TWI269990B - Quality prognostics system and method for manufacturing processes with generic embedded devices - Google Patents

Abstract

A quality prognostics system and a method for predicting the product quality during manufacturing processes with generic embedded devices are disclosed. The quality prognostics system includes a GED, a measurement collection plan and a remote host, wherein the GED is installed in the production equipment, and is linked to a pluggable quality prognostics module, and the measurement collection plan is connected to a measurement equipment for collecting actual quality data of product from measurement equipment. The pluggable quality prognostics module is used to consider the current production equipment parameters sensed during the manufacturing process and several previous quality data collected from the measurement equipment to predict the quality of the product in the future.

Description

1269990 IX. Description of the Invention "Technical Fields of the Invention" The present invention relates to a quality prediction system and method for a manufacturing process, and more particularly to an application-type embedded device (Generic Embedded Device) that can be embedded in a device. GED) Quality Prediction System and Method for Production Process [Prior Art] _ General semiconductor and TFT-LCD factory product manufacturing and inspection processes are usually carried out separately 'that is, after processing by the production machine first, Then send it to the test machine for testing. And based on cost considerations, most of the tests use the method of sampling to determine the quality of the product. Therefore, if a batch of products has problems in the manufacturing process, you must wait until At the time of testing, it will be discovered. 'At this time, the process equipment may have produced several batches of defective products. At present, most of the methods are to monitor the process parameters of the production equipment, and judge whether the quality of the produced products is abnormal according to this parameter. The common disadvantage is that when the Ludang monitoring system detects an abnormality in the process parameters, Good products have already been produced. And since most of the current semiconductor and TFT-Lcd plants use batch production, if the product is found to be abnormal, the amount of scrapped is usually not one or two, and the whole batch is not only reduced. The yield also has a serious impact on the production and production costs of the factory. Therefore, manufacturers are not trying to find a way to predict the quality of the next batch of products. 曰目=How many scholars have predicted how to predict product quality or equipment The process is not the same as the research. These studies include a proposal to promote a semiconductor 1269990 Ή Π repair: for: * * to infer from the production materials that the product does not 1 ==:: network as electricity_machine The immediate fault monitoring is based on the technique to determine the parameters in the plasma remnant process: STM is used as an example to use the Fuzzy Associative Memory (FAM) to extract the thoughts and cooperate with the hierarchical synergy. Job management strategy, developed a knowledge, face-to-face SMT quality prediction and control, a shape of human-machine-based ancient "species use to obtain wafer defect density ^ 欠 + - rough w J. Find the time-of-day clustering problem of the Japanese yen defect cluster. The time and cost of the collection of the 乂 乂 枓 温度 温度 温度 预测 预测 预测 预测 预测 预测 预测 GaAs GaAs GaAs GaAs GaAs GaAs GaAs GaAs GaAs GaAs GaAs GaAs GaAs GaAs GaAs GaAs GaAs GaAs GaAs GaAs The effect on the lifetime prediction of Si〇2 film. Open the door again. 'The above research mainly focuses on how to use the current news to estimate the production that is occurring in the production. ^ ^ ^ This * anomaly, rather than predicting the next t in °, in general, most studies used Estimating the input parameters and estimating the applicability of the wood structure are limited to specific machines. _The existing patent application, applied to the semiconductor industry "Me 廿 system f〇r controlling chemicai p〇iighing this heart (1) rem 〇 Val" patent (US patent before (four) M% 542 - control semiconductor wafer process grinding thickness System, this system consists of three main parts: wafer grinder, media thickness measuring device, and a grinding speed control system. This special (four) thick comparison method to analyze the grinding speed required, basically The actual sacrifice α and the mouth-to-mouth detection value are required to provide the information needed for the grinding control. The virtual thickness measurement method cannot be used to estimate the film thickness value, and the result must be known after 1269990 to the measurement. The "Run-to-run control over ^semiconductor processing tool based upon mirror image target" patent (U.S. Patent No. 6,625, 5 13) proposes a method for using a database parameter model to compare semiconductor manufacturing tools. The change in wear during the production process, and the parameter setting according to the comparison result, so that the processed product is not caused by the tool wear. Often, the disadvantage of this method is that when the process parameters become more, the type and difficulty of establishing the database parameter model will be relatively improved, and it is not allowed to use the appropriate prediction model according to different device characteristics without elasticity. As for other existing patent applications, the "Method of monitoring and/or controlling a semiconductor manufacturing apparatus and a system" patent (U.S. Patent No. 6,616,759), which is applied to the semiconductor process, proposes a method for monitoring semiconductor process equipment and predicting process results. Method: This method takes the sensor data and process result measurement values of the process equipment, and uses the partial least-square method (Partial Least Square Method) to calculate new parameter settings. However, this method can only be based on existing parameters. The data predicts the value of the product being produced and cannot further predict the quality of the product to be produced in the future. "Method for predicting temperature, test wafer for use in temperature prediction, and method for evaluation The lamp heating system patent (U.S. Patent No. 6,666,577) proposes a method for predicting the wafer process temperature. This method utilizes two different coating layers on the wafer to predict the wafer Ϊ 269990 method. Suitable for a specific type of process temperature. The predictive machine of this patent lacks versatility. On the other hand, Xi Zhixi; Φ female also 丨^ The quality prediction system of the production process adopts the centralized type 2: Γ: The 4, /, of each equipment (including the production machine and the measuring machine) The measured data is transmitted to a central host for analysis, monitoring, and profit. # This architecture has two major drawbacks: - When the central host: 2 o'clock, the entire analysis and prediction system will be shut down. Second, this kind of operation will consume the network bandwidth resources. In a semiconductor or TFT_LCD factory, there may be a beta and a beta. If the state parameters or measurement data of the device are transmitted to the central host, The network bandwidth it occupies will be considerable. In addition, the inventor of the case in the Republic of China Patent No. 1225606 certificate number reveals a universal embedded device that can extract and transmit information for customized applications. This universal intrusion device can be installed inside various information devices. Information used to capture, collect, manage, and analyze equipment. However, this general-purpose embedded device does not disclose an application module having a quality prediction function for a production process. Therefore, it is very urgent to develop a product-predictive system and method for applying general-purpose embedded devices, which are embedded in the device to extract device parameters and analyze the data so that it can be immediately before the product is produced. 'According to the current process parameters of the production machine itself and the quality inspection data of the previous batches of products, the quality of the next batch of products is predicted, and it has versatility, thereby improving the shortcomings of the prior art. SUMMARY OF THE INVENTION The main object of the present invention is to provide a quality prediction system and method for manufacturing a manufacturing process using a universal embedded device 1269990, -, s Tian (1) Mountain, and a common type of 友 in the device. The input device can be used according to the current process parameters of the production port itself and the previous batches of production, and the purpose of the product, and the mouth mouth quality test of the mouth of the mouth. Information, pre-listing the quality of the next batch of products, because of the quality of both products and the effectiveness of the machine and the rate of coverage f Availahintp, ^ text. Availablhty, which further enhances the competitiveness of the manufacturing industry. According to the above object of the present invention, a quality prediction system using a general-purpose embedded device 2 production process is provided. According to a preferred embodiment of the present invention, the quality prediction system for the production process of the general-purpose embedded device includes at least: a first embedded type device, a measurement data mining plan (Data c〇iiecti〇n and Remote Host, wherein the first universal embedded device is installed in the production machine and is connected with a pluggable quality prediction module; the measurement data acquisition plan is connected to the measuring machine And is responsible for describing the data type required by the remote host and the pluggable quality prediction module, and collecting the actual quality detection value of the product by the self-measuring machine; the remote host is responsible for processing by the first general type The embedded device transmits the returned data, and displays and stores the predicted result and the abnormal shape. The first general-purpose embedded device includes at least: the first data acquisition plan (Data

