200928843 九、發明說明： 【發明所屬之技術領域】 本發明是有關一種虛擬量測系統，特別是一種具有新穎性偵 測與自適性學習的虛擬量測系統。 【先前技術】 於先進半導體生產製程中，為確保合乎品質規格的高良率和 高穩定性產出。皆會插入乙片以監控為目的之測試晶圓，進行週 〇期性ασ貝里測，然而，測試晶圓在尚未量測之前，無法即時察覺 .機台發生性能漂移的問題，這可能導致產出有瑕疵的晶圓，導致 最終產品作廢，造成極高的成本損失。 目珂的晶圓廠和面板廠大部分仍是採用統計製程管制 (Statistical Process Control，SPC)技術來監控重要的製程參數，並以 例行測機、機台預防保養等方法提高製程穩定性，改善製程良率。 然而，測機或後續檢驗與量測的結果無法立即產生，故這種監控 ❹=式對於無預警式H統良率損失而言，當製程巾發生異常狀況 π就，可忐造成大$數目的基板報廢，嚴重地影響到生產成本 與產能。 為了提升設備的製程能力’遂有網路式診斷(e_Diagn〇stics)、 網路式預防保養(e-Maintenance)和虛擬量測(Virtual MetiOl〇gy， 谓)等觀心被提出，其中虛擬量測的概念便*以網路式預兆偵測 之技術’結合先進半導體技術，來達成即m測之目標，克 服上述問題。VM沒有實際的量測行為，而是使用生產過程中所 記錄的製程參數資料，有效地去翻每—片晶_品質。藉由生 200928843 產過程所收集的製裎參數資料，VM的預測能力可以立即得知晶 圓的品質和設備異常現象，並且可以達到即時監控，實現逐片 (Wafer-To-Wafer)檢測品質的目標，並將良好的晶圓繼續送往下一 步製程處理。 在已公開的台灣發明專利TW200717592「量測方法、虛擬量 測系統以及電腦可讀取媒體」（美國同案專利20070100487)，提出 了〆種適用於半導體製造的虛擬量測系統與方法，主要是利用類 神經模糊法則預測量測的結果。但是神經元架構及相關參數的決 ❹定缺乏糸統化的準則、區域性極值的發生及其難以解釋的神經模 如都是疾人5后病的缺點。 另外在已公開的台灣發明專利TW200619886「即時預估量測 系疵内整合製程資訊以及用以預測一虛擬量測工具中至少一個輸 出的方法」（美國同案專利20060129257)，則提出了 一種用以運作 /製糕工具的即時預估量測系統架構，包括一資料擷取系統、一 虡採*測系統、一錯誤偵測與分類系統以及一先進製程控制系 Q鍊^雖然以上的兩件發明技術都揭露了—種虛擬量測系統及方 法，俱並未揭示應用於CVD設備的技術，也未提出適應性學習的 技『 . 在已核准的台灣發明專利U67012「生產製程之品質預測系 疵與方法」，其中提出了一種應用於生產製程之品質預測的系統 與方法’基本上是利用類神經網路預測量測結果，利用權重移動 乎均’藉由先前實際量測值與模型預測值推估下一批產品之品 質；雖然，本件發明技術還包含有自我搜尋裝置與自我調適裝置， 200928843200928843 IX. Description of the Invention: [Technical Field] The present invention relates to a virtual measurement system, and more particularly to a virtual measurement system with novel detection and adaptive learning. [Prior Art] In the advanced semiconductor manufacturing process, high yield and high stability output are guaranteed to meet the quality specifications. All of them will be inserted into the test wafer for monitoring purposes, and the weekly ασ Berry measurement will be performed. However, the test wafer cannot be detected immediately before the measurement is performed. This may cause the performance drift of the machine. Producing flawed wafers results in the end product being voided, resulting in extremely high cost. Most of the fabs and panel factories that are witnessing still use statistical process control (SPC) technology to monitor important process parameters, and improve process stability by routine testing and machine maintenance. Improve process yield. However, the results of the tester or subsequent inspection and measurement cannot be immediately generated. Therefore, this type of monitoring ❹= for the loss rate of the unwarranted H system, when the process towel has an abnormal condition π, it can cause a large amount of $. The scrapping of the substrate seriously affects production costs and capacity. In order to improve the process capability of the device, there are network diagnostics (e_Diagn〇stics), network-based preventive maintenance (e-Maintenance) and virtual measurement (Virtual MetiOl〇gy, verb), etc. The concept of measurement will be based on the technology of networked precursor detection combined with advanced semiconductor technology to achieve the goal of m measurement and overcome the above problems. The VM does not have actual measurement behavior, but uses the process parameter data recorded in the production process to effectively turn each slice-quality. Through the data of the enthalpy parameters collected during the production process of 200928843, the predictive power of the VM can immediately know the quality of the wafer and the abnormality of the equipment, and can achieve real-time monitoring to achieve Wafer-To-Wafer quality. Target, and send good wafers to the next process. In the published Taiwan invention patent TW200717592 "measurement method, virtual measurement system and computer readable medium" (American Patent No. 20070100487), a virtual measurement system and method suitable for semiconductor manufacturing are proposed, mainly The results of the measurement are predicted using a neuro-fuzzy-like rule. However, the determination of neuron architecture and related parameters lacks the standard