基于机器学习的无组织排放颗粒物粒径测量

Particle Size Measurement of Fugitive Emissions Based on Machine Learning

摘要: 针对多波长后向散射的无组织排放颗粒物的粒径测量问题，搭建了基于BP神经网络和粒子群算法(PSO)的粒径预测模型。根据Mie散射理论，在颗粒物粒径1~40 μm范围内，对四种波长激光下的后向散射光强比值进行了仿真，分别应用于两种预测模型进行训练和预测。预测结果显示，BP神经网络预测的尺寸参数D和分布参数k 的相对误差分别在±13%，±6%以内，而PSO预测的相对误差分别在±2%，±4%以内。最后通过平均粒径10 μm的聚苯乙烯颗粒在水悬浮液中的标定实验提取了四波长激光下的灰度值比值，通过测量系统相关系数I₀K₁ 修正后用于两种预测模型进行粒径预测。结果表明，BP神经网络预测耗时较短，但容易出现收敛精度不高等问题，有时不能得到较为准确的反演结果，而PSO预测精度较高，但耗时稍长。

Abstract: Aiming at the problem of particle size measurement of fugitive emission particles with mul-ti-wavelength backscattering, a particle size prediction model based on BP neural network and par-ticle swarm optimization (PSO) is proposed. According to Mie scattering theory, in the range of par-ticle size 1~40 μm, the ratio of backscattered light intensity under four wavelengths of laser is sim-ulated, and applied to the two prediction models for training and prediction. The prediction results show that the relative errors of size parameter D and distribution parameter k predicted by BP neural network are within ±13% and ±6% respectively, while the relative errors predicted by PSO are within ±2% and ±4% respectively. Finally, the ratio of gray value under four-wavelength laser is extracted through the calibration experiment of polystyrene particles with an average particle size of 10 μm in water suspension, and the correlation coefficient of the measurement system I₀K₁ is corrected for the two prediction models for particle size prediction. The results show that the BP neural network prediction takes a short time, but it is prone to the problem of low convergence ac-curacy, and sometimes it can not get more accurate inversion results, while the PSO prediction ac-curacy is high, but it takes a little longer.