One of the main drawbacks of soft-sensors based on
neural networks is that prediction may become equipment-
specific, i.e., only applicable for the training
conditions and equipment used for calibration. Generalizing
the results to other equipment and plants should
therefore be approached with caution. This limitation is
important to the Chilean Pisco industry, since a variety
of still designs are used, sometimes even within a single
plant. To assess generalizing potential, we tested the