Lim, Shi Xuen (2023) The formulation of a transfer learning pipeline for the classification of the wafer defects. Masters thesis, Universiti Malaysia Pahang Al-Sultan Abdullah (Contributors, Thesis advisor: Ismail, Mohd Khairuddin).
|
Pdf
ir.The formulation of a transfer learning pipeline for the classification of the wafer defects.pdf - Accepted Version Download (130kB) | Preview |
Abstract
In a semiconductor manufacturing process, a semiconductor wafer may have defects which are unacceptable due to its complexity in manufacturing process. Defect detection in wafer is vital in avoiding yield loss in end product, which is often achieved by visual judgement using an optical microscope. This often causes misjudgment and inconsistency result between different personnel. Automated processes have been used commonly in recent years, with the judgement done by using conventional image processing algorithm. However, limitations such as robustness and difficulty in setting up the parameters required for image processing algorithm encourages the investigation in using Deep learning classification in detecting the wafer defects. Deep learning classification produces excellent and robust results in classifying defects in wafer images, but challenges in data scarcity and hyperparameter tuning hampers its implementation in the industry. To combat this challenge, Transfer learning (TL) is investigated to utilize pretrained weights from other models, which in turns produces similar excellent results while improving the efficiency of implementation. Thus far, there are still limited studies that investigate the classification of wafer defects using TL combined with a classical Machine learning (ML) pipeline. Thus, this study will aim to explore 17 types of TL models, and classify the features extracted using 3 different ML algorithms, namely Support Vector Machine (SVM), k-Nearest Neighbor (kNN) and Random Forest (RF). The ML classifiers were tuned via a 5-fold cross-validation technique through grid search approach. The input images were split into a stratified ratio of 60:20:20 ratio as training, validation and testing set respectively. Different combinations of TL-ML pipeline were evaluated and compared to obtain the best performing pipeline by various performance measures such as the classification accuracy, confusion matrix, precision, sensitivity and F1 score. It is observed that the ResNet101v2 model pairing up with an optimized SVM pipeline is able to achieve the best classification accuracy of 95% for training, validation and testing data.
| Item Type: | Thesis (Masters) |
|---|---|
| Additional Information: | Thesis (Master of Science) -- Universiti Malaysia Pahang – 2023, SV: Dr. Ismail Mohd Khairuddin, NO. CD: 13508 |
| Uncontrolled Keywords: | semiconductor wafer, transfer learning (TL) |
| Subjects: | T Technology > TA Engineering (General). Civil engineering (General) T Technology > TS Manufactures |
| Faculty/Division: | Institute of Postgraduate Studies Faculty of Manufacturing and Mechatronic Engineering Technology |
| Depositing User: | Mr. Nik Ahmad Nasyrun Nik Abd Malik |
| Date Deposited: | 03 Sep 2024 02:07 |
| Last Modified: | 03 Sep 2024 02:07 |
| URI: | http://umpir.ump.edu.my/id/eprint/42467 |
| Download Statistic: | View Download Statistics |
Actions (login required)
 |
View Item |