Statistical approaches to fault analysis in multivariate process control

After a brief review of some statistical approaches to multivariate process control, we present a technique for determining root causes when information is available on likely out of control scenarios or fault types. We utilize linear dimension reduction techniques such as principal component analysis or partial least squares to limit the number of latent variables to study. While using historical control data is important in establishing control means and limits, these data often have less structure for dimension reduction than do data which come from known fault types. If these latter data are available, the expanded data set can be analyzed for dimension reduction, using the in control data to set limits in the reduced set. When a sequence of points is then seen to be beyond the control limits, the distance to the nearest known fault type is measured. If the dimensions can be reduced to two, these can be plotted as well. The new problem is classified into one of the existing fault types when its distance to it becomes smaller than a pre-specified criterion. If it remains out of control, but fails to approach an existing fault type, a new fault paradigm is created. Our approach is demonstrated on a simulated chemical process.