基於FDC之半導體晶圓切割設備零組件壽命預測架構設計

Design of an FDC-Based Framework for Component Life Prediction in Semiconductor Dicing Equipment

在半導體製造中，設備稼動率與製程穩定性高度仰賴關鍵零組件的健康狀態，零組件一旦突發失效，往往造成停機、良率下降與交期延誤等連鎖損失。傳統設備維護多仰賴固定週期保養或人員經驗判斷，常出現「尚未損壞即提前更換」或「已接近失效卻未及時處理」的兩難。本研究以晶圓切割設備（Dicing Equipment）為應用場域，提出一套以故障偵測與分類（Fault Detection and Classification, FDC）為核心的零組件壽命預測架構，使傳統被動維修得以轉向預測性維護（Predictive Maintenance）。該架構透過蒐集設備運行時的多源感測資料（如滾珠螺桿振動、電流訊號與設備運作參數），先以快速傅立葉轉換（FFT）將時域訊號轉換為頻域特徵並加以影像化，再以混合式模型進行預測：以支援向量迴歸（SVR）進行短期健康趨勢評估，並以深度學習模型進行長期劣化的異常分類，據以推估零組件的剩餘壽命（Remaining Useful Life, RUL）。本研究闡述其「資料收集—頻域特徵萃取—混合模型—即時監測」的四階段流程，並以一組示範性情境說明健康指標與RUL的判讀方式，同時提出對應的AI智慧製造人才培育模式。本架構的意義在於降低突發停機風險、優化維護時機，並建立可落地且能銜接教學的半導體設備預知保養應用模式。

In semiconductor manufacturing, equipment availability and process stability depend heavily on the health of critical components; a sudden component failure often triggers cascading losses such as downtime, yield decline, and delivery delays. Conventional maintenance relies on fixed-interval servicing or operator experience, leading to the dilemma of either premature replacement before failure or delayed response to imminent failure. Using wafer dicing equipment as the application domain, this study proposes a component life-prediction framework centered on Fault Detection and Classification (FDC), enabling a shift from reactive repair toward predictive maintenance. The framework collects multi-source sensing data during equipment operation (e.g., ball-screw vibration, current signals, and operating parameters), applies the Fast Fourier Transform (FFT) to convert time-domain signals into frequency-domain features and imagifies them, and then adopts a hybrid model for prediction: Support Vector Regression (SVR) for short-term health-trend assessment and a deep learning model for long-term degradation classification, thereby estimating the Remaining Useful Life (RUL) of components. A four-stage pipeline-data collection, frequency-domain feature extraction, hybrid modeling, and real-time monitoring-is elaborated, and an illustrative scenario explains how the health index and RUL are interpreted. A corresponding AI smart-manufacturing talent-cultivation model is also proposed. The framework aims to reduce unexpected downtime, optimize maintenance timing, and establish a deployable, education-linked paradigm for predictive maintenance of semiconductor equipment.