of serigraphy, the occurrence of regional extremum and the unexplained neurological model are all shortcomings of the disease. In addition, in the published Taiwan invention patent TW200619886 "Integrated process measurement system for real-time estimation measurement system and method for predicting at least one output of a virtual measurement tool" (US Patent No. 20060129257), a use is proposed. The real-time predictive measurement system architecture of the operation/fabrication tool includes a data acquisition system, a measurement system, an error detection and classification system, and an advanced process control system Q chain. The invention discloses a virtual measurement system and method, which does not disclose the technology applied to the CVD equipment, and does not propose the technology of adaptive learning. In the approved Taiwan invention patent U67012 "Quality prediction system of the production process" "疵 and method", which proposes a system and method for quality prediction of production process 'basically using neural network to predict measurement results, using weights to move both' by previous actual measurements and model predictions The value estimates the quality of the next batch of products; although, the invention also includes a self-searching device and self-adapting device, 200928843
伙原始數據(即製程參數)中萃取代表性 機制以及提昇量崎度的方法。 般而σ T導體業製程所產生的資料量非常龐大，表格綱 要(Table Schema)亦相當繁雜，因此都需要利用資料庫系統進行資 料儲存與㈣纽1料歧大的資料量若不經碱理會造成建 Q 置虛擬量測系統效率不彰的結果。 虛擬量_速度取決於幾_鍵：⑴观與重姻練最隹夕 虛擬量測模型的速度(2)虛擬量測模型的預測速度⑶具代表性: 特徵(Featu㈣的萃取與資料量縮減。在過去的贿中，設備參數 種類的選取多獅經驗法則、或是簡單的觀察觸。然而，以這 樣直觀非量化崎轉取方式所触徵不太可祕預測準確 率最佳化，即使麵量測模型建立的非f好，但準確率還是會直 接受所選取的特徵限吿丨丨，&钮、土法~ θA method for extracting representative mechanisms and increasing the amount of sag in the raw data (ie, process parameters). The amount of data generated by the σ T conductor industry process is very large, and the Table Schema is quite complicated. Therefore, it is necessary to use the database system for data storage and (4) the data volume of the New Zealand material is not basic. The result is that the built-in virtual measurement system is inefficient. The virtual quantity _ speed depends on a few _ keys: (1) the speed of the virtual measurement model of the view and the re-marriage practice (2) the prediction speed of the virtual measurement model (3) is representative: the feature (Featu (four) extraction and data reduction). In the past bribes, the choice of equipment parameters is more than the lion's rule of thumb, or a simple observation. However, the use of such an intuitive non-quantitative rugged approach is less predictable, and the accuracy is optimized even if The non-f established by the measurement model is good, but the accuracy rate is directly limited by the selected feature, & button, soil method ~ θ
本發明的目的之—在提出—種具有新紐躺與自適性學 用於蒐集CVD設備的製程資料並進 習的虛擬量測系統，包括彳： 一製程資料蒐集模組，戶 行統計製程檢额動情處理以麵絲f料的正確性和完替 性，接剌龍域分分析(KPCA)萃取代表性細倾並調^ 200928843 度範圍以取得特徵資料集合，以供後續的模型建立以及品質預測 之用； 一虛擬量測模組(GA-SVR Module)，用以訓練及建置一品質 預測模型’係利用基因演算法(Genetic Algorithms)對支援向量回歸 模型(Support Vector Regression Model，SVR Model)進行最佳化之 模型參數選取’在訓練階段可自動地完成支援向量回歸模型(SVR model)參數的最佳化選擇’並且建置SVR模型(SVR model)以作為 〇預測品質的品質預測模型，以便在對CVD設備進行虛擬量測的過 程中，依據即時性的製程資料輸出品質預測的結果； 一新穎性偵測模組(Novelty Detection Module)，在訓練過程中 利用支持向量資料描述(Support vector data description, SVDD)建 立一新穎性偵測模型(Novelty Detection Model),而在虛擬量測過 程中對萃取獲得的特徵資料集合進行分類以及辨識是否具有新的 特徵資料出現，並且將新穎性辨識的結果作為後續之虛擬量測可 信度與是否重新訓練新穎性镇測模型和品質預測模型的判斷依 @據；以及 —增量學習模組(Incremental Learning Module)，係在新穎性 偵测模組發現新的特徵資料出現時啟動這個增量學習模組，對新 賴性偵測模型及品質預測模型$行重新訓練，使重新訓練後的新 祕偵峨賴品質测翻可以適應動鍾化 確的預測。 兄做出/丨'月 _本發明的目的之—在減少訓練資料量，提昇虛擬量測系統的 虛擬量測速度。本發明的製程資料蓋集模組採用核主成分分析 200928843 (KPC A)進行特徵萃取的動作， ^ 用非線性映射(N〇niineai. 間，亡間先行映射至更高維度的特徵空 錢趨錄^制；料，用域分分析_ W工間中將座標轉換並同時進行維度縮減。不但可以〜 擬量測的精度，也可以大幅減少輸人的特徵數目，進而達 虛擬量測的目的。 ❹ 錢树明之詳細·無佳實關1配合圖輯細說明 如下，其内容足以使任何熟習侧技藝者了解本發明之技術内容 並據以實施，錄據本綱#所揭露之内容及圖式，任何熟習相 關技藝者可輕易地理解本發明相關之目的及優點。 【實施方式】 本發明以下列舉之實施例僅用於說明本發明之目的與較佳 的實施例，並非用以限制本發明之範圍。 