Collection Plan), the first device driver (EqUipment Driver), the first communication agent (Communication Agent), the application interface (Application Interface), and the first data collection manager (Data Collection Manager), wherein the first data capture The plan is used to describe the data type required by the remote host and the pluggable quality prediction module, and generate a first data collection report (Data Collection Report); the first device driver is used to retrieve the data according to the first data Reporting to obtain the process parameter data of the production machine; the first communication agent is used to Ϊ269990 to process the remote host and the pluggable quality prediction module is connected to the pluggable quality prediction module. In order to enable the pluggable product, the predictive module transmits the remote data through the application interface via the first communication agent; the first data retrieval manager is responsible for processing all the internal parts of the first universal private device. Message delivery to process production machine, pluggable = quality prediction module and remote host output input to the first universal: two, toast eight armed The information acquisition data acquisition plan is located in the first general-purpose embedded device or a Y-gene general-purpose intrusion device, and the equivalent data acquisition plan is located in the first universal t-input device. At the time of 'generating a second data acquisition plan, the first device driver i then according to the second data acquisition plan, the self-measurement machine collects the actual quality h measurement value of the product, the first data acquisition manager and the processing amount The measuring machine, the pluggable tamper prediction module and the remote host output the input to the first universal embedded device. When the equivalent data acquisition plan is located in the second general-purpose embedded device, the second universal embedded device further includes at least a measurement data acquisition plan, a first shame driver, and a second data capture system. Manager. The measurement data acquisition plan is to describe the heart of the remote host and the pluggable quality prediction module and generate a second data capture report; the second device driver is used to capture the plan according to the first data. The measuring machine collects the actual quality detection value of the product; the first communication agent is used to process the requirements of the remote host; the second data retrieval manager is responsible for processing all the information transmission inside the second universal embedded device. The processing of the measurement machine, the pluggable quality prediction module, and the remote host output signals to the second general-purpose embedded device. The pluggable quality prediction module includes at least: a estimating mode device (Means), a 10 1269990, and a prediction mode device, -, and a setting, wherein the estimating mode device is driven by the manufacturing process of the production machine. Estimate the media γ,, ^ and the production quality of one batch of products produced in the production machine is based on the first type of neural network, fuzzy method I* ^ 4 wood survey and other The estimation of a group consisting of technology with estimation ability is preliminarily established. The pre/post plate device is based on the current estimated product of this batch of products. The predicted value of the next batch of products is predicted by the Φ and Wu detection values of at least one of the previous batches collected by the test machine. J: This prediction mode is a total of 筈 1 ^ 1^ It is selected from the group consisting of weighted moving averages, the second type of neural network ports, and other predictive capabilities. The method for predicting a group of people consisting of the r method is based on the above-mentioned purpose of the present invention, and provides a method for predicting the quality of the application of the universal type of embedding, + * production and production. A preferred embodiment in accordance with the present invention includes at least a training phase and an operation (4) of a quality prediction method for a production process using a general-purpose embedded device. This training phase includes at least the following data r general-purpose embedded device to collect the process of the production machine. The general-purpose embedded device is installed in the production machine, and the system is installed in the production machine. ; the first universal embedded device transmits

.^"Beef material to the remote host; use the measurement data to calculate the gauge to search for at least one sleeve measured by the measuring machine ~ Libei (four) take ~ m ghosts from I where 筮 -, sm 刖At least one actual quality test value of the batch product, embedded: pass type: the in-line device is installed in the measuring machine; the second universal type self-searching two-pronged chest (four) - the general 35 Hawthorn M formulates and sends a set The best weight and function data for the first and the actual b self-search (four) the remaining (four) process parameters of the 'happiness" value to select the most appropriate 1269990, the combination of the good weight and the function data required by the estimation method or prediction method, The operational phase is based on the pluggable ": estimation/predictive accuracy. The following steps are included: · According to the most:: 2: in the township module, the operation phase to the estimation model Newton-heavy, the function data is used for function setting; the eight steps are used to make use of the production by 篦_ Process parameter data of the machine: pass: type embedded device, the estimated quality value of one batch of products, and the estimated value obtained by the production machine is producing the first-class neural network, a fuzzy fuzzy two-order estimation mode step Use the estimation method selected from the technical street of 由, 相, _, data exploration and other populations with estimation ability; provide the prediction mode step, so as to use the current estimation Twist α actual quality test value to predict; 22nd quality value 'plus the predicted quality value of the batch product of the previous batch product, wherein the pre-mode step system is selected from the weighted moving average, the second type of neural network and A prediction method for a group of other algorithms with predictive capabilities. Therefore, 'the application of the present invention can be put into the universal embedded device inside the equipment (production machine and measuring machine), so that the current process parameters of the production machine itself and the previous few can be used before the product is produced yet. The quality inspection data of the batch products predicts the quality of the next-batch products to effectively improve the product quality and the use efficiency and proper rate of Xinfengtai, thereby enhancing the competitiveness of the manufacturing industry. [Embodiment] Please refer to FIG. 1 , which illustrates the architecture of the general-purpose embedded device of the present invention, which includes three main parts: a remote host 80 , a universal embedded device 90 , and a universal embedded device . 50 and a pluggable quality prediction module 40 interconnected with a universal embedded device, wherein the universal embedded device % is installed in the raw 12 1269990 ^ machine 20, and the universal embedded device 5 is installed in the amount Test machine %. • The remote host 80 and the universal embedded devices 9 and 5 are connected via a network 60 such as Ethernet or Ethernet. The three main sections are explained below. The back-end host 80 is mainly responsible for receiving the process parameters and the quality prediction analysis result data transmitted by the general-purpose embedded device 9 and receiving the quality inspection data of the products transmitted by the universal embedded device 50, and The data is stored in the database. The remote host 8〇 of the present invention is not responsible for the analysis and prediction of the state information of the production machine, but distributes the function to the quality prediction module of the general-purpose embedded device 9 of the production machine 20 4〇 to deal with, this architecture can avoid the problem of the entire analysis and monitoring system stall due to the central host failure. At the same time, this architecture only transmits a small amount of analysis and prediction results back to the remote host for storage, or as an integrated evaluation of the overall device status and performance, and does not require sending a large amount of process parameter data for each production machine. The remote host is 8〇, so the problem of network data transmission and network congestion can be greatly reduced. The universal embedded device 90 can be embedded in the interior of the production machine 2, and has three main functions: one is to capture the process parameter data of the production machine 20, and send it to the pluggable quality prediction module 40. Analysis, monitoring, and prediction; second, the prediction and analysis results are transmitted to the remote host 8 through the communication agent 82; third, the remote host 80 can perform relevant parameters on the production machine 2 through the universal embedded device 90. Data inquiry. In addition, the dynamically loadable pluggable quality prediction module is responsible for monitoring and predicting the state of the machine 20 and proposing quality predictions for the production process.壬, 口 The general-purpose embedded device 50 mainly uses the measurement data acquisition plan 58 from 13 1269990 • The measuring machine 30 collects the actual quality detection value of the product. Since the structure of the 9〇 is equivalent to the general-purpose embedded device 5〇, therefore, the two-two-quality prediction system can directly extract the measurement data into the device of the second type, and the general type is omitted. The rear-loading device W is placed in a universal type embedding: referring to FIG. 2 and FIG. 2', FIG. 2 is a schematic structural view showing the general-purpose embedded government of the present invention. The general-purpose embedded system and the 9-inch structure shown in Fig. i are roughly equivalent to the general-purpose embedded device 7〇 shown in Fig. 2, except that the universal embedded device 5G does not need to be connected to the application. Modules, without the application of Lu interface 74. The system function requirements of the universal embedded device f 70 and its module functions are described as follows: The system function requirement of the universal embedded device 70 of the present invention is that it must be able to break into various devices, such as a production machine. Station, measuring machine, transport equipment, etc., and must meet SEM such as semiconductor or electronics industry [Organization to meet some specifications requirements of EEC (Equipment Engineering Capability)" such as E132 (Authentication and Authorization; A&A) , E120 ^(Common Equipment Model ; CEM) > E125 (Equipment Self

