EVALUATION OF ARTIFICIAL NEURAL NETWORKS EFFECTIVENESS FOR UNFOLDING GAMMA-SPECTRUM OF 137CS

Development of machine learning methods for spectrum processing is one of the most promising ways for gamma- spectrometry automation and accuracy improvement. Effectiveness of fully connected and convolution neural networks for quantitative γ-spectrometry analysis using scintillation detector NaI(Tl) and lead shielding is presented in the article. Semi-synthetic spectrums were used for the models training; the semi-synthetic spectrums are in channels additions of random spectrums measured at a short duration. The analysis shows advantages of artificial neural networks compare to the common analytical method of spectrum unfolding. The mean square error of activity evaluation is 2–4 times lower than the common method if measuring time is equal to 100 s. In highly standardized conditions of measuring, the advantages of convolution neural networks appear with increasing radiation source activity. Validation with sources not used in training of neural networks has shown fully connected and convolution neural networks can have advantages over the standard method when activity of γ-radiation source is relatively high.