Studies On Applications Of Flow Industry Process Monitoring Based On ICA And BSS

Flow industry is an essential part of our national economy. As the process involves high temperature, high pressure and high risk, the monitoring techniques become the most important part of the production process. With the help of modern signal processing techniques, this paper studies process monitoring methods within the application of data-driven multivariate statistical process monitoring methods in the flow industry.This paper mainly discusses the methods of Independent Component Analysis (ICA) and Blind Source Separation (BSS). ICA searches for the underlying statistical independent components in the data and explores the inherent characteristics. In the past 10 years, ICA has become a hot topic and attracted widely attention, which reflects the powerful vitality and broad application prospects. Blind Source Separation is to get the source signal only by analyzing the received signals without any information about the source and the circumstances of transmission channel. BSS and ICA are very similar in the object of analysis, data model, research methods etc., but there are also some distinctions in theory and engineering between them. They are developing with mutual reference and complementarity.According to the characteristics of the process monitoring, this paper studies some parts of the ICA and BSS theory. The major contributions are as follows:1. Several aspects of ICA theory are descriped such as background, modeling and the ways to compute independent components. Actual examples are given to show the non-gaussian distribution of the process variables in flow industry. It is also proved that the nonlinear property of the system affects the distribution of the process data. Therefore, it is neccessary to utilize ICA method in the process monitoring application.2. According to the vector expansion of ICA theory, a dynamic independent subspace monitoring method is studied by considering the self-correlation of the variables of the process, composing the time-series subspace and updating its data with the lapse of time, which makes the process monitoring method available. The simulation results of TE process reveal the effectiveness of this method.3. Two new monitoring methods are proposed by involving generalized eigen- decomposition (GED) used in the BSS problem: high order statistic matrices of the data are used though recursive GED to search for the independent components and achieve the process monitoring; on the other hand, temporal covariance matrix pencil with time-lagged window is also adopted to make use of the time series correlation, by which the new process monitoring method shows its superiority of sensitivity and robustness.4. Non-negative Matrix Factorization (NMF) is a developing matrix factorization techniquethese years. A new monitoring method based on NMF is also proposed in this paper. It is proved to be effective through a simulate application upon TE process.