Study On Particle Size Distribution Inversion Of Dynamic Light Scattering Based On Generalized Regression Neural Network

Dynamic light scattering technique is an effective method widely used in the measurement of nano and submicron particles.This method obtains particle size distribution by inverting the electric field autocorrelation function.However,the first kind of Fredholm integral equation needs to be solved in the inversion process,which is a typical ill conditioned problem.Small noise may cause serious deviation of the inversion result.For this reason,various inversion algorithms have been proposed,but as the noise level increases,the accuracy of the inversion results often cannot be guaranteed.With the wide application of neural network,inversion algorithms of dynamic light scattering have been expanded.Among them,generalized regression neural network(GRNN)has the advantages of high accuracy,few parameters,simple structure,easy implementation,etc.,but its calculation accuracy and speed are affected by the training sample generation strategy and the selection method of smoothing parameter.Therefore,the purpose of this paper is to improve the accuracy of inversion results and inversion speed by studying the generation strategy of training samples and the selection method of smoothing parameter,using GRNN as the inversion algorithm of particle size distribution.The main research content includes:1.In order to accurately invert the particle size distribution of dynamic light scattering,an inversion algorithm of particle size distribution based on fixed training samples and leave one out cross validation GRNN is proposed(FT-CV-GRNN).This algorithm uses a fixed training sample generation strategy to generate training samples,leave one out cross validation method to optimize smoothing parameter,and can effectively invert particle systems within a fixed particle size range.The inversion of simulated and experimental data show that compared to Tikhonov,the peak position of the inversion results of the algorithm is more accurate,but it is more sensitive to the inversion of wide and narrow distribution,mainly showing that the inversion results of wide and narrow distribution have larger distribution errors.2.In response to the issue that the accuracy of FT-CV-GRNN inversion results is greatly affected by fixed training samples,an inversion algorithm of particle size distribution with adaptive training samples and smoothing parameter optimization GRNN is proposed(Tikhonov-PGRNN).The algorithm uses Tikhonov to invert particle size distribution as prior information for adaptive generate training samples,and uses one-dimensional search method to obtain the optimal smoothing parameter.The inversion of simulated and experimental data show that compared to Tikhonov,the algorithm has stronger noise resistance and the inversion results are better.3.In response to the problem of redundant training samples and long inversion time caused by human experience affecting the selection of smoothing parameter in Tikhonov-PGRNN,an inversion algorithm of particle size distribution with reducing training samples and adaptively optimizing smoothing parameter GRNN is proposed.The algorithm utilizes an improved Tikhonov based adaptive training sample generation strategy to reduce the number of training samples,and uses the grey wolf optimization algorithm to adaptively solve the optimal smoothing parameter.The inversion of simulation and experimental data show that the algorithm can shorten the calculation time on the premise of ensuring the accuracy of inversion result,especially in the inversion of bimodal distribution.So far,there are few studies on particle size distribution inversion of dynamic light scattering based on GRNN,and the inversion results are usually unimodal distribution and particle size.Therefore,this paper studies particle size distribution inversion of dynamic light scattering using GRNN,which provides references for the subsequent accurate measurement of particle size distribution through dynamic light scattering.