首先請參閱「第1圖」，係為本發明所提出之運用於化學氣 相沉積設備的虛擬量測系統的一較佳實施例，包括有： 一製程資料蒐集模組20’首先在CVD設備10裝設有數個感 測器負責蒐集製程資料，這些製程資料係經由錯誤偵測與分類 (Fault Detection and Classification)處理蒐集獲得的原始製程資料 (FDC Raw Data)，這些原始製程資料會被存入第一資料庫η成為 歷史製程資料，而製程資料蒐集模組2 0將會擷取這些歷史製程資 料進行資料前處理和統計製程檢驗(Statistical Process ControlThe object of the present invention is to provide a virtual measurement system with a new lie and self-adaptability for collecting process data of a CVD device, including: a process data collection module, and a household statistical process check The estrus treatment is based on the correctness and completeness of the surface material, and the KPCA extracts the representative fine pitch and adjusts the range of 200928843 degrees to obtain the feature data set for subsequent model establishment and quality prediction. A virtual measurement module (GA-SVR Module) for training and building a quality prediction model 'Support Vector Regression Model (SVR Model) using Genetic Algorithms Optimized model parameter selection 'automatically complete the optimization of the support vector regression model (SVR model) parameters in the training phase' and build the SVR model (SVR model) as the quality prediction model for predictive quality. In order to perform virtual measurement on the CVD equipment, the quality prediction result is output according to the instantaneous process data; a novelity detection module (Novelty Detection M) Odule), using the support vector data description (SVDD) to establish a Novelty Detection Model during the training process, and classifying the feature data sets obtained by the extraction in the virtual measurement process and Identify whether there is new feature data, and use the result of novelty identification as the subsequent virtual measurement credibility and whether to re-train the novelity of the township model and the quality prediction model; and - incremental learning mode The Incremental Learning Module activates the incremental learning module when the novelty detection module finds new feature data, and retrains the new detection model and the quality prediction model. The new secret detective depends on the quality of the test can be adapted to the prediction of the clock. The brother made / 丨 'month _ the purpose of the invention - to reduce the amount of training data, improve the virtual measurement speed of the virtual measurement system. The process data cover module of the present invention adopts nuclear principal component analysis 200928843 (KPC A) for feature extraction, and uses non-linear mapping (N〇niineai., the first to map to the higher dimension of the feature empty money trend Recording system; material, using domain analysis _ W work in the coordinate conversion and dimensional reduction at the same time. Not only can the accuracy of the quasi-measurement, but also can greatly reduce the number of features of the input, and thus achieve the purpose of virtual measurement ❹ The details of Qian Shuming and the details of the map are as follows, and the content is sufficient for any skilled person to understand the technical content of the present invention and implement it according to the contents and drawings disclosed in this section. The objects and advantages of the present invention are readily understood by those skilled in the art. The embodiments of the present invention are intended to illustrate the objectives and preferred embodiments of the present invention. Scope of the Invention First, please refer to "FIG. 1", which is a preferred embodiment of the virtual measurement system for chemical vapor deposition equipment proposed by the present invention, including: The data collection module 20' first installs a plurality of sensors in the CVD apparatus 10 to collect process data, and the process data is collected through the Fault Detection and Classification process (FDC Raw Data). ), the original process data will be stored in the first database η as historical process data, and the process data collection module 20 will retrieve the historical process data for data pre-processing and statistical process control (Statistical Process Control)
