Applying the Metapattern Mechanism to Time Sequence Analysis

This paper reports an application of the metapattern mechanism (Shen, Ong, Mitbander and Zaniolo 1994) to analyzing time sequence data for semiconductor process control and fault-detection. The challenge of this task includes the high-dimensionality and large quantity of data, several types of uncertainties in measurements, and lack of effective machine learning tools for time sequence analysis. Using metapatterns, we effectively find a nice angle to view the sequences and compare positive with negative sequences. Such views and comparison helped us to develop a new technology called time-sequence fault-detector. The main idea behind the detector is to learn a typical positive sequence for each control parameter from the classified sequences, along with thresholds for allowed deviations for values, durations, and time shifts. The typical sequence and the thresholds are then used on-line as a basis to detect faults in new and unclassified sequences incrementally. At the present, this detector is realized as a new action for metapatterns (composed from previous metapattern actions), and performs well on the example sequences that are currently available to us. Ultimately, this detector will perform on-line on Motorola's semiconductor plasma machines to detect deficiencies of operations on wafers in real-time.