基於遷移學習之晶圓瑕疵分類方法與資料集適應性比較

NCUIR > School of Management at National Central University > Graduate Institute of Information Management > Electronic Thesis & Dissertation > Item 987654321/98377

Please use this identifier to cite or link to this item: https://ir.lib.ncu.edu.tw/handle/987654321/98377

Title:	基於遷移學習之晶圓瑕疵分類方法與資料集適應性比較
Authors:	李東鑫;Lee, Tong-Hsin
Contributors:	資訊管理學系
Keywords:	深度學習;晶圓瑕疵
Date:	2025-07-26
Issue Date:	2025-10-17 12:42:19 (UTC+8)
Publisher:	國立中央大學
Abstract:	隨著半導體製程技術的高度發展，晶圓生產過程中微小的瑕疵可能嚴重影響產品良率與製程穩定性。晶圓瑕疵圖（Wafer Defect Map）作為識別製程異常的重要依據，已逐漸從傳統人工判讀與規則式演算法，轉向以深度學習為核心的自動化分類技術。然而，晶圓缺陷資料通常具備標記成本高、樣本稀少及類別不平衡等特性，對模型訓練形成挑戰。為此，本研究提出一套結合漸進式遷移學習策略之晶圓瑕疵分類架構，旨在提升模型於稀少樣本下的學習效能與實務應用潛力。本研究以VGG16與 VGG19為骨幹模型，分別於公開合成資料集MixedWM38進行預訓練，並遷移至實際資料集 WM811K進行微調。實驗設計涵蓋模型訓練穩定性與不同訓練資料量兩項主軸，採對照實驗比較漸進式與單一遷移學習策略之效能差異，並透過集成學習評估最終模型穩定性與準確性。評估指標包括模型訓練過程的準確率、損失值以及最終模型的準確率、精確率、召回率、 F1分數及混淆矩陣等。實驗結果顯示，漸進式遷移學習在各種訓練條件下皆表現出較高的分類效能與訓練穩定性，特別在小樣本情境中展現明顯優勢。相較於單一遷移策略，漸進式方法能有效促進模型特徵遷移與收斂速度，並提高對少數類別的識別能力。整體而言，本研究驗證了漸進式遷移學習於晶圓瑕疵圖分類任務之可行性，並提供具體的訓練策略與資料使用指標，作為後續應用於半導體製程品質監控之參考。;With the rapid advancement of semiconductor manufacturing processes, even minor defects on wafers can significantly affect product yield and process stability. As a critical tool for identifying process anomalies, wafer defect maps (WDMs) have increasingly shifted from traditional manual inspection and rule-based algorithms to deep learning-based automated classification techniques. However, challenges such as high labeling costs, limited data availability, and severe class imbalance often hinder the training of effective models. To address these issues, this study proposes a wafer defect classification framework that incorporates a progressive transfer learning strategy to enhance model performance under data-scarce conditions. The proposed approach employs VGG16 and VGG19 as backbone models, initially pre-trained on the publicly available synthetic dataset MixedWM38, and subsequently fine-tuned on the real-world industrial dataset WM811K. The experimental design comprises two main parts: training stability analysis and data volume evaluation. A series of controlled experiments are conducted to compare the performance of progressive transfer learning and single-step transfer learning strategies. Ensemble learning is further applied to improve overall model robustness and accuracy. Performance indicators include accuracy and loss during training process, and accuracy, precision, recall, F1 score, confusion matrix for the final model performance. Experimental results demonstrate that the progressive transfer learning strategy consistently outperforms the single-step approach across different training conditions, particularly under low-sample scenarios. The proposed method effectively accelerates convergence, enhances feature transferability, and improves the recognition of minority classes. Overall, this study validates the feasibility of progressive transfer learning in wafer defect classification tasks and offers practical training strategies and data utilization guidelines for future applications in semiconductor process quality monitoring.
Appears in Collections:	[Graduate Institute of Information Management] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	22	View/Open

社群 sharing

Loading...