摘要
Wafer probing tests are crucial for smart semiconductor manufacturing to identify known good dies for integrated circuit (IC) packaging. Probe cards, composed of printed circuit boards (PCBs), serve as the interface between fabricated wafers and probing equipment. As IC feature sizes shrink, maintaining data integrity during wafer probing becomes increasingly challenging. However, full inspection of all signal channels on probe card PCBs is often impractical due to time and cost constraints. As a result, selective channel inspection is commonly employed in practice. Focusing on realistic needs, this study proposes the UNISON decision framework to support effective inspection of PCB channels. The framework further enables the prediction of unmeasured channels, ensuring inspection quality and enhancing information coverage for advanced quality control. In particular, the framework integrates a Denoising Variational Autoencoder (DVAE)-based forecasting model for unmeasured channels and a simulation module to explore trade-offs between prediction accuracy and inspection time. An empirical study was conducted at a leading semiconductor testing company to validate the framework. The proposed DVAE-based imputation method was benchmarked against eight baseline approaches: four statistical methods (mean imputation, linear interpolation, multivariate linear regression, and MICE) and four machine learning models (K-Nearest Neighbors, Random Forest, MissForest, and XGBoost). Across all evaluation metrics—Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Euclidean distance—the proposed method consistently outperformed the baselines. These results demonstrate the practical viability and robustness of the developed framework in improving decision quality for selective PCB channel inspection in smart semiconductor manufacturing.