Rapid radionuclide identification is a crucial step in preventing the loss, smuggling, terrorist attacks, and radioactive contamination involving hazardous materials. Most current identification methods rely on gamma spectra as the primary analytical tool. However, due to limitations in spectral statistics, these methods require extended processing times to giving results, making them slow, less accurate, and poorly generalized for low-count-rate applications. Emerging radionuclide identification methods now utilize nuclear pulse peak sequence for analysis. However, these methods often fail to fully capture the features of nuclear pulse peak sequence, which limits the identification accuracy.

This study aims to propose a novel radionuclide identification method to overcome the limitations of spectral statistics and enhance the speed and performance of radionuclide identification.

A two-dimensional convolutional neural network (2D-CNN) utilizing nuclear pulse peak sequences was employed in this study. Firstly, four radioactive sources (¹³⁷Cs, ⁶⁰Co, ¹⁵⁵Eu and ²²Na) were used to collect low-count-rate nuclear pulse peak sequences for each single source, mixed sources and environmental backgrounds at varying source distances using a NaI(Tl) detector in the laboratory. Then, the collected sequences were preprocessed through fixed-length segmentation, min-max normalization, and two-dimensional matrix mapping to generate multiple nuclear pulse peak sequence datasets with different matrix sizes. Subsequently, a 2D-CNN model was developed to optimize the convolution kernel size and padding method using 10-fold hierarchical cross-validation to enhance feature extraction from nuclear pulse peak matrices. Finally, radionuclide identification capabilities of the model were tested on datasets with four simple and easily distinguishable sequences, five-category sequences, and eight complex-category sequences, and its performance was compared against three other methods: BPNN+PCA (Back Propagation Neural Network + Principal Component Analysis), SVM+PCA (Support Vector Machine + Principal Component Analysis) and 2D-CNN+spectrum.

The 2D-CNN radionuclide identification results show that with only 300 nuclear pulse sequence points, it achieves an accuracy of 99.61% on four easily distinguishable sequence sets. For five category sequence sets, an accuracy of over 95% is achieved with just 400 pulse sequence points. Moreover, for eight complex category sequence sets, a stable recognition accuracy is attained with 400 pulse sequence points. Additionally, comparative experiments with different models indicate that the 2D-CNN achieves accuracies of 100%, 98.61% and 84.45% for classifying four, five and eight category sequence sets, respectively. This performance significantly surpasses that of the BPNN+PCA, SVM+PCA, and 2D-CNN+gamma spectrum methods, and it also outperforms these models in single-source generalization.

The 2D-CNN model proposed in this study demonstrates feasibility in automatically extracting features from fixed-length nuclear pulse peak sequences. It effectively extracts pulse sequence features within a 40 cm detection range and performs rapid radionuclide identification. This method exhibits advantages in both accuracy and generalization, making it suitable for rapid radionuclide identification tasks with low count rates.