Data Based Statistical Process Monitoring For Complex Industrial Processes

Process safety is a cornerstone of the new generation of industial revolution nowadays.In order to ensure process safety,process monitoring methods have been widely used in industrial processes.Recently,process monitoring becomes one of the basic technologies during the new generation of industrial revolution to ensure intelligent manufacturing.With the development of measurement and data storage techniques,a large amount of data can be collected from industrial sites,which makes data-based process monitoring approaches become quite popular.Among various kinds of data-based methods,multivariate statistical analysis methods have caught strong attention because of their merits in handling high-dimension and highly correlated data.Traditional statistical process monitoring approaches always make some idealized assumptions,for example,modeling data is sufficient,modeling data should be regular,measured variables are purely linear correlated or nonlinear correlated,and process data are collected from steady operating conditions(there is no dynamic behavior).However,the above assumptions may be invalid as modern industrial processes are becoming much more complex,resulting in that the monitoring performance of traditional statistical process monitoring methods would be jeopardized.Hence,how to design reliable statistical monitoring schemes for practical industrial process with complex characteristics becomes a critical problem.As two major manufacturing types of modern industry,batch process and large-scale continuous process always operate in complex conditions where the aforementioned assumptions cannot hold.As a result,this dissertation analyzes the complex process characteristics of batch processes and large-scale continuous processes and developes statistical process monitoring approaches for them to solve practical problems without idealized assumptions.(1)In practice,it is sometimes difficult to obtain enough modeling data from some batch processes(i.e.bio-pharmarcy processes),resulting in small data problem.For these batch processes,a generalized time slice is first constructed for exploring process information,based on which a step-wise sequential phase partition algorithm is developed,which achieves phase identification by iteratively evaluating the process correlations among each generalized time slices.Besides,the monitoring models developed upon each phases can be updated with the accumulation of new data,thereby reliable monitoring performance can be ensured.(2)In practice,the operation of different batches may not be strictly the same,which means that different batches may have different durations and different batch trajectories,generating irregular process data.Since the crtical element of irregular batches is irregular phases,an irregular phase identification algorithm is proposed,which evaluates time-wise variable correlations for each batch and thus can obtain irregular phases.For online monitoring,two regions are distinguished which avoids the false alarms caused by uncorrect identification of irregular phases and thus improves monitoring performance.(3)Considering that measured data of large-scale processes are collected from different operation units which have different operating patterns,variable correlations are always hybrid that linear correlations and nonlinear correlations coexist.For analyzing hybrid correlations,this dissertation first proposes the concept that linear correlations and nonlinear correlations should be separately treated.Based on this concept,a linearity evaluation and linear variable subset partition algorithm is proposed to decompose the process into different linear subsets and nonlinear subset.Then,a hierarchical modeling and monitoring strategy is proposed,which not only explores the local linearity but also considers the global nonlinearity of process,thereby enhancing monitoring performance.(4)For large-scale processes with high-dimension variables and hybrid correlations,fault information would be buried and fault may be encapsulated with linear patterns and nonlinear patterns,resulting in that it is difficult to extract accurate fault features.Therefore,a hybrid fault characteristics decomposition algorithm is proposed to decompose the complex fault characteristics into different subsets,which makes the extraction of accurate fault feature much easier.Then,a distributed fault modeling strategy is proposed,in which different diagnosis models are designed to closely describe different types of fault patterns and thus informative fault features can be extracted to improve the fault diagnosis accuracy.Besides,fault diagnosis results are calculated in a probabilistic way,which not only tells the fault types but also reveals the extents of samples affiliated to each fault class.(5)Industrial processes always subject to frequent changes of operating conditions due to various factors,i.e.alternation of control strategies.Since large-scale processes consist of several operation units,workshop and so on,their dynamic behaviors appear to be much more complex.Considering that static and dynamic information reveal different process characteristics,a sparse slow feature analysis based distributed and concurrent monitoring strategy is proposed,which firstly decomposes the whole process into different blocks by developing a sparse slow feature analysis algorithm and then develops a distributed monitoring system to check the local and global process chacteristics from static and dynamic aspects.The above methods provide some novel ideas for solving the problems brought by the complex process characteristics of practical processes.Their effectivenesses are verified by some numerical examples and practical industrial processes.Finally,based on current work of this dissertation,some future research topics are discussed.