Managing Massive Volumes of Runnability Data

摘要

This paper offers a new approach to the processing of three-dimensional da a. The approach is based on the application of multivariate methods to analyse the relationships between a number of specified factors and the responses. X-proFiles is a user-friendly PC-based software system to apply statistical design and evaluation to huge amounts of experimental data with complex mutual correlations. The ultimate goal is to find the optimal settings for the factors which actually influence the responses. The multivariate data analysis makes it possible to discover hidden relationships between many parameters in the data block. The tools for preprocessing the original data make extensive use of visualisation and fitting. The intermediate processing results make it possible to build a very user-friendly interface to the algorithms. The software can be used in an industrial tool. The study was also a statistical approach to runnability problems in newsprint paper production. The multivariate data analysis tool, X-proFiles, was used to construct models for predicting runnability and paper strength properties. The results of this study may be summarised as follows: 1. X-proFiles was found to be a convenient tool for processing large volumes of cell data. It uses combinations of numerical methods and visualisation tools, which allow a rapid and precise interpretation of the relationships between interdependent characteristics and quality properties. 2. Multiple linear regression may give results which are as good as those of the more advanced multivariate methods (principal component regression, partial least-squares analysis and neural network partial least-squares analysis), and it is much easier to introduce these results to end-users. 3. Our work has also shown that PCR is a reliable method for reducing the dimensionality of highly co-linear data sets in a comp ex production process. 4. By applying these methods to the huge amount of statistical data on paper reels, we obtain an unambiguous and quantitative figure for the i npact of the predictors, such as deviations in the moisture profile and small ho es, on the frequency of web breaks. 5. These methods also generated models to predict rheological properties, such as tearing resistance, using parameters which are continuously measured on-line, such as the moisture profile and the grammage.