Statistical Process Monitoring Methods For Complex Processes

Process safety and product quality are two important issues that should be paid great attentions by modern industrial processes. As one of the key technologies in the process system engineering area, process monitoring can be employed to solve those two aspects. According to the wide use of the distribution control system in industrial processes, large amounts of data were collected, which have greatly accelerated the development of data-based process monitoring methods along the past decade. Particularly, the multivariate statistical process control (MSPC) method has gained greatly attentions both in industry and academy. and has become a hot spot in process monitoring area.However, the traditional MSPC method is limited to Gaussian, linear, stationary and single mode processes. Based on the existing research works, this dissertation proposes several efficient monitoring methods for different complex processes, which are summarized as follows.(1) According to the complex distribution of the process data, a two-step independent component analysis and principal component analysis (ICA-PCA) based information extraction strategy is proposed for process monitoring, which is sequently improved by support vector data description (SVDD) and factor analysis (FA). A new SVDD reconstruction based method is proposed to address the difficulty of non-Gausian fault diagnosis problem, which can be considered as a complement of reconstruction-based methods for fault diagnosis. Additionally, a mixed index is proposed for fault identification, which can also improve the performance of process monitoring.(2) For nonlinear process monitoring, the statistical local approach (LA) is introduced upon the traditional kernel PCA modeling structure, which can effectively eliminate the restriction of Gaussian distribution for the process data. Due to the offline modeling and online implementation difficulties of the existing methods, a new viewpoint is proposed for nonlinear process monitoring, which is based on linear subspace integration and Bayesian inference. Compared to the existing nonlinear methods, the new method can both improve the monitoring performance and reduce the algorithm complexity.(3) In order to improve the monitoring performance for time-varying and multimode processes, three new methods are proposed. A local least squares support vector regression (LSSVR) based method is proposed, which can greatly enhance the real-time performance of the method; A robust nonlinear external analyais is proposed, which extends the conventional linear external analysis to nonlinear processes, and simultaneously improves the robustness of the method to noises and outliers; A two-dimensional process monitoring method is also proposed, which greatly alleviates the lean of the monitoring method to process knowledge and experiences.(4) Due to the research status on non-Gaussian dynamic processes, few works have been reported. In this dissertation, a new monitoring method is proposed for these special processes, which is based on subspace model identification (SMI) and local approach (LA). In contrast to other methods, the new proposed SMILA method is more efficient in monitoring non-Gaussian dynamic processes.(5) The batch process is considered to be more complicated than continuous processes. Due to the lack of the monitoring research work on multimode batch process monitoring, a Bayesian inference based method is proposed for this special kind of processes, which is also extended for monitoring multiphase batch processes.Finally, Conlcusions and future research studies of the process monitoring ares are illustrated.