A Unified Framework for Outlier Detection in Trace Data Analysis

摘要: Process trace data (PTD) is an important type in semiconductor manufacturing and has a very large aggregate volume. While mining statistical analysis play key role the quality control of wafers, existence outliers adversely affects applications benefiting from PTD analysis. Due to complexities resultant outlier patterns, this paper proposes unified detection framework which takes advantages complexity reduction using entropy abrupt change cumulative sum (CUSUM) method. To meet practical needs analysis, two-step algorithm taking into account related domain knowledge developed, its effectiveness validated by real sets production example. The experimental results show that proposed method outperforms Fast Greedy Algorithm (FGA) Grubb's test, two commonly used techniques for univariate data.