Soft Sensor Development Considering Process Time-delay Estimation

Chemical industrial plants have exhibited significant strong nonlinearity and time-varying behavior.In order to implement efficient monitoring and control strategies,soft sensors have been widely applied to estimate product quality-related variables that are difficult to measure online(i.e.primary variables).Nowadays,with the increasing complexity of process working conditions,much higher requirements have been accordingly put forward by the industrial circles on improving the reliability and accuracy of soft-sensing technology.In the real processes,the acquisition of the primary variable is often restricted by device cost,instrument reliability or technical bottlenecks etc.,which shows great measurement delay.Although soft sensor modeling field has been moving towards adaptive era,the time-delay information is unfortunately overlooked in the modeling procedure.Aiming at further improving the prediction accuracy of the traditional soft sensor methods,this thesis not only deals with the time-varying and nonlinearity characteristics of industrial processes,but also considers the time-delay property hidden in the modeling dataset.On the basis of current soft sensor research results and the time difference Gaussian process regression algorithm,several adaptive soft sensor approaches that involve time-delay estimation are studied in this thesis.The main contents of the thesis are listed as follows:1.Considering the problems of timing mismatch and variable drift of modeling data,a time difference Gaussian process regression(TDGPR)soft sensor is proposed based on fuzzy curve analysis(FCA).The time-delay parameters are achieved through offline estimation and then utilized for variable sequence reconstruction.When new query samples are available,the TDGPR model is employed for online prediction of the current primary variable value.2.Aiming at the aging issue of traditional global time difference(TD)model,an adaptive modeling method is proposed based on selective ensemble of local time difference Gaussian process regression(LTDGPR)models.First,the data for modelling is reconstructed by time delay and dynamic information extracted from the database.Then,an adaptive localization step based on statistical theory is taken on the time-differenced reconstructed dataset,and LTDGPR model set can thus be obtained.For the new input sample,the primary variable dynamic drift prediction value with certain time difference is achieved through a selective ensemble of LTDGPR models which have strong generalization ability.Finally,the online primary variable prediction result is acquired on the basis of TD model theory.3.In view of the stage characteristics of process nonlinearity and time-delay feature,a method of moving window time difference Gaussian process regression(MWTDGPR)is proposed based on local time-delay reconstruction(LTR).The method characterizes in that the moving window method and the TD strategy are combined to adapt to both gradual and abrupt changes in local process characteristics,and during this procedure the mapping relationship on the local window are corrected and reconstructed simultaneously,which provides a feasible framework for other modeling problems existing in nonlinear time-varying processes with time-delay.The validity and accuracy of the above research work are further verified through data simulations of the real industrial processes.The simulation results adequately indicate that adaptive soft sensor modeling considering time-delay estimation is of great significance for the economic benefits,safe and stable operation of chemical processes.