Description, ESD), E134 (Data Collection Management, DCM), etc. A&A provides a method of authentication and authorization when communicating and connecting with equipment in the factory; CEM aims to provide a generalized static model to describe the structure of the physical device and the attributes it should have; ESD is an abstraction The Metadata model of the device, which is used to describe the units of the production machine or measuring machine, Types, Structure, State Models, Events. ), Alerts, and Exceptions 14 1269990 (Exceptions), etc.; * DCM|lJ is used to describe the abstract capture plan model. - When the pluggable quality prediction module 4G predicts the _ 〇' value of the next-batch product, the general-purpose embedded oscillating 7 〇 must be able to transmit the predicted quality value to the remote host 80 . On the other hand, the 'general-purpose embedded device 7〇 must be able to receive the best weight and function data transmitted from the remote host 8〇 to the universal dragon-type device 70 to set the pluggable quality pre-_ In addition, the universal embedded device 7〇 must provide a common application interface, so that it can be loaded according to the characteristics of different production machines or measuring machines. The pluggable quality prediction module 4〇. As shown in Fig. 2, the general (four) human device 7 () includes: a communication agent 72, a data retrieval manager 76, a data mining plan 78, and a data acquisition report. η, device driver 75 and application interface 74 and other functional modules. Communication agent?! is the general-purpose rear-entry device 70 external communication module, which is responsible for processing the remote host 8〇, and reply if necessary To the remote host 8. According to the 1£132 specification, the action that the client calls when communicating with the communication agent 72 must be implemented as a simple object connection agreement (Simple 0bject Access p_c〇i; s〇Ap). Sub-delivery mechanism, if necessary The agent 72 can also support the wireless communication function. The data manipulation manager 76 is the core of the universal embedded device, which is responsible for handling all the general-purpose drum-type installations. 7〇 Internal message transmission to process the production machine. 20 or measuring machine 纟 3 〇, pluggable quality prediction mode, component and remote host 8 〇 output input to the universal embedded device 7Q message The daring (Gu) specification provides (4) (4) The design of the manager 26 is the same as that of the remote host. The 心理 心理 _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ After the (4) of the human device 7Q, after obtaining the (4) 15 1269990 acquisition plan 78, the data retrieval manager 76 can generate a data acquisition report 73 based on the data acquisition plan 78, and then request the device driver. • 75 to retrieve the data of the production machine or measuring machine according to the requirements of the data retrieval report 73 and fill in the data acquisition report 73. Then, the data extraction manager 76 sends the data through the communication agent 72. The capture report 73 is sent to the remote host 80. The sample pluggable quality prediction module 4 can also transmit a data capture plan 78 to the data capture manager 76 via the application interface 74, and the data capture manager 76 will use the device capture program 78 via the device driver. 75: Extract the data of the required φ ΐ production machine or measuring machine, and fill in the data retrieval report 73. The data retrieval manager 76 will send the data retrieval report 73 via the application interface 74 to the insertable The quality prediction module 40 is used for analysis, monitoring, or prediction. The data acquisition plan 78 is responsible for describing events (Events) and data tracking (Trace Data) required by the remote host and the pluggable quality prediction module 40. With the Exceptions data type, and the data capture plan π can support such as Diagnostics, Health M〇nit〇ring, Utilization Tracking and Process Control (process c〇ntr) 〇 1) A variety of different data collection needs. The data capture report 73 is responsible for defining the extended markup language (xml) message format of various types of data including events, data tracking and exceptions, and the same as the timestamp and time resolution (Resolution). Format and specifications. The data capture report 73 sends the data to the remote host 80 via the correspondent agent 72 in accordance with the specifications of the XML message structure defined by semiei 28. The protocol of the correspondent agent 72 is the same as that of the remote host 8's communication agent 8 2 (as shown in Fig. 1). The device driver 75 is connected to the data and output of the controlled device via a standard serial interface (such as rs_232, rs_485 16 1269990, etc.), an Ethernet interface, or an analog digital converter. Here, SEMI's CEM (E120) specification provides a conformance vocabulary specification that describes the equipment frame. The ESD (E125) specification provides a standard data structure to describe the unit, type, structure, and state model of the device. , events, alerts, and exceptions. Moreover, in order to achieve the purpose of dynamically loading the pluggable quality prediction module 40, it is necessary to design a general application interface 74 to load application modules having various functions to achieve interchangeability (Interchangeability). Regionalization (L〇calizati〇n), Customization, Decentralization and Empowerment. One of the characteristics of this quality prediction system is that it can select an appropriate pluggable quality prediction module according to the requirements of various production machines. In order to allow the user to easily select and establish the required quality prediction system, the present invention designs the pluggable quality prediction module 40 into a dynamically loadable manner, and the user can select the characteristics and requirements according to the device. The selected modules are selected from the pluggable application module library and loaded into the general-purpose embedded device 70 for performing the analysis, monitoring, prediction, and the like. The pluggable quality prediction module can receive the remote host 8 through the application interface 74, and then send the incoming message to the relevant client, and the pluggable quality prediction module 40 detects the controlled device. When an exception occurs, the exception message can also be transmitted to the remote host 80 via the application interface 74 and the communication agent 72. In order to fb dynamically load the pluggable quality prediction module 4, a universal application interface 74 must be designed in the general-purpose embedded device 70 to load modules with various application functions. The functions and data formats of various quality predictions are defined first, and then different methods (Methods) of 17 1269990 are defined according to different functions and data types for callable application module calls. In addition, if the user wants to develop his own pluggable application module, he/she only needs to use the method provided by the application interface 74 to perform data receiving and transmitting work with the universal mourning device 4. In order to be able to dynamically load the universal type of enter-in device, it is necessary to package the designed pluggable application module into a type such as dll (7) 10 (10) 卜 (Link Library), ActiveX, etc., and add a pluggable application module. In the library. Compared with the current general monitoring system such as SPC (Statistic Prence, APC (AdVanced Process Control), etc., the present invention can be based on the current process parameters of the equipment (production machine) itself before the product is produced. The quality inspection data of the previous batches of products predicts the quality of the next batch of products. Please refer to Figure 3, which is the main structural diagram of the pluggable quality prediction module of the present invention, wherein the pluggable quality The prediction module is composed of the estimation mode device 100 and the prediction mode device 200. This quality prediction system can be used as both virtual measurement and quality prediction, and the estimation mode device can be used as a virtual measurement. The two-layer architecture allows the quality prediction system of the present invention to be appropriately combined according to the actual needs of the factory, and thus is more flexible in practical application. In addition, the estimation mode device is used in the first embodiment. The estimation method is replaceable and can be selected according to the physical and parameter characteristics of the actual production equipment: ^ artificial intelligence, statistical methods or mathematical algorithms, such as the nerves Roads, fuzzy theory, data mining techniques, etc. As for the prediction method used in the prediction mode device, it can be a weighted moving average, a neural network or other algorithm with predictive ability. Please refer to Figure 1 and 4, FIG. 4 is a structural diagram showing the quality prediction system of the self-searching device and the self-adaptation device of the present invention, and the estimation mode device 1 Ο Ο, estimation accuracy rate, and prediction accuracy rate in 18 1269990 Evaluation (1) 10 2 'set 13 °, prediction mode device ^ 'set 21G#, located in the pluggable quality prediction module '〇中' and select and set the interface and prediction mode self-adjustment device 410 / position estimate = soil type The self-adapting device of the mountain is located in the remote host 80, and the original data pre-processing device 1 1 0 can be used in the &<彳', type embedded device 9〇. Module 40 or General First, the production machine 2 〇 寂 寂 杳々 杳々 # # 资料 资料 资料 资料 资料 资料 资料 资料 资料 资料 资料 资料 资料 资料 资料 资料 资料 资料 资料 资料 资料 资料 资料 资料 资料 资料 资料 资料 资料 资料 资料 资料 资料 资料 资料 资料 资料Process parameters and bean subtraction There are a wide variety of data, and it is necessary to estimate the part of the data to be used in the 咸 咸 1 1 1 1 1 ' ' ' ' ' ' ' ' 原始 原始 原始 原始 原始 原始 原始 原始 原始 原始 原始 原始 原始 原始 原始 原始 原始 原始 原始 原始 原始 原始 原始 原始 原始Process parameters or sensors (4) Which data is selected as the input parameter is different according to the production machine Γ ΓΛ and the selected estimation method or prediction method. The process parameters and sensor data obtained by this 7 preparations may be There are different = formats, the original data pre-processing device 110 must be able to handle these different 1 and convert to the specific data format that the estimation mode device just needs. To the estimation 2: the original poor material pre-processing package is 11 〇 processing The input data transmitter is second. The purpose of the estimation mode device 1〇0 is to use the production data to estimate the quality of the product being produced, and obtain the estimation to the ^2/: 2# estimation f value input to the _ mode device 2〇0' ,, 〇^ ^ The purpose is to use the estimated negative value (y) of the estimated products of the current evaluation, and the actual products of the previous batches of products obtained by the I machine 30

