An Incremental Clustering-Based Fault Detection Algorithm for Class-Imbalanced Process Data

Jueun Kwak, Taehyung Lee, Chang Ouk Kim

Research output: Contribution to journalArticle

21 Citations (Scopus)

Abstract

Training fault detection model requires advanced data-mining algorithms when the growth rate of the process data is notably high and normal-class data overwhelm fault-class data in number. Most standard classification algorithms, such as support vector machines (SVMs), can handle moderate sizes of training data and assume balanced class distributions. When the class sizes are highly imbalanced, the standard algorithms tend to strongly favor the majority class and provide a notably low detection of the minority class as a result. In this paper, we propose an online fault detection algorithm based on incremental clustering. The algorithm accurately finds wafer faults even in severe class distribution skews and efficiently processes massive sensor data in terms of reductions in the required storage. We tested our algorithm on illustrative examples and an industrial example. The algorithm performed well with the illustrative examples that included imbalanced class distributions of Gaussian and non-Gaussian types and process drifts. In the industrial example, which simulated real data from a plasma etcher, we verified that the performance of the algorithm was better than that of the standard SVM, one-class SVM and three instance-based fault detection algorithms that are typically used in the literature.

Original languageEnglish
Article number7123674
Pages (from-to)318-328
Number of pages11
JournalIEEE Transactions on Semiconductor Manufacturing
Volume28
Issue number3
DOIs
Publication statusPublished - 2015 Aug 1

All Science Journal Classification (ASJC) codes

  • Electronic, Optical and Magnetic Materials
  • Condensed Matter Physics
  • Industrial and Manufacturing Engineering
  • Electrical and Electronic Engineering

Fingerprint Dive into the research topics of 'An Incremental Clustering-Based Fault Detection Algorithm for Class-Imbalanced Process Data'. Together they form a unique fingerprint.

  • Cite this