In semiconductor fabrication, abnormal behavior may first appear in a small subset of serial numbers before it is reflected in lot-level yield metrics. We present a monitoring framework that detects measurement item outliers using Z-scores, aggregates them into a serial-level cumulative-risk score, provides exploratory views of lot- and site-level trends, and ranks high-risk serials for follow-up. The approach is evaluated on an industrial semiconductor manufacturing dataset comprising 14,142 unique serials (Nserial = 14,142). Because most TestResult labels are PASS, label-based yield shows little variation. In this setting, label-based yield alone is not informative enough for early monitoring, so we use outlier-based yield as the primary metric, defined as the proportion of serials with cumulative risk below the threshold (R(s)<τ, where τ=10). A sensitivity study of the outlier threshold κ (σ-multiplier) shows that yield varies widely, from 61.66% at κ=3 to above 99% at κ≥7. This result shows the trade-off between detection sensitivity and inspection workload. Case studies of top-ranked serials show two recurring patterns: cumulative risk is driven either by isolated extreme deviations or by the accumulation of moderate deviations across multiple items. These results support the use of the proposed score for inspection prioritization.
Ryu et al. (Sun,) studied this question.