Melt Index Modeling And Optimization Via Sparse Bayesian Learning In Propylene Polymerization Process

Melt index(MI)is an important quality index that determines the grade of polypropylene products.Accurate prediction of the melt index of propylene polymerization can shorten the time of grade switching,save material waste,reduct energy consumption,improve production efficiency and increase product profits.In the industrial production process,the melt index is obtained by manual sampling and offline analysis.It is difficult to meet the requirements of online quality monitoring and control of industral processes.Facing the complicated polymerization reaction mechanism,the disturbance and the noise accompanying in industrial production,it is difficult for the traditional melt index model to obtain good prediction accuracy and robustness.The sparse Bayesian learning method derives the posterior distribution of unknown variables by obtained samples based on Bayesian reasoning.The complexity of the model is reduced by the sparsity constraint,which has good application potential in industrial process quality prediction,especially for small sample problems.Based on the previous research work,this paper proposes a variety of effective melt index prediction models in the sparse Bayesian probabilistic learning framework for the challenges of sample label minority,variable coupling,complex process nonlinearity,chaos characteristic,and time-varying property.This work focus on online intelligent optimal prediction of the melt index of the propylene polymerization process.The main work and innovation of this paper are as follows:1.Considering the coupling problem of propylene polymerization process variables,a melt index prediction model based on t-distribution stochastic neighborhood embedding is proposed in this work.The distribution of observation data in the sample neighborhood is used to reduce the feature dimension of the model.A low-dimensional feature matrix is constructed to overcome the problem of information redundancy caused by high correlation of process operating variables and reducing the impact of disturbance and noise on the melt index prediction in the polypropylene process.By applying it to real data of industrial propylene polymerization,experimental results prove the effectiveness of the proposed model.2.Considering that the industrial production process has a small number of sample labels and a large number of unlabeled sample data,a semi-supervised regression method based on neighborhood kernel density estimation is proposed to make the full use of unlabeled sample information to improve the accuracy of the melt index prediction in the propylene polymerization process.Compared with the traditional melt index prediction methods,the model makes the information integration of unlabeled data under the Bayesian probability framework.The kernel function method is used to estimate the distribution of unlabeled samples in the neighborhood and construct a function with a small number of sample labels.The mapping is derived from the posterior distribution of the melt index by Bayes'theorem.The maximum likelihood estimation of the model parameters is obtained to improve the accuracy of the model prediction,and a sparse constraint is introduced to avoid overfitting.The experimental results of real industrial process data show that the method has better prediction accuracy than the previous melt index prediction models.3.Considering the non-linearity of the complex process of propylene polymerization,parameters directly affect the prediction accuracy of the melt index model.An intelligent melt index prediction model based on chaotic artificial bee colony optimization is proposed in this work.The improved chaos artificial bee colony method was used to optimize the kernel function parameters of the model,and chaos mapping was introduced to enhance the algorithm's convergence ability and optimization efficiency.Then an intelligent optimal prediction model for the melt index of the propylene polymerization process was obtained.The multiple comparison methods are used to investigate the chaotic maps and CABC's structure.Applying this model to the real industrial production process,the results show that the chaotic artificial bee colony Bayesian regression model has good prediction performance and generalization ability.4.Considering the chaotic characteristics of the propylene polymerization process,the melt index time series has a long-range correlation.A semi-supervised prediction model of the melt index based on chaos analysis and co-training is proposed.The chaotic characteristics of the melt index series are analyzed,and the chaos feature matrix the melt index is extracted by phase space reconstruction to establish a prediction model.The co-training method is applied to make full usage of unlabelled data with a chaos-based predictor and a feature-based predictor.A hybrid diffriential evolution bee colony algorithm is further introduced to optimize the melt index prediction model.The experimental results show that the proposed semi-supervised intelligent optimal prediction model based on chaos analysis has better prediction accuracy than other models,and has potential for application in the melt index prediction of propylene polymerization process.5.Considering the time-varying property of the polypropylene industrial process,equipment worn out and disturbutions in operating conditions leading to the mismatch of the melt index model,with an online correct strategy,a melt index prediction method based on particle filtering is proposed in this work.The state transfer equation is constructed for the prediction model,controling prediction error gradually reduced with the iterations,an optimal model is obtained by particle filtering algorithm.The prediction model is further updated in real time by an online correction strategy.Compared with other melt index models,the advantage of this model is to obtain probability results of the model parameters through the state transfer equation,and to update the model in real time according to the prediction performance.This method is applied to the real industrial data,and it proves that the proposed model has good accuracy and robustness in the propylene polymerization process.