Research On Modeling Method For Soft Sensor Of Batch Processes With Irregular Data

Batch processes are the main production methods in the biopharmaceutical,fine chemical,microelectronics,food,and other industries,and are widely used in the production of high value-added products such as medicine,biological products,semiconductor devices.In batch processes,acquiring online measurement data of process variables,analyzing the status of process variables,and implementing online monitoring and optimization control of the production process effectively improves the safe operation and production optimization of batch processes.Restricted by the online measurement technology,there are process variables that are not easy to measure in batch processes,which directly affects the online monitoring and optimal control of batch processes.By establishing a mathematical model between hard-to-measure process variables and easy-to-measure process variables in batch processes,soft sensor can estimate the hard-to-measure process variables online and has become an important approach for online measurement of hard-to-measure process variables.However,because of factors such as operation process,production environment,and measurement instrument performance decay,the batch process data acquired have irregular characteristics,mainly represented by unequal-length batch process data caused by batch switching and multiphase characteristic caused by frequent switching operation stages.If these irregular data are directly used in the soft sensor of batch processes,it will seriously affect the accuracy of the data-driven soft sensor.Meanwhile,the existing soft sensor modeling methods for batch processes usually only consider the non-linearity of the batch process data and do not consider the high dimensional and dynamic characteristics of the batch process data,which reduces the accuracy of the soft sensor modeling.Therefore,it is of great theoretical significance and application value to develop the soft-sensor modeling method for batch processes with irregular data.Based on the detailed analysis of the irregular data characteristics and soft sensor technology of batch processes,this dissertation studies the modeling method for soft sensor of batch processes with irregular data,mainly completing the following research work:1.To solve the unequal-length problem of the batch process data,a Kernel Dynamic Time Warping(KDTW)-based method for batch-process-data synchronization is proposed.This method first projects the unequal-length batch process data into the high-dimensional feature space through the kernel function;then constructs the synchronization performance evaluation index(SPCI)to obtain the optimal kernel parameters and the optimal synchronization path in the high-dimensional feature space;finally,DTW is used to synchronize the unequal-length batch process data.The experimental results of the penicillin fermentation process and the semiconductor etch process show that compared with the DTW method,the synchronization results of the unequal-length batch process data based on KDTW has higher accuracy and noise robustness,which can provide consistent process data for soft-sensor modeling of batch processes.2.To solve the multiphase-data problem of batch processes,a method for partitioning batch processes data based on Sequential Kernel Fuzzy Partitioning(SKFP)is proposed.This method combines information entropy and kernel trick to construct an objective function with nonlinear constraints and noise robustness.By constructing the inner membership and sequential membership,the batch process is divided into phases.Using the proposed evaluation index of phase partitioning to automatically obtain the optimal number of phases.The experimental results of the numerical simulation and the penicillin fermentation process demonstrate that the proposed method can effectively address the sequential partitioning of nonlinear process data.The obtained phase partitioning results have higher accuracy and noise robustness.3.To address the problem of soft sensor modeling of process data with highly nonlinear,high dimensional,and dynamic characteristics,based on improved Just-in-time Learning and Relevant Vector Machine,an adaptive soft sensor,termed IJITL-RVM,is proposed.The IJITL-RVM integrates the IJITL algorithm and the RVM algorithm into a unified online modeling framework with the ability to perform adaptive updating and dynamic modeling.First,to enhance the performance of online prediction,an IJITL is designed to select modeling data based on the support vector data description(SVDD)algorithm and the kernel function.Based on the comprehensive consideration of the strong nonlinearity and high dimensionality of process data,the IJITL can adaptively and accurately select the modeling data.Afterward,a local model is established by using the RVM for online prediction.The experimental results of the numerical simulation example,the UCI datasets,and the penicillin fermentation process suggest that compared with the RVM-based soft sensor and the JITL-RVM-based soft sensor,the IJITL-RVM-based soft sensor has a more stable and accurate prediction performance with higher robustness to noise.4.Based on the irregular data processing and the IJITL-RVM method,a soft-sensor modeling framework for batch processes with irregular data is constructed.First,the KDTW is applied to synchronize the unequal-length batch process data;then the SKFP is used to divide the synchronized batch process data into multiple phases,and an online phase identification strategy based on sequential kernel distance is constructed;finally,the modeling sample dataset based on the adjacent phases is constructed,and the JIITL-RVM is used to establish the soft sensor model to realize the online measurement of process variables.The experimental results of the industrial-scale penicillin fermentation process and industrial-scale chlortetracycline fermentation process show that the proposed soft sensor model for the batch process with irregular data effectively realizes the online measurement of process variables with high accuracy.In the research of the modeling method for soft sensor of batch processes with irregular data,the proposed irregular data processing method can provide a new way for the irregular data processing of batch processes;the proposed soft-sensor modeling method can provide an effective and highly accurate modeling method for soft sensor with nonlinear,high-dimensional and dynamic feature data,and can be extended and applied to the field of regression prediction in machine learning;the proposed soft-sensor modeling method based on irregular data processing can not only provide theoretical support for improving the accuracy of soft sensor of batch processes but also promote the industrial application of soft-sensor methods and technologies of batch processes,which will be used in biopharmaceuticals,fine chemicals,microelectronics,food,and other industries.