Research On Distributed Process Monitoring Method Based On Novel Multi-block Modeling Algorithms

With the increasingly complex structure and constantly expanding scale of modern process industries,the operating condition of industrial processes must be monitored in real time to ensure production safety and improve product quality.Due to the widespread application of distributed control systems and advanced computer technology in industrial process systems,a large number of rich process data can be collected and stored.How to mine feature information that reflects the process operating condition from these data for process monitoring has became a very challenging question.Data-driven monitoring methods,especially multivariate statistical process monitoring,have developed rapidly under this background.Distributed process monitoring method,which is an MSPM method,can effectively reduce the complexity of the model and improve the model interpretation ability.It is usually the default approach for large-scale industrial process monitoring.Although distributed process monitoring methods have received more and more research in recent years and have achieved fruitful results,However,many problems are worthy of in-depth research,such as how to extract correlation information between sub-blocks and quantify relevance and so on.Based on the existing research,this paper takes the correlation between sub-blocks as the research focus,and proposes two new distributed process monitoring methods as follows:1)Given that the correlation between sub-blocks is ignored for all distributed monitoring methods in the non-Gaussian process,a novel distributed process monitoring method based on MBMICA(Multi-block Modified ICA)algorithm is proposed for non-Gaussian process monitoring.First,the MICA(modified ICA)algorithm can be obtained by modifying the orthogonal process for the Fast ICA algorithm,and then the unmixed vectors of the entire measured variable space can be used to guide the extraction of independent components in each sub-block,MBMICA is then derived which can consider all the measured variables as a whole.Finally,the TE simulation verify the feasibility of the MICA algorithm and the superiority of MBMICA method compared to other distributed process monitoring methods.2)Aiming at the problem that the existing distributed monitoring methods cannot quantify the correlation between sub-blocks,a new distributed process monitoring method based on MBCCA(Multi-block CCA)algorithm is proposed.This method first obtains the latent variables of each sub-block by maximizing the sum of squares of canonical correlation coefficient of all subblocks.These latent variables can explicitly represent global features and local features;then regression model is used for latent variables containing global feature information,which extracts the residual information that reflects the difference between the changes of each sub-block;then the monitoring statistics are constructed from the residual information of each sub-block and the latent variables containing local feature information,and merged into a set of probabilistic monitoring indicators.Finally,the effectiveness and superiority of the proposed method are verified through the simulation of TE process.The work of this paper is summarized at length,and the future research and direction are prospected for the field of distributed process monitoring.