Modeling And Monitoring For Industry Processes Based On Just-in-time Learning

In order to improve the product quality,ensure the safe and efficient operation of the processes,it is essential to measure the key quality variables reflecting the product performance and mine valuable process information from massive data to realize the online monitoring and control of the industrial processes.Due to the severe measuring environment,expensive measuring instruments and other factors,the key quality indicators of industrial processes are usually difficult to measure online or there is a serious lag in the measurements.In this situation,by analyzing the relationship between process variables and key quality variables,soft sensor technology has been utilized to deduce the mathematical model between them and perform the prediction of the hard-to-measure key variables based on the easy-to-measure process variables.Unlike the key quality variable which is difficult to measure online,a large amount of process data can be easily sampled and collected with the rapid development of data acquisition,transmission,storage and processing technology.In this way,data-driven monitoring technology based on the collected process data is an effective means to ensure the safe and efficient operation of the process.In the past decades,soft sensor modeling and process monitoring technology have not only received wide attention in academia,but also been widely applied in practical industrial processes.Due to its large internal structure and complicated reaction mechanism,industrial processes often exhibit many characteristics,such as multimodalities,dynamics,and nonlinearities,etc.Therefore,in this thesis,soft sensor modeling and process monitoring for complex nonlinear processes are studied based on just-in-time learning through in-depth analysis of complex characteristics of industrial processes and data.Our research works are listed as follows:(1)A probabilistic modeling method based on just-in-time learning is proposed for nonlinear processes in which the input variables of training and test data contain missing values.In this probabilistic just-in-time learning(P-JITL),Bayesian Gaussian Process Latent Variable Model is firstly utilized to derive the variational distribution of the latent variables when there are missing values in process data.Then,the symmetric Kullback-Leibler divergence is applied to measure the dissimilarity between two probability distributions for selecting the relevant samples in JITL framework.Finally,a nonlinear regression model between the mean values of the variational distribution for the latent variables and the output variables is established.The proposed P-JITL framework can effectively deal with the missing values in input data and select the relevant samples more accurately.(2)Because the single-layer nonlinear latent variable model can only extract limited process characteristics,a Deep Gaussian Process model based on stacked multi-layer Bayesian Gaussian Process Latent Variable Model is applied to mine detailed features and highly nonlinear ones hidden in process data.Instead of only selecting the highest latent information for soft sensor modeling,different process data characteristics embedded in each hidden layer are fully exploited.Based on the symmetric Kullback-Leibler divergence,several local regression models are established based on the latent variables from each hidden layer and the output variable.At last,the prediction capabilities of these local models are integrated to make predictions for the output values of the test data.The proposed method can extract richer nonlinear process information and takes into account the randomness of process data.(3)For multimodal industrial processes,a data preprocessing method is put forward to eliminate the multimodal data characteristics that affect the monitoring performance.Firstly,the process data matrix is augmented based on the serial correlation of process variables.Then,the process operating modes are divided by some kind of pattern recognition method,and the differences of covariance structures of different modal data are eliminated by zero-phase component analysis.In the way,the influence of linear correlation between variables on sample similarity calculation is reduced,and the multi-modal process data follows an approximate single-modal distribution.Finally,a fault detection model for nonlinear processes is established based on Auto-associative Kernel Regression method,and the square prediction error statistic is further modified to improve the detection ability of the model for small disturbances.(4)A fault detection method based on adaptive Auto-associative Kernel Regression is proposed for process data containing outliers.Firstly,robust pre-whitening of the training data set containing outliers is considered to improve the accuracy of sample similarity.In order to reduce the influence of outliers on model parameter estimation and avoid fault data with high similarity to outliers being misjudged as normal ones,a truncation function is constructed to exclude outliers from data reconstruction calculation of fault samples.Finally,in order to further improve the detection ability of the model for small disturbances,all residual variables involved in the construction of squared prediction error statistic are weighted according to the fault information they contain,and the new monitoring statistic of the model is obtained.