基於隱馬可夫鏈及生產排程之預測性設備健康度

Abstract

設備健康狀況的精準評估與管理，對於監控設備利用率至關重要，尤其考慮現代工廠的大量硬體投資。過往針對設備健康評估的研究多取決於生產期間產生的製程訊號，而這樣的依賴限制了在實際生產進行之前，將設備健康狀況考慮納入長期生產調度和規劃的可能性。本研究提出使用基於隱馬可夫模型及生產排程之預測性設備健康度(PEHMM)框架進行設備健康監測與預後分析，主要引入高斯混合隱馬可夫模型(GM-HMM)學習並預測製造設備的健康狀態行為變化。PEHMM包括兩個主要階段：離線行為學習和線上預後分析階段。在離線行為學習階段中，PEHMM使用歷史製造資料以非監督式學習方式訓練GM-HMM，不仰賴預先標記資料，以捕捉更全面的設備運作狀態。在線上預後分析階段，相較於傳統模型多仰賴虛擬量測(VM)或實時感測器資料，PEHMM可基於擬定之生產排程預測設備運行效能。 本研究透過兩個個案針對PEHMM進行驗證，包括軸承系統及離子磨蝕刻 (IME)系統。此系列實驗透過MES和FDC資料訓練非監督的行為模型，而不依賴虛擬量測之結果。這意味著即使無法取得實時的FDC數據，PEHMM仍能有效地預測設備運作狀況，並透過Hotelling T2分解分析特徵的貢獻。實驗結果驗證了PEHMM作為先進製程控制(APC)實用工具的適用性，並期能支援批次調度、設備利用率和產量監控等重要決策流程，有助於提高製造效能。

The meticulous estimation and management of equipment health are pivotal for monitoring equipment utilization, especially given the substantial hardware investments in contemporary factories. The majority of research in equipment health estimation hinges on process signals generated during physical operations. This reliance limits the integration of equipment health considerations into long-term production scheduling and planning, as physical operations have not yet occurred. This research proposes the Predictive Equipment Health Monitoring using Hidden Markov Models and Production Scheduling (PEHMM) framework, employs Gaussian Mixture Hidden Markov Models (GM-HMM) to learn and predict the operational behavior of manufacturing equipment. PEHMM includes two primary phases: offline behavior learning and online prognostic phase. During the offline behavior learning phase, PEHMM uses historical manufacturing data to train the GM-HMM in an unsupervised manner, capturing a comprehensive range of equipment operating states. This phase is instrumental in unraveling the complex dynamics of equipment without relying on pre-labeled data. The online prognostics phase pivots to real-time prediction, applying insights from offline analysis to predict equipment health with or without real-time sensor data. The PEHMM framework was tested in two case studies involving bearing systems and the Ion Mill Etching (IME) system. MES and FDC data were used to develop an unsupervised behavioral model, which marks a departure from conventional approaches that predominantly rely on virtual metrology (VM). The PEHMM framework was effective in predicting equipment health even when real-time FDC data was not available, emphasizing its suitability as a practical tool in advanced process control (APC). The contribution of features is analyzed through Hotelling T2 decomposition. The framework is expected to improve decision-making processes in areas such as lot scheduling, equipment utilization, and production yield monitoring, contributing to more efficient and informed manufacturing operations.