Direction of Arrival Estimation for Radionuclides Based on Neural Networks Approach

A study aimed at enhancing Direction of Arrival (DOA) estimation for gamma-emitting nuclides by employing deep neural networks. The precise determination of DOA is important in Homeland Security (HLS) applications and ensures enhanced safety measures during Decontamination and Decommissioning (D&D) processes. Our approach involves considering the complete energy spectrum and implementing data augmentation techniques on the recorded data. This strategy results in our neural network models surpassing conventional beamforming (BF) methods and demonstrating superior robustness in various background scenarios. The standout performer among our models is the 3-layer Convolutional Neural Networks (CNN), which improves error reduction by up to 40% (for 60 Co and 241 Am nuclides). Traditional approaches, including BF, exhibit inherent limitations regarding accuracy and sensitivity to noise and variations in background radiation. In recent years, the emergence of data-driven techniques, particularly leveraging CNNs and Gated Recurrent Unit (GRU) models, has showcased promise in elevating the accuracy of DOA estimation. This research presents a reliable and data-driven method for precise DOA estimation and opens avenues for its potential applications in nuclear security and advancing safety practices in D&D scenarios.