Research On Quality-related Process Monitoring For Large-scale Complex Industrial Processes

The modern industrial process towards large-scale,automation,and informatization.As an essential technique,process monitoring has become an indispensable part of the modern industrial system for guaranteeing product quality,avoiding safety accidents,and decreasing ecosystem pollution.Because of being equipped with precise industrial sensors and wide application of data storage and management technologies,it is easier to collect large amounts of data reflecting process functioning state than before.Under this circumstance,data-driven process monitoring has emerged.Specifically,the multivariate statistical process monitoring(MSPM)method has become a hot research point in academies and industries due to its advantages in processing multivariate data.Although many traditional MSPM methods have been reported,there are still some issues in handling modern large-scale industrial processes.In practice,however,the whole process is usually composed of multiple operational units.Meanwhile,there are complex relationships within the same units and between different units,if roughly selecting all variables to construct a global model,the local information may be neglected.Based on existing works,this dissertation takes the block model as a framework and conducts some research for large-scale industrial processes with nonlinear and dynamic characteristics.The main construction is summarized as follows:(1)Considering the relationship between variables within the same operation unit,the process variables are divided into different sub-blocks based on prior knowledge,and the dynamic multiblock partial least squares(DMBPLS)model is used for fault detection.Firstly,the augmented matrices technique is introduced to improve the interpretability of the dynamic characteristic.Secondly,block division not only mines local information of quality-related in the process but also overcomes the problem of high dimension brought by the augmented matrices.The superior anomaly detection capability of DMBPLS is demonstrated by a real papermaking wastewater treatment process.(2)Considering that the block division using prior knowledge is not always available,the dissertation proposed a new block modeling strategy based on data.According to the variable importance in the projection(VIP)value,all the process variables are firstly partitioned into quality-related and quality-independent spaces.Then,combined with the augmented matrices technique,the dynamic concurrent partial least squares model is constructed within two spaces to decompose the original process space,comprehensively.Finally,fusing all statistics through support vector data description to provide an overall indication.The experiment shows that the proposed model can effectively enhance the monitoring performance.(3)Considering the prevalent nonlinear characteristic of data,the improved VIP(IVIP)is structured by inserting mutual information into the VIP technique,which makes the block result more accurate and reasonable.Meanwhile,to improve the performance of model for incipient faults,the multivariate exponentially weighted moving average technique is firstly utilized to preprocess the original data.Then,the kernel principal component analysis model is established in quality-related and quality-independent spaces for monitoring.The proposed method not only effectively reveals the relationship between variables and achieves incipient faults detection accurately.The applications of a numerical simulation system and the Tennessee Eastman process indicate that the proposed model greatly decreases the fault detection delay.