Research On Dynamic Soft Sensor Modeling Of Complex Chemical Process

The key variables of chemical process,such as important process or product quality parameters,are generally eager for online real-time estimation owing to the needs of closed-loop control and real-time optimization,but it is difficult to measure them through hardware sensors due to economic or technical reasons.The development of soft sensor technology provides support for the realization of this demand.A large amount of process data is stored in real time according to time stamp,which provides raw materials and driving force for data-driven soft sensor modeling.Chemical processes often show complex characteristics such as large delay,nonlinearity and time variation.How to effectively describe the nonlinear dynamic characteristics of chemical process modeling objects to build dynamic soft sensor models has become an important topic widely concerned by researchers.Aiming at the core issue of dynamic soft sensor modeling of key variables in chemical process,this paper focuses on the research theme of "feature analysis of process data of modeling objects → construction of functional latent variables in modeling space → selection of dynamic nonlinear feature characterization methods → mining of key dynamic feature parameters and expansion of unlabeled data → exploration of data and knowledge fusion strategies",and aims to improve the prediction accuracy and stability of soft sensing models,A set of dynamic soft sensor modeling method is proposed,which is verified by public simulation data experiment and industrial field data application.It mainly includes the following aspects:Firstly,aiming at the characteristics of process noise interference,frequent dynamic changes,and serious multicollinearity that are common in chemical processes,a nonlinear auto regressive dynamic soft sensor modeling method based on nonlinear principal component feature extraction and wavelet denoising is studied.An online nonlinear autoregressive dynamic neural network modeling method with external input,which has the function of weight suppression and noise filtering,is proposed.Firstly,nonlinear feature information of sample data is extracted by kernel principal component feature transformation to the maximum extent,and multiple collinearity among auxiliary variables is eliminated.Secondly,wavelet denoising is used to filter the noise contained in each principal component information,thus simplifying the complexity of the data set.Then,a nonlinear autoregressive model integrating external input information is used to model dynamic soft sensor based on sliding time window.Finally,the model was applied to the industrial field data application verification of solvent dehydration process in PTA plant after being verified by Tennessee Eastman process simulation data experiment.Secondly,a new idea of dynamic soft sensor modeling is sought for the hidden danger of gradient dispersion and gradient explosion in traditional dynamic neural network based on gradient learning method.A dynamic soft sensor modeling method based on improved orthogonal sparse echo acoustic state network is proposed.Drawing on the idea that the state space equation characterizes the dynamic characteristics through state recursion,the classical echo state network has the excellent characteristics of both dynamic recursion and nonlinear characterization.Focusing on the inherent reservoir ill posed problem of the classical echo state network,an effective method to overcome the multicollinearity of the classical echo state network and improve its robustness is studied.The verification of Tennessee Eastman process simulation data and industrial field data of propylene polymerization process shows that the indirect orthogonalization and sparsity of the reservoir can effectively enhance the prediction accuracy and stability of the model on the premise of improving the operation efficiency of the model.Thirdly,for the fact that dynamic neural network or based on feedback recursive structure or based on state recursive update to depict the nonlinear dynamic characteristics of the modeling object but ignore the problem of characterizing the key dynamic characteristic parameters such as the delay time of each auxiliary variable relative to the dominant variable and the time of the transition process in the process of dynamic soft sensor modeling,we focus on the in-depth analysis of the generation mechanism of the dynamic characteristic parameters among the key variables of the chemical process and research appropriate parameter mining methods.An improved distributed differential evolution intelligent optimization algorithm,which is verified by CEC benchmark function to have faster convergence speed,higher convergence accuracy and more robust convergence stability than other intelligent optimization algorithms,is proposed to meet the requirements of dynamic characteristic parameter mining among key variables of chemical process.Fourthly,to solve the problem that the simple data driving can only mine the potential correlation and causality between the operation modes and variables contained in the data,but can not effectively integrate the empirical knowledge in the modeling object field,a dynamic soft sensor modeling method based on data and knowledge fusion is proposed.Firstly,the threshold setting method of cumulative mutual information is proposed to screen auxiliary variables subjectively and objectively.Secondly,the key dynamic parameters are optimized and the unlabeled data within the interval of sparse sampling time is further mined based on the improved differential evolution algorithm.The propylene polymerization production process is taken as a representative to study the process mechanism,and the dynamic soft sensor model is constructed by analyzing various influencing factors related to the dominant variable melt index and selecting the key process variables for necessary mechanism transformation.Finally,the proposed method is applied to the industrial field data of propylene polymerization process in an actual polypropylene plant for application verification and obtained more accurate and more stable prediction results.