Intelligent Modeling And Multi-objective Optimization Of Polyester Fiber Polymerization Process

Polyester fiber,also known as polyester,is widely used in daily life because of its excellent textile properties.However,the production mechanism of polyester fiber is complex,there are many chemical reactions and many different production stages,so the quality of polyester fiber products is easily affected by many factors.Therefore,in recent years,how to efficiently produce polyester fiber and improve the quality of polyester fiber products have become the focus of many scholars.The polymerization process is the first stage of polyester fiber production,which plays an important role for the final fiber product.Intently focused on this polymerization process,this dissertation conducted a series of modeling and optimization research by utilizing intelligent modeling,and multi-objective optimization algorithms.The purpose of this dissertation is to fully understand the production situation of polyester fiber polymerization process,to establish the dynamic mechanism of polyester fiber polymerization process,data driven and hybrid model,the real-time predict important performance indexes,including melt viscosity,diethylene glycol content percentage,degree of polymerization,the average molecular weight and the value of the esterification rate based on the mechanism of polyester fiber polymerization process and huge amounts of data collected.The major contributions of this dissertation are:(1)The static mechanism model cannot reflect the change of the concentration of each reactant in real time in the polymerization process of polyester fiber.To solve the problem,this dissertation studied deeply the mechanism,and added time dimension to the static mechanism to establish the dynamic mechanism model in the polymerization process of polyester fiber.The model can not only reflect the relationship between steady process and fiber properties,but also reflect the state switch when the process changes.At the same time,because of the high complexity of the dynamic model,these equations can not be solved directly by conventional methods,a numerical method,called MOL method,is selected to solve this problem.At the end,the dynamic model and the numerical method are verified by comparison experiments.(2)Based on the problems of redundancy and dynamic nonlinearity of massive data collected by DCS,extreme gradient boosting decision tree combing with bidirectional gates long short-term self-attention(WBi CG-LSTM-SEA)algorithm is presented.Firstly,Xgboost is used to select the input variables according to the relation between input and output,and the input variables which are higher correlation with output variables are selected as the input of soft sensor modeling.It then acts as an encoder to weigh the selected input variables based on their importance scores.Next,the encoded input variables are normalized and then sent to the bidirectional converted gates LSTM to extract dynamic features hidden in the process data.Finally,the proposed WBi CG-LSTM-SEA soft senser framework is applied to predict the melt intrinsic viscosity in the polymerization process,and the experimental results verified its effectiveness.(3)To extract the nonlinear,dynamic and spatial correlation features of data collected from DCS in the polymerization process of polyester fiber and handle the instability problem of soft sensor model,a parallel interaction spatio-temporal constraint variational autoencoder(PISTCVAE)soft sensor framework is proposed.In PIST-CVAE model,the constrained variational autoencoder(CVAE)is first proposed.Secondly,a parallel interaction mechanism(PIST)is introduced,which can effectively extract spatio-temporal features from input samples.Finally,a regression prediction is finished by using the low-dimensional nonlinear features extracted from PIST.The effectiveness of CVAE and PIST-CVAE is validated in the polymerization process of polyester fiber.The results show that CVAE is able to reconstruct the inputs with higher accuracy.Meanwhile,the proposed method has a more accurate estimation for the melt intrinsic viscosity quality index than other methods.(4)There are many important quality indexes of polymerization process,and they are interrelated with each other.To deal with these problems,a multi-output soft sensor hybrid modeling method is proposed.In the algorithm,convolutional neural network,long short-term memory and self-attention mechanism are used to assist canonical correlation analysis.And then it is applied for simultaneous estimation of several important quality indexes in industrial polymerization process.The algorithm is divided into three modules,namely nonlinear feature extraction module,regularization module and regression module.Because most methods based on canonical correlation analysis assume the gaussianity of the measured signal and the linear relationship between variables.However,in the actual industrial process,the process data is usually obtained by complex nonlinear mapping.Inspired by this,convolutional neural network and long short-term memory network together constitute a nonlinear feature extraction module.Then,the self-attention mechanism is selected to form a regularization module to prevent model overfitting.Finally,the weighted nonlinear features are sent to canonical correlation analysis to complete the regression prediction.The proposed algorithm is not only tested through the actual polyester plant,but also compared with existing multi-output methods in which statistical analysis and sensitivity analysis are also conducted.The results show that the proposed method has good generalization performance.(5)It contains four conflicting quality indicators in polymerization process.however,most existing literature considers only two performance indexes,which can not fully reflect the real situation of the polyester fiber polymerization process.To solve this problem,an adaptive highdimensional multi-objective optimization algorithm based on projection distance(PARVEA)is presented,and then the four-objective optimization model of the polymerization process is established.In this algorithm,the projection distance is introduced to the original angular penalty distance to accelerate the convergence of the optimal solution.At the same time,the reference vector is updated adaptively based on the distribution of optimal solutions from next generation,which increases the diversity of optimal solutions.Finally,the proposed algorithm is applied to the four-objective optimization problem of the polymerization process and the DTLZ1-DTLZ7 test problems.The results show that compared with other multi-objective algorithms,the PARVEA algorithm not only has a faster convergence speed,but also better diversity of solutions.