Checking, SPC Check)以確保原始資料的正碟性和完整性，接著利 10 200928843 用核主成分分析(KPCA)萃取出特徵資料並調整尺度範圍； 一第二資料庫30，用以儲存訓練過程和虛擬量測過程中的資 料(包括製程資料蒐集模組20萃取所得的特徵資料集合)以及虛擬 量測系統的組態資訊(configuration setting)，訓練完成的品質預測 模型和新穎性偵測(Novelty Detection)模型。 一虛擬量測模組(GA-SVR Module)40，甩以訓練及建置一品 貝預測模型’係利用基因演异法(Genetic Algorithms)對支援向量回 ❹歸模型(Support Vector Regression Model, SVR Model)進行最佳化 的模型參數選取，在訓練階段可自動地完成支援向量回歸模型 (S m〇del)參數的最佳化選擇，並且建置s vr模型（SVR model)，而這個§^模型就是本發明中作為預測品質的品質預測 模型，以使在對CVD設備進行虛擬量測的過程中，依據即時性的 製程貧料輸出品質預測的結果，而這個品質預測結杲、SVR模型 及其參數都被送至第二資料庫30之中儲存；以及 ❾一新穎性偵測模組(Novelty Detection Module)50,在訓練過程 中々·!用支才寸向 1 資枓描述(SUpp0rt vector data description, SVDD) L立新穎性偵測模型(N〇veity Detection Model)然後被送至第 一貝枓庫30之中儲存，而在虛擬量測過程中對萃取獲得的特徵資 斗集合進行分類以及辨識是否具有新的特徵資料出現，並且將新 ^性辨識的結果作為後續之虛擬量測可信度與是否重新繼新賴 測模型和品f預職型的觸依據，在虛擬制過程中冤集 ^即時性的製程資料經由製程轉荒減組2G進行鱗製程檢 ㈣貝斜減理’再映射至相同的做間與尺度難 么 200928843 由^♦型進行新穎性辨識，辨識結果若是落於超球體 (換5之就疋未辨識出新的特徵資料)，就將即時性的製程 型然嶋品細值，觸結果若是落於超球體夕^ 之’，尤、辨識出新的特徵資料）’赌新的特徵資料送至增量學習 、且p X便#惰雜制模型及SVR^(即品質测模型)進行 重新训練， 一增I學習模組(Increme麻1 Learning Module)60，在新穎性 ❹齡齡5G發_的特徵資料時被啟動，依騎雛偵測模組 )斤4見之新的4寸徵賁料對新穎性偵測模型及svr模型(即品 預測模型)進行重新訓練。 、 利用本發明仕「第1圖」所揭露的虛擬量測系統，'對CVD 私進仃虛婦_方法包括：爾驗以及虛擬制過程兩大 部份〇見在配合「第2圖」及「第3圖」說明如次： _ 4練過(見「第2圖」），其中實線表示訓練的路徑， 虛線表示訓練資料的傳遞路徑，训練的過程包括： L &貝㈣集模組2Q的處理過程：製程資料t集模组 録CVD設備1G裝置有數個感·，這些感·的較佳例子包 括但不限疋於.TEOS流量感測器、沉積壓力感測器、沉積時間 感測器、溫度感測器與P]y[保養__彳等，如進行製程責料 的荒集，這些製程資料係經由錯誤債測與分類(F滅Detecti〇nand Classification)所荒集獲得的原始製程資料(FDC D_,這些原 始製程貧料會被存入第-資料庫u成為歷史製程資料，而製程資 料荒集模組Μ的處理步驟包括1先雛前述的歷史製程資料 200928843 21;進行資料前處理22;和統計製裎檢驗(Siatistical pr〇eess cQnttQlChecking, SPC Check) to ensure the authenticity and integrity of the original data, then extract the feature data and adjust the scale range using Nuclear Principal Component Analysis (KPCA); a second database 30 for storing the training process And the data in the virtual measurement process (including the feature data set extracted by the process data collection module 20) and the configuration setting of the virtual measurement system, the quality prediction model and the novelty detection (Novelty) Detection) model. A virtual measurement module (GA-SVR Module) 40, which trains and builds a product prediction model, uses the Genetic Algorithms to support the Vector Regression Model (SVR Model). Optimized model parameter selection, the optimization of the support vector regression model (S m〇del) parameters can be automatically completed in the training phase, and the s vr model (SVR model) is built, and the §^ model In the present invention, as a quality prediction model for predicting quality, in the process of performing virtual measurement on a CVD apparatus, the quality prediction result, the SVR model and the quality prediction result thereof are based on the instantaneous process yield quality prediction result. The parameters are sent to the second database 30 for storage; and a Novelty Detection Module 50, during the training process, is used to describe the asset description (SUpp0rt vector data description) , SVDD) The L〇veity Detection Model is then sent to the first beta library 30 for storage, and the feature acquisition set obtained by the extraction is performed during the virtual measurement process. Class and identification whether new feature data appears, and the result of the new identification is used as the follow-up virtual measurement credibility and whether to re-take the new measurement model and the pre-position of the product f, in the virtual process The process data of the 冤 ^ 即时 即时 即时 即时 即时 即时 即时 即时 即时 即时 即时 四 四 四 四 四 四 四 四 四 四 四 四 四 四 2009 2009 2009 2009 2009 2009 2009 2009 2009 2009 2009 2009 2009 2009 2009 2009 2009 2009 2009 2009 2009 2009 2009 2009 2009 2009 2009 2009 2009 Falling in the supersphere (when you change the 5, you don't recognize the new feature data), you will immediately use the process value of the process, and if the result falls on the