Authors - Mohammad Arafat Ullah Abstract - Fault Detection and Classification (FDC) plays a critical role in semiconductor manufacturing by identifying defective wafers before subsequent processing stages, thereby reducing manufacturing cost, material waste, and production time. Traditional Statistical Process Control (SPC)-based FDC systems are widely used in semiconductor fabrication; however, machine learning techniques can significantly improve defect detection and process monitoring efficiency. In this research, the SECOM semiconductor manufacturing dataset collected from Kaggle was analyzed using multiple machine learning approaches. Several classification techniques including custom Support Vector Machine (SVM), kernel-based SVM, custom K-Nearest Neighbor (KNN), and Random Forest were implemented and compared for defective wafer detection. In addition, pseudo time-series semiconductor signals were reconstructed from static process features. Exponentially Weighted Moving Average (EWMA) smoothing and temporal feature extraction were then applied for signal-based fault analysis. Experimental results show that SVM-based approaches achieved strong classification performance on the SECOM dataset, while temporal signal reconstruction provided additional insight into semiconductor process behavior. The study presents a comparative analysis between conventional feature-based learning and reconstructed temporal feature-based learning for semiconductor fault detection applications.
Open Zoom