應用LightGBM模型於機台之預診斷與健康管理;Application of the LightGBM Model to Equipment Prognostics and Health Management

NCUIR > School of Management at National Central University > Graduate Institute of Industrial Management > Electronic Thesis & Dissertation > Item 987654321/97272

Please use this identifier to cite or link to this item: https://ir.lib.ncu.edu.tw/handle/987654321/97272

Title:	應用LightGBM模型於機台之預診斷與健康管理;Application of the LightGBM Model to Equipment Prognostics and Health Management
Authors:	陳玥瑄;Chen, Yueh-Hsuan
Contributors:	工業管理研究所
Keywords:	智慧製造;機器學習;LightGBM;預診斷與健康管理;Smart Manufacturing;Machine Learning;LightGBM;Prognostics and Health Management
Date:	2025-07-01
Issue Date:	2025-10-17 11:04:31 (UTC+8)
Publisher:	國立中央大學
Abstract:	近年來，隨著工業4.0與智慧製造的演進，設備管理逐漸朝向即時預測的預診斷與健康管理系統發展。透過物聯網感測技術與機器學習演算法整合，機台可即時蒐集運行資料，並藉由分析模型預測未來可能發生之異常事件，協助企業提前採取維護行動，降低停機風險與成本。本研究以A公司塗佈機之捲取A軸設備為對象，提出以LightGBM為基礎之異常預測模型，並延伸至異常風險分級與維護策略建構。針對資料不平衡問題，採用Tomek Link技術進行欠採樣，兼顧異常樣本之辨識能力與模型泛化性。模型訓練與驗證階段以異常類別召回率為評估指標，並選定最適參數組合，包含樣本權重倍率1.0、決策樹數量100、學習率0.1、L1正規化係數0.3。測試資料集異常類別召回率達100%、F1-score為96%、整體準確率99.95%；推論資料集異常類別召回率100%、準確率99.48%，惟異常類別精確率僅25.58%。研究結果顯示模型雖具潛在異常辨識能力，實務應用仍需搭配現場情境進行調整與驗證。特徵重要性分析以張力預覽變數對預測貢獻度最高，基於此結果，本研究設計三級預警制度，結合產線燈號提示與張力參數優先檢查建議，協助現場人員根據預測結果，採取更具針對性的應變行動，進而強化決策判斷與維修效率。;In recent years, advancements in Industry 4.0 and smart manufacturing have accelerated the development of predictive diagnostics and health management (PHM) systems with real-time anomaly prediction capabilities. By integrating IoT sensing technologies with machine learning algorithms, equipment can collect operational data in real time and utilize predictive models to identify potential anomalies, enabling enterprises to take proactive maintenance actions and reduce downtime risks and costs. This study investigates the winder A-axis of a coating machine at Company A and proposes an anomaly prediction model based on LightGBM, further extending to risk grading and maintenance strategies. To address data imbalance, Tomek Link under-sampling was applied to enhance anomaly recognition and maintain generalization. The model was trained with a focus on maximizing anomaly class recall, and the final parameters were configured as follows: scale pos weight of 1.0, 100 estimators, a learning rate of 0.1, and L1 regularization of 0.3. Experimental results show that the model achieved 100% anomaly class recall, a 96% F1-score, and 99.95% accuracy on the test dataset. On an unseen inference dataset, recall remained at 100% with 99.48% accuracy, while precision for anomalies was only 25.58%. These results suggest strong potential for high-sensitivity detection, though practical deployment requires adaptation to field conditions. Feature importance analysis revealed that the tension preview variable contributed the most to prediction. Based on this, a three-level warning system combining visual alerts and prioritized tension inspection was designed to help frontline operators take targeted actions and improve maintenance efficiency.
Appears in Collections:	[Graduate Institute of Industrial Management] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	1	View/Open

社群 sharing

Loading...