supersphere, you can identify the new one. Feature data) 'Bet new feature data sent to incremental learning, and p X will # inertial model and SVR^ (ie quality measurement model) for retraining, an increase in I learning module (Increme Ma 1 Learning Module 60), when the novelty age 5G hair _ characteristic data is activated, according to the riding detection module), the new 4 inch levy material for the novelty detection model and the svr model (ie, the product) Predictive model) for retraining. Using the virtual measurement system disclosed in the "Picture 1" of the present invention, the method of "individual CVD private 仃 _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ "Picture 3" shows the following: _ 4 practiced (see "Figure 2"), where the solid line indicates the training path, and the dotted line indicates the transmission path of the training data. The training process includes: L & Bay (four) set The processing of the module 2Q: the process data t set module recording CVD equipment 1G device has several senses, and preferred examples of these feelings include but not limited to .TEOS flow sensor, deposition pressure sensor, deposition Time sensor, temperature sensor and P]y [maintenance__彳, etc., such as the waste of process responsibilities, these process data are ruined by false debt measurement and classification (F Detecti〇nand Classification) The obtained raw process data (FDC D_, these original process poor materials will be stored in the first-database u become historical process data, and the process data collection module Μ processing steps include the first historical process data 200928843 21 ; pre-processing data 22; and statistical test (Siatistical) Pr〇eess cQnttQl
Checking, SPC Check)23，以確保原始資料的正確性和竽整性；接 者進行核主成分分析(KPCA)24 ;進而萃取出特徵資彳#g ;再作調 主尺度範圍26，而上述過私中產生的資料也都會被送至第二資料 庫30中儲存； 2·虛擬量測模組40的處理過程包括：對來自於製程資料 蒐集模組20中已完成調整尺度範圍的特徵資料集合進行格式化 0 41，开>成一卽丨練資料組，调用基因演算法48選擇的SVR模型 參數43 ;訓練支援向量機(Support Vector Machine,SVM)44 ;測試 文谖向里機45 ’進彳亍適應度评估46 ’在符合—預設之終止條件時 對SVR模型進行最佳化處理47並儲存至第二資料庫3〇之中，以 便在虛擬制過程中以這個SVR難作為本發_品質預測模 型，以及在不符合一預設之終止條件時運行基因演算法48重新對 SVR模型芩數進行最佳化選擇’再使用重新選擇好的模型參 數重覆前述43至48的步驟’直至用以預測品質的SVR模型被建 ®置完成，而另-較佳的實施例，更包括利用一預設的測試資料421 送入新建置完成的SVR模型，再經由前述的步驟45至步驟你進 行測試； 3‘ 新顆性偵測挺組50的處理過程包括：對來自於夢释資 料荒集模組20中已完成調整尺度範圍的特徵資料集合進行於式 化51 ;形成一訓練資料組52 ;調用透過基因演算法％所撰擇的 SVDD模型參數5：?，§)丨丨練SVDD模型54 ;測試svDD汽刑55 . 進行適應度評估56 ;在符合一預設之終止條件時對模型進 13 200928843 行最佳化處理57並儲存至第_ _ , ^ _ 乐—貝斜庫之中，以便在虛擬量測 〇 Μ作為本發明的新類性偵測模型，·以及在 不f、預叹之'，、止條件化運行基因演算法58重新對SVDD模 ^數1ΓΤ取佳化選擇，再使用重新選擇好的SVDD模型參數重 夜月)之至）8的少驟’直至新藉性偵測模型（即模型)被 建置完成，㈣-齡的實關，更包括_—預設的測試資料 52!送入新建置完成的新穎性偵測模型(即svdd模型），再經由前 述的步驟：)5至步驟58進行測試； p ^ 壬至此為止就已經完成了品質預測模型和新顆 性摘測模型_練與建置，然後就可以被獅於對㈣設備ι〇 進行品質預測。 二、虛擬量測過程(見「第3圖」），其中實線表示虛擬量測 的路徑，虛線表示量測過程中即時性的製程資料的傳遞路徑，虛 操量測的過程包括： i.氣耘資科蒐集模組20的處理過程包括：擷取CVD設備 0 1〇在生產過程中即時性的製程資料27並調用訓練過中所儲存在 第二貢料庫30的組態設定(c〇nfigurati〇n Setting鳩案(内含在訓練 階段時由製程資料蒐集模組20所產生的各種组態設定，如資料前 處理，KPCA分析以及調整尺度範圍等相關處理的設定值），在相 同的組態設定條件下進行資料前處理22 ;接著利用核主成分分析 (KPCA)24萃取出代表性特徵資料25 ;將代表性特徵資料25映射 炱與訓練過程中相同的座標軸空間與尺度範圍26 ;然後直接傳送 炱新穎性偵測模組50，進行新穎性辨識； 14 200928843 置完成的處理過程包括：__程中建 且兀成的新砸谓測模型（即為訓練過 SVDD模型)57，對 τ巳每直凡成的取佳化 圍代表性特徵麵2== 26 域標軸如與尺度範 内(換言之就是未辨識顏性辨識’辨識結果若是落於超球體 荒集模⑽處理(Γ^Γ 細)，就將經過前述製程資料 虛擬量概二步驟26)之㈣'陶程—送至 新的特η， 务是落於超球體外(換言之就是辨識出 ❹，二亡新娜徵資料送至增量學f模組60; I板夏測模'組40的處理過程：利用訓 品質預測模型(即最佳化SVR 脾 R置凡成的 2〇處理(步驟^牛㈣、、—)U —賢料兔集模組 測後值，、隹二：：)之即時性的製程資料27轉換為虛擬量 則钛值進叩輸出品質預測的結果； 4丄增量學習模組60的處理過程：本發明的増量學習模組 6〇只啕在辨識出新的特徵資土、，主 '、、” 咖错才會姐動，峨過前述製程資 ❾㈣處理(步驟22至步驟26)之即時性的製程資料π， 曰刀々？运至新賴性_模組之重新訓練機制63，和—瓢 輪64，綱麵鞋細輸偵測模組 |心里顺組40 ’使用經過前述製程資料荒集模組％處理 至步驟26)之即時性的製程資料27，對新嶺性偵測二 虛擬量職組4Q進行減訓練，仙產生新⑽酿侧 杈型和SVR模型(即品質預測模型）（過程如前述「第 以達到自適性的功效。 」斤不） 依據本發明所揭露之技術，當系統在虛擬量測過程中因為發 15 200928843 見·^的何彳^資4而被通知該重新訓練模型以面對已發生變化的設 備參數後，之前所建構的SVR翻和新穎性偵測模型就會依照= 7開也重新料。為提升制速度並逐步累積新穎性偵測模組5〇 ，、虛被里測棋組4〇 |習的知識，本發明除了採用核主成分分析 (KPCA)進辟_度_，更咖f量學細㈣ 、技1藉由SVM所採用之結構風險最小化的原理，將〇奶設 備10運轉時所荒集的新特徵資料集合（即由新賴模組所發現的新 ❹特徵資料），結合之前訓練集合所推論出的少數支持向量(3卿⑽ Vector)構絲__合，重_輯麵侧额％與虛擬量 測模組40分別產生新的新紐偵測模型及svr模型(即品質預測 知土）如此-來’里新訓練後產生的新穎性偵測模型與品質偵測 模型，就能適應動態變化的環境，做出精確的預測。 