時間序列對比學習於晶片製程瑕疵預測;A Time Series Contrastive Learning for Defect Prediction on IC Manufacturing

NCUIR > School of Management at National Central University > Graduate Institute of Information Management > Electronic Thesis & Dissertation > Item 987654321/98226

Please use this identifier to cite or link to this item: https://ir.lib.ncu.edu.tw/handle/987654321/98226

Title:	時間序列對比學習於晶片製程瑕疵預測;A Time Series Contrastive Learning for Defect Prediction on IC Manufacturing
Authors:	王星皓;Wang, Hsing-Hao
Contributors:	資訊管理學系
Keywords:	深度學習;時間序列;對比式學習;知識蒸餾;瑕疵偵測;Deep learning;time series;contrastive learning;knowledge distillation;fault detection
Date:	2025-07-03
Issue Date:	2025-10-17 12:31:07 (UTC+8)
Publisher:	國立中央大學
Abstract:	本研究提出一種時間序列對比學習架構（Time Series Contrastive Learning for Fault Detection and Classification, TSCL-FDC），用於解決晶圓製程中異常偵測與瑕疵分類的挑戰。由於半導體製程高度複雜且多變，傳統監督式學習在面對多變量時間序列資料、資料不平衡與異常訊號微弱等問題時，效果受限。本研究融合時域與頻域資訊，建構一個具備多領域學習能力的對比學習模型，並結合知識蒸餾機制，使模型在只使用單一時域資訊時仍能保持高度預測效能。實驗採用來自真實晶片封裝製程的 Assembly 資料集進行驗證，該資料集包含電流、電壓、變形與頻率等感測器量測訊號。結果顯示，TSCL-FDC 模型在準確率、精確率、召回率與 F1 分數等多項指標上皆優於傳統對比學習與其他基準方法，並展現出優異的類別區辨能力與異常偵測精度。個案研究亦證實該模型能有效應用於實際工業環境，提升製程監控與瑕疵預測能力。本研究證明將頻域訊號整合入對比學習架構，並透過知識蒸餾方式對多種輸入形式進行壓縮，能有效提升模型在智慧製造領域中對高維序列資料的處理能力，具高度實務應用潛力與延伸價值。 ;This study proposes a Time Series Contrastive Learning architecture for Fault Detection and Classification (TSCL-FDC) to address the challenges of anomaly detection and defect classification in wafer manufacturing processes. Due to the high complexity and variability of semiconductor manufacturing, traditional supervised learning methods often struggle with multivariate time series data, data imbalance, and subtle anomaly signals. To tackle these issues, we integrate both time-domain and frequency-domain information to construct a contrastive learning model with cross-domain learning capabilities. Additionally, a knowledge distillation mechanism is incorporated to ensure the model maintains high predictive performance even when only time-domain inputs are available. Experiments were conducted using a real-world Assembly dataset from chip packaging processes, which includes sensor signals such as current, voltage, deformation, and frequency. Results show that the proposed TSCL-FDC model outperforms traditional contrastive learning methods and other baselines across multiple metrics, including accuracy, precision, recall, and F1-score. The model demonstrates strong class discrimination and anomaly detection capabilities. Case studies further confirm its practical applicability in industrial environments, enhancing process monitoring and defect prediction. This research demonstrates that integrating frequency-domain signals into contrastive learning architectures, along with compressing multiple input modalities via knowledge distillation, significantly improves the model′s ability to handle high-dimensional sequential data in smart manufacturing applications. The proposed method offers substantial practical potential and extensibility for real-world industrial use.
Appears in Collections:	[Graduate Institute of Information Management] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	7	View/Open

社群 sharing

Loading...