基於圖離差網路模型之設備健康狀態預測框架

Abstract

智慧製造已成為近年來半導體產業中的焦點。先進製程控制（APC）系統對精進製造技術具有重要貢獻，而設備故障預測及健康管理（PHM）則是實現APC製程監測功能的關鍵，因此設備健康狀態預測任務具有極高的實用價值。然而在預測上存在一系列挑戰，最明顯的是能夠代表設備狀態的資訊往往難以取得，使得模型的建構及訓練受到阻礙。常見的線性多項式推估方法，無法滿足設備損耗的複雜行為。即使轉而使用設備感測器資料，也將面臨多變量時間序列複雜的時間特性及變數間交互作用的不確定性等問題，這使得傳統的統計特徵萃取方法在健康狀態預測任務中表現出明顯的局限性。更甚者，當預測更長期的狀態趨勢時，由於缺乏任何設備運作資訊，這類模型幾乎無法發揮作用。 為此本研究提出一基於深度學習模型的創新設備健康狀態預測框架，通過預測未來時間點的品質量測值，對設備健康狀態間接推論，以克服前述挑戰。模型始於處理複雜多變量時間序列，採用基於資訊聚合及注意力機制的圖離差網路（GDN），捕捉設備感測器資料之間的交互作用，同時藉由圖結構的自我學習能力，在訓練過程發現變數的潛在關聯性，萃取出代表整體系統的特徵圖。接著交由閘門循環單元（GRU）進一步萃取與時間順序相關之特徵。在架構中特別將由虛擬量測（VM）模型所預測之虛擬品質資料納入考量，確保有關製程與品質的時間特性皆被保留於最終特徵中，從而得到更全面的預測結果。以本框架為基礎執行之PHM，將可協助APC系統建立潛在故障反應及預警機制，減少產線意外中斷的風險，提升設備維護計畫的準確性，為半導體製造系統的效率及穩定做出貢獻。 本研究以半導體化學機械研磨（CMP）製程的公開資料集進行案例分析。針對CMP製程的材料移除率（MRR）進行預測，並與傳統時間序列模型及其他基線模型進行效能比較，經實驗證實本框架無論在品質全檢或抽樣檢測情況下皆能展現更佳的效能，並可藉由模型學習之圖結構確立感測器資料對預測結果的貢獻程度。

Smart manufacturing has emerged as a key focus in the semiconductor industry in recent years. Advanced Process Control（APC）systems play a critical role in enhancing manufacturing technologies, with Prognostics and Health Management（PHM）being key to enabling APC's process monitoring capabilities. Consequently, the task of predicting equipment health states holds significant practical value. However, this task faces numerous challenges, the most prominent being the difficulty in acquiring information that effectively represents equipment conditions, which hinders the construction and training of models. Conventional linear polynomial estimation methods are inadequate for capturing the complex behavior of equipment degradation. Even when utilizing equipment sensor data, challenges arise from the intricate temporal characteristics of multivariate time series and the uncertainties in interactions between variables, highlight the limitations of traditional statistical feature extraction methods in health prognostics tasks. Furthermore, when predicting long-term trends, the absence of operational information renders such models nearly ineffective. This study proposes an innovative deep learning-based framework for health prognostics. By forecasting quality measurement values at future time points, the framework infers equipment health states indirectly, addressing the aforementioned challenges. The model begins by processing complex multivariate time series, utilizing a Graph Deviation Network（GDN）based on information transfer and attention mechanism to capture the interactions among equipment sensor data. Through the self-learning capability of the graph structure, the model uncovers the underlying relationships between variables and extracts feature maps that represent the overall system. Subsequently, these features are further processed by Gated Recurrent Units（GRU）to extract temporal characteristics. Within this framework, virtual quality data predicted by the Virtual Metrology（VM）model is specifically integrated, ensuring that both process and quality temporal characteristics are preserved in the final features, leading to more comprehensive prediction results. PHM implemented based on this framework can assist APC systems in establishing potential failure responses and early warning mechanisms, reducing the risk of unexpected production line disruptions, improving the accuracy of equipment maintenance plans, and contributing to the efficiency and stability of semiconductor manufacturing systems. The case study using a publicly available dataset from the semiconductor Chemical Mechanical Polishing（CMP）process is conducted in this research. The framework predicts the Material Removal Rate（MRR）of the CMP process and compares its performance with traditional time-series models and other baseline models. Experimental results confirm that our proposed framework outperforms under both full inspection and sampling conditions, Moreover, the graph structure learned by the model can identify the contribution of sensor data to the prediction outcomes.