Multivariate real-time monitoring using principal component analysis and projection of latent structures

The chemical, oil semiconductor, mining, and paper and pulp industries, as well as most other modern industries, are rapidly expanding the range of measurements used to characterize raw materials, process steps, intermediates, and final products. In addition, process variables such as temperature, pressure, and flow become more and more numerous, providing extra information on the state of the process. However, no matter how many sensors are added, the underlying sources of variations that drive the process behavior remain constant. On the other hand, the measuring frequency of these data has increased from once a day or once an hour to once, ten, or even hundred times a minute. This gives masses of data that are swelling in both directions (more variables and more observations). These data need to be properly stored, managed, and analyzed to provide diagnostics about the state of the process as well as an understanding of the relationships between different parts of the process and the product properties. Projection methods such as PCA (principal components analysis) and PLS (projection to latent structures) compress process data to informative summaries of the original variables. These can be shown on line, providing an efficient, graphically oriented suite of diagnostics regarding the "health" and performance of the process. The power of these methods is illustrated by two examples of industrial data in the mining industry.