A Semantic Mixture-of-Experts Framework for Defect Rate Prediction Using Data Fusion in High-Mix Low-Volume Manufacturing

Key Points

• This research aims to develop a framework utilizing unstructured text for predicting defect rates in manufacturing environments.
• Proposed a semantic mixture-of-experts framework to handle data heterogeneity and clashes.
• Converted unstructured text from ERP and MES systems into numerical data using TF-IDF.
• Integrated quantile regression with Conditional Value-at-Risk for better risk management.
• Designed a gating network to identify hidden regimes in manufacturing processes.
• Achieved approximately 16.8% reduction in mean absolute error compared to baseline models.
• Successfully identified high variability processes and assigned specialized expert models for predictions.
• Provided robust risk management guidelines applicable in manufacturing settings.

Abstract