、、 而依據本發明的較佳實施例之一(見「第4圖」），前述的增 量學習模組(I職職tal Leammg M〇她)6〇，還具有一統計製雜 制(S加iStical Process Contol，SPC)的檢驗機制，用以對經過前述製 程資料荒集模組20處理(步驟22至步驟26)之即時性的製程資料 27進行檢驗，檢驗通過之即時性的製程資料& |分別被送至— 新穎性侧模組之重新爾機制63，和—SVR模型之*新 機制64，再分別重新送至前述的新穎性侧模组％及虛擬量測 模組40進行訓練，用以重新產生新顯性偵測模型及svr模型(即 品質預測模型），未檢驗通過者則啟動一驗證機制(ν_ύ〇η)65 , 如可用以通知現場工程_行機妓備的魏校驗，雛賴異 16 200928843 綜上所述，本發明所揭露之運用於化學氣相沉積設倩的虛擬 量測系統，利用基因演算法和支援向量回歸(ga_svr)之分析方法 建立品質預測模型，並輔以新穎性偵測模組5〇，使得本發明的虛 擬量測系統可以達_態適麟糾，並減提供準確之預 測結果的功效。另-方面，本明所揭露的虛擬量測线，採用核 主成分分析(KPCA)進行特徵萃取的動作，不但可以提高虛擬量測 的精度，也可以大幅減讀人的特徵數目，進而達壯速虛擬量 ^ 測的目的。 Ο 雖然本發明以前述之較佳實施例揭露如上，然其並非用以限 之本毛明’任何熟纽-領域之技藝者，在不脫離本發明之精神 和範圍内’所為之更動與顯，均縣發明之專·護範圍，因 此本發明之專娜護翻馳本朗書_之_請翔範圍所界 定者為準。 [圖式簡單說明】 第1圖’為本發明之线架構的—較佳實施例。 第2圖’為本發明之虛擬量測系統的運作流程圖，顯示進行刺練 是作流程圖，顯不進行虛擬 第3圖，為本發明之虛擬量測系統的 量測的過程。 第4圖，為本發明之系統架翻另—較佳實施例。 【主要元件符號說明】 10 CVD設備 11第一資料庫 17 200928843 20製程資料蒐集模組 21歷史製程資料 22 資料前處理 23 統計製程檢驗 24核主成分分析(KPCA) 25特徵資料 26調整尺度範圍 ^ 27即時性的製程資料 30第二資料庫 40 虛擬量測模組(GA-SVR Module) 41 格式化 42 訓練資料組 43調用SVR模型參數 44 訓練支援向量機(Training Support Vector Machine, SVM) 4 5測試支援向量機 ® 46適應度評估 47對SVR模型進行最佳化 48基因演算法 50 新穎性偵測模組(Novelty Detection Module) 51 格式化 52 訓練資料組 53 調同SVDD模型參數 54 訓練SVDD模型 18 200928843 55 測試SVDD模型 56 適應度評佑 57對SVDD模型進行最佳化 58基因演算法 60增量學習模組 61 統計製程控制(StatisticalProcess Control, SPC)的檢驗機制 62被檢驗迫過的即時製程貢料 0 63 新穎性偵測模組之重新訓練機制 64 SVR模型之重新訓練機制 65 驗證機制(vaiidBtioii) ❹ 19Checking, SPC Check) 23 to ensure the correctness and refinement of the original data; the receiver performs nuclear principal component analysis (KPCA) 24; and then extracts the characteristic asset #g; and then adjusts the main scale range 26, and the above The data generated in the private database will also be sent to the second database 30 for storage; 2. The processing process of the virtual measurement module 40 includes: the feature data from the process data collection module 20 that has completed the adjustment scale range. The set is formatted 0 41, open > into a training data set, the SVR model parameter 43 selected by the gene algorithm 48 is called; the support vector machine (SVM) 44; the test text machine 45' The fitness assessment 46' optimizes the SVR model 47 when it meets the pre-set termination condition and stores it in the second database, so that it is difficult to use this SVR in the virtual process. The _quality prediction model, and running the gene algorithm 48 when not meeting a predetermined termination condition, re-optimizing the SVR model parameters. 'Reusing the reselected model parameters and repeating the steps 43 to 48 above 'until The SVR model with the predicted quality is built, and the other preferred embodiment further includes using a preset test data 421 to send the newly created SVR model, and then proceeding through the aforementioned steps 45 to The process of processing the 3's new detection set 50 includes: formulating a set of feature data from the range of the adjusted scale of the sleep release data collection module 20; forming a training data set 52 Calling the SVDD model parameters selected by the gene algorithm %5:?, §) practicing the SVDD model 54; testing the svDD steaming 55. performing the fitness assessment 56; the model is met when a predetermined termination condition is met. Progress 13 200928843 Line optimization 57 and stored in the _ _ , ^ _ Le-Bei oblique library, in order to use the virtual measurement as the new type of detection model of the present invention, and Pre-sighing ',, the conditional running gene algorithm 58 re-selects the SVDD module 1 and then uses the reselected SVDD model parameter to the next few months. Sexual detection model (ie model) is completed, (four)-aged Reality, including _-preset test data 52! Send the newly created novelity detection model (ie svdd model), and then test through the above steps:) 5 to 58; p ^ 壬The quality prediction model and the new sex measurement model have been completed, and then the quality of the equipment can be predicted by the lion. Second, the virtual measurement process (see "Fig. 