Research On Soft Sensors For Batch Processes Based On Mode Analysis

Batch process is an important production method,which is widely used in chemical,pharmaceutical,microelectronic.Restricted by sensing technology,some key process variables are difficult to measure directly,which seriously affect the process control of batch processes.Soft sensors realize online measurement of process variables(primary variables)by establishing the model of primary variables and secondary variables,utilizing the online measurement results of secondary variables to estimate the primary variables.Due to multiple switching of operating procedures during the production of batch processes,batch processes have muti-mode characteristics,and there are significant differences in data characteristics between each mode,which influences the accuracy of soft sensors.The existing soft sensor modeling methods for batch processes mostly consider the overall perspective of batch processes without fully considering the differences in data characteristics between each mode.The established soft sensor models are difficult to reflect the relationship of primary variables and secondary variables accurately.These methods ignore the influence of the parameter quantity and risk points on the risk of the soft sensor model,which affects the accuracy of soft sensor model for batch processes;Meanwhile,the existing online estimation methods for primary variables in batch processes do not fully consider the temporal characteristics of batch process data,nor do they fully utilize delayed measurement information with high accuracy and strong correlation with primary variables,which affects the accuracy of online estimation of primary variables.As a result,considering the characteristic of multi-mode batch processes,it is of great theoretical significance to develop soft sensors for batch processes based on mode analysis.Based on the detailed analysis of the characteristic of multi-mode batch processes and process data,this dissertation studies the soft sensor of batch processes based on mode analysis,mainly completing the following research work:1.The multi-mode characteristics of batch processes affect the accuracy of soft sensing modeling and online estimation of primary variables.Based on the mode analysis of batch processes,a mode processing method for batch processes is constructed.Sequential Constrained Fuzzy c-Means(SCFCM)is used to realize offline mode partitioning of batch processes.An online mode identification method for batch processes based on double boundary support vector data description(DBSVDD)is proposed.A mode recognition model for batch processes is constructed using DBSVDD;Then,an online mode identification strategy for batch processes is established,which fully considers the timing characteristic and abnormal data of batch processes.The experimental results show that the constructed mode processing method for batch processes can realize reasonable mode partitioning and online mode identification,laying the foundation for soft sensor modeling and online estimation of primary variables in batch processes.2.The number of parameters and risk points in batch processes affect the accuracy and robustness of soft sensor models,leading to the presence of risks in soft sensor models.A model risk assessment method of hybrid soft sensor model for batch process is proposed.Using BIC to assess the influence of the quantity of samples and model parameters on the risk of soft sensor models,and SRM to evaluate the influence of risk points on the risk of soft sensor models.Based on these,the risk of soft sensor models is reasonably evaluated;A hybrid model structure of soft sensor using weight fusion is presented,and according to the risk assessment results of each model,the value of the weight in the multimode hybrid model is determined,and the soft sensor modeling for multi-mode batch processes is realized.The experimental results show that the hybrid model of batch processes has high accuracy and robustness.3.There is delayed measurement information in batch processes,and using delayed measurement information to correct the estimation results of primary variables can effectively improve the accuracy of online estimation.Therefore,introducing covariance cross fusion into cubature Kalman filter,a state estimation method based on covariance cross fusion-cubature Kalman filtering(CI-CKF)is proposed;On this basis,the CI-CKF method is used to fuse the delayed measurement information into the online estimation of the state variables of the batch processes,so as to realize the online estimation of the batch processes.Numerical simulation and the experimental results of penicillin process show that the established online estimation method of batch processes can fully utilize the measurement information and has high accuracy of state estimation.4.The implementation process of soft sensor for batch processes based on modal analysis is given.Firstly,the soft sensor hybrid modeling method combining model risk assessment is used to build the soft sensor model of batch processes;Then,the online mode identification method based on DBSVDD is used to identify mode of online batch process,and the corresponding model is selected for the online estimation of primary variables.At the same time,the CI-CKF state estimation method is used to integrate the delayed information of batch processes with the established soft sensor model to estimate the state variables online;The experimental results show that the established soft sensor for batch processes can realize online measurement of primary variables for batch processes accurately.The online mode identification method for batch processes in this dissertation can reasonably identify the mode of online data;The proposed model risk assessment method for batch process can reasonably evaluate the risk of soft sensor models;The constructed online estimation method for batch processes can accurately estimate the primary variables;The soft sensor of batch processes based on mode analysis can realize online estimation of primary variables in batch processes,and the estimation results have high accuracy and robustness.