考虑传统动态光散射颗粒粒度分布测量用的反演算法复杂、精度不够、抗噪能力差, 本文基于大数据思想, 提出了一种动态光散射颗粒分布软测量方法。该方法通过调节颗粒粒度分布形状参数获得大量自相关函数及其对应颗粒分布的数据; 使用这些数据对子学习机进行训练。最后, 针对训练数据维数较高的特点对传统Bagging算法进行改进, 并利用改进的Bagging集成算法集成子学习机以提高软测量模型的精度及泛化能力。通过模拟单峰数据和对300 nm标准粒径进行软测量开展了验证实验。结果表明, 该方法能够较好地测量出不同动态光散射颗粒分布的峰值及分布宽度, 模拟单峰数据测量峰值精度可达1 nm, 300 nm和503 nm, 标准粒径测量精度分别可达3 nm和4 nm, 优于一般的反演算法。该软测量方法为动态光散射颗粒分布测量开辟了新的途径。

As the traditional inversion algorithms for particle size distribution measurement by dynamic light scattering show complex computation, lower accuracy and poorer anti-noise capacity, this paper proposes a soft sensing method for particle size distribution based on improved Bagging algorithm by using idea big data. The data of autocorrelation function and particle sizing distribution were obtained by changing the parameters of particle distribution shape. Then the learning machines were trained by the data. Finally, the traditional Bagging algorithm was improved on the basis of the character of high dimensional data. The improved Bagging strategy was used to aggregate the machines for bettering the model accuracy and its generalization performance. A validation experiment was performed by simulating the single peak data and soft sensing for the standard particles with a diameter of 300 nm. Experiment results demonstrate that the proposed method predicts the peak position and the width of particle sizing distribution accurately, and the best accuracy of peak position measurement is 1 nm. Meanwhile, the accuracies for standard particles with diameters of 300 nm and 503 nm are 3 nm and 4 nm, respectively. The proposed method provides a new way for the particle size distribution measurement in dynamic light scattering.