201135475 六、發明說明： 【發明所屬之技術領域】 本發明係有關-種短時離散小波轉換_神經娜的監控系統 及方法，尤指-種可以對物件生產設備的各工作系統進行整合監控 者。 【先前技術】 _。㈣卫業的快速發展’使得各種精密加•產業急遽增加，尤其是 ，圓、半導體與面板的生產設備所需的品管要求亦不斷地日趨精細與 嚴格。其中生產设備之電力供應系統、物件缺陷檢測系統、管路系統 以及移動輸送系統、動力系統、機械電控系統以及驅動系統的運轉狀 心白攸關成。。產出的品質良劣，因此，如何建構一套可以有效整合並 改。生產a備之電力供應、物件缺陷制、管路供應、移動輸送、機 械電控以及㈣㈣統運概態的監控系統，實已成為各晶圓、半導 •體與面板生產廠商所欲急於克服的難題與挑戰。 依據目前所知，如本國發明專利公嶋肅27_號『透明基板 端面之檢查裝置及檢查方法』、發明專利公開第2觀則號『缺陷 檢查方法』、發明專利公開第200813421號『表面檢查裝置』、發明專 利公開第200804758『表面檢查裝置』、發明專利公開第2_號 『表面檢查裝置』、發明專利公開第2_號『表面檢查裝置及表 面檢查方法』以及發明專利公開第2_5538『缺陷檢測裝置』，等七 件專利前案时以短咖散小波轉換技術來檢測物件之表面缺陷，而 疋以办像處理比雜術’或是喊光元件所收集的反射訊號來判斷反 201135475 t«的絲度分佈來敏物件絲是否有缺陷。由於影像處 技術解讀歸織費較歸_毅，卿必舰 f比對待測影像與基準影像之間的異同，如此方能比對出物== 驟相 。陳所在，如此雛其使得影像處理崎技術較無法檢測製程步 對較快的晶圓生產線上。 夕 此外，光強度的檢财式軸不馳影像處理，而可加快整體的 檢測速度，惟，其僅能針對物件表面粗縫度、平滑度做檢測而已，其 鲁無=對物件表面之微裂痕包括裂痕形貌以及裂痕深度做進一步精確的 判疋。再者’上料猶案於晶圓制時需令生產線停止，再以機械 手臂將待·圓取至旋轉平台上進行檢測，藉由旋轉平台的轉動角度 使晶圓表面得以被掃瞒光線予以掃猫，因而造成檢測時間與工時的ς 加而且會嚴重影響晶圓製造產線的產出效能。故而上述專利前案確 貝有再改良的必要性。 依據目刖所知的短時離散小波轉換技術，主要係應用於影像壓縮 鲁與影像還原處理、類神經網路或是其他的技術領域上。一般離散小波 轉換主要包括小波分解與小波重建兩卿分，小波分解的步驟係將原 始資料中的一致性資料與高度變化性資料分解成低頻與高頻兩種獨立 的貪料；另小波重建的部分係將低頻訊號經過低頻合成濾波之後的訊 號’與尚頻訊號經過高頻合成濾波之後的訊號予以相加整合，即可還 原成原始的訊號。 由於實際硬體架構的緣故，一般離散小波轉換技術多將訊號分段 的區間設定為較為綿密的點來做離散小波轉換，但是實際上應用於線 4 201135475 卩祕料⑽無法做到點對闕訊號處理， 間做轉換處理以取巧而耗費較長的時 用，因此，-如 分析結果無法做為自動辨識之 研九所提出離散小波轉換技術皆無 晶圓檢測領域上。供田… ㈣綠上即時的 音混淆，^纽 果财讀震鋪起之噪 / …、法辨識。更何況人眼的視野無法察覺單晶❹曰太陽 能晶圓中所包含之料剥、麻私+ 見早日日次夕日日太1% 及胸n = 且晶圓微裂痕會隨著製程的流程以201135475 VI. Description of the invention: [Technical field to which the invention pertains] The present invention relates to a monitoring system and method for short-term discrete wavelet transform _ neurona, and in particular to an integrated monitoring system for each working system of an object production device . [Prior Art] _. (4) The rapid development of the health industry has made the various precision and industrial industries increase rapidly. In particular, the quality control requirements required for the production equipment of round, semiconductor and panel are constantly becoming more sophisticated and stricter. Among them, the power supply system of the production equipment, the object defect detection system, the piping system, and the mobile transportation system, the power system, the mechanical electronic control system, and the driving system are all in a state of mind. . The quality of the output is good, so how to construct a set can be effectively integrated and changed. The production of power supply, object defect system, pipeline supply, mobile transportation, mechanical and electronic control, and (4) (4) general monitoring system have become the eager to overcome the wafer, semi-conductor body and panel manufacturers Challenges and challenges. According to the current knowledge, such as the national invention patent publication 嶋 27 27_ No. "Inspection device and inspection method for the end surface of the transparent substrate", the invention patent disclosure second view number "defect inspection method", invention patent publication No. 200813421 "surface inspection "Device", Patent Publication No. 200804758 "Surface Inspection Apparatus", Invention Patent Publication No. 2_ "Surface Inspection Apparatus", Invention Patent Publication No. 2_ "Surface Inspection Apparatus and Surface Inspection Method", and Invention Patent Publication No. 2_5538" The defect detection device, in the case of seven patents, uses short-wave-scattering wavelet transform technology to detect the surface defects of the object, and the image processing is compared with the reflection signal collected by the hybrid or the flashing component to judge the anti-201135475 The silky distribution of t« is sensitive to whether the object wire is defective. Because the interpretation of the technical interpretation of the image is more _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ Chen, so it makes image processing technology more difficult to detect the faster step on the wafer production line. In addition, the detection axis of the light intensity is not image processing, but can speed up the overall detection speed. However, it can only detect the roughness and smoothness of the surface of the object, and it has no objection to the surface of the object. Cracks include crack morphology and crack depth for further precise determination. In addition, when the material is processed, the production line needs to be stopped, and then the robot arm will take the round to the rotating platform for inspection. The surface of the wafer can be bounced by the rotation angle of the rotating platform. Sweeping the cat, which results in increased inspection time and man-hours and can seriously affect the output performance of the wafer manufacturing line. Therefore, the above patents have confirmed the necessity of further improvement. According to the known short-term discrete wavelet transform technology, it is mainly applied to image compression and image restoration processing, neural network or other technical fields. The general discrete wavelet transform mainly includes wavelet decomposition and wavelet reconstruction. The steps of wavelet decomposition are to decompose the consistent data and highly variability data in the original data into two independent greed materials: low frequency and high frequency. In part, the signal after the low-frequency signal is subjected to low-frequency synthesis filtering and the signal after the high-frequency synthesis filtering is added and integrated, and then the original signal is restored. Due to the actual hardware architecture, the general discrete wavelet transform technology sets the interval of the signal segmentation to a relatively dense point for discrete wavelet transform, but it is actually applied to the line 4 201135475 卩 secret material (10) can not do point-to-point Signal processing, and the conversion processing is tricky and time consuming. Therefore, if the analysis result cannot be used as the automatic identification, the discrete wavelet transform technology proposed in the field of waferless detection. For the field... (4) The real tone of the green is confused, ^Nu Guocai read the shock of the noise / ..., the law identification. What's more, the human eye's field of vision can't detect the single-crystal ❹曰 solar energy. The material contained in the wafer is stripped and smuggled. + See the early day and the next day, the sun is too 1% and the chest n = and the wafer micro-crack will follow the process flow.
、、υ繼、·，只蔓延’而此一缺陷則會導致功率上的損失。 第一 3二類=網魅要係以簡易元件所組成，單—神經元結構如 系统為目Γ細平仃式架構為運作模式，其元件以仿效生物神經 權番柏Γ斤以其元件間之連接方式決定網路之功能，元件間連結 2值可被調整，並被加以訓練以實行其特殊功能。一般而言，對特 疋认而達成目標輸出時，類神經網路須經由調整及訓練過程。網路 :調整是基於輸_標值之味反覆，直_路輸出與目標相符。 般而吕，訓練網路需要許多輸入目標對資料。類神經網路已被訓練 於執仃捕知’並朗於不同領域，例如型騎識，認定分類，語 音視覺及控制系統。 應賴私鱗的代表性專利如糊發明第號『智慧型醫 療執仃系統』u騎型細541微『賴經網路贼之純波浪量測 裝置』。料翻結構軸具有__路計算魏，惟，其並無離散 臣皮轉換之功此建置’而且麵構在晶圓或面板生產設備的整體檢測 監控的用途上’因此，該等_結财定無法充分揭露本發明的整體 技術特徵’確實穌發明係為分屬二财同的技術領域。 5 201135475 【發明内容】 本發明主要目的在於將離散小波轉換與類神經網路技術做一整 合，以應用在晶圓、半導體或是面板相關精密工業的生產設備當中， 並能有效整合及改善生產設備之電力供應、物件缺陷檢測、管路供應 以及移動輸送'機械電控、動力系統以及驅動等系統的運轉狀態，藉 以達到整合性監控、即時量測以及對網路觀值進行調㈣多重功效 目的，不僅可以降低生產設備的維護成本，並可主動掌控生產設備的 籲即日守狀况以提刖獲悉即將發生的問題所在資訊，進而採取因應的雛 或保護措施’以大幅降低設備機件故障的機率，故而可以提升生產設 備的運轉效能，以增進成品之生產品質的良率。 、為達成上述侃’本發明制之技術手段細複數鋪測器分別 感測至少-物件生產設備之複數個工作系統的運作狀態，而可產生複 f個工作職的電壓峨’並以資觸取暨觀介面擷取各該電壓訊 紐位訊號’再以離散小波轉換手段料電壓訊號訊號區分 •=數鋼長度的區間，並對每一區間做離散小波轉換，而可於每-區間取得一低頻係數與一高頻係 up „ & θ 並對母一该低頻係數與該高頻係 =:爾鼻分析，進而輸出各該區間能量振幅驟變的特徵 ::將規則表與特徵值進行比對，據此得以輸出正常=二 實施方式】 壹·本發明之基本技術特徵 6 201135475 - 1.1本發明應用與特點 請配合參看第-、二圖所示，本發明主要係應用在物件生產設備 (10)之各工作系統⑼的監控用途上’尤其是針對物件生產設備⑽ 之生產機台、氣體管路线、㈣管路纽、電源供應祕、電控系 統、移動輸送系統、動力系統、驅動系統以及物件缺陷檢測系統…等 等工作系統(11)進行全面整合的監控。具體言之，係將離散小波轉換 與類神經網路技術做—整合而顧在晶圓、半導體或是面板相關精密 參工業的生產設備上，並能有效全面整合及改善生產設備之生產機台、 氣體管路、㈣管路、電祕應祕、電控系統、移動輸送系統、動 力系統、驅動系統以及物件缺陷檢測系統等工作系統（11)的運轉狀 態，藉以達到整合性監控、即時量測以及對網路權重值進行調整等多 重功效目的，不僅可以降低生產設備的維護成本，並可主動掌控生產 設備的即時狀況以提前獲悉即將發生的問題所在資訊，進而採取因應 的調校或保護措施，以大幅降低設備機件故障的機率，故而可以提升 0生產設備的運轉效能以及提升物件生產的品質。 1. 2本發明的基本技術特徵 请配合參看第一、二圖所示，為達上述功效，本發明基本技術特 徵係包括複數個感測器（20)、一資料擷取暨轉換介面（4〇)、—離散小 波轉換手段(31)以及一類神經網路監控手段(32)。 其中係以複數個感測器（20)分別感測至少一物件生產設傷（1〇)之 複數個工作系統(11)的運作狀態’而可產生複數個工作周期的電壓訊 號’並以資料擷取暨轉換介面（40)擷取各電壓訊號後轉換為數位訊 201135475 =離::波轉換手段(31)將各電复數個同長 f對母—區間做離散小波轉換，柯於每-區間取得_攸 頻=^與=_數’鱗每—該低頻係數與該高頻係數做區間能量 的6异力斤’進而輸出各區間能量振幅驟變的特徵值至類神經網路於 控手段⑽中，_經網路監控手段㈣包含有—建立有—簡表: 在線監控模式，而可將規啦與·值進行崎，俾 或是異常訊號。 Φ 貳·本發明技術特徵之具體實施 2· 1感測器與各工作系統 第—、二圖所示，於—種具體實施例中，本發明工作系統 匕電力供應系統、氣體管路系統、液體管路系統、移動輸送 系統生產輸送帶、機械手臂）、生產機台、驅動系統(如pLc或⑽ 控織）、動力純(如健馬達或是馬達）以及物件缺陷檢測系統(如 晶圓、面板檢測機台），進而得以達壯面性整合之監控目的。 此外’感測器（20)主要係用以感測生產設備之各工作系統(⑴的 運作狀態’電壓感湘⑽，制以感測電力供應系統之電力線路 的電壓狀g、絲探針陣列’係做為物件缺陷檢測系統的光電轉換器、 電荷搞合元件CCD係做為物件缺陷檢測祕的光電轉換器、互補式金 屬氧化層半導體⑽S係做為物件雜檢測祕的光電轉換H、位置感 測器（2G)PSD係做為物件缺陷檢測系統的光電轉換器、壓力感測器 ⑽）’係用以感測液體;|；路系統之壓力狀態、流量感測器⑽，係用 以感測液體官路祕之流量狀態、三軸加速度計，制以感測移動輸 送系統之移動座標位置、距離感測器（2〇)，係用以感測生產機台之物 201135475 _件數量或距離、氣體成份分析儀係用以感測氣體管路系統之成份漢声 狀態’以及角度感測器⑽，侧以感測動力系統之運㈣度 等。 2.2資料擷取暨轉換介面 請參看第-、二圖所示’本發明資料擷取暨轉換介面⑽的且體 實施例包含-TCP/IP界面或是RS232順85界面，以及一訊號轉換電 路，並以TOVM面或是RS232/RS485界面來擷取各感測器⑽的各 電壓訊號’再以訊號轉換電路將各電壓訊號轉換為複數個工作周期的 數位式訊驗傳輸雜散小雜換手段⑶Η，以進行離散轉 的運算。 、 2. 3離散小波轉換手段 2. 3.1離散小波轉換手段的具體實施 請參看第-、二圖及第八騎示’離散小波轉換手段(31)主要係 蔣各電壓訊號施以短時離散小波轉換，藉以娜各工作系統於狀態骤 變時產生的舰值。於-種較為具體雜實施例巾，離散小波轉換手 鲁段(31)可以雙值遽波器組(Dyadlc Filter Banks)之方式來建構離散小波 轉換的夕重分析架構。上述雙值遽波器組所擷取之電壓訊號，係使用 -組低通舰器响及__其歧的高_波㈣]分解成兩組頻寬 不同的訊號’且由低通渡波器⑻制後的分支，並使用同一型該低通 濾波器/ψ]再重覆處理複數次。 上述具體實施例中，可以6組同一型低通滤波器响重覆處理六 次，以取得第六階層之低頻係數(eM)，並由高通渡波器别取得第一 階層之高頻係數cDl。 201135475 離散小波轉換手段(31)可以内建之Matlab或是C++軟體模式對該 第一階層之高頻係數cDl及第六階層之低頻係數(cA6)進行區段能量 的分析。其中，離散小波轉換手段(31)可用包含一内建有該Matlab或 是C++軟體模式的電腦（30)，或是一微處理器，並以電腦（3〇)或是微 處理器對第-階層之高頻係數eD1及第六階層之低頻係數(eA6)進行 能量區段分析，以辨別訊號中的微弱突變。 另一方面，離散小波轉換手段(31)將由各感測器（2〇)所收集到之 各電壓訊號分別置於數個緩衝區中，並將每—電壓訊號内之資料串列 力叫分段為複數烟長度的區段，並對緩舰之每—區段之電壓訊號 =以離散小波轉換處理’且離散小波轉換手段⑻依據—窗函數來取^ :=區&長度’以決定掃_析度的大小，而離散小波轉換手段 與固函數之具體實_，倾时限脈衝響麟波器來達成上述 ^衝^衝區之具體實施例可以是内建糊⑽或是微處理器, υ继,··, only spread ‘and this defect will result in loss of power. The first 3 and 2 categories = the network charm should be composed of simple components, the single-neuron structure such as the system for the purpose of the fine-grained architecture as the operating mode, its components to emulate the biological nerve rights of the cypress cymbal with its components The connection method determines the function of the network, and the value of the link between components can be adjusted and trained to perform its special functions. In general, when the target output is specifically identified, the neural network must undergo an adjustment and training process. Network: The adjustment is based on the value of the input_standard value, and the straight_channel output matches the target. As usual, the training network requires a lot of input target data. Neural networks have been trained to capture and understand and to be used in different fields, such as type riding, classification, speech vision and control systems. Representative patents that should rely on private scales, such as the invention of the invention, the "smart medical treatment system" u ride type fine 541 micro "Like the network thief's pure wave measurement device". The material-turning structure axis has the __ way to calculate Wei, but it has no discrete material conversion and this is built and 'the surface structure is used for the overall inspection and monitoring of the wafer or panel production equipment'. Therefore, these _ knots Cai Ding cannot fully disclose the overall technical characteristics of the present invention. 5 201135475 SUMMARY OF THE INVENTION The main object of the present invention is to integrate discrete wavelet transform with neural network-like technology for application in wafer, semiconductor or panel related precision industrial production equipment, and to effectively integrate and improve production. Equipment power supply, object defect detection, pipeline supply and mobile transport 'mechanical electronic control, power system and drive system, etc., to achieve integrated monitoring, real-time measurement and adjustment of network views (4) multiple functions The purpose is not only to reduce the maintenance cost of the production equipment, but also to take the initiative to control the situation of the production equipment to improve the information of the upcoming problems, and then take the corresponding chick or protection measures to greatly reduce equipment failures. The probability of the production equipment can be improved to improve the production quality of the finished product. In order to achieve the above-mentioned technical means of the invention, the fine and plural detectors respectively sense the operation state of at least the plurality of working systems of the object production equipment, and can generate voltages of the work of the other f jobs. Take the cum view interface and take the voltage signal signal of each of the 'wavelength signal signals by the discrete wavelet transform method to distinguish the length of the steel length, and perform discrete wavelet transform for each interval, and obtain it in each interval. A low frequency coefficient and a high frequency system up „ & θ and the mother and the low frequency coefficient and the high frequency system are analyzed, and then the characteristics of the energy amplitude of each interval are suddenly changed: the rule table and the eigenvalue are Performing the comparison, according to which the output is normal = two embodiments] 壹 · The basic technical features of the present invention 6 201135475 - 1.1 The application and characteristics of the present invention, please refer to the first and second figures, the invention is mainly applied to the production of objects The monitoring use of each working system (9) of the equipment (10) is especially for the production machine of the object production equipment (10), the gas pipeline route, (4) the pipeline, the power supply secret, the electric control system, the mobile transportation system. The power system, the drive system, the object defect detection system, etc., and the working system (11) are fully integrated and monitored. Specifically, the discrete wavelet transform and the neural network-like technology are integrated into the wafer and semiconductor. Or the production equipment of the panel-related precision ginseng industry, and can effectively integrate and improve the production equipment, gas pipelines, (4) pipelines, electric secrets, electronic control systems, mobile transportation systems, power systems, The operating state of the working system (11) such as the driving system and the object defect detecting system can achieve the multiple functions of integrated monitoring, real-time measurement, and adjustment of the network weight value, thereby not only reducing the maintenance cost of the production equipment, but also Take the initiative to control the real-time status of the production equipment to learn the information of the upcoming problems in advance, and then take the appropriate adjustment or protection measures to greatly reduce the probability of equipment failure, so it can improve the operating efficiency of the production equipment and improve the production of objects. 1. The basic technical features of the invention, please refer to the first, As shown in the figure, in order to achieve the above effects, the basic technical features of the present invention include a plurality of sensors (20), a data acquisition and conversion interface (4〇), a discrete wavelet conversion means (31), and a neural network. The monitoring means (32), wherein the plurality of sensors (20) respectively sense the operation state of the plurality of working systems (11) of the at least one object to produce the injury (1), and the plurality of working cycles can be generated. The voltage signal 'takes the data acquisition and conversion interface (40) and takes each voltage signal and converts it into digital information. 201135475 = away: :wave conversion means (31) will make each of the plurality of electrical lengths of the same length f to the mother-interval discrete wavelet Conversion, Ke obtains the _攸 frequency=^ and =_number's scales per-interval. The low-frequency coefficient and the high-frequency coefficient do the interval energy of 6 different force's and then output the characteristic values of the sudden changes in the energy amplitude of each interval to In the neural network control method (10), _ via the network monitoring means (4) contains - the establishment of - profile: online monitoring mode, and can be used to carry out the singularity, ambiguity or abnormal signal. Φ 贰 · The specific implementation of the technical features of the present invention 2 · 1 sensor and each working system shown in the first and second diagrams, in a specific embodiment, the working system of the present invention 匕 power supply system, gas pipeline system, Liquid piping system, mobile conveyor system production conveyor belt, robotic arm), production machine, drive system (such as pLc or (10) control weaving), power pure (such as motor or motor) and object defect detection system (such as wafer , panel inspection machine), in order to achieve the purpose of monitoring the integration of strong surface. In addition, the sensor (20) is mainly used to sense the working systems of the production equipment ((1) the operating state of the voltage sense (10), the voltage shape of the power line for sensing the power supply system, the wire probe array 'As a photoelectric converter for the object defect detection system, the CCD system is used as a photoelectric converter for object defect detection, and a complementary metal oxide semiconductor (10) S system is used as a photoelectric conversion H for the object detection. The sensor (2G) PSD is used as the photoelectric converter of the object defect detection system, the pressure sensor (10)) is used to sense the liquid; the pressure state of the road system, the flow sensor (10) is used to Sensing the flow state of the liquid official road, three-axis accelerometer, sensing the moving coordinate position of the moving conveyor system, distance sensor (2〇), used to sense the production machine's object 201135475 _ pieces Or the distance, the gas component analyzer is used to sense the component of the gas pipeline system and the angle sensor (10), and the side to sense the power system (four degrees). 2.2 Data Capture and Conversion Interface Please refer to the data capture and conversion interface (10) of the present invention as shown in the first and second figures, and the physical embodiment includes a TCP/IP interface or an RS232 smooth 85 interface, and a signal conversion circuit. And using the TOVM surface or the RS232/RS485 interface to capture the voltage signals of each sensor (10)', and then converting the voltage signals into a plurality of duty cycles by the signal conversion circuit to transmit the spurious small miscellaneous means (3) Η to perform the discrete rotation operation. 2. 2. Discrete wavelet transform means 2. 3.1 Discrete wavelet transform means the specific implementation, please refer to the first, second and eighth riding 'discrete wavelet transform means (31) mainly rely on Jiang each voltage signal to apply short-term discrete wavelet Conversion, by the value of the ship generated by each of the working systems in the sudden change of state. In the case of a more specific heterogeneous embodiment, the discrete wavelet transform hand segment (31) can construct a discrete wavelet transform analysis system by means of the Dyadlc Filter Banks. The voltage signal drawn by the above-mentioned two-value chopper group is decomposed into two sets of signals with different bandwidths by using a group of low-pass ship sounds and __high-waves (four) of the difference ′′ and by a low-pass waver (8) Post-production branch, and use the same type of low-pass filter / ψ] to repeat the processing multiple times. In the above specific embodiment, six sets of the same type low pass filter can be repeatedly processed six times to obtain the low frequency coefficient (eM) of the sixth level, and the high frequency coefficient cD1 of the first level is obtained by the high pass waver. The 201135475 discrete wavelet transform method (31) can analyze the segment energy of the first-level high-frequency coefficient cD1 and the sixth-level low-frequency coefficient (cA6) by the built-in Matlab or C++ software mode. The discrete wavelet transforming means (31) may comprise a computer (30) having a built-in Matlab or C++ software mode, or a microprocessor, and a computer (3〇) or a microprocessor pair - The high frequency coefficient eD1 of the hierarchy and the low frequency coefficient (eA6) of the sixth hierarchy perform energy segment analysis to identify weak mutations in the signal. On the other hand, the discrete wavelet transform means (31) places the voltage signals collected by the respective sensors (2〇) in a plurality of buffers, and points the data in each of the voltage signals. The segment is a segment of the length of the plurality of cigarettes, and the voltage signal of each segment of the slow ship = the discrete wavelet transform process 'and the discrete wavelet transform means (8) according to the - window function to take ^ : = zone & length ' to determine Sweeping the magnitude of the resolution, and the discrete wavelet transform means and the solid function of the solid state, the tilting time pulse sounding the pulsar to achieve the above-mentioned ^ punching and punching area specific embodiment may be built-in paste (10) or microprocessor
2.3.2離散小波轉換手段的公式推導 號之關聯性計算，以下為 連續小波轉換是小波函數與欲處理訊 波函數之定義如下式： 連續小波轉_為_麵函數,、物式以)定義式 〜(0 士 (？） (2) 輸，Μ觸數，賴小波轉_咖)定義為： 201135475 C(a,b)= (3) /W為電壓訊號’在(2)式中’ α和&都是實數或連續。其離散型式為： α = 〇0 , 〇〇 > 1, b = nb0a^, bQ>〇 (4) 離散時域指標，w:離散時域倍數指標。 (3)式可改寫為： 匕 _„6。） (5) 本發明離散式小波轉換之實現架構稱為次頻帶濾波（Subband Filtering)或是偏移樹狀遽波(Dyadic Filtering)。經由將％替換為 2,可以將（5)式重新改寫為如式6 : ⑹ Ψ],Λη^ = 2 2w[2~Jn~k] 〇] = 2 2#[2H] 次頻帶濾波所使用濾波器定義如下： g[«]，g[«]:南通 decomposition and reconstruction quadrature mirror filters. λ[«]，λ[«]:低通 decomposition and reconstruction quadrature mirror filters. 在（6)式中，小波及濾波器函數的關聯如下： ⑻ =Ts\n-2k](f>._l[n] η 6 =Σ咖-琴“Μ 第一級是分析或分解，其函數示於在（9)式。 201135475 = Σ如-2灸]<丨 η ⑼ di 以下為離散小波轉換之分析（分解)雜示意以數 ,Σ h[n-2k] Σ Hln-2k] 學式來表示：2.3.2 Correlation calculation of formula derivation number of discrete wavelet transform means, the following is the definition of continuous wavelet transform is a wavelet function and the function of the signal to be processed is defined as follows: continuous wavelet transform _ is _ plane function, and the formula is defined by The formula ~ (0 士 (?) (2) input, Μ touch number, Lai Xiaobo _ _ coffee) is defined as: 201135475 C (a, b) = (3) / W is the voltage signal 'in (2) Both α and & are both real or continuous. The discrete forms are: α = 〇0 , 〇〇 > 1, b = nb0a^, bQ> 〇 (4) Discrete time domain metric, w: discrete time domain multiple metric. The equation (3) can be rewritten as: 匕_„6.) (5) The implementation architecture of the discrete wavelet transform of the present invention is called Subband Filtering or Dyadic Filtering. % is replaced by 2, and equation (5) can be rewritten as equation 6: (6) Ψ], Λη^ = 2 2w[2~Jn~k] 〇] = 2 2#[2H] Filter used for sub-band filtering The definition is as follows: g[«],g[«]: Nantong decomposition and reconstruction quadrature mirror filters. λ[«],λ[«]: low-pass decomposition and reconstruction quadrature mirror filters. In (6), wavelet and filtering The correlation of the function is as follows: (8) =Ts\n-2k](f>._l[n] η 6 =Σ咖-琴"Μ The first level is analysis or decomposition, and its function is shown in (9). 201135475 = Σ如-2灸]<丨η (9) di The following analysis (discrete) of discrete wavelet transforms is represented by the number, Σ h[n-2k] Σ Hln-2k]
其中dWhere d
⑹以及。以和⑵的頻率範圍包含高頻和低頻部分， 使用db l〇，j、波函數。資料緩衝和判斷級用於收集局部的瞬間變化和 低頻成分。 、下為短時小波轉換之基本定義，反射光轉換成電壓訊號為咖）， 在即時彳5號處理時電壓訊號須分段至於緩衝區中如下所示· /-1 X,·⑻={x〇 · Ζ)，.·.，χ(/. 1 - 1)}， μ{π) % 0<n<L-l 0， ehe或其他窗函數 而分段後之電壓訊號再以離散數位小波轉換加以處理。 yi(m) = x.(m)w(n-m)i m=〇, ι, 2,... L~\ 之後對每一段短時小波轉換系數計算區間能量。 Α = ^(〇Α(η))2 , D = ^{cD(n))2 n=0 n=0 在第六圖中，輸入的電壓訊號係以區段作處理，其中f為即時次料 區。此一流程必需有卜1(先前)和/ + 1(之後）緩衝資料區。資料處理區 大小為80個取樣點。 12 201135475 、 2.4類神經網路監控手段 凊參看第九圖所不，所知的類神經網路係以簡易元件所組成，並 以平行式架構為運作模式，其元件以仿效生物神經系統為目的，所以 疋件間之連接方式決定網路之功能，元件間連結權重值可被調整，並 被加以訓練以實行其特殊功能。—般而言，對特钱人而達成目標輪 出時’類神經網路須經由調整及訓練過程，網路之調整是基於輸出與 目標值之比較反覆，直到網路輸出與目標相符。—般而言，訓練網路 鲁需要許多輸入目標對資料，類神經網路已被訓練於執行複雜功能，並 應用:不同領域’例如型態辨識，認定分類，語音視覺及控制系統。 〜請參看第-、二圖所示，於一麵神經網路監控手段⑽的具體 貫知例巾&括-建構於上述電腦（3〇 )内的類神經網路軟體模組及— ^輸出單元(33)，此類神經網路軟體模組更包含—資料庫⑽，此 貝料庫(34)用以記錄正常訊號以及異常訊號資料，上述資訊輸出單元 (33)則將正常訊號與異常訊號資料予以輸出，同時類神經網路軟體模 組將關聯值予以計算以建立_線性與非線性的邊界條件，再將經計算 結果如正常訊號或是異常訊號資料記錄在上述的資料庫⑽中，以供 使用者隨時輪出查證之用。 /、 再5月參看第二、三圖所示，類神經網路監控手段(32)更包含一在 _練模式’用以將特徵值與—期望值進行誤差計算，並判斷誤差值 疋否小於等於預設值，是，則 ^ 郷線【減式，否，則進行網路權 4後再次進人在線訓練模式，並重覆上述步驟。其中，上述 指的資料庫係建置在電腦⑽之記憶元件内，另資訊輸岭元⑽的 m 13 201135475 -具體實施侧可以是顯示幕、蜂鳴n或是列印機。 參•本發明具體實施的運作 凊參看第-至三圖及第八圖所示，首先將各感測器⑽裝設在物 纽產設備⑽之各工作系統⑼當中，以產生複數個電壓訊號，經 貧料擷取暨轉換介面⑽擷取各電壓訊號後轉換為數位式訊號後傳輸 至離政j /雄換手段(31)巾以進行離散小波轉換運算，由於離散小波 轉、手4又(31)疋以雙值濾波器組(Dyadic Fiiter ^肪⑼方式來建構離散 _ 1波轉換的夕重分析架構’所以雙值濾、波器組得以摘取上述之電壓訊 唬並使用組低通遽波器刷及一組與其正交的高通遽波器别分解 成兩組頻寬不_訊號’且由低通·器⑽制後的分支，並使用同 一型該低通驗ϋ/ψ]再處職6次，以取得第六_之低頻係數 (CA6)，並由尚通濾波器4]取得第一階層之高頻係數CD1。 離散小波轉換手段(31)可以内建之Matlab或是C++軟體模式對該 第-P皆層之高頻係數cD1及第六階層之低頻係數(cA6)進行區段能量 ►的分析’以辨別訊號中的微弱突變，如此即可獲取各工作系統⑼於 運作狀態骤變時所產生的特徵值，如電力線路因電壓驟降所產生的特 徵值，或是物件檢測系統因檢測出物件缺陷所在而產生的特徵值等。 此時，類神經網路監控手段(32)將已建立之規則表與上述特徵值 進行比對’以綱各卫作系統(11)於的運作狀態是否正常，當判斷結 果為正常時則輸出正常訊號；反之，當判斷結果為不正常時(如發現電 力供應系統之電力供應發生糕料輯況；或是檢_物件有裂痕 的情況），則輸出是異常訊號，同時將判斷結果記錄在資料庫(⑷中。 201135475 另 方面’再將特徵值與-期望值進行誤差計算，並判斷誤差值是否 小於等於麟值，是，則酬在線監控模式，否，則進行網路權重值 調整後再錢人在_賴式，並重紅齡驟，如此即可主 物件生產設備⑽之各工作系統⑽的即時狀況，藉以提前獲悉㈣ 發生的問題所在資訊，進而採取輯的調校或保護措施，如此即可大 幅降低物件生產設備（10)機件的故障機率。 肆.本發明實驗例分析 •本發明實驗與案例分析所使用之類神經網路型態為前饋倒傳遞式 網，工作系統（11)則為晶圓微裂痕檢測機台，請參看第六圖所示，盆 係為類神經網路鱗過程_㈣分析誤差值，（由目標值減去輸出值) :疋為w ’分析過財可見誤差值快速下降。再請參看第四圖所示， 圖係為^始晶圓微裂痕檢測機台以光學掃裝置所取得的電壓訊號曲線 胃參看第五圖所示’其係為經短時離散小波轉換所得的特徵值， 籲=^區間㈣為正常部分，而較低區_為微裂痕所在的區間。請 參看第七圖所示，其係為類神經網路訓練完畢後之結果輸出值，黑色 線^分為預設目標值，紅色之圈繞部分為訓練後之輸出值，可見案 結果與馳完全符合，振幅丨部分對照至上圖較高關信號(正 刀）’振幅2部分對照至上圖較低區間信號(微裂痕部分），由此可 Z本發贿物晴驗手段(32)確實可針對本發財驗例達成自 動辨識功能，進而對微裂痕區間輸出一不同值。 伍•結論 因此藉由上述技娜徵建置，本發明確實可以㈣散小波轉換 201135475 _與類神經網路技術做一整合，以應用在晶圓、半導體或是面板相關精 密工業的生產設備，並能有效整合及改善生產設備之電力供應、物件 缺陷檢測、管路供應以及移動輸送、機械電控、動力系統以及驅動等 系統的運轉狀態，藉以達到整合性監控、即時量測以及對網路權重值 進行凋14夕重功效目的，不僅可以降低生產設備的維護成本，並可 主動掌控生產設備的即時狀狀提前獲悉即將發生的問題所在資訊， 進而採取因應的調校或保護措施，以大幅降低設備機件故障的機率， 鲁故而可以提升生產設備的運轉效能，以增進物件生產品質的良率。 以上所述’僅為本發明之一可行實施例，並非用以限定本發明之 專利範圍’凡舉依據下财請補範騎述之内容、特徵以及其精神 而為之其他魏的等效實施’皆應包含於本發明之專郷圍内。本發 明之方法及錢構’除上述優點外，並深具產業之_性，可有效改 善習用所產生之缺失，而且所具體界定於申請補之舰，未見 於同類物品’故而具實用性與進步性，已符合發明專利要件，羡依法 »具文提出申請，謹請鈞局依法核予專利，以維護本申請人合法之權 盈0 【圖式簡單說明】 第一圖係本發明基本架構之實施示意圖。 第二圖係本發明硬體設備之實施示意圖。 圖 第三圖係本發明運算控制流程之示意 第四圖係本發明由物缺陷檢測系統取得之訊號波形示意圖 第五圖係本發_離散小波轉換後之特徵鱗示意圖。 201135475 -第六圖係本發明類神經網路訓練過程之曲線變化示意圖。 第七圖係本發明類神經網路訓練輸出結果之對照示意圖。 第八圖係本發明經離散小波轉換手段之解析架構示意圖。 第九圖係本發明單一神經元網路之實施架構示意 【主要元件符號說明】 (10)物件生產設備 (20)感測器 (31)離散小波轉換手段 (33)資訊輸出單元 (11)工作系統 (30)電腦 (32)類神經網路監控手段 (40)資料擷取暨轉換介面 m 17 ·(6) and. The frequency range of sum and (2) contains the high frequency and low frequency parts, using db l〇, j, wave function. Data buffering and decision levels are used to collect local transients and low frequency components. The following is the basic definition of short-time wavelet conversion, and the reflected light is converted into a voltage signal for the coffee.) In the instant processing of the 5th, the voltage signal must be segmented into the buffer as shown below. /-1 X,·(8)={ X〇· Ζ),.·.,χ(/. 1 - 1)}, μ{π) % 0<n<Ll 0, ehe or other window function and the segmented voltage signal is then converted by discrete digit wavelet Handle it. Yi(m) = x.(m)w(n-m)i m=〇, ι, 2,... L~\ Then calculate the interval energy for each short-wavelet transform coefficient. Α = ^(〇Α(η))2 , D = ^{cD(n))2 n=0 n=0 In the sixth figure, the input voltage signal is processed in segments, where f is instant Material area. This process must have buffer 1 (previous) and / + 1 (after) buffer data areas. The data processing area is 80 sample points. 12 201135475, 2.4 types of neural network monitoring means 凊 See the ninth figure, the known neural network is composed of simple components, and the parallel architecture is the operating mode, the components of which aim to imitate the biological nervous system. Therefore, the connection between the components determines the function of the network, the link weights between components can be adjusted, and trained to perform its special functions. In general, when a target is turned out for a special person, the neural network needs to be adjusted and trained. The network adjustment is based on the comparison of the output and the target value until the network output matches the target. In general, training networks require a lot of input target-to-data, and neural networks have been trained to perform complex functions and apply: different areas such as type identification, classification, speech vision and control systems. ~ Please refer to the first and second pictures, the specific knowledge of the neural network monitoring means (10) and the neural network software module constructed in the above computer (3〇) and - ^ The output unit (33), the neural network software module further comprises a data library (10) for recording normal signals and abnormal signal data, and the information output unit (33) is for normal signals and The abnormal signal data is output, and the neural network software module calculates the associated value to establish a linear and non-linear boundary condition, and then records the calculated result such as a normal signal or an abnormal signal data in the above database (10). In order to allow users to turn around for verification at any time. /, in May, as shown in the second and third figures, the neural network monitoring method (32) further includes a _ training mode to calculate the error between the eigenvalue and the expected value, and determine whether the error value is less than Equal to the preset value, yes, then ^ 郷 line [minus, no, then enter the online training mode again after the network right 4, and repeat the above steps. Among them, the above-mentioned database is built in the memory component of the computer (10), and the information is transmitted to the Lingyuan (10) m 13 201135475 - the specific implementation side can be a display screen, a buzzer n or a printing machine. Referring to the operation of the specific implementation of the present invention, as shown in the first to third and eighth figures, each sensor (10) is first installed in each working system (9) of the material processing equipment (10) to generate a plurality of voltage signals. After the poor material extraction and conversion interface (10), the voltage signal is converted into a digital signal and then transmitted to the political j/male replacement means (31) towel for discrete wavelet conversion operation, due to the discrete wavelet rotation, the hand 4 (31) The dual-value filter bank (Dyadic Fiiter (9) method is used to construct the discrete-time analysis structure of the discrete _1 wave conversion'. Therefore, the dual-value filter and the wave group can extract the above-mentioned voltage signal and use the group low. The pass filter and a set of high-pass choppers orthogonal thereto are decomposed into two sets of branches whose bandwidth is not _signal and made by the low-pass (10), and the same type of low pass test/ψ is used. ] 6 times to obtain the sixth low frequency coefficient (CA6), and the high frequency coefficient CD1 of the first level is obtained by the Shangtong filter 4]. The discrete wavelet transform means (31) can be built in Matlab or It is the C++ software mode that goes into the high-frequency coefficient cD1 of the first-P layer and the low-frequency coefficient (cA6) of the sixth level. The analysis of the line segment energy ► to identify the weak mutations in the signal, so that the characteristic values generated by each working system (9) when the operating state is suddenly changed, such as the characteristic value generated by the voltage drop of the power line, or It is the characteristic value generated by the object detection system due to the detection of the object defect. At this time, the neural network-like monitoring means (32) compares the established rule table with the above-mentioned characteristic values to the respective security system ( 11) Whether the operation status of the operation is normal, when the judgment result is normal, the normal signal is output; otherwise, when the judgment result is abnormal (if the power supply system of the power supply system is found to have a cake condition; or the inspection_object has In the case of a crack, the output is an abnormal signal, and the judgment result is recorded in the database ((4). 201135475 Another aspect is to calculate the error between the characteristic value and the expected value, and determine whether the error value is less than or equal to the value of the column, yes, In the online monitoring mode, if not, the network weight value will be adjusted after the money is in the _ Lai style, and the age is red, so that the main object production equipment (10) The immediate situation of the system (10), in order to be informed in advance (4) the information of the problem occurred, and then take the adjustment or protection measures, so that the probability of failure of the object production equipment (10) can be greatly reduced. 实验. Analysis of the experimental example of the present invention • The neural network type used in the experiment and case analysis of the present invention is a feedforward inverted transmission network, and the working system (11) is a wafer micro-crack detection machine. Please refer to the sixth figure, the basin system is Class-like neural network scale process _ (4) Analyze the error value, (subtract the output value from the target value): 疋 is w 'Analyze the error and the error value decreases rapidly. Please refer to the fourth figure, the picture is ^ starting crystal The circular micro-crack detection machine uses the voltage signal curve obtained by the optical scanning device. See the fifth figure, which is the characteristic value obtained by short-time discrete wavelet transform, and the interval (4) is the normal part, and the lower part is lower. Zone _ is the interval where the micro-crack is located. Please refer to the seventh figure, which is the output value of the neural network after training. The black line is divided into the preset target value, and the red circle is the output value after training. Fully compliant, the amplitude 丨 part is compared to the higher signal (positive knives) of the above figure. The amplitude 2 part is compared to the lower interval signal (micro-crack part) of the above figure, so that the bribe clearing means (32) can be used. An automatic identification function is achieved for the present financing example, and a different value is output for the micro-crack section. Wu·Conclusion Therefore, the invention can be implemented by the above-mentioned technology, and the invention can be integrated with the neural network technology to apply to the production equipment of wafer, semiconductor or panel related precision industries. It can effectively integrate and improve the power supply of the production equipment, object defect detection, pipeline supply, and the operation status of mobile transportation, mechanical electric control, power system and drive system, so as to achieve integrated monitoring, real-time measurement and network The weight value is used for the purpose of reducing the maintenance cost of the production equipment, and can not only reduce the maintenance cost of the production equipment, but also take the initiative to control the immediate status of the production equipment to learn the information of the upcoming problem in advance, and then take the corresponding adjustment or protection measures to Reduce the probability of equipment failure, and thus improve the operating efficiency of production equipment to improve the quality of the production quality of the object. The above description is only a possible embodiment of the present invention, and is not intended to limit the scope of the patent of the present invention. The equivalent implementation of other Wei according to the content, characteristics and spirit of the following 'All should be included in the special area of the present invention. In addition to the above advantages, the method and the structure of the invention have the advantages of industry, can effectively improve the lack of use, and are specifically defined in the application for repairing the ship, which is not found in the same kind of article, so it is practical and Progressive, has met the requirements of the invention patent, and has filed an application according to law. Please ask the bureau to approve the patent according to law to maintain the legal rights of the applicant. 0 [Simple description] The first picture is the basic structure of the invention. Schematic diagram of the implementation. The second figure is a schematic diagram of the implementation of the hardware device of the present invention. Figure 3 is a schematic diagram of the operation control flow of the present invention. The fourth figure is a schematic diagram of the signal waveform obtained by the object defect detection system of the present invention. The fifth figure is a schematic diagram of the feature scale after the present-discrete wavelet transform. 201135475 - The sixth figure is a schematic diagram of the curve change of the neural network training process of the present invention. The seventh figure is a control diagram of the output of the neural network training of the present invention. The eighth figure is a schematic diagram of the analytical architecture of the discrete wavelet transforming method of the present invention. The ninth diagram is a schematic diagram of the implementation of the single neuron network of the present invention [the main component symbol description] (10) object production equipment (20) sensor (31) discrete wavelet transform means (33) information output unit (11) work System (30) computer (32) type neural network monitoring means (40) data acquisition and conversion interface m 17 ·