APPLICATION OF ONE-DIMENSIONAL CONVULSIVE NEURAL NETWORKS FOR IDENTIFICATION OF WEAK RADIOACTIVE SIGNALS IN DYNAMIC MONITORING SYSTEMS

Abstract

The paper addresses the urgent problem of automated detection of low-activity moving sources of ionizing radiation in dynamic radiation monitoring systems. The limitations of classical threshold-based signal processing algorithms, characterized by low sensitivity and high false alarm rates under low signal-to-noise ratio conditions, are analyzed. A novel method for identifying radiation anomalies based on a One-Dimensional Convolutional Neural Network (1D CNN) is proposed. A mathematical model and software algorithm for generating synthetic time series of a scintillation detector have been developed, taking into account the stochastic nature of radiation detection (Poisson statistics), vehicle passage geometry, and shielding effects. The developed neural network model was trained and tested on the generated datasets. The effectiveness of the proposed approach has been experimentally confirmed: the classification accuracy for medium-activity sources reached 98.5%, and for extremely weak signals visually indistinguishable from noise, it remained at 83.5%. The obtained results demonstrate the potential of applying deep learning methods to enhance the reliability of radiation security systems.