3"), where the solid line indicates the path of the virtual measurement, and the dotted line indicates the transmission path of the instantaneous process data during the measurement process. The process of virtual operation measurement includes: i. The processing process of the gas collection module 20 includes: taking the CVD equipment 0 1 即时 the process data 27 in the production process and calling the configuration settings stored in the second tributary 30 in the training (c 〇nfigurati〇n Setting file (including the various configuration settings generated by the process data collection module 20 during the training phase, such as data pre-processing, KPCA analysis, and adjustment of the scale range and other related processing settings), in the same The data is pre-processed under the configuration setting conditions 22; then the representative feature data 25 is extracted by the kernel principal component analysis (KPCA) 24; the representative feature data 25 is mapped and the same coordinate axis space and scale range is 26 in the training process. Then, the novelty detection module 50 is directly transmitted for novelty identification; 14 200928843 The completed processing process includes: __ Cheng Zhongjian and the new 砸 砸 砸 ( ( ( ( ( ( ( 训练 训练 训练 训练Type) 57, for the τ 巳 every straightforward to take the optimization of the representative feature surface 2 == 26 domain standard axis and scale within the standard (in other words, the identification of unidentified sensibility identification) if the results fall in the supersphere The mold (10) processing (Γ^Γ fine), will be passed through the aforementioned process data virtual quantity two steps 26) (four) 'Tao Cheng' to the new special η, is to fall outside the supersphere (in other words, identify the ❹, The second death Xinna sign data is sent to the incremental learning f module 60; I board summer test mode 'group 40 processing process: using the training quality prediction model (that is, optimizing the SVR spleen R set to 2 〇 processing (steps ^Niu (four),, -) U - yin rabbit set module measured value, 隹二::) The instantaneous process data 27 is converted into a virtual quantity, then the titanium value is the result of the output quality prediction; The processing of the quantity learning module 60: the measurement learning module 6 of the present invention only recognizes the new feature soil, and the main ',,' coffee error will be moved, and the processing of the aforementioned process (four) is processed ( Step 22 to step 26) of the immediate process data π, 曰 々 运 运 运 运 运 运 模组 模组 模组 模组 module re-training mechanism 63, and - scoop wheel 64 The platform shoe fine transmission detection module|Heali Shun group 40' uses the process data 27 which is processed by the above-mentioned process data waste module% to step 26), and the new ridge detection second virtual quantity group 4Q For the reduction training, the new (10) side-sucking type and the SVR model (ie, the quality prediction model) are generated (the process is as described above to achieve the effect of self-adaptability.) According to the technique disclosed by the present invention, when the system is virtual During the measurement process, because the retraining model was notified to face the changed device parameters, the previously constructed SVR and novelty detection model will be followed according to the report. = 7 is also open again. In order to improve the speed of the system and gradually accumulate the novelty detection module 5〇, and the knowledge of the virtual game group, the invention is in addition to the nuclear principal component analysis (KPCA). Quantitative (4), Technology 1 By the principle of minimizing the structural risk adopted by SVM, the new feature data set (ie, the new ❹ characteristic data discovered by the Xinlai module) will be collected when the milking equipment 10 is running. The new support vector and the svr model are generated respectively