Comparative analysis of multi-nuclide spectrum recognition methods based on neural network

To address the limitations of traditional peak-finding approaches in complex radiation environments, this paper conducts a systematic comparative analysis of multi-nuclide energy spectrum recognition technology based on neural network methods. Eight nuclides from the industrial nuclide library are selected for the study. The energy spectrum dataset is constructed through the Monte Carlo simulation method, and the performance of four typical neural network models (BP, CNN, ResNet and LSTM) in nuclide identification is compared. The experimental results show that the ResNet model exhibits optimal performance under both expected label formats (1 × 8 and 1 × 1024). Its nuclide recognition accuracy rate is as high as 87.6 %, the average error of relative activity prediction is the lowest at 0.14, and it significantly outperforms other models in weak peak recovery and anti-interference ability. CNN and BP models perform second best in complex tasks, while LSTM models have relatively limited performance due to the indistinct characteristics of energy spectrum sequences. In addition, ResNet demonstrated excellent stability in both high and low activity ranges, verifying its practical application potential in complex radiation fields. This study provides a reference for model selection in the field of nuclide identification and promotes the optimization of neural networks in energy spectrum analysis.