Segmented Dynamic Time Warping Based Signal Pattern Classification

Abstract

The semiconductor manufacturing process is divided into fabrication process and packaging process. Fabrication process is a core process for manufacturing semiconductors and consists of about 700 unit processes. This unit process accumulates vast amounts of data, and many manufacturing companies apply data-based algorithms to manufacturing systems to improve process yield and quality. Data generated during the semiconductor manufacturing process from the process equipment is called fault detection and classification (FDC) trace data, and this data has time-series characteristics of different patterns depending on the sensor type or the recipe. Therefore, it is necessary to develop a classification algorithm appropriate to the signal pattern for process monitoring. In this paper, we develop segmented dynamic time warping technique which is specialized for process signal classification. Generally, it is known that dynamic time warping (DTW) has superior classification performance for time series data. However, there is a limit to classification that reflects the characteristics of semiconductor process signals. Therefore, we developed a classification algorithm for process signal data through segmented DTW using maximum overlap discrete wavelet transform (MODWT) and random sample consensus (RANSAC), and validated it.