by combining a few support vectors (3 Qing (10) Vector) with the previous training set (3 Qing (10) Vector), the weight _ face side % and the virtual measurement module 40 respectively. (That is, quality prediction knows the soil) So - the novelity detection model and quality detection model generated after the new training can adapt to the dynamic environment and make accurate predictions. According to one of the preferred embodiments of the present invention (see "Fig. 4"), the aforementioned incremental learning module (I job tal Leammg M〇 her) 6〇 also has a statistical system ( The S-plus iStical Process Contol (SPC) inspection mechanism is used to inspect the process data 27 that is processed by the process data collection module 20 (steps 22 to 26), and to verify the immediacy of the process data. & | respectively sent to - the novel side module re-mechanism 63, and - SVR model * new mechanism 64, and then separately sent to the aforementioned novel side module % and virtual measurement module 40 Training to regenerate the new dominant detection model and the svr model (ie, the quality prediction model), and the unverified passer initiates a verification mechanism (ν_ύ〇η) 65, which can be used to notify the on-site engineering Wei verification, the younger than the same 16 200928843 In summary, the virtual measurement system disclosed in the present invention for chemical vapor deposition is used to establish quality prediction using genetic algorithm and support vector regression (ga_svr) analysis method. Novel detection model 5〇, so that the virtual measurement system of the present invention may be suitable Lin _ correct state, and reduce the accuracy of the predicted results of efficacy. On the other hand, the virtual measurement line disclosed in the present invention uses the Kernel Principal Component Analysis (KPCA) for feature extraction, which not only improves the accuracy of the virtual measurement, but also greatly reduces the number of features of the person. The speed of the virtual amount ^ test purpose. Although the present invention has been disclosed in the foregoing preferred embodiments, it is not intended to limit the scope of the present invention to those skilled in the art and without departing from the spirit and scope of the invention. The scope of the invention of the county is inspected. Therefore, the scope of the invention is determined by the scope of the book. BRIEF DESCRIPTION OF THE DRAWINGS Fig. 1 ' is a preferred embodiment of the wire architecture of the present invention. Fig. 2' is a flow chart showing the operation of the virtual measuring system of the present invention, showing that the sniping is a flowchart, and the virtual drawing is not shown, which is the measuring process of the virtual measuring system of the present invention. Figure 4 is a perspective view of a system frame of the present invention. [Main component symbol description] 10 CVD equipment 11 first database 17 200928843 20 process data collection module 21 historical process data 22 data pre-processing 23 statistical process test 24 core principal component analysis (KPCA) 25 feature data 26 adjustment scale range ^ 27 Instant Process Data 30 Second Library 40 Virtual Measurement Module (GA-SVR Module) 41 Format 42 Training Data Set 43 Call SVR Model Parameters 44 Training Support Vector Machine (SVM) 4 5 Test Support Vector Machine® 46 Fitness Assessment 47 Optimizing the SVR Model 48 Gene Algorithm 50 Novelty Detection Module 51 Format 52 Training Data Set 53 Tuning SVDD Model Parameters 54 Training SVDD Model 18 200928843 55 Testing SVDD Model 56 Fitness Evaluation 57 Optimizing SVDD Model 58 Gene Algorithm 60 Incremental Learning Module 61 Statistical Process Control (SPC) Inspection Mechanism 62 is tested for immediate process Tribute 0 63 Retraining mechanism of novelty detection module 64 Retraining mechanism of SVR model 65 Verification mechanism (vaiidBtioii) ❹ 1 9