Soft Sensing For Complicated Nonlinear Processes Based On Just-in-time Learning

The online process monitoring and control have been playing a more and more important role in improving process productivity and ensuring process safety,which are heavily dependent on the measurement of key process quality variables.In many situations,these key variables may be very difficult to measure online due to reasons like severely measuring environment,expensive analyzers and large measuring delays.In this way,soft sensing has been developed to predict the hard-to-measure key variables by those easy-to-measure process variables.In the past decades,soft sensor modeling has become a hot topic in process control fields.Also,a lot of successful applications have been reported in industrial applications.With complicated structures and complex physicochemical reactions,most industrial processes are intrinsically characterized with nonlinearities,time-variants,dynamics,etc.In this thesis,we mainly focus on the soft sensor modeling for complex nonlinear processes based on just-in-time learning.Our research works are described as follows.1)A new just-in-time learning framework by taking the output information for sample selection and feature representation is proposed for nonlinear process soft sensing.Most traditional sample selection techniques only consider the input information of samples,without any reference to the output information.In this paper,a new sample selection method in the supervised latent structure is proposed,in which the latent variables are highly related to the output.In this way,relevant samples can be selected more correctly for local modeling.After that,to enhance the performance of JITL local model,both sample relevance and variable relevance are taken into consideration for feature extraction.Thus,double locally weighted principal component regression(D-LWPCR)is proposed to extract output-related feature for soft sensor regression model.2)A JITL-based locally weighted kernel principal component regression(LWKPCR)is proposed to solve the nonlinear and time-variant problems of industrial processes.First,the original input variables are mapped into a high-dimensional nonlinear space to deal with variable nonlinear problems.Then,feature representations are carried out in this high-dimensional nonlinear space.By selecting relevant samples that have the most similar pattern with the query sample and taking their similarity for weight assignment,the locally weighted technique is used for dealing with process nonlinearities and time-variants simultaneously.By combination of kernel trick and weighting technique,LWKPCR can provide better prediction accuracy for output prediction.3)A spatio-temporal adaptive soft sensor framework,which is based on the moving window and just-in-time learning techniques under the time difference framework,is proposed for soft sensing of nonlinear and time-varying processes with variable drifting.To reduce the influence of variable drifts,time difference model is used to calculate the first-order time differences of inputs and outputs and build regression model on them.However,global time difference model cannot tackle the process nonlinear and time-varying problems.In this way,temporal adaptive moving window technique is used to track the newest process state and spatial adaptive locally weighted partial least squares is used to deal with process nonlinearity.4)For nonlinear processes with random noises and uncertainties,a weighted probabilistic principal component analysis(WPPCA)is proposed for feature representation.To take the random noises and data uncertainties into consideration,probabilistic latent variable modeling is necessary for feature extraction.Moreover,by utilizing sample weights into PPCA,WPPCA can extract nonlinear features for these processes.Hence,WPPCA-based soft sensor can improve the prediction accuracy.Since the proposed WPPCA cannot guarantee that the extracted feature are output-related,weighted probabilistic principal component regression(WPPCR)is further proposed.In WPPCR,the latent variables are determined by both input and output variables.Thus,the extracted feature are more related for output prediction.5)Two soft sensor approaches,probabilistic JITL(P-JITL)and semi-supervised JITL(S-JITL)are proposed for nonlinear processes with missing data in inputs and nonlinear processes with limited output labels,respectively.In P-JITL,variational Bayesian principal component analysis(VBPCA)is first utilized to handle missing values and extract Gaussian posterior distributions for latent variables.Then,symmetric Kullback-Leibler(SKL)divergence is creatively employed to measure the dissimilarity of two distributions for relevant sample selection in the JITL framework.At last,a nonlinear regression model,Gaussian process regression(GPR),is carried out to model the nonlinear relationship between the output and the extracted latent variables.In this way,the proposed P-JITL is capable of dealing with missing data and selecting relevant samples more accurately.The novel S-JITL framework is based on semi-supervised weighted probabilistic principal component regression(SWPPCR).In the S-JITL framework,traditional Mahalanobis distance and a new proposed scaled Mahalanobis distance are used for similarity measurement and weight assignment.By selecting the most relevant labeled and unlabeled samples and assigning them with corresponding weights,a local SWPPCR can be built to estimate the output variables of the query sample.6)A novel weighted linear dynamic system(WLDS)is proposed for feature extraction in nonlinear dynamic processes.For these processes,data relationships are represented by a first-order Markov chain of latent state equation and state emission equation.In order to approximate the two nonlinear relationships,two kinds of weights are proposed for local linearization of the nonlinear state evolution and state emission relationships in WLDS.Then,a weighted log-likelihood function is designed for parameter estimation using expectation-maximization algorithm.Thus,WLDS can extract the nonlinear dynamic features for subsequent output prediction.