A New Bisphenol-A Soft Sensor Technology Based On Multi-models

The object in the process of bisphenol-A production often has complex characteristics,such as multivariate,nonlinear and multiple conditions.It is difficult to describe its process characteristic by single data-driven model,and prediction accuracy of the established model is also poor.It is a difficult problem for improving the prediction accuracy of the model.The application of multi-models modeling method provides an effective way to solve this problem in the soft sensor modeling world.In order to improve the prediction accuracy of the soft sensor model in complex industrial process,this paper mainly studies the soft sensor modeling methods based on multi-models.These multi-models modeling methods improve the prediction accuracy by completing missing sample data and improving clustering methods.This paper compares the single model and the multi-models to the crystal unit in the process of bisphenol-A production,whose data is obtained by the actual industry.The main research results are as follows:(1)Sample data loss in industrial process is widespread and severe.It can increase the number of training samples and improve the prediction accuracy of the established multi-models if the missing data can be complemented effectively.This paper proposes an improved k-nearest neighbor data complement algorithm to the neighbor selection bias problem of missing data in k-nearest neighbor completion algorithm.The improved algorithm assigns different weights to neighbor sample according to the different distances between selected nearest neighbor samples and the missing sample,which effectively solves the bias of k-nearest neighbor completion algorithm in the selection of neighbor.Sub-models are built up by Gaussian process regression after using the k-means clustering method and the final multi-models are calculated by using the "switching" sub-models fusion mode.350 sets of data are trained from C303 crystallization unit during the production of bisphenol-A.Another data sets of 50 groups are used for testing.The simulation results show that the average relative error of completion multi-model is 1.18%,which means it shows higher prediction accuracy compared to the single model and non-completion multi-model.(2)Clustering is an important method in soft sensor modeling based on multi-models.And it can effectively improve the prediction accuracy of multi-models based on the accurate classification of sample data.This paper proposes an improved extended search clustering algorithm,which is aiming at the shortcomings of traditional clustering methods that rely on data space distribution and prior knowledge.The influence of the sample density is taken into account,which is suitable to the distribution of samples in various shapes.In addition,different search distances are used according to the density of each sample point.The threshold is used to introduce different clustering methods for different density points.Sample data is clustered by respectively using improved extended search clustering algorithm,search clustering algorithm and k-means clustering algorithm.Multi-models are built up by Gaussian process regression and fusion mode of “switching” according to the results of clustering.300 sets of data are trained from V304 dissolved tank during the production of bisphenol-A.Another data sets of 50 groups are used for testing.The simulation results show that the average relative error of the soft sensing model based on the improved extended search clustering algorithm is 1.2%,which means it has higher prediction accuracy than the other comparison methods.(3)In this paper,a new clustering algorithm based on shuffled frog leaping algorithm and fuzzy C-means clustering is proposed,which solves the shortcomings of fuzzy C-means clustering algorithm.The optimization mechanism of shuffled frog leaping algorithm is studied firstly,and an improved shuffled frog leaping algorithm is proposed to solve the problems of falling into the local optimum and the not ideal convergence effect.The optimal clustering center is obtained by applying the improved shuffled frog leaping algorithm to fuzzy C-means clustering algorithm.The samples are clustered by using the optimal clustering center.The system output is obtained by “Weight summation” fusion mode after building the corresponding sub-models for each sample subset by the Gaussian process regression.300 sets of data are trained from V304 dissolved tank during the production of bisphenol-A.Another data sets of 50 groups are used for testing.The average relative error of the simulation results show that the clustering soft sensor model based on the combination of improved shuffled frog leaping algorithm and fuzzy C means algorithm is 0.85%,which means the method can improve the prediction accuracy of the model.