Incorporating orthogonal multivariate methods to increase effectiveness

Traditional multivariate projection methods including PLS and PLS-DA attempt to find covariation between sets of data. In the case of PLS, the two sets of data are most often process variables and metrology variables. In the case of PLS-DA, the sets of data are usually from different chambers or before and after a maintenance or scrap event. The powerful thing about PLS and its extensions is that it not only provides information about how the variables are correlated, it also provides information about variation within the individual sets of data. The problem with this is that these two types of information are confounded with each other making it difficult to understand which variation is correlated (predictive) and which variation is not (orthogonal).