隨著工業4.0革命改變傳統製造流程，製造業在生產過程中廣泛部署感測器以即時收集製程監控參數，旨在降低最終產品的瑕疵率。製造業重視故障檢測與分類（FDC）任務，以找出關鍵的製程錯誤原因。近年來，研究人員開始探索監督式對比學習（SCL），通過將原始數據轉換為具辨別性的特徵表示來優化模型，從而提高下游分類器的準確性。然而，類別不平衡的工業製程資料集、長序列製程資料的獨特性、以及SCL在處理特徵相似但屬於不同類別數據時的挑戰，導致模型訓練難以達到最佳效果。為了解決這一問題，我們提出了基於中心點集中特性的對比學習（C2CL）架構，在SCL的基礎上引入各類別樣本群的共同特徵作為附加資訊。實驗結果證實，基於C2CL架構訓練的編碼器能生成具類間分離和同類內集中特性的特徵表示。我們通過全面的實驗研究，證明了所提出的模型在所有評估指標上均優於目前最先進的基準模型，尤其在真實電子零組件供應商資料集上，F1-score指標提升了5.98%。;With the revolution of Industry 4.0 transforming traditional manufacturing processes, sensors are widely deployed in the manufacturing industry to collect real-time process monitoring parameters, aiming to reduce the defect rate of the final product. The industry focuses on fault detection and classification (FDC) tasks to identify key process error causes. In recent years, researchers have explored supervised contrastive learning (SCL), optimizing models by transforming raw data into discriminative feature representations, thereby improving downstream classifier accuracy. However, issues such as class imbalance in industrial process datasets, the unique nature of long sequence process data, and the difficulty of SCL in handling similar features that belong to different classes hinder optimal model training. To address this problem, we propose a Centroid-Concentrated Contrastive Learning (C2CL) framework, which introduces the common features of each class sample group as additional information based on SCL. Experimental results demonstrate that the encoder trained under the C2CL framework can generate feature representations with inter-class distraction and intra-class concentration. Through comprehensive experimental investigation, we prove that our proposed model outperforms the current state-of-the-art benchmark models across all evaluation metrics, especially achieving a 5.98% improvement in the F1-score on a real-world electronic component supplier dataset.
