在工業4.0浪潮驅動下，晶圓代工產業加速導入智慧製造技術，透過感測器佈建與資料驅動模型，實現對製程設備與產品品質的即時監控。隨著製程複雜度提升與良率要求日益嚴苛，缺陷偵測與分類（Fault Detection and Classification, FDC）成為關鍵技術。然而，實際製程中產生的高維時間序列資料，常伴隨噪音干擾、特徵重疊與類別不平衡，使傳統統計與機器學習方法難以穩定運作。本研究提出一套整合多模態特徵與語意對比策略的瑕疵辨識方法，將時間域與頻率域訊號特徵共同納入分析架構中，進一步透過樣本間的語意關係學習，提升異常樣本的可辨識性與類別分離度。為強化結構表徵，亦引入資料分群觀點輔助模型學習資料內部結構與變異趨勢。實驗顯示，本方法能有效降低誤判與漏報風險，在多種時間序列資料集上展現出良好的準確性與穩健性，提供一項具實務可行性的FDC智慧監控解決方案。;Under the momentum of Industry 4.0, the semiconductor foundry industry is rapidly adopting smart manufacturing technologies to achieve real-time monitoring of equipment conditions and product quality through sensor deployment and data-driven models. As process complexity increases and yield requirements become more stringent, Fault Detection and Classification (FDC) has emerged as a critical enabler in modern fabs. However, time-series data generated from manufacturing sensors are often high-dimensional, noisy, and affected by class imbalance and overlapping features, making traditional statistical or machine learning methods inadequate for reliable detection. This study proposes a defect identification approach that integrates multi-modal feature representations and semantic contrastive strategies. By jointly analyzing signals in both time and frequency domains, and learning semantic relationships between samples, the method enhances anomaly separability and representation clarity. A clustering-based perspective is further introduced to capture latent data structures and process variation patterns. Experimental results demonstrate improved accuracy and reduced false detections across diverse time-series datasets, validating the approach as a practical and scalable solution for FDC in semiconductor process monitoring.
