negative relative error is below 10 %. In addition, the model achieves an accuracy of over 97 % in separating overlapping peaks between 276.40 keV, 279.54 keV, 284.31 keV, 356.01 keV, and 364.49 keV, with the maximum separation error in peak counts being 16 %. The model demonstrates a certain degree of generalization and anti-interference capability when applied to characteristic peak mixed spectra and drift spectra that have not been encountered during training. Finally, the ablation experiments of Attention-Unet demonstrated the effectiveness of the Attention improvement. The Attention-Unet approach simplifies the spectral analysis process, providing a technical foundation and research perspective for the application of deep learning methods in spectrum analysis.