Study on Data Preprocessing for Machine Learning Based on Semiconductor Manufacturing Processes.

Area of Science:

• Semiconductor Manufacturing
• Machine Learning
• Data Science

Background:

• Semiconductor manufacturing generates diverse, imbalanced data from various sensors.
• Data preprocessing is crucial for accurate yield prediction and cost reduction.
• Existing methods often struggle with imbalanced datasets and lengthy processing times.

Purpose of the Study:

• To evaluate machine learning analysis and preprocessing techniques for predicting semiconductor product quality.
• To enhance product yield and decrease manufacturing costs through improved defect prediction.
• To address data challenges like diverse units, imbalanced classes, and high dimensionality.

Main Methods:

• Utilized the SECOM dataset for analysis.
• Applied preprocessing techniques: missing value imputation, dimensionality reduction, resampling (oversampling/undersampling), and feature scaling.
• Implemented and compared six machine learning models, focusing on geometric mean (GM) for evaluation.
• Ensured data integrity by preventing data leakage through early train/test set separation.

Main Results:

• Oversampling methods, excluding KM SMOTE, effectively balanced the dataset classes.
• The combination of Support Vector Machines (SVM), Adaptive Synthetic Sampling (ADASYN), and MaxAbs scaling achieved the highest performance.
• The optimal model yielded an accuracy of 85.14% and a geometric mean (GM) of 72.95%.

Conclusions:

• The proposed preprocessing and machine learning approach significantly improves defect prediction in semiconductor manufacturing.
• Feature reduction techniques shorten training and prediction times, enhancing practical applicability.
• The study demonstrates the effectiveness of SVM with ADASYN and MaxAbs scaling for imbalanced semiconductor data.