Bay detection value (less one, V port ~, & ···, U ' to predict the quality of the next batch of products to be predicted, (U pre-then mode device 200 and estimation mode device 1〇〇) Similarly, according to the characteristics of the production machine a * ' σ 20 itself, according to the characteristics of the production machine a * ' σ 20, the search mode, such as the movement, the appropriate prediction method to establish the pre-sand-dry, neural network Etc. Just send the estimated quality value (exempt) to the company, and the actual product of this batch of products is recognized by the estimated accuracy evaluation device 130, 俨 η. The actual 〇 mouth quality detection value U) In contrast, the choice of 士士, 隹念玄^ is shown. At the same time, the forecast mouth Zeng Yi,,,, 乂 侍 侍 侍 侍 侍 侍 准确 & & & & & & & & & & & & & & & & & & & & & Recalling F set 21Γ) ” The actual quality test value of this batch of products is estimated to be 4, 210, and the index is estimated. The evaluation is based on +/), to obtain the prediction accuracy rate τ 1 ancient estimation / prediction accuracy τ There are many kinds of characteristics of the object to be evaluated, which can be evaluated according to the desired evaluation of "μαρε" and the maximum error, such as the average absolute percentage error. x Error), etc. can be used to assess the accuracy of the indicator. 0 as the estimation/prediction of the present invention, because of the wide variety of production equipment of various devices, such as semiconductor 盥 TFT T rn e, it is used for various kinds of chickens.趴 趴 趴 。 。 。 。. Therefore, the present invention and the predictive mode device 200, the user of the basin and the sputum, pass through the selection interface 120, according to the production machine, the 蛀 圪 圪 、 、 、 、 、 、 、 、 、 、 、 、 、 、 The characteristics of 20, to choose a variety of different artificial Hui, statistical methods, or mathematics to play a win-win system y 堪 η 于 肩 肩 肩 肩 肩 肩 肩 肩 肩 肩 肩 肩 肩 肩 肩 肩 肩 肩 肩 肩 肩 肩 肩 肩 肩 肩 肩 肩 肩 肩 肩Fang, Ben & cat's ^ η method and prediction method. With the use of machine types, plus, ... some prediction algorithms are gradually increasing the number of measurement algorithms more and more, the applicability of the pre-97 prediction structure for the use of the building It will be more and more s H: The purpose of face 120 is to assist the user to select the appropriate estimation or prediction method and to capture the initial value associated with g* Μ /dr Α4» 卫 疋. After the user completes the selection and setting, the quality prediction system can be executed. In addition, the self-searching device 3 of the estimation mode and the self-searching device 310 of the prediction mode, and the self-adapting device of the estimation mode and the prediction mode self 20 1269990 ^the device 410 is another item of the present invention. The manpower and time of the estimation/prediction architecture. The main purpose of the point is to reduce the construction of the ground - a estimation model or a prediction model - the map does not require a self-search device for efficient algorithm selection. 'Heart' This invention provides a built-in different neural network to establish the estimation mode. From: collision: The user 10 selects the number of hidden layers set by the reverse transmission class 10, each inch, and 300. According to the number of users, etc., automatically search for a mother:: Layer 2: number range... User! 0 uses fuzzy theory to establish the combination of the push function and the weight value. If the possible attribution function == is set, the device 3 is self-searched. . If the user 10 uses the weight to move out the best attribution function. Similarly, the search device 31 will combine the self-search weights according to the :: remaining mode set by the user 10. Therefore, the user 1 〇 finds the best type of number, self-searching device 300 and parameter range or function parameter or function combination and weight value. usually. : The number of generators and the characteristics of the sensor are also different. To establish a η characteristic to establish the estimation model and source and time, so self-search::::: usually need to invest a considerable amount of manpower and the purpose of my technology device is to reduce the manpower demand 'by one by one The model that has been pre-built is::: The setting of the estimation/prediction system when it is transplanted to the new machine = the invention is shortened. W杈 time can be generated in a large amount. In the process of production, it may increase the time and parts of the parts, and the clothes will be retired, which may cause the equipment to shift. Otherwise, it may be used to repair and replace parts. The characteristics of the platform are not the same as before. The prediction architecture should have the ability to self-adjust. The self-adapting device 4i〇 from the 21 1269990 device and prediction mode is a 14-point problem. By monitoring the estimation/pre-test, then entering a also «set 400/410 can always know the quality = the rate of evaluation indicators, the self-adjusted pre-solution is the situation of the estimation/prediction mode device 100/200, and according to the actual situation The equipment demand sets a limit of M% (such as 9G% to 99%). The default rate of the next time (such as]a $) push/predict system accuracy has been continuous η - person to 5 - person) below this pre- When the lower limit is overturned*, the self-adapting device 400/410 breaks up and cooperates with the self-searching device to establish a new Μσ, and a new one is based on the recent equipment parameters. Oral quality prediction system, 俾#, / _ 'Enable estimation/predictive accuracy rate can be recovered, 'predicting the sampling range of training materials and test data, by: the use of the person "wide: self in the new training Adaptation mechanism, choose to happen... 吏 1 with wide. Examples of this description ^ΜΑ^ΙΙΦΦ, Shenglian, Ge 2 long-term accuracy rate before the 5 〇 pen-funded for 4 training materials (2) pen) and test data (25 pens). The self-^ of the prediction mode in Fig. 4 indicates that the self-adaptation device 41() is set to use the dotted line, and the selected prediction mode is determined to be =, and the prediction system constructor can be based on the six-eight wood/day疋 Do you want to add this self-predictive mode device to use weight shift ^ :. For example, the face ^ makes the self-mouthed her ^ because the special test mode of the different method itself self-adjusts "set 4: · fruit ^ easy to highlight, so consider not to add the pre-construction method, then join the self, I ^ ^ nerve The result of the network as a predictive mode is more accurate. 将 0 will reduce the prediction error, and the pre-established node number range setting can be used for the search time of the previous adaptive device. The sputter machine saved from the 4::::ft-lcd factory is used as an application example to illustrate the present invention. The 5th picture of the shame of shame, the bean helmet-shicheng ^ n Ώ /, the TFT_LCI of the application example The machine is plated with 22 1269990 machine schematic diagram. There are 16 modules inside the machine, and each module has 5 scoops of action. The purpose is to apply a layer of film on the glass (raw material), and the metal film The function is used as the color border of the LCD panel. The direction of the arrow in the figure is the processing sequence of the process. First, the glass is processed by the entrance and exit: the °杈 group is processed in the order of the makeup group. The first 8 modes are completed. Cylinder f 'connector rotation module rotates the glass to number 9 The group continues to process, the most Γ = the entrance is sent out, and each module is in vacuum during the operation: the negative shell pumping line module is equipped with a vacuum sensor for direct handling. It is responsible for conveying inert gas. The module is provided with a gas concentration sensor: ^ The actual state of the gas concentration. For the control of the positive and negative ion electric field ^ 2 two electric field, there is also a sensor for detecting voltage, current and power: The actual processing of the private 'will record a series of sensor data. Raw ^ in the process of ◎ record process parameters (sensor) data, glass finish plus (four) mother hundred pieces of finished products random sampling two pieces sent to the measuring machine , measure and record the film thickness of the sputter layer. Please refer to Figure 6, which shows the quality prediction process of this application example. #测测质量质量质量质量的质量质量质量;···,3^) 'The test parameters of the product of the batch of products that are currently estimated by the production machine are estimated by the production machine. ^ Negative value (everything can be predicted) The predicted quality value of the next batch of products to be produced (U. The neural network estimation model used in the application example is established by using a mechanism. The system developer only needs to input the possible number of hidden layers, the number of each hidden layer node, the type of transfer function, and the termination condition. According to this (4), the mechanism will automatically find the best neural network estimation model 23 1269990 value', but the time to search for the best model is related to the setting. If you want to shorten the search time, you can count the number.妒来今田整> Number of search range size B# Μ · -r- ^ ^ ^, 〇 and 疋# range can reduce the search to the right, right does not consider the time factor, then increase the ginseng 1 better neural network Road model. , I δ will have the opportunity to find another 'this application example is to use the moving average to establish the pre-this moving average is to use two quantity, work sorrow, value (3) in the "two H of the first two batches actually paid Quality inspection = °:: Measure the quality of the next batch of products (5)). This application example (4) compares the early moving average and the weighted moving average prediction modes to compare the prediction effects, wherein the weight value range η # ma consumes P and also uses the self-search mechanism (device) to search for the most Good prediction mode. The weighted moving average formula is as follows: = Wiy{ + w^jy^j + w^2y^2 where M^., VV/_/, VVi·-2 is the weight value. The user only needs to set the range of weight parameters. The self-searching mechanism will test various possible weight combinations and find the best weight combination based on the predicted value (the difference between the u and the actual quality detection value Xin$.) This application example uses the evaluation index to judge the prediction accuracy. The evaluation index of this application example can use the average absolute percentage error (the heart grabs Abs〇iute)

