CCL-FDC：用於工業4.0之缺陷檢測與分類的協同對比學習架構;CCL-FDC: Collaborative Contrastive Learning for Fault Detection and Classification in Industry 4.0

NCUIR > School of Management at National Central University > Graduate Institute of Information Management > Electronic Thesis & Dissertation > Item 987654321/98286

Please use this identifier to cite or link to this item: https://ir.lib.ncu.edu.tw/handle/987654321/98286

Title:	CCL-FDC：用於工業4.0之缺陷檢測與分類的協同對比學習架構;CCL-FDC: Collaborative Contrastive Learning for Fault Detection and Classification in Industry 4.0
Authors:	劉冠甫;Liu, Kuan-Fu
Contributors:	資訊管理學系
Keywords:	工業4.0;深度學習;對比學習;缺陷檢測與分類（FDC）;Industry 4.0;deep learning;Contrastive Learning;Fault Detection and Classification (FDC)
Date:	2025-07-16
Issue Date:	2025-10-17 12:35:03 (UTC+8)
Publisher:	國立中央大學
Abstract:	由於感測與控制技術的重大進展，在工業4.0時代，缺陷檢測與分類（Fault Detection and Classification, FDC）已成為製造過程中的一項關鍵任務。近年來，對於對比式學習（Contrastive Learning, CL）技術的關注顯著增加，其目標是將資料轉換為具判別性的嵌入式表示。特別是在時間序列應用中，CL 能夠直接處理複雜的原始數據，以優化後續的檢測與分類效能。然而，由於工業環境的複雜性與多樣性，在工業時間序列場景中應用對比式學習仍面臨諸多挑戰。本研究提出了一種新穎的工業時間序列協同對比式學習框架 CCL-FDC，以解決工業應用中的多項難題。CCL-FDC 框架基於所提出的協同質心概念，同時訓練時間域與頻率域編碼器，以從時間與頻率兩個維度中同步提取並融合特徵，進而生成高度具判別力的嵌入式表示。我們在真實工業數據集上進行了大量實驗，驗證 CCL-FDC 在各項評估指標上相較於現有先進方法的優越性。我們亦透過案例研究展示該框架在智慧製造缺陷分類上的實用性。;Owing to the great advances of sensing and cybernetic technologies, in industry 4.0, Fault Detection and Classification (FDC) has become an essential task in manufacturing processes. Recently, significant attentions have been put forward to the contrastive learning (CL) technique, which aims to transform data into discriminative embedding representations. Especially on time-series application, CL could directly tackle the complicated raw data in models to optimize downstream detection and classification performance. However, due to the complexity and diversity of industry environment, applying contrastive learning in industrial time-series scenarios still faces various obstacles. In this study, we propose a novel Collaborative Contrastive Learning on FDC, CCL-FDC framework, which solves several challenges in industry applications. Based on the introduced concept of collaborative centroid-orientation, the CCL-FDC framework trains both time- and frequency-based encoders to generate highly discriminative embedding representations by extracting and integrating both features from time and frequency dimensions concurrently. Extensive experiments are conducted on real industrial datasets to demonstrate the superiority of CCL-FDC over the state-of-the-art methods in terms various evaluation metrics. We also use a case study to show the practicability of the proposed framework on fault classification in smart manufacturing.
Appears in Collections:	[Graduate Institute of Information Management] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	34	View/Open

社群 sharing

Loading...