Abstract: The poor energy resolution of NaI(Tl) detector makes it difficult to accurately identify radionuclides based on the acquired gamma spectrum. In order to improve the identification accuracy, this paper proposed a hybrid PCA-FCN identification model based on the advantages and disadvantages of existing research methods and models. Based on the randomization strategy, a robust gamma energy spectrum datasets was constructed through empirical measurements and Monte Carlo simulations. The datasets was utilized to train the model and conduct validation experiments. Results indicate that the PCA-FCN model achieved radionuclide identification performance factors of 0.982 3 for average precision (A_P) and 0.980 1 for the F₁ score, markedly surpassing the performance of PCA models, FCN models, and traditional full-energy peak analysis methods. The identification accuracy can still be maintained under varying spectral complexities and statistical fluctuations. This conclusion shows the potential of the PCA-FCN model and the random sample generation strategy for future applications in quantitative measurement of radioactivity.

Key words: NaI(Tl) detector, radionuclide identification, gamma spectrum, principal component analysis, fully connected network