| In modern industrial production,any abnormality may lead to catastrophic accidents,resulting in serious product quality fluctuations and even endangering people’s lives and property.Therefore,effective industrial process monitoring is very important for ensuring the safety of operations and product quality.At the same time,with the rapid development of modern distributed control systems,industrial internet of things and sensor communication technologies,a large amount of valuable industrial data are collected,transmitted and stored.These underutilized industrial data are rich in process information,in stark contrast to the urgent needs of modern industrial process monitoring.To alleviate this contradiction,data-driven process monitoring methods have been widely developed in recent years.Among them,multivariate statistical process monitoring(MSPM)can achieve effective monitoring,optimization and control of industrial processes without excessive prior knowledge and complex process mechanism models.Therefore,it has become a research hotspot in the field of intelligent control.In general,the practical performance of data-based methods largely depends on the completeness and accuracy of data itself.However,the acquired process data are not always complete due to the multi-rate sampling among sensors,transmission delays,and missing data,etc.On the other hand,industrial process data are always autocorrelated,cross-correlated,quality-related and noise contaminated,which makes it difficult for traditional MSPM to maintain optimal performance under ideal assumptions.To this end,an effective MSPM model established for various data characteristics is a common concern to researchers.This work originates from a practical production need of "safety,stability and quality benefits".By analyzing various data characteristics,the quality-related information based complex industrial process monitoring strategies with incomplete data issue are studied under probabilistic frameworks.The main research results include the following parts:(1)Aiming at the problems of quality-related,autocorrelation and noises in process data,a probabilistic dynamic dual-latent variable(PDDLS)model and its fault detection scheme are proposed.The proposed model introduces two types of dynamic latent variables to take care of quality-related and quality-unrelated dynamic information within process data,which are then used for new statistics.Compared with the static single latent variable structure,PDDLS can provide more reasonable description for process information,and can extract the complete process dynamic information from the quality-related and quality-unrelated aspects.Hence,better fault detection performance can be obtained.(2)Aiming at the multi-rate problem of process data,the previous model PDDLS is modified by a multi-rate strategy to meet the task of multi-sampling rate process monitoring(MRPDDLS).Specifically,MRPDDLS establishes the evolution equation between the dual-latent variables and data subsets with different sampling rates simultaneously,which then successfully transforms the multi-rate problem into considering constrained relationships among the sub-datasets.In addition,the solution of the model parameters is performed by local maximum likelihood estimation after improving the expectation maximization(EM)algorithm.Finally,MRPDDLS can consider the cross-correlation of multi-rate variables and the auto-correlation of samples in one model without discarding any data information.(3)Aiming at the problem of the coexistence of dynamic and static characteristics and the difficulty in obtaining quality variables within industrial processes,a concurrent dual-latent variable(CDLV)model and its semi-supervised extension are proposed.First,CDLV introduces two sets of quality-related latent variables,one of which uses the first-order Markov process to learn quality-related dynamic information.The other one can be designed by independent Gaussian for quality-related static information.In such a structure,the quality-related information can be maintained from both dynamic and static features,thus improving the performance of fault detection based on quality related information.On this basis,the CDLV model is further extended into a semi-supervised form(Ss DLV),which can provide a reasonable solution for process monitoring designs with insufficient quality variables.(4)Aiming at the completely random missing of variable values in process data,a linear dynamic time-varying parameter structure(LDVPS)is proposed,which provides a flexible and effective probabilistic modeling idea for irregular missing data issues.By establishing the probability conversion relationship between each measurement variable and dynamic latent variable,LDVPS realizes the main information extraction within incomplete data from available variables and the established Markov process.Among them,the "time-varying" means that LDVPS can adjust parameter size according to the conditions of sample variables at each moment,and then provide adaptive posterior estimation for feature latent variables.The above methods proposed new solutions for quality-related information based complex industrial process monitoring with incomplete data issues,which were verified by simulation and real industrial process cases,respectively.Finally,on the basis of summarizing the research of this paper,future work is prospected. |
