Dimension-Reductional Mapping Based Industrial Process Modeling And Monitoring

Industry 4.0 and "Made in China 2025" bring more opportunities and challenges to the modern industry.The modem industrial automation systems are becoming more and more complex,informatic and intelligent.Process modeling and monitoring is the key technology for the modern industry to ensure process safety,improve product quality,and reduce energy consumption and pollution.The existing traditional data-driven modeling and monitoring methods are based on some simple assumptions(such as independent identical distribution,linearity,and steady condition),so that they are multifacetedly challenged by the complexities of the practical industrial processes,which can be systematically concluded as data related complexity and variable related complexity.The former one mainly refers to outliers,missing values,multi-rates,and big data volume;the latter one mainly refers to cross-correlations,nonlinear relationships,constrained relationships between process and quality data,and time-varying characteristics.From the perspective of manifold learning technique,this dissertation proposes a systematics set of dimension-reductional mapping based industrial process modeling and monitoring,aiming at the above two issues.The main research area includes the following six parts:(1)Due to the nonlinear relationships and cross-correlations among variables in production quality prediction,the self-learning kernel regression(SLKR)model is proposed,which also considers the constrained relationships between process and quality data.Different from tradition kernel-based regression methods,which need to artificially select the form and parameter of a kernel function,the SLKR model can self-learn a kernel space from specific training data through a specially designed semi-definite programming(SDP)problem.In the learned kernel space,the manifold structure of the input variables is guaranteed to be unfolded for the dimension reduction,while the regression relationships between the projected input variables and the output variable are maximized.The SLKR model based quality prediction has a high prediction accuracy.(2)The information of local proximity is more likely to disclose the intrinsic relationship and feature among data.Traditional data-driven quality prediction models are constructed to define the outer shape of the data using generalized models.They do not provide any insight into the local proximity that relates the data samples.Due to this,the neighborhood preserving embedding regression(NPER)model is proposed,which is constructed from local information.Further,the sparse neighborhood preserving embedding regression(SNPER)model is proposed to ensure the sparsity of regression model when facing P>N data,utilizing the elastic net regularization(EN).Finally,inspired by the just-in-time learning,the local-weighted sparse neighborhood preserving embedding regression(LW-SNPER)model is proposed to deal with the time-varying characteristics.The LW-SNPER model based quality prediction has a guaranteed performance.(3)For the quality-related fault detection problem,the supervised self learning kernel(S-SLK)is proposed for process modeling and fault detection.S-SLK extends MVU to a supervised form through considering the information contained in the constrained relationships between process and quality data.Further,the generalized semi-supervised self learning kernel(GSS-SLK)model is proposed to make the best use of data with arbitrary missing of quality measurements,considering the preciousness of them.The GSS-SLK model can also handle multi-rate processes.(4)Due to the big data volume in process monitoring,the multi-level maximum variance unfolding(MLMVU)model is proposed for process modeling and fault detection.High computational complexity and storage requirements limit the scalability of the traditional MVU.The MLMVU model orderly divides data to multi-levels,thus can significantly reduces computational complexity and storage requirements on the sacrifice of only limited accuracy.The simulation based on TE process data with high sampling rate demonstrates the effectiveness of the proposed algorithm,and further presents the intrinsic significance of big data in industrial process monitoring.(5)When abnormal operation of the process is detected,accurate information of fault type is needed to make right countermeasures.In order to solve the fault classification problem with class constraint,two novel supervised maximum variance unfolding algorithms,SMVU1 and SMVU2,are developed respectively.The SMVUs models can extract the most related information of fault type through the introduction of class constraint,while the manifold structure of the input variables is guaranteed to be unfolded for the dimension reduction.(6)As to the outliers,all the above models are accordingly designed so that the models are more adaptive and robust against modeling outliers.For the MVU method and its extensions,the rigorous equality constraints are relaxed to obtain more robust models,so that a few outliers won’t influence the modeling much.The NPE method and its extensions are naturally insensitive to outliers,since they are constructed through minimizing local reconstruction error.As to the cross-correlations and nonlinear relationships in the variable related complexity,this dissertation proposes a systematics set of dimension-reductional mapping based industrial process modeling and monitoring from the perspective of manifold learning technique.The proposed models are constructed with local proximity,thus the global nonlinear structure can be expediently approximated by local linearization.