Percentage Error (MAPE) and maximum error (MaxErr〇r) are two error metrics for performance evaluation of the quality prediction model as follows: ΜΑΡΕ^±±Λ〇/ο n 4

MaxError = Μαχ\Α. - Ff\ 24 1269990 with medium, n actual value, clever · predictive value, n: number of samples - MAPE error means that the value of the && private value is closer to zero, indicating the model The better predictive power; "deductive private represents the biggest difference between the actual value and the predicted value. Please refer to Chapter 7 4+ Cut ..., Figure ' to show the prediction results of this application example. After screening, the amount of data used by the 曰t and inverted transmission neural networks was 85. m _5 〇 资 乍 = 乍 贝 , 51 51 51 51 51 51 51 51 51 51 51 51 51 51 51 51 51 51 51 51 51 51 51 51 51 51 51 51 51 51 51 51 51 51 51 51 Ben: Test the junction with an example... APE value is 2 〇536%, and the maximum error value is (10) (equivalent to 5.8% of the standard film thickness value of 7GGnm). In addition, the quality prediction architecture proposed by the present invention can use other combinations to construct quality predictions in addition to using the neural network _ estimation mode and the weight moving average (WMA) prediction mode shown in the above example. system. ^ a class of neural network (NN) estimation mode and neural network (NN) prediction mode; or a combination of single-class neural network (CNN) estimation prediction mode. The operation flow of the quality prediction method for the production process of the general-purpose enter-type device of the present invention will be described below. Referring to Fig. 1 and Fig. 8', Fig. 8 is a flow chart showing the quality prediction method of the production process of the universal embedded device (GED) according to the present invention. The quality prediction method of the production process of the invention (IV) using the embedded type device provides a training phase, an operation phase and a self-adjustment phase. In the training phase, the universal embedded device 50 (step 5) and the universal embedded device 90 (step _) are first initialized, wherein the universal money type 9 () is installed in the production machine: The universal embedded device 50 is installed in the measuring machine 3〇, the universal embedded device 90 is installed with the pluggable quality prediction module 4〇, and the universal embedded 25 1269990 input type is set to 5 Measurement data acquisition plan 5 8. Next, the general-purpose post-input device 90 is used to periodically collect process parameter data for the production machine 2 (steps 610). Then, the general-purpose embedded device 9 transmits the process parameter data to the remote host 80 (^62). At the same time, using the general-purpose embedded device 5's measurement data acquisition plan 58 to periodically collect at least one actual quality detection value of the at least one previous batch of products measured by the self-measuring machine 3 (step 510) . Then, the general-purpose embedded device 50 transmits the actual quality detection value to the remote host 80 (step 52A). In succession, the remote host 80 performs step 71〇 and step 72〇 to store the actual quality detection value and the process parameter data to the f-seed respectively. After inputting the actual quality detection value and the process parameter data (step 722), the remote host 8〇 performs a self-searching step to train the pluggable quality prediction mode, the group 4〇 (step 730), and the towel formulates the concurrent ^ most The weights and function data are given to the general-purpose embedded device 90 (step 740) 'where the self-searching step selects the optimal weight and function data required for the estimation method or the prediction method based on the process parameter data and the actual quality detection value'. The combination 'to increase the estimation/prediction accuracy. Then, the system enters an operation phase, which is mainly performed in the pluggable quality prediction module 4G. First, 'initialize the pluggable quality prediction module 40 and start the online product f prediction (step 63 ()) 'where the general-purpose embedded device 90 is based on the optimal weight and & ^ f A #, ^ ' The pluggable quality prediction module 40 advances the work. After inputting the actual quality detection value and the process moxibustion data from the instant collection (step 632), you can select λ + Q #表枉> (step "〇", you can insert two: "Quality prediction module 4" The operation of the start-up port 0 prediction module 40 is mainly based on the estimation mode step and the prediction mode step. (4) The input device 90 instant nursery estimation mode step is produced by the general ghost collection. The process parameter data of the machine 20, 26 1269990, is estimated to obtain the estimated quality value of the batch product which is being produced in the production machine 20 4: square a - medium estimation, and the basin exploration "19th class-type neural network , model _ theory, data using the current estimation method of Doosan* 筏 。. The prediction mode steps are profitable. Quality II!: The estimated quality value of this batch of products plus the real two = measured value of the previous batch of products To predict the predicted quality values of the next batch of products, where predictions are used, for example, weighted moving averages, second type neural networks, or other prediction methods that predict the algorithm of this force. The receiver's general-purpose embedded device 90 transmits the prediction result (predicted to the ring-end host 80 (step 65G), and the remote host 8G stores the prediction result to the database (step 750). At the same time, 'go step 66() Determining whether the estimation/predictive accuracy is a succession—the preset number of times (for example, i to 5 times) is lower than a predetermined lower limit value (for example, 9〇% to 99%), and if the result of step 660 is no, Then continue to operate (step 64 〇 If the result of step 660 is yes, then enter the self-adjustment phase, that is, evoke the remote host (9) to perform self-adaptation mechanism (step 670). This self-adaptation mechanism is based on the latest process parameter data and actual The quality detection value is to retrain the pluggable quality prediction wedge set 40 (step 760), and another set of optimal weights and function data is selected and sent to the universal embedded device 90 (step 770), wherein the self-search The step is based on the process parameter data and the actual quality detection value. Then, the pluggable quality pre-rigid module 40 is initialized and the online quality prediction is restarted (step 68〇), and the operation phase is again entered. In addition, the present invention is pre-processed. The training phase of the method can also be performed in the first universal type of input device, so that the device data need not be transmitted to the remote host, so as to eliminate the problem of insufficient network data traffic and bandwidth. Please refer to the figure! FIG. 9 is a flow chart showing the quality pre-processing method of the production process of the general-purpose dispersing 27 1269990-input device according to another preferred embodiment of the present invention. A general-purpose embedded device 99 is used for cross-connection. /, Product machine port 20 and measuring machine 3〇 in the training phase '#First initialization pass_ 8〇〇), general-purpose embedded device 9 俶曰, A ^ u Female Ling has a pluggable quality prediction module 4〇 and a poor material extraction plan 