| 摘要: | 本研究以工業4.0與智慧製造興起之背景為出發點，製造業中確保生產連續性與降低維護成本成為關鍵課題，而隨著物聯網與感測器在各產業中普及，企業能以更低成本收集大量數據，並透過大數據與人工智慧方法，制定主動式預測性維護(Predictive Maintenance, PdM)策略。
A公司塗佈機具備大量感測器數據，其資料集具有時間序列性、高維度與異常樣本不平衡問題，而本研究針對這些特性，採用自動編碼器(Autoencoder)搭配極限梯度提升決策樹(eXtreme Gradient Boosting, XGBoost)，提出一數據驅動之異常預測模型。先以自動編碼器對原始高維特徵進行壓縮並提取關鍵特徵，再將瓶頸層輸出作為XGBoost模型輸入進行監督式訓練，透過其二階導數、貪婪演算法，學習各特徵與異常之間的複雜非線性關係，使模型在異常辨識上有一定程度之準確性與穩定性。
本研究完整呈現資料前處理流程與模型建構過程，透過系統化的架構測試與超參數調整，最終選擇最佳參數組合：學習率0.005搭配決策樹數量800、最大深度3與調整異常樣本權重為6，並於測試集上獲得99.99%的整體準確率，其餘各項評估指標皆大於0.92，在精確率與召回率間取得最佳平衡，後續模型於推論階段也具備良好判別力。而為了強化實務應用性，進一步設計健康指數(Health Score)計算方法作為設備健康狀態衡量指標，並制定三級維護策略，能依據不同程度採取相應措施，有效支援預測、監測與決策支援三大功能，實現預診斷與健康管理(PHM)完整概念。
;This study is motivated by the emergence of Industry 4.0 and smart manufacturing, where ensuring production capacity and reducing maintenance costs have become critical challenges for the manufacturing industry. With the widespread adoption of IoT and sensor technologies across industries, enterprises can now collect large volumes of data at low cost and leverage big data analytics and artificial intelligence techniques to implement proactive Predictive Maintenance (PdM) strategies.
Focusing on a coating machine from Company A, which generates a high-dimensional, time-series dataset with a significant class imbalance, this research proposes a data-driven anomaly prediction model based on the integration of an Autoencoder and eXtreme Gradient Boosting (XGBoost). The Autoencoder is first used to compress the original high-dimensional features and extract latent representations, which are subsequently used as input for the supervised XGBoost model. Through second-order optimization and an exact greedy algorithm, XGBoost effectively captures the complex nonlinear relationships between features and anomalies, yielding reliable and accurate anomaly detection performance.
This study presents a comprehensive workflow including data preprocessing and model construction. Through systematic architecture testing and hyperparameter tuning, the optimal parameter combination was determined, consisting of a learning rate of 0.005, 800 decision trees, a maximum tree depth of 3, and an anomaly sample weight adjustment of 6. The model achieved an overall accuracy of 99.99% on the test dataset, with all other evaluation metrics exceeding 0.92, demonstrating an optimal balance between precision and recall. To enhance practical applicability, we designed a Health Score calculation method to serve as an indicator of equipment condition and established a three-tiered maintenance strategy to effectively support prognostics, monitoring, and decision-making, thereby realizing the complete concept of Prognostics and Health Management (PHM). |
