High-Sensitivity Cesium Aerosol Sensing for Nuclear Severe-Accident Monitoring by Integrating Laser-Induced Plasma RGB Imaging with Convolutional Neural Network (CNN) Analysis

Rapid identification of radioactive cesium (Cs) aerosols is important for nuclear severe-accident monitoring. However, conventional analytical methods often require extended measurement times or complex instrumentation. Here, we present a sensitive and compact sensing approach that integrates laser-induced plasma RGB imaging with a convolutional neural network (CNN). A detection configuration was established in which laser irradiation generated plasma from Cs-containing aerosols within a flowing gas stream, and the resulting emission was directly captured using a CMOS camera. Rather than resolving individual emission lines, the CNN learned subtle Cs-dependent variations in the spatial and RGB intensity distributions of the plasma images. To simplify training under limited-data conditions, the model was designed to address a binary classification task, distinguishing Cs-negative conditions (normal) from Cs-positive conditions (abnormal). Predictions from 100 consecutive laser shots were aggregated to provide a sensing decision within 5 s. Using a criterion requiring Cs-positive classification in at least 99% of independent measurements, the operational limit of detection (LOD) was determined to be 0.03 μg/m3, approximately one order of magnitude lower than values reported for the closest comparable laser-based cesium aerosol measurements. These results demonstrate that plasma RGB imaging combined with a CNN algorithm can provide a compact, highly sensitive, and real-time platform for cesium aerosol monitoring.