Time series data analytics : clustering-based anomaly detection techniques for quality control in semiconductor manufacturing

Abstract

Optimizing their manufacturing systems and processes whilst ensuring a low production cost is important for Analog Devices, Inc. (ADI). Therefore, detecting anomalies on production lines and alerting on out-of-control processes is crucial. Although Statistical Process Control (SPC) methods have been implemented in the past and have proven to be efficient, the company seeks improvements using machine learning. The Machine Health Project is one of the data analytics-based projects under way at ADI to implement such improvements. Anomaly detection techniques can be effective in improving the quality control on semiconductor production lines. Sets of data collected from semiconductor manufacturing machines, such as a plasma etcher, can be analyzed to control the fabrication process and test the efficiency of machine learning algorithms. This thesis focuses on cluster analysis for outlier detection, and provides a univariate strategy to find potential anomalous behaviors in the data when a given parameter is known as relevant. If a more thorough analysis of the data is needed, a multivariate clustering analysis can also be computed. In addition, decomposition-based algorithms are presented. These rely on techniques such as the STL and SAX representations of time series, and provide a visual computation of time series discords. In this thesis, these methods are implemented, and their results are compared. Recommendations are provided as to how to best utilize the outputs of these outlier detection algorithms.