58. Next, Putian, s t0 .,... Use the universal embedded device 99 to come to Zhou Process parameters of the production machine 2G and the periodicity, the detection value (steps). κι _ at least one actual quality of the product, and then, the general-purpose mortal-type device 99 performs step 71 〇 and the top of the step, so as not to store Actual quality test value and process parameter data to the database. After the quality check and process parameter data (step 722), the general type funeral 99 performs the self-searching step, and sets the 7M, the work system-w-eight port as the predictive group 40 (step 730), and "sends one. The best weight and function data of the group is given to the pluggable quality prediction module of the universal embedded type 99 (step μ 苴, according to the process parameter data and actual quality: self-search step or prediction method) The weight 盥 function is used to measure the accuracy. The combination of the number of 枓 枓 以 以 增加 以 以 以 以 以 以 以 以 以 以 以 以 以 以 以 以 以 以 以 以 以 以 以 以 以 以 以 以 以 以 以 以 以 以 以 以 以 以The group 40 starts the online quality prediction (step 63〇), wherein the universal, and the core 99 are based on the optimal weight and function data, and the plug-in & input device performs function setting. After inputting the _set: number data ((4), the pluggable quality prediction module 4 28 28 1269990 (step 64 〇), the pluggable quality prediction module 4 〇 - mode step and prediction mode step The production machine system that is based on the estimation of the arrogant device 99 is used to estimate the _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ The use of, for example, the first type of neural network. The value of the quality value of its exploration or other estimator's technology is estimated, and the data is estimated by the current estimated product value of the batch. The pre-mode step is a quality test value to predict the actual batch mode of the next batch of products. For example, the weight shift is used. /, the forecast target: ®: the sundial A t music class Network or other - predictive method of predicting the algorithm of this force. Then, the general-purpose embedded 奘詈QQ captures the soil and enters the eight-eighth transmission of the prediction result (predicted product step-by-step main =) 8 () (gate step 650) 'Remote host 80 then stores the predicted results to the database Step 75)). At the same time, step 660 is performed to determine whether the estimation/prediction accuracy is connected: T is set, the column is: 1... times) is lower than the preset lower limit value (for example: 90%). If the result of step 660 is no, then continue. Operation (step 64 〇); If the cow: 60 _ fruit is 疋, then enter the self-adjustment stage, that is, evoke the self-adjusting machine I:, 67G). This self-adaptation mechanism selects and develops another set of optimal weights and function data for general-purpose embedding based on the latest process parameter data and actual mouth-and-mouth quality detection values to retrain the pluggable quality prediction module. The pluggable quality prediction module 4〇 (step 77〇), wherein the self=step is based on the process parameter data and the actual quality detection value. Then, the pluggable quality prediction module 40 is initialized and re-introduced, and then enters the operation phase again. It can be seen from the above-described preferred embodiments of the present invention that the application of the invention can be easily embedded with the production process equipment without the need for an existing data acquisition system. Or spend money to construct a set of data • = • device; embedded in each device a universal embedded device for data extraction and analysis, to avoid the failure of centralized predictive monitoring system caused by = predictive monitoring function The problem of designing the quality prediction function module into a dynamic loading mode, the user can select the required module from the quality prediction function module library to load into the universal embedding according to the characteristics and requirements of the device. In the device, for equipment analysis, monitoring, and forecasting; only a small amount of analysis, _ prediction results are transmitted back to the remote host for storage, or as an integrated assessment of the overall device status and performance, no need to A large amount of state parameter data of the device is sent to the far 2 host, thus greatly reducing the problem of network data transmission and network congestion. Before the production, the quality of the next batch of products can be predicted according to the current process parameters of the production machine and the quality inspection data of the previous batches of products, so that the production of defective products can be greatly reduced, and the production capacity of the factory is not only improved. Compared with the yield of the product, the production cost is also reduced, and the competitiveness of the factory is further improved. Further, the present invention can be used as both virtual measurement and quality prediction, and has versatility. Further, the present invention has a self-searching mechanism and a self-adapting mechanism, which can effectively estimate/predict accuracy. The present invention has been described above with reference to a preferred embodiment. However, it is not intended to limit the scope of the invention, and various modifications and improvements can be made without departing from the spirit and scope of the invention. Therefore, the scope of protection of the present invention is defined by the scope of the patent application. Field 30 1269990 [Simplified description of the drawings] The following is a more complete understanding of the present invention and its advantages in order to more fully understand the following drawings, in which: Figure 1 shows the application of the present invention. Schematic diagram of the general-purpose quality prediction system. 2 is a schematic view showing the structure of the present invention. The main knot φ of the input quality prediction system Fig. 4 is a schematic diagram showing the structure of the quality prediction system of the right white washing device of the present invention, Fig. Y inch and the self-networking machine, Fig. 5 is a drawing Schematic diagram of the device of the application example of the invention. < Machine money machine Figure 6 is a schematic diagram showing the flow of the mouth of the mouth. Figure 7 is a graph showing the results of the application of Xiyi and a π. Fig. 8 is a flow chart showing the method of predicting the quality of the production process of the production process using the universal pen-in device according to the preferred embodiment of the present invention. _ Fig. 9 is a flow chart showing the quality of the production process of the embedded device according to the present invention, and the general application and the top method of the preferred embodiment. [Main component symbol description] 10 User 2〇 30 Measuring machine 4〇 50, 70, 90 Universal embedded 裴 58 Measurement data acquisition report Production machine Pluggable quality prediction module Set 1269990

60 Network 73 Data Capture Report 75 Device Driver 78 Data Capture Plan 82 Communication Agent 90 Universal Embedded Device 100 Estimation Mode Device 110 Raw Data Pre-Processing Device 120 Selection and Settings Interface 130 Estimation Accuracy Evaluation Device 200 Predictive mode device z communication agent 74 application interface 76 data retrieval manager 80 remote host 210 300 310 400 410 500 510 520

600 610 620 Predictive Accuracy Assessment Device Estimation Mode Self-Searching Scenario Prediction Model Self-Searching Predictive Model Self-Adjustment Attack Prediction Mode Self-Adjustment Swing Initialization General-purpose embedded Duntage periodically collects products The actual σ is... By Lt, the quality detection value is transmitted, the actual quality detection value is initialized, and the general type of enemy type is periodically collected. The process parameter data transmission process parameter data 630 of the production machine is initialized and the pluggable quality prediction module is initialized. Start online quality prediction 632 Enter the actual quality detection value and process parameter data from the instant t set 32 1269990 640 Operation ^ 650 Transfer prediction result ^ 660 Estimate/predictive accuracy is the continuous preset number lower than the preset lower limit value 670 Self-adaptation mechanism 680 Restart online quality prediction 710 Store actual quality detection value 720 Store process parameter data 722 Enter actual quality detection value and process parameter data φ 730 Training pluggable quality prediction module 740 Send optimal weight and function data 750 Storage Forecast Results 760 Retrain Pluggable Quality Forecast Another group of preferred transmission right group 770 and the weight function information embedded device 800 initializes the general type

33

Claims (1)

1269990 X. Patent application scope, k A quality prediction system for a production process of a general-purpose embedded device (Generic Embedded ^ Device·' GED), comprising at least: a first universal embedded device mounted on a production machine The first universal embedded device is coupled to a pluggable quality prediction module; and a remote host is configured to process the data transmitted by the first universal embedded device. And displaying and storing the predicted result and the _ abnormal condition; and a data collection plan (Data Collection Plan) connected to a measuring machine, wherein the measuring data capturing plan is responsible for describing the remote host and the The data type required by the plug-in quality prediction module is used to collect the actual quality detection value of the product from the measuring machine; wherein the first universal embedded device includes at least: a first data acquisition plan for Describe the data type required by the remote host and the pluggable quality prediction module, and generate a first φ data acquisition report (Dat a Collection Report); a first device driver (Equipment Driver) for obtaining a process parameter data of the production machine according to the first data acquisition report; a first communication agent (Communication Agent) for Processing the requirements of the remote host; an application interface is interconnected to the pluggable quality prediction module, so that the pluggable quality prediction module passes through the application interface through the communication agent, To transmit data to the remote host; and a first data capture manager (Data c〇Uecti(10) M_ge〇' is responsible for processing all message passing inside the first universal embedded device to process the production machine The pluggable quality prediction module and the 13⁄4 remote host output the information input to the first universal private device; , wherein the pluggable quality prediction module comprises at least: - an estimated mode device (Means) ) 'Using the process parameter data of the production machine to estimate the quality value of the batch of products being produced in the production machine Wherein the estimation mode device is determined by a method of estimation::: and the estimation method is selected from the group consisting of: a first type of neural network, a group of people; and a method of estimating a force-force group = mode The device utilizes the current estimated value of the batch of products, plus the at least one actual product f detection value of at least one of the previous batches collected by the measuring machine, to produce a predicted quality value. Where the prediction mode is: and the prediction method is selected from a weighted moving average, a first type of neural network, and other _ groups. The composition of the pre-U method - 35 1269990 group includes at least: - the original processing device of the original "secondary", used to convert the data entered by the production machine into a specific data format. Enter the data. 3. If the application for the dedicated production process group includes at least: The general-purpose embedded defect quality prediction system described in item 1 of the scope of interest, wherein the pluggable quality pre-modeling %, a Suitable device, wherein when the production machine is operated - the time is estimated / the accuracy of the prediction is reduced to a preset accuracy, or the characteristics of the production machine are changed due to the repair and replacement of the parts, the self-destruction The appropriate mechanism will be activated to correct and meet the new requirements for the production machine and the estimation/predictive accuracy. 4. The quality of the production process for the application of the universal embedded device as described in the scope of the patent application. a prediction system, wherein the estimation mode device is provided with a virtual metrology function. 5 · The quality of the production process of the general-purpose embedded device as described in claim 1 a prediction system, wherein the measurement data acquisition plan is located in the first general-purpose embedded device for generating a second data retrieval report, the first device driver extracting a report according to the second data, The measuring machine collects the actual quality detection value of the product, and the first data extraction manager is responsible for processing the measuring machine, the pluggable quality prediction module and the remote host output input to the first universal type The message of the rear-loading device. 36 1269990 6·The application of the general-purpose embedded device & the quality prediction system of the 玍 転 転 如 如 如 如 如 如 如 如 如The second transmission will be stored in the database by the first-send-data query request, and the general-purpose embedded device will be sent back to the remote host after obtaining the data. · As stated in the patent pending 聋巳 之 之 应用 通用 通用 通用 通用 通用 通用 通用 通用 通用 通用 ° ° ° ° ° ° ° ° ° ° ° ° ° ° ° ° ° ° ° ° ° ° ° ° ° The structure of the second general-purpose embedded device is equivalent to the first general-purpose embedded device, wherein the second universal embedded device further comprises: a data acquisition plan for describing the remote end a data type required by the host and the pluggable quality prediction module, and generating a second data capture report; a second device driver for extracting an advertisement according to the second data from the measurement The machine collects the actual quality detection value of the product; - the second communication agent handles the request of the remote host; and a second data retrieval manager who is responsible for processing the second universal plastic input device All internal messages are transmitted to process the measurement machine, the delta pluggable quality prediction module, and the remote host output input to the second universal embedded device. 37 1269990 8. The quality prediction system for a general-purpose embedded-mounted production process as described in claim 7 of the scope of the patent application, wherein the remote host is to be the first-general-type embedded device and the second The data transmitted by the universal embedded device exists in the database, and can send a data query to the universal embedded device and the second universal embedded device, and the first universal embedded device and the first The second general-purpose embedded device returns the remote host after obtaining the data. • The quality prediction system for the production process of the universal embedded device as described in claim 1, wherein the remote host develops the estimation method or the prediction method through a self-search device A combination of the best weights of a group and the function data. 10. The quality prediction system for a production process using a general-purpose embedded split as described in claim 9 wherein the remote host transmits the set of optimal weights and function data to the first general-purpose embedded device To perform the function setting of the callable quality prediction module. η. The quality prediction system for a production process of a general-purpose embedded device as described in the scope of the patent application, wherein the remote host sends a quality prediction shoulder to the quality prediction function module, and the quality prediction The functional wedge group responds to the quality prediction query request to reply the relevant result to the remote host. 12. The quality prediction system for the production process of the general-purpose embedded device 38 1269990, as described in claim 1, wherein the application interface dynamically loads the appropriate one according to the characteristics of the controlled device. The pluggable quality prediction module 13 is the quality prediction system for the production process of the universal embedded device as described in claim 1, wherein the data retrieval report is responsible for defining events, data tracking ( Trace DaU) Extensible Markup Language (XML) message format for various types of data with exceptions; and formats and specifications for defining Timestamps and Resolutions. 14. The quality prediction system for a production process using a general-purpose embedded device according to claim 1, wherein the device driver is via a standard serial interface, an Ethernet interface or a An analog digital converter is connected to the production machine or the measuring machine. 15. A quality prediction method for a production process using a general-purpose embedded device, comprising at least: providing a training phase, wherein the training phase includes at least: collecting a process of a production machine using a first general-purpose embedded device Parameter data, wherein the first universal embedded device is installed in the production machine, the first universal embedded device is provided with a pluggable quality prediction module; and the first universal embedded device transmits process parameter data To 39 1269990 a remote host; use ^ a test data; jig and μ, the needle to draw the leaf stroke to collect at least one actual product of at least one of the previous batch products measured from a measuring machine Quality detection value; Transmit actual quality check |, Sub-double value to the remote host; and the remote host enters thousands, - Schedule a self-search step to develop and send a set of optimal weights and function data仏Jr ^ ^ Beiding卞,,, 口5hai first general-purpose embedded device, in which the self-search step is based on process parameter data and actual quality detection values. To pick out _ |, which required a conjecture of a prediction method or methods The best weight of the group and the combination of the |^ ^ / A mountain number to increase the estimation/prediction accuracy; and the provide-operation phase, wherein the operation phase is mainly performed in the pluggable quality prediction module. The operation phase includes at least: f performing function setting according to the optimal weight of the group and function data; and providing a step of estimating mode to utilize the production machine of the first universal type of intrusion device Process parameter data to estimate a derived quality value of a batch of products being produced in the production machine, wherein the estimation mode step uses the estimation method and the estimation method is selected from a first type of nerve The network, the touch, the one-shot exploration, and other techniques of estimating the ability to provide a predictive mode step to utilize the current estimated quality value of the < a batch of products, plus The pre-production product is not qualified; [: human measurement value] predicts a predicted quality value of the next batch of products, wherein the ^^ measurement mode step uses the prediction method, and the prediction method is curved Λ, , Weight of 1,269,990 a weighted moving average, and a second neural network with other predictive ability of the method depends isobutyl group consisting of one. 16. The method for predicting the quality of a production process using a general-purpose embedded device as described in claim 15 of the patent application, further comprising: performing a raw data pre-processing step whereby the original process to be rotated by the production machine The parameter data is converted into an input material having a specific data format, wherein the original data pre-processing step is performed in the pluggable product pre-module module. . 17. The quality prediction method for a production process using a general-purpose embedded device as described in claim 15 of the patent application, wherein the estimation mode step X is provided with a virtual measurement function. /, 18· As described in claim 15 of the scope of application of the general-purpose embedded, the quality prediction method of the production process of the device, at least include: Formula% provides an evaluation index to evaluate the quality of the production process Accuracy. The system for predicting the quality of a production process using a universal embedded gas device as described in claim 18, wherein the evaluation index is selected from the group consisting of an average absolute percentage error and a maximum error. . 2〇·Assessing the application of general-purpose embedding as described in the application of the scope of the patent, the quality prediction method of the production process of the type 41 12699900 device, and the method of moving the average weight of the building The weight is moved, and ^ ^ Λ ... sentence is used to obtain the actual products of the two batches of products obtained by the 5 Hai® measuring machine. The actual value of the wooden mouth is / / /, and the value is negative. υ, plus the estimated mouth value (yes), to predict the mass value of the batch δ hai hai pre-deletion (U, β weight moving average formula is .~ ^, two less...= + + U-gate is the weight value. • As applied in the general scope of the application, the application of the general-purpose embedded name, the quality prediction method of the production process, at least include: Provide a self-adjustment stage, so as to estimate / When the prediction accuracy is continuously reduced to less than a predetermined lower limit value, the first universal embedded device evokes the remote host to perform a self-adaptation mechanism, wherein the self-peripheral L mechanism is based on the latest Process parameter data and actual quality test value To select another set of optimal weights and function data, so that the pluggable oral quality prediction module re-functions according to the other set of optimal weights and function data. 22. For example, claim 21 The method for predicting quality of a production process using a general-purpose embedded salty method, wherein the preset number of times is between 1 and 5 times. 23· Applying universal embedding as described in claim 21 of the patent application scope The method for predicting the quality of the production process, wherein the preset lower limit value is between 90% and 99%. 42 1269990 ★ 24· Application general-purpose embedded as described in claim 15 The method for predicting the quality of the production process of the device, wherein the measurement data acquisition plan can be located in the first general-purpose embedded device. 25· Applying the universal embedded device as described in claim 15 The quality prediction method of the production process, wherein the measurement data acquisition plan is located in a second general-purpose embedded device, and the structure of the second universal embedded device is equivalent to the first universal embedded device The second universal type of intrusion device is installed in the measuring machine. 26. A quality prediction method for a production process for applying a universal embedded device, comprising at least: providing a training phase, wherein the training phase is at least include: Using a general-purpose intrusion device to collect process parameter data of a production machine and at least an actual quality detection value of at least one pre-batch product measured from a measuring machine, wherein the universal embedded device is Interconnected with the production machine and the measuring machine respectively, the universal embedded device is equipped with a pluggable quality prediction module and a measurement data acquisition plan; the universal embedded device enters, Self-search steps to develop and send a set of optimal weights and functions to be cautious μ ^ ^ 默贝枓 to the pluggable quality prediction module 'after the self-search step & 乂·乡乡 according to process parameters and The actual quality detection value is used to select the combination of the best weight and the function data required by the estimation method or a prediction method 43 1269990 to increase the prediction accuracy; and #/ provide an operation phase, wherein The operation phase is mainly performed in the pluggable mouth prediction module, and the operation phase includes at least: ^ function setting according to the optimal weight and function data of the group Providing an estimation mode step, by using the process parameter data of the production machine collected by the universal embedded device, to estimate a quality value of a batch of products being produced by the production machine, Wherein the estimation mode step uses the estimation method, and the estimation method is selected from the group consisting of a first type of neural network, a fuzzy theory, a data arbor and other techniques with estimation capabilities. a group; and providing a predictive model step to utilize the estimated quality value of the batch of products currently estimated, plus the actual quality of the previous batch of products: the measured value to predict a predicted quality of the next batch of products The value, wherein the pre-step method uses the prediction method, and the prediction method is selected from the group consisting of a weighted moving average, a second type of neural network, and other predictive algorithms. The production process of the device is carried out as a raw • The general-purpose embedded quality prediction method described in item 26 of the patent application scope The Bayer prediction method further comprises: a Γ processing step, wherein the input pre-processing step of converting the man-made material input by the production machine into a specific data format is based on the plug-in quality prediction 44 1269990 26 The 嵌入式 type embedded measuring method, wherein the estimating mode step has the function of preparing a virtual measurement of the quality of the production process of the device of the patent application. Provide an evaluation indicator to assess the accuracy of the system. 29. If the quality of the production process of the device for applying for the patent range is predicted, the general-purpose method for the application of the method is more than at least: the quality prediction system of the production process is as described in item 29 of the patent application scope. A quality prediction method for a production process using a general-purpose embedded device, wherein the evaluation index is selected from the group consisting of an average absolute percentage error and a maximum error. 3 1 · A quality prediction method for a production process using a general-purpose embedded device as described in claim 26, wherein the prediction method uses the Lu weight moving average, and the weight moving average is adopted by the measuring machine The actual quality test values (h, D, plus the estimated quality value (free)) of the first two batches of products obtained by the station to predict the predicted quality value (meaning +/) of the next batch of products, The formula for the weighted moving average of the sea weight is: L = + w / brother · · y + where is the weight value. 32 · The quality prediction method of the production process using the universal embedded device as described in claim 26 of the patent application scope, More than at least: 45 1269990 Box self-adjustment stage 'by the time when estimating/predicting the Bureau of Punishment: value, evoking the universal type of access: exhibiting a self-adapting mechanism' where the self-tuning: the latest process parameters and眚^口@^ is based on the number of shells and the quality detection value to select the other group of the best weight and function data 'to make the pluggable quality prediction module according to the other '' and finally Soil weight The function data is re-configured. 33. The quality prediction method for the production process of the general-purpose embedded device described in claim 32, wherein the preset number is between 1 and 5 times. 4. The quality prediction method for the production process using the universal type septic device as described in claim 3, wherein the preset lower limit value is between 90% and 99%.