Independent Component Analysis Based Industrial Process Abnormal State Monitoring And Diagnosis Methodologies

It is significantly important to monitor and diagnose abnormal states in petrochemical and other process industries, both from academia and engineering perspectives. It is practical that independent component analysis (ICA) which is recognized as a class of effective multivariate statistical analysis techniques has been increasingly utilized in industrial process control and related fields. Neverthelss, industrial production processes still demand for more effective abnormal state monitoring and diagnostic techniques to improve operation performances. Motivated by this challang, this thesis conducts a systematic investigation on online monitoring and diagnosis of abnormal states in industrial processes based on independent component analysis, successfully resulting in novel methods and techniques. The main contents conclude following aspects:Firstly, an improved ICA approach to process abnormal state monitoring is presented. To avoide limitations of traditional selections of key independent components, we introduce a novel approach to quantifying system deviations, which can reveal differences between the original and reconstructed independent components in terms of signal recovery so as to optimize the number of dominant independent components. Additionally, in order to overcome the high sensitivity of kernel density estimation to disturbances on the ICA statistics, Box-Cox transformations are invoked to transforme the ICA statistics into Gaussian distribution data before calculating the control limits conventionally. Based on these ideas, an improved ICA abnormal state monitoring method is created. Case studies consisting in continuous and batch processes are carried out to demonstrate the effectiveness of the contribution.Secondly, a singular vector recognition based ICA abnormal state monitoring method is developed. In response to the issue of continuous process operational mode switching or batch process stage changing, relative change indics of singular vectors are introduced to indicate state transitions of processes, taking into account of the changes in the data amplitudes and angles. Accordingly, the whole process can be divided into alternating forms of stable and transitional sub-processes. Therein, the stable sub-processes are monitored with the improved ICA process monitoring method, while the transitional sub-processes are delt with combinatory linear models of the two adjacent stable sub-processes, in which the weighting coefficients are inversely proportional to the distance from the sampling time to the adjacent stable sub-processes. Therefore, a multi-mode or multi-stage process abnormal state monitoring method is established, which can significantly reduce the false alarm rate and avoid certain false negative rate, along with improving the reliability of ICA based monitoring approaches by distinguishing different process running states. The benefits of the proposed methods are illustrated through applications to experimental continuous and batch process cases.Thirdly, we introduce a process abnormal state monitoring and diagnosis method based on reconstructed contributions of combined ICA indices which can reduce the number of statistics, epitomize the abnormal state influences and avoid conflicting diagnosis. It is acknowledged that fault reconstruction can resolve false diagnose problems of traditional contribution plot methods caused by smearing effects. Thus, with reconstructed contributions of combined indices available, a comprehensive process abnormal state diagnosis system can be established, which could significantly simplify the tasks in terms of calculation, monitoring and analysis involved, effectively improving the accuracy of the abnormal state diagnosis. Relavent theoretical proofs and exemplary verifications are provided here.Subsequently, a process abnormal state diagnosis method based on an integration of ICA and conditional probability Petri nets is proposed. Initially, to avoid the backward reasoning limitations of fuzzy Petri nets, a kind of conditional probability Petri nets is build. In this context, we established a two-layer ICA-CPPN process abnormal state monitoring and diagnosis architecture. The lower ICA monitoring models could capture local dynamics of the process and provide quantitative statistical information for the upper CPPN, while the upper CPPN models could characterize the overall dynamic movements of abnormal states in each local area of ICA models providing in-depth diagnostic information. In addition, because the process is divided into several operating units, the computational burden is reduced and the modeling accuracy is improved for the ICA layer. Moreover, the places associated with the upper CPPN layer are greatly reduced due to the replacement of single variable knowledge representation by the local combined indices, avoiding the state explosion problems.Finally, resultant conclusions of the contribution are presented along with promising future researches.