Branislav Kveton, University of Pittsburgh; and and Denver Dash, Intel Research
In the past decades, high-volume manufacturing processes have grown increasingly complex. If a failure in these systems is not detected in a timely manner, it often results in tremendous costs. Therefore, the demand for methods that automatically detect these failures is high. In this work, we address the problem of automatic excursion detection based on parametric tests. Overlooking the complexity of wafer fabrication processes, we propose two structurally simple excursion detection models: Naive Bayes classifier and boosted decision stumps. We apply these models in the domain of semiconductor manufacturing, compare them to off-the-shelf classification techniques, and show significant gains in the precision and recall of detected excursions. These results encourage our future work that should primarily focus on increasing the recall of our models.