Monitoring Process And Product Performance On-Line Using New Powerful Multivariate Statistical Methods
ABSTRACT:
With process computers routinely collecting data from on-line sensors on hundreds to thousands of variables every few seconds, large databases are accumulated in industry. The exploitation of these data is a critical component in the successful operation of any industrial process. Establishing methods whereby we can learn from our past experiences to eliminate problems in our process or in our operating procedures, monitor our current operations in real time, and find improved conditions for both our present and new products is of paramount importance in competitive environment today. New methods can monitor process performance on-line and assess product quality long before off-line product lab data become available. Successful applications are reported by many diverse industries such as pharmaceuticals, semiconductor manufactures, steel producers, pulp and paper producers, polymer plants and refineries
Up to now, rarely something has been done with these data, mainly because of the nature of such data: 1) The size of the data sets is overwhelming. 2) The data are highly correlated (many variables being collinear) and non-causal in nature. 3) The information contained in any one variable is often very small due to the low signal / noise ratios. 4) There are often missing measurements on many variables. In order to utilize these databases, an empirical modelling method must be able to deal effectively with all these difficulties.
Some relatively new and very promising approaches (based on latent variables) for troubleshooting, on-line process monitoring, fault diagnosis and equipment maintenance based on multivariate projection methods has gained rapid acceptance by industry. These methods are capable of utilizing massive amounts of correlated data (450 variables and 500 observations being a typical example) and compress the information in this data down into low dimensional latent variable spaces in which analysis of the process and interpreting the results are much easier. Simple, easy to interpret multivariate charts derived from the methods are attractive to the plant engineers and operators. The concepts of process analysis, monitoring, fault detection and diagnosis using multivariate projection methods will be presented. Examples from successful industrial applications will be used to illustrate the power of these methods in analyzing, monitoring and diagnosing operational problems.
About the Speaker
Dr. Theodora Kourti is the research manager of the McMaster Advanced Control Consortium (MACC) and an adjunct professor with the Dept. of Chemical Engineering at McMaster Univerity. MACC member companies are 20 of the largest multinationals including Shell, Dupond, Kodak, Rohm and Haas etc. Dr. Kourti has an extensive experience with Multivariate Statistical Methods and their applications to Industry, and she has been involved in more than 70 major industrial projects in North America and Europe. These are either off-line or on-line applications for batch and continuous processes, in diverse industries such as Chemicals, Pharmaceuticals, Semiconductor, Mining, Pulp and Paper, Petrochemicals, Photographic and Steel Industry. She also worked with Exxon Research and Engineering (USA) and Esso Rotterdam Refinery (The Netherlands) where she applied multivariate methods for Abnormal Situation Detection and provided training for plant personnel. She has published extensively in the area of Multivariate Projection Methods.
References
Some relevant background references will be posted here shortly.
