The precise characterization of the size distribution of radioactive aerosol particles is essential for environmental radiation protection. The screen diffusion battery (SDB) method has been extensively utilized for the sizing of radioactive aerosols. Nevertheless, the accurate and rapid inversion remains a significant challenge within this methodology. This study proposes several multilayer perceptron neural network models to tackle this issue. The performance and applicability of models were assessed using simulated and laboratory measurement data. Results showed that the model achieved accuracy of above 89 % at the error noise level of 10 % in the peak shape classification task. For unimodal distribution parameters, the MAPE (Mean Absolute Percentage Error) of AMD (Activity Median Diameter) and (GSD, Geometric Standard Deviation) were maintained below 1.5 % and 3.4 % at noise level of 10 %. For bimodal distribution parameters, the MAPE of all parameters were below 10.5 % at noise of 10 %. These results demonstrate the model's exceptional predictive accuracy and robust noise immunity. Comparison results in radon chamber showed a good linear correlation between AMD (obtained by the SDB system using MLP) and CMD (Count Median Diameter, measured by SMPS) with R2 of 0.888, verifying the practical measurement capability of MLP method. The prediction results can be generated almost instantaneously in the millisecond range, which presents potential for real-time and large-scale measurements. Moreover, compared with the conventional method, this MLP method does not require manual selection of initial iteration parameters and will not produce unstable oscillating solutions. These findings can significantly enhance the efficiency and reliability of radioactive aerosol size distribution analysis, supporting improved environmental monitoring, radiation risk assessment and safety protocols in nuclear facilities.
扫码关注我们
求助内容:
应助结果提醒方式:
