Efficient selection of signatures for analog/RF alternate test

This work proposes a generic methodology for selecting meaningful subsets of indirect measurements (signatures). This allows precise predictions of the DUT performances and/or precise pass/fail classification of the DUT, while minimizing the number of necessary measurements. Two simple figures of merit are provided for ranking sets of signatures a priori, before training any machine learning model. These two figures evaluate the quality of each signature based on its Brownian distance correlation to the target specifications, and on its local distribution in the proximities of the pass/fail decision boundaries. The proposed methodology is illustrated by its direct application to a DC-based alternate test for LNAs.