In the rMM approach, two local models centered in the response
space in 1
y = 0.5% and 2
y = 5% were utilized, and the width
parameter in the response validity functions (4) was = 2.5. The
locations and width of the local model regions were selected so that
the first model could handle the normal operation conditions with
low copper values and the second model the unusually high copper
levels related to process failures. The nominal recursive forgetting
factor in rMM was 0 = 0.95 and the forgetting factor in rPLS was
= 0.95. The number of factors in both static and rMM OSC deflation
was f = 1. More OSC factors were also tested, but the modeling
results were not improved. The number of PCA latent variables that
were used to invert the deflated predictor covariance matrix (15)
was selected so that the condition number of Rˆi
tto(k) was not too
high. With eight latent variables the maximum condition number
was kept around 106, which ensured that no numerical problems
occurred when inverting the matrix.
T