EVIDENCE FROM AN IC PACKAGING FOUNDRY BY USING A TWO-PHASE CLUSTERING METHODOLOGY

ABSTRACT

Clustering is to group objects together so that they are as homogenous as possible within the same cluster while most distinct in different clusters. This paper uses a two-phase clustering methodology that integrates the self-organizing maps (SOM) algorithm in the first phase with the k-means algorithm and the minimum spanning tree-based (MST-based) clustering in the second phase. The MST-based clustering is used because it is efficient to solve tree-type problems and tends to be less sensitive to the geometric shape of data. Two types of data transformations including min-max normalization and z-score normalization are employed to deal with the situation where magnitudes of real-life data differ sharply. We compare clustering results in terms of Davies-Bouldin (DB) value and Wilk's lambda value. According to the results by using the data of Wire Bond machines from a Taiwanese IC packaging foundry, we find that applying the k-means algorithm in the second phase to the data with min-max normalization is better in terms of jointly considering DB value and Wilk's lambda value. Despite that applying the MST-based clustering in the second phase does not outperform the k-means algorithm; however, we find that the former prevails over the latter in terms of detecting outliers especially when normalized data are used.

Keywords:

• clustering
• self-organizing maps
• minimum spanning tree
• IC packaging