A Novel Strategy of the Data Characteristics Test for Selecting a Process Monitoring Method Automatically

作者：Zhang, Shumei; Wang, Fuli; Zhao, Luping; Wang, Shu^*; Chang, Yuqing

来源：Industrial & Engineering Chemistry Research, 2016, 55(6): 1642-1654.

DOI：10.1021/acs.iecr.5b03525

• INDEPENDENT COMPONENT ANALYSIS
• ASSESSING MULTIVARIATE NORMALITY
• NONLINEAR PROCESSES
• FAULT-DETECTION
• DYNAMIC PROCESSES
• NEURAL-NETWORKS
• KERNEL PCA
• IDENTIFICATION
• KICA

Process monitoring of industrial processes has attracted broad attention in recent years. Many approaches, such as PCA, KICA, and so forth, have been proposed and successfully applied in many fields. However, the assumption of data characteristics for each method has significantly limited its application. PCA is more suitable for Gaussian processes, whereas ICA is always used in non-Gaussian processes. ICA-PCA has been proposed to analyze some complex processes that contain both Gaussian and non-Gaussian information, and Kernel-based methods (i.e., KPCA, KICA, and kernel ICA-PCA) have been shown to be more effective in nonlinear processes. Therefore, data characteristics should be known to select the most suitable monitoring method. However, prior knowledge is in fact hard to obtain in most industrial processes. For coping with the problem, a novel strategy that can test the data characteristics has been proposed in this paper to select an appropriate monitoring method automatically. A novel nonlinearity coefficient is first introduced to determine whether the process is linear or nonlinear. Then, the improved F-straight method is used to determine the distribution characteristic of the data. The most suitable method is then selected according to the test results of the data characteristics. The feasibility and efficiency of the proposed method has been demonstrated by two numerical case studies. Finally, the proposed method has been successfully applied to a real industrial process of the annealing furnace system.

• 出版日期2016-2-17
• 单位东北大学