基于Transformer的放射性气溶胶子体扣除算法

Transformer-Based Algorithm for Radioactive Aerosol Sub-Body Deduction

摘要: 在人工放射性气溶胶浓度监测过程中，氡及其子体的信号干扰是一个严重的问题。传统方法在适应环境本底变化方面表现出明显的局限性，这些方法往往难以准确捕捉因环境变化引起的信号时变特性及其在不同时间尺度上的特征变化。本文引入了Transformer架构到气溶胶人工放射性浓度计算方法中，利用其多头自注意力机制和特征融合模块来优化信号处理，有效区分氡子体信号及其变化。同时，模型通过多尺度注意力机制增强了对信号动态特性的适应性和鲁棒性。此外，Transformer架构中的特征融合模块动态调整特征表示可根据输入信号的实时变化，进一步提升监测的准确性。实验结果显示，引入Transformer架构后，模型的整体性能得到显著提升，尤其在复杂和动态的环境中处理信号效果突出，使用该方法不仅提高了检测的准确性，而且有效地降低了误报和漏报率。在区分氡子体信号和人工放射性信号的任务中，Matthews相关系数(MCC)达到了0.85，整体准确率为92%。

Abstract: In the monitoring process of artificial radioactive aerosol concentration, signal interference from radon and its progeny poses a significant challenge. Traditional methods exhibit clear limitations in adapting to changes in environmental background, often struggling to accurately capture the time-varying characteristics of signals caused by environmental changes and their feature variations across different time scales. This paper introduces the Transformer architecture into the calculation method for aerosol artificial radioactivity concentration, utilizing its multi-head self-attention mechanism and feature fusion module to optimize signal processing and effectively differentiate radon progeny signals and their variations. At the same time, the model enhances adaptability and robustness to the dynamic characteristics of signals through a multi-scale attention mechanism. Additionally, the feature fusion module in the Transformer architecture can dynamically adjust feature representations based on real-time changes in the input signal, further improving monitoring accuracy. Experimental results show that after introducing the Transformer architecture, the overall performance of the model is significantly improved, especially in processing signals in complex and dynamic environments. This method not only increases detection accuracy but also effectively reduces false positives and missed detections. In the task of distinguishing radon progeny signals from artificial radioactive signals, the Matthews correlation coefficient (MCC) reached 0.85, with an overall accuracy of 92%.