Supply chain operations drive the planning, manufacture, and distribution of billions of semiconductors a year, spanning thousands of products across many supply chain configurations. The customizations span from wafer technology to die stacking and chip feature enablement. Data quality drives efficiency in these processes and anomalies in data can be very disruptive, and at times, consequential. Developing preventative measures that automate the detection of anomalies before they reach downstream execution systems would result in significant efficiency gain for the organization. The purpose of this research is to identify an effective, actionable, and computationally efficient approach to highlight anomalies in a sparse and highly variable supply chain data structure. This research highlights the application of ensemble unsupervised learning algorithms for anomaly detection on supply chain demand data. The outlier detection algorithms explored include Angle-Based Outlier Detection, Isolation Forest, Local Outlier Factor and K-Nearest Neighbors. The application of an ensemble technique on unconstrained forecast signal, which is traditionally a consistent demand line, demonstrated a dramatic decrease in false positives. The application of the ensemble technique to the sales-order netted demand forecast, a signal that is irregular in structure, the algorithm identifies true anomalous observations relative to historical observations across time. The research team concluded that assessing an outlier is not limited to the most recent forecast’s observations but must be considered in the context of historical demand patterns across time.

Creative Commons License

Creative Commons Attribution-Noncommercial 4.0 License
This work is licensed under a Creative Commons Attribution-Noncommercial 4.0 License